1//===- AddressSanitizer.cpp - memory error detector -----------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file is a part of AddressSanitizer, an address basic correctness
10// checker.
11// Details of the algorithm:
12// https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm
13//
14// FIXME: This sanitizer does not yet handle scalable vectors
15//
16//===----------------------------------------------------------------------===//
17
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/Twine.h"
28#include "llvm/Analysis/GlobalsModRef.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/EHPersonalities.h"
47#include "llvm/IR/Function.h"
48#include "llvm/IR/GlobalAlias.h"
49#include "llvm/IR/GlobalValue.h"
50#include "llvm/IR/GlobalVariable.h"
51#include "llvm/IR/IRBuilder.h"
52#include "llvm/IR/InlineAsm.h"
53#include "llvm/IR/InstVisitor.h"
54#include "llvm/IR/InstrTypes.h"
55#include "llvm/IR/Instruction.h"
56#include "llvm/IR/Instructions.h"
57#include "llvm/IR/IntrinsicInst.h"
58#include "llvm/IR/Intrinsics.h"
59#include "llvm/IR/LLVMContext.h"
60#include "llvm/IR/MDBuilder.h"
61#include "llvm/IR/Metadata.h"
62#include "llvm/IR/Module.h"
63#include "llvm/IR/Type.h"
64#include "llvm/IR/Use.h"
65#include "llvm/IR/Value.h"
66#include "llvm/MC/MCSectionMachO.h"
67#include "llvm/Support/Casting.h"
68#include "llvm/Support/CommandLine.h"
69#include "llvm/Support/Debug.h"
70#include "llvm/Support/ErrorHandling.h"
71#include "llvm/Support/MathExtras.h"
72#include "llvm/Support/raw_ostream.h"
73#include "llvm/TargetParser/Triple.h"
74#include "llvm/Transforms/Instrumentation.h"
75#include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h"
76#include "llvm/Transforms/Instrumentation/AddressSanitizerOptions.h"
77#include "llvm/Transforms/Utils/ASanStackFrameLayout.h"
78#include "llvm/Transforms/Utils/BasicBlockUtils.h"
79#include "llvm/Transforms/Utils/Local.h"
80#include "llvm/Transforms/Utils/ModuleUtils.h"
81#include "llvm/Transforms/Utils/PromoteMemToReg.h"
82#include <algorithm>
83#include <cassert>
84#include <cstddef>
85#include <cstdint>
86#include <iomanip>
87#include <limits>
88#include <sstream>
89#include <string>
90#include <tuple>
91
92using namespace llvm;
93
94#define DEBUG_TYPE "asan"
95
96static const uint64_t kDefaultShadowScale = 3;
97static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
98static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
99static const uint64_t kDynamicShadowSentinel =
100 std::numeric_limits<uint64_t>::max();
101static const uint64_t kSmallX86_64ShadowOffsetBase = 0x7FFFFFFF; // < 2G.
102static const uint64_t kSmallX86_64ShadowOffsetAlignMask = ~0xFFFULL;
103static const uint64_t kLinuxKasan_ShadowOffset64 = 0xdffffc0000000000;
104static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 44;
105static const uint64_t kSystemZ_ShadowOffset64 = 1ULL << 52;
106static const uint64_t kMIPS_ShadowOffsetN32 = 1ULL << 29;
107static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa0000;
108static const uint64_t kMIPS64_ShadowOffset64 = 1ULL << 37;
109static const uint64_t kAArch64_ShadowOffset64 = 1ULL << 36;
110static const uint64_t kLoongArch64_ShadowOffset64 = 1ULL << 46;
111static const uint64_t kRISCV64_ShadowOffset64 = kDynamicShadowSentinel;
112static const uint64_t kFreeBSD_ShadowOffset32 = 1ULL << 30;
113static const uint64_t kFreeBSD_ShadowOffset64 = 1ULL << 46;
114static const uint64_t kFreeBSDAArch64_ShadowOffset64 = 1ULL << 47;
115static const uint64_t kFreeBSDKasan_ShadowOffset64 = 0xdffff7c000000000;
116static const uint64_t kNetBSD_ShadowOffset32 = 1ULL << 30;
117static const uint64_t kNetBSD_ShadowOffset64 = 1ULL << 46;
118static const uint64_t kNetBSDKasan_ShadowOffset64 = 0xdfff900000000000;
119static const uint64_t kPS_ShadowOffset64 = 1ULL << 40;
120static const uint64_t kWindowsShadowOffset32 = 3ULL << 28;
121static const uint64_t kEmscriptenShadowOffset = 0;
122
123// The shadow memory space is dynamically allocated.
124static const uint64_t kWindowsShadowOffset64 = kDynamicShadowSentinel;
125
126static const size_t kMinStackMallocSize = 1 << 6; // 64B
127static const size_t kMaxStackMallocSize = 1 << 16; // 64K
128static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
129static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
130
131const char kAsanModuleCtorName[] = "asan.module_ctor";
132const char kAsanModuleDtorName[] = "asan.module_dtor";
133static const uint64_t kAsanCtorAndDtorPriority = 1;
134// On Emscripten, the system needs more than one priorities for constructors.
135static const uint64_t kAsanEmscriptenCtorAndDtorPriority = 50;
136const char kAsanReportErrorTemplate[] = "__asan_report_";
137const char kAsanRegisterGlobalsName[] = "__asan_register_globals";
138const char kAsanUnregisterGlobalsName[] = "__asan_unregister_globals";
139const char kAsanRegisterImageGlobalsName[] = "__asan_register_image_globals";
140const char kAsanUnregisterImageGlobalsName[] =
141 "__asan_unregister_image_globals";
142const char kAsanRegisterElfGlobalsName[] = "__asan_register_elf_globals";
143const char kAsanUnregisterElfGlobalsName[] = "__asan_unregister_elf_globals";
144const char kAsanPoisonGlobalsName[] = "__asan_before_dynamic_init";
145const char kAsanUnpoisonGlobalsName[] = "__asan_after_dynamic_init";
146const char kAsanInitName[] = "__asan_init";
147const char kAsanVersionCheckNamePrefix[] = "__asan_version_mismatch_check_v";
148const char kAsanPtrCmp[] = "__sanitizer_ptr_cmp";
149const char kAsanPtrSub[] = "__sanitizer_ptr_sub";
150const char kAsanHandleNoReturnName[] = "__asan_handle_no_return";
151static const int kMaxAsanStackMallocSizeClass = 10;
152const char kAsanStackMallocNameTemplate[] = "__asan_stack_malloc_";
153const char kAsanStackMallocAlwaysNameTemplate[] =
154 "__asan_stack_malloc_always_";
155const char kAsanStackFreeNameTemplate[] = "__asan_stack_free_";
156const char kAsanGenPrefix[] = "___asan_gen_";
157const char kODRGenPrefix[] = "__odr_asan_gen_";
158const char kSanCovGenPrefix[] = "__sancov_gen_";
159const char kAsanSetShadowPrefix[] = "__asan_set_shadow_";
160const char kAsanPoisonStackMemoryName[] = "__asan_poison_stack_memory";
161const char kAsanUnpoisonStackMemoryName[] = "__asan_unpoison_stack_memory";
162
163// ASan version script has __asan_* wildcard. Triple underscore prevents a
164// linker (gold) warning about attempting to export a local symbol.
165const char kAsanGlobalsRegisteredFlagName[] = "___asan_globals_registered";
166
167const char kAsanOptionDetectUseAfterReturn[] =
168 "__asan_option_detect_stack_use_after_return";
169
170const char kAsanShadowMemoryDynamicAddress[] =
171 "__asan_shadow_memory_dynamic_address";
172
173const char kAsanAllocaPoison[] = "__asan_alloca_poison";
174const char kAsanAllocasUnpoison[] = "__asan_allocas_unpoison";
175
176const char kAMDGPUAddressSharedName[] = "llvm.amdgcn.is.shared";
177const char kAMDGPUAddressPrivateName[] = "llvm.amdgcn.is.private";
178const char kAMDGPUBallotName[] = "llvm.amdgcn.ballot.i64";
179const char kAMDGPUUnreachableName[] = "llvm.amdgcn.unreachable";
180
181// Accesses sizes are powers of two: 1, 2, 4, 8, 16.
182static const size_t kNumberOfAccessSizes = 5;
183
184static const uint64_t kAllocaRzSize = 32;
185
186// ASanAccessInfo implementation constants.
187constexpr size_t kCompileKernelShift = 0;
188constexpr size_t kCompileKernelMask = 0x1;
189constexpr size_t kAccessSizeIndexShift = 1;
190constexpr size_t kAccessSizeIndexMask = 0xf;
191constexpr size_t kIsWriteShift = 5;
192constexpr size_t kIsWriteMask = 0x1;
193
194// Command-line flags.
195
196static cl::opt<bool> ClEnableKasan(
197 "asan-kernel", cl::desc("Enable KernelAddressSanitizer instrumentation"),
198 cl::Hidden, cl::init(Val: false));
199
200static cl::opt<bool> ClRecover(
201 "asan-recover",
202 cl::desc("Enable recovery mode (continue-after-error)."),
203 cl::Hidden, cl::init(Val: false));
204
205static cl::opt<bool> ClInsertVersionCheck(
206 "asan-guard-against-version-mismatch",
207 cl::desc("Guard against compiler/runtime version mismatch."), cl::Hidden,
208 cl::init(Val: true));
209
210// This flag may need to be replaced with -f[no-]asan-reads.
211static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
212 cl::desc("instrument read instructions"),
213 cl::Hidden, cl::init(Val: true));
214
215static cl::opt<bool> ClInstrumentWrites(
216 "asan-instrument-writes", cl::desc("instrument write instructions"),
217 cl::Hidden, cl::init(Val: true));
218
219static cl::opt<bool>
220 ClUseStackSafety("asan-use-stack-safety", cl::Hidden, cl::init(Val: true),
221 cl::Hidden, cl::desc("Use Stack Safety analysis results"),
222 cl::Optional);
223
224static cl::opt<bool> ClInstrumentAtomics(
225 "asan-instrument-atomics",
226 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden,
227 cl::init(Val: true));
228
229static cl::opt<bool>
230 ClInstrumentByval("asan-instrument-byval",
231 cl::desc("instrument byval call arguments"), cl::Hidden,
232 cl::init(Val: true));
233
234static cl::opt<bool> ClAlwaysSlowPath(
235 "asan-always-slow-path",
236 cl::desc("use instrumentation with slow path for all accesses"), cl::Hidden,
237 cl::init(Val: false));
238
239static cl::opt<bool> ClForceDynamicShadow(
240 "asan-force-dynamic-shadow",
241 cl::desc("Load shadow address into a local variable for each function"),
242 cl::Hidden, cl::init(Val: false));
243
244static cl::opt<bool>
245 ClWithIfunc("asan-with-ifunc",
246 cl::desc("Access dynamic shadow through an ifunc global on "
247 "platforms that support this"),
248 cl::Hidden, cl::init(Val: true));
249
250static cl::opt<bool> ClWithIfuncSuppressRemat(
251 "asan-with-ifunc-suppress-remat",
252 cl::desc("Suppress rematerialization of dynamic shadow address by passing "
253 "it through inline asm in prologue."),
254 cl::Hidden, cl::init(Val: true));
255
256// This flag limits the number of instructions to be instrumented
257// in any given BB. Normally, this should be set to unlimited (INT_MAX),
258// but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
259// set it to 10000.
260static cl::opt<int> ClMaxInsnsToInstrumentPerBB(
261 "asan-max-ins-per-bb", cl::init(Val: 10000),
262 cl::desc("maximal number of instructions to instrument in any given BB"),
263 cl::Hidden);
264
265// This flag may need to be replaced with -f[no]asan-stack.
266static cl::opt<bool> ClStack("asan-stack", cl::desc("Handle stack memory"),
267 cl::Hidden, cl::init(Val: true));
268static cl::opt<uint32_t> ClMaxInlinePoisoningSize(
269 "asan-max-inline-poisoning-size",
270 cl::desc(
271 "Inline shadow poisoning for blocks up to the given size in bytes."),
272 cl::Hidden, cl::init(Val: 64));
273
274static cl::opt<AsanDetectStackUseAfterReturnMode> ClUseAfterReturn(
275 "asan-use-after-return",
276 cl::desc("Sets the mode of detection for stack-use-after-return."),
277 cl::values(
278 clEnumValN(AsanDetectStackUseAfterReturnMode::Never, "never",
279 "Never detect stack use after return."),
280 clEnumValN(
281 AsanDetectStackUseAfterReturnMode::Runtime, "runtime",
282 "Detect stack use after return if "
283 "binary flag 'ASAN_OPTIONS=detect_stack_use_after_return' is set."),
284 clEnumValN(AsanDetectStackUseAfterReturnMode::Always, "always",
285 "Always detect stack use after return.")),
286 cl::Hidden, cl::init(Val: AsanDetectStackUseAfterReturnMode::Runtime));
287
288static cl::opt<bool> ClRedzoneByvalArgs("asan-redzone-byval-args",
289 cl::desc("Create redzones for byval "
290 "arguments (extra copy "
291 "required)"), cl::Hidden,
292 cl::init(Val: true));
293
294static cl::opt<bool> ClUseAfterScope("asan-use-after-scope",
295 cl::desc("Check stack-use-after-scope"),
296 cl::Hidden, cl::init(Val: false));
297
298// This flag may need to be replaced with -f[no]asan-globals.
299static cl::opt<bool> ClGlobals("asan-globals",
300 cl::desc("Handle global objects"), cl::Hidden,
301 cl::init(Val: true));
302
303static cl::opt<bool> ClInitializers("asan-initialization-order",
304 cl::desc("Handle C++ initializer order"),
305 cl::Hidden, cl::init(Val: true));
306
307static cl::opt<bool> ClInvalidPointerPairs(
308 "asan-detect-invalid-pointer-pair",
309 cl::desc("Instrument <, <=, >, >=, - with pointer operands"), cl::Hidden,
310 cl::init(Val: false));
311
312static cl::opt<bool> ClInvalidPointerCmp(
313 "asan-detect-invalid-pointer-cmp",
314 cl::desc("Instrument <, <=, >, >= with pointer operands"), cl::Hidden,
315 cl::init(Val: false));
316
317static cl::opt<bool> ClInvalidPointerSub(
318 "asan-detect-invalid-pointer-sub",
319 cl::desc("Instrument - operations with pointer operands"), cl::Hidden,
320 cl::init(Val: false));
321
322static cl::opt<unsigned> ClRealignStack(
323 "asan-realign-stack",
324 cl::desc("Realign stack to the value of this flag (power of two)"),
325 cl::Hidden, cl::init(Val: 32));
326
327static cl::opt<int> ClInstrumentationWithCallsThreshold(
328 "asan-instrumentation-with-call-threshold",
329 cl::desc("If the function being instrumented contains more than "
330 "this number of memory accesses, use callbacks instead of "
331 "inline checks (-1 means never use callbacks)."),
332 cl::Hidden, cl::init(Val: 7000));
333
334static cl::opt<std::string> ClMemoryAccessCallbackPrefix(
335 "asan-memory-access-callback-prefix",
336 cl::desc("Prefix for memory access callbacks"), cl::Hidden,
337 cl::init(Val: "__asan_"));
338
339static cl::opt<bool> ClKasanMemIntrinCallbackPrefix(
340 "asan-kernel-mem-intrinsic-prefix",
341 cl::desc("Use prefix for memory intrinsics in KASAN mode"), cl::Hidden,
342 cl::init(Val: false));
343
344static cl::opt<bool>
345 ClInstrumentDynamicAllocas("asan-instrument-dynamic-allocas",
346 cl::desc("instrument dynamic allocas"),
347 cl::Hidden, cl::init(Val: true));
348
349static cl::opt<bool> ClSkipPromotableAllocas(
350 "asan-skip-promotable-allocas",
351 cl::desc("Do not instrument promotable allocas"), cl::Hidden,
352 cl::init(Val: true));
353
354static cl::opt<AsanCtorKind> ClConstructorKind(
355 "asan-constructor-kind",
356 cl::desc("Sets the ASan constructor kind"),
357 cl::values(clEnumValN(AsanCtorKind::None, "none", "No constructors"),
358 clEnumValN(AsanCtorKind::Global, "global",
359 "Use global constructors")),
360 cl::init(Val: AsanCtorKind::Global), cl::Hidden);
361// These flags allow to change the shadow mapping.
362// The shadow mapping looks like
363// Shadow = (Mem >> scale) + offset
364
365static cl::opt<int> ClMappingScale("asan-mapping-scale",
366 cl::desc("scale of asan shadow mapping"),
367 cl::Hidden, cl::init(Val: 0));
368
369static cl::opt<uint64_t>
370 ClMappingOffset("asan-mapping-offset",
371 cl::desc("offset of asan shadow mapping [EXPERIMENTAL]"),
372 cl::Hidden, cl::init(Val: 0));
373
374// Optimization flags. Not user visible, used mostly for testing
375// and benchmarking the tool.
376
377static cl::opt<bool> ClOpt("asan-opt", cl::desc("Optimize instrumentation"),
378 cl::Hidden, cl::init(Val: true));
379
380static cl::opt<bool> ClOptimizeCallbacks("asan-optimize-callbacks",
381 cl::desc("Optimize callbacks"),
382 cl::Hidden, cl::init(Val: false));
383
384static cl::opt<bool> ClOptSameTemp(
385 "asan-opt-same-temp", cl::desc("Instrument the same temp just once"),
386 cl::Hidden, cl::init(Val: true));
387
388static cl::opt<bool> ClOptGlobals("asan-opt-globals",
389 cl::desc("Don't instrument scalar globals"),
390 cl::Hidden, cl::init(Val: true));
391
392static cl::opt<bool> ClOptStack(
393 "asan-opt-stack", cl::desc("Don't instrument scalar stack variables"),
394 cl::Hidden, cl::init(Val: false));
395
396static cl::opt<bool> ClDynamicAllocaStack(
397 "asan-stack-dynamic-alloca",
398 cl::desc("Use dynamic alloca to represent stack variables"), cl::Hidden,
399 cl::init(Val: true));
400
401static cl::opt<uint32_t> ClForceExperiment(
402 "asan-force-experiment",
403 cl::desc("Force optimization experiment (for testing)"), cl::Hidden,
404 cl::init(Val: 0));
405
406static cl::opt<bool>
407 ClUsePrivateAlias("asan-use-private-alias",
408 cl::desc("Use private aliases for global variables"),
409 cl::Hidden, cl::init(Val: true));
410
411static cl::opt<bool>
412 ClUseOdrIndicator("asan-use-odr-indicator",
413 cl::desc("Use odr indicators to improve ODR reporting"),
414 cl::Hidden, cl::init(Val: true));
415
416static cl::opt<bool>
417 ClUseGlobalsGC("asan-globals-live-support",
418 cl::desc("Use linker features to support dead "
419 "code stripping of globals"),
420 cl::Hidden, cl::init(Val: true));
421
422// This is on by default even though there is a bug in gold:
423// https://sourceware.org/bugzilla/show_bug.cgi?id=19002
424static cl::opt<bool>
425 ClWithComdat("asan-with-comdat",
426 cl::desc("Place ASan constructors in comdat sections"),
427 cl::Hidden, cl::init(Val: true));
428
429static cl::opt<AsanDtorKind> ClOverrideDestructorKind(
430 "asan-destructor-kind",
431 cl::desc("Sets the ASan destructor kind. The default is to use the value "
432 "provided to the pass constructor"),
433 cl::values(clEnumValN(AsanDtorKind::None, "none", "No destructors"),
434 clEnumValN(AsanDtorKind::Global, "global",
435 "Use global destructors")),
436 cl::init(Val: AsanDtorKind::Invalid), cl::Hidden);
437
438// Debug flags.
439
440static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
441 cl::init(Val: 0));
442
443static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
444 cl::Hidden, cl::init(Val: 0));
445
446static cl::opt<std::string> ClDebugFunc("asan-debug-func", cl::Hidden,
447 cl::desc("Debug func"));
448
449static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
450 cl::Hidden, cl::init(Val: -1));
451
452static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug max inst"),
453 cl::Hidden, cl::init(Val: -1));
454
455STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
456STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
457STATISTIC(NumOptimizedAccessesToGlobalVar,
458 "Number of optimized accesses to global vars");
459STATISTIC(NumOptimizedAccessesToStackVar,
460 "Number of optimized accesses to stack vars");
461
462namespace {
463
464/// This struct defines the shadow mapping using the rule:
465/// shadow = (mem >> Scale) ADD-or-OR Offset.
466/// If InGlobal is true, then
467/// extern char __asan_shadow[];
468/// shadow = (mem >> Scale) + &__asan_shadow
469struct ShadowMapping {
470 int Scale;
471 uint64_t Offset;
472 bool OrShadowOffset;
473 bool InGlobal;
474};
475
476} // end anonymous namespace
477
478static ShadowMapping getShadowMapping(const Triple &TargetTriple, int LongSize,
479 bool IsKasan) {
480 bool IsAndroid = TargetTriple.isAndroid();
481 bool IsIOS = TargetTriple.isiOS() || TargetTriple.isWatchOS() ||
482 TargetTriple.isDriverKit();
483 bool IsMacOS = TargetTriple.isMacOSX();
484 bool IsFreeBSD = TargetTriple.isOSFreeBSD();
485 bool IsNetBSD = TargetTriple.isOSNetBSD();
486 bool IsPS = TargetTriple.isPS();
487 bool IsLinux = TargetTriple.isOSLinux();
488 bool IsPPC64 = TargetTriple.getArch() == Triple::ppc64 ||
489 TargetTriple.getArch() == Triple::ppc64le;
490 bool IsSystemZ = TargetTriple.getArch() == Triple::systemz;
491 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64;
492 bool IsMIPSN32ABI = TargetTriple.getEnvironment() == Triple::GNUABIN32;
493 bool IsMIPS32 = TargetTriple.isMIPS32();
494 bool IsMIPS64 = TargetTriple.isMIPS64();
495 bool IsArmOrThumb = TargetTriple.isARM() || TargetTriple.isThumb();
496 bool IsAArch64 = TargetTriple.getArch() == Triple::aarch64 ||
497 TargetTriple.getArch() == Triple::aarch64_be;
498 bool IsLoongArch64 = TargetTriple.isLoongArch64();
499 bool IsRISCV64 = TargetTriple.getArch() == Triple::riscv64;
500 bool IsWindows = TargetTriple.isOSWindows();
501 bool IsFuchsia = TargetTriple.isOSFuchsia();
502 bool IsEmscripten = TargetTriple.isOSEmscripten();
503 bool IsAMDGPU = TargetTriple.isAMDGPU();
504
505 ShadowMapping Mapping;
506
507 Mapping.Scale = kDefaultShadowScale;
508 if (ClMappingScale.getNumOccurrences() > 0) {
509 Mapping.Scale = ClMappingScale;
510 }
511
512 if (LongSize == 32) {
513 if (IsAndroid)
514 Mapping.Offset = kDynamicShadowSentinel;
515 else if (IsMIPSN32ABI)
516 Mapping.Offset = kMIPS_ShadowOffsetN32;
517 else if (IsMIPS32)
518 Mapping.Offset = kMIPS32_ShadowOffset32;
519 else if (IsFreeBSD)
520 Mapping.Offset = kFreeBSD_ShadowOffset32;
521 else if (IsNetBSD)
522 Mapping.Offset = kNetBSD_ShadowOffset32;
523 else if (IsIOS)
524 Mapping.Offset = kDynamicShadowSentinel;
525 else if (IsWindows)
526 Mapping.Offset = kWindowsShadowOffset32;
527 else if (IsEmscripten)
528 Mapping.Offset = kEmscriptenShadowOffset;
529 else
530 Mapping.Offset = kDefaultShadowOffset32;
531 } else { // LongSize == 64
532 // Fuchsia is always PIE, which means that the beginning of the address
533 // space is always available.
534 if (IsFuchsia)
535 Mapping.Offset = 0;
536 else if (IsPPC64)
537 Mapping.Offset = kPPC64_ShadowOffset64;
538 else if (IsSystemZ)
539 Mapping.Offset = kSystemZ_ShadowOffset64;
540 else if (IsFreeBSD && IsAArch64)
541 Mapping.Offset = kFreeBSDAArch64_ShadowOffset64;
542 else if (IsFreeBSD && !IsMIPS64) {
543 if (IsKasan)
544 Mapping.Offset = kFreeBSDKasan_ShadowOffset64;
545 else
546 Mapping.Offset = kFreeBSD_ShadowOffset64;
547 } else if (IsNetBSD) {
548 if (IsKasan)
549 Mapping.Offset = kNetBSDKasan_ShadowOffset64;
550 else
551 Mapping.Offset = kNetBSD_ShadowOffset64;
552 } else if (IsPS)
553 Mapping.Offset = kPS_ShadowOffset64;
554 else if (IsLinux && IsX86_64) {
555 if (IsKasan)
556 Mapping.Offset = kLinuxKasan_ShadowOffset64;
557 else
558 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
559 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
560 } else if (IsWindows && IsX86_64) {
561 Mapping.Offset = kWindowsShadowOffset64;
562 } else if (IsMIPS64)
563 Mapping.Offset = kMIPS64_ShadowOffset64;
564 else if (IsIOS)
565 Mapping.Offset = kDynamicShadowSentinel;
566 else if (IsMacOS && IsAArch64)
567 Mapping.Offset = kDynamicShadowSentinel;
568 else if (IsAArch64)
569 Mapping.Offset = kAArch64_ShadowOffset64;
570 else if (IsLoongArch64)
571 Mapping.Offset = kLoongArch64_ShadowOffset64;
572 else if (IsRISCV64)
573 Mapping.Offset = kRISCV64_ShadowOffset64;
574 else if (IsAMDGPU)
575 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
576 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
577 else
578 Mapping.Offset = kDefaultShadowOffset64;
579 }
580
581 if (ClForceDynamicShadow) {
582 Mapping.Offset = kDynamicShadowSentinel;
583 }
584
585 if (ClMappingOffset.getNumOccurrences() > 0) {
586 Mapping.Offset = ClMappingOffset;
587 }
588
589 // OR-ing shadow offset if more efficient (at least on x86) if the offset
590 // is a power of two, but on ppc64 and loongarch64 we have to use add since
591 // the shadow offset is not necessarily 1/8-th of the address space. On
592 // SystemZ, we could OR the constant in a single instruction, but it's more
593 // efficient to load it once and use indexed addressing.
594 Mapping.OrShadowOffset = !IsAArch64 && !IsPPC64 && !IsSystemZ && !IsPS &&
595 !IsRISCV64 && !IsLoongArch64 &&
596 !(Mapping.Offset & (Mapping.Offset - 1)) &&
597 Mapping.Offset != kDynamicShadowSentinel;
598 bool IsAndroidWithIfuncSupport =
599 IsAndroid && !TargetTriple.isAndroidVersionLT(Major: 21);
600 Mapping.InGlobal = ClWithIfunc && IsAndroidWithIfuncSupport && IsArmOrThumb;
601
602 return Mapping;
603}
604
605namespace llvm {
606void getAddressSanitizerParams(const Triple &TargetTriple, int LongSize,
607 bool IsKasan, uint64_t *ShadowBase,
608 int *MappingScale, bool *OrShadowOffset) {
609 auto Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan);
610 *ShadowBase = Mapping.Offset;
611 *MappingScale = Mapping.Scale;
612 *OrShadowOffset = Mapping.OrShadowOffset;
613}
614
615ASanAccessInfo::ASanAccessInfo(int32_t Packed)
616 : Packed(Packed),
617 AccessSizeIndex((Packed >> kAccessSizeIndexShift) & kAccessSizeIndexMask),
618 IsWrite((Packed >> kIsWriteShift) & kIsWriteMask),
619 CompileKernel((Packed >> kCompileKernelShift) & kCompileKernelMask) {}
620
621ASanAccessInfo::ASanAccessInfo(bool IsWrite, bool CompileKernel,
622 uint8_t AccessSizeIndex)
623 : Packed((IsWrite << kIsWriteShift) +
624 (CompileKernel << kCompileKernelShift) +
625 (AccessSizeIndex << kAccessSizeIndexShift)),
626 AccessSizeIndex(AccessSizeIndex), IsWrite(IsWrite),
627 CompileKernel(CompileKernel) {}
628
629} // namespace llvm
630
631static uint64_t getRedzoneSizeForScale(int MappingScale) {
632 // Redzone used for stack and globals is at least 32 bytes.
633 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
634 return std::max(a: 32U, b: 1U << MappingScale);
635}
636
637static uint64_t GetCtorAndDtorPriority(Triple &TargetTriple) {
638 if (TargetTriple.isOSEmscripten()) {
639 return kAsanEmscriptenCtorAndDtorPriority;
640 } else {
641 return kAsanCtorAndDtorPriority;
642 }
643}
644
645namespace {
646/// Helper RAII class to post-process inserted asan runtime calls during a
647/// pass on a single Function. Upon end of scope, detects and applies the
648/// required funclet OpBundle.
649class RuntimeCallInserter {
650 Function *OwnerFn = nullptr;
651 bool TrackInsertedCalls = false;
652 SmallVector<CallInst *> InsertedCalls;
653
654public:
655 RuntimeCallInserter(Function &Fn) : OwnerFn(&Fn) {
656 if (Fn.hasPersonalityFn()) {
657 auto Personality = classifyEHPersonality(Pers: Fn.getPersonalityFn());
658 if (isScopedEHPersonality(Pers: Personality))
659 TrackInsertedCalls = true;
660 }
661 }
662
663 ~RuntimeCallInserter() {
664 if (InsertedCalls.empty())
665 return;
666 assert(TrackInsertedCalls && "Calls were wrongly tracked");
667
668 DenseMap<BasicBlock *, ColorVector> BlockColors = colorEHFunclets(F&: *OwnerFn);
669 for (CallInst *CI : InsertedCalls) {
670 BasicBlock *BB = CI->getParent();
671 assert(BB && "Instruction doesn't belong to a BasicBlock");
672 assert(BB->getParent() == OwnerFn &&
673 "Instruction doesn't belong to the expected Function!");
674
675 ColorVector &Colors = BlockColors[BB];
676 // funclet opbundles are only valid in monochromatic BBs.
677 // Note that unreachable BBs are seen as colorless by colorEHFunclets()
678 // and will be DCE'ed later.
679 if (Colors.empty())
680 continue;
681 if (Colors.size() != 1) {
682 OwnerFn->getContext().emitError(
683 ErrorStr: "Instruction's BasicBlock is not monochromatic");
684 continue;
685 }
686
687 BasicBlock *Color = Colors.front();
688 Instruction *EHPad = Color->getFirstNonPHI();
689
690 if (EHPad && EHPad->isEHPad()) {
691 // Replace CI with a clone with an added funclet OperandBundle
692 OperandBundleDef OB("funclet", EHPad);
693 auto *NewCall =
694 CallBase::addOperandBundle(CB: CI, ID: LLVMContext::OB_funclet, OB, InsertPt: CI);
695 NewCall->copyMetadata(SrcInst: *CI);
696 CI->replaceAllUsesWith(V: NewCall);
697 CI->eraseFromParent();
698 }
699 }
700 }
701
702 CallInst *createRuntimeCall(IRBuilder<> &IRB, FunctionCallee Callee,
703 ArrayRef<Value *> Args = {},
704 const Twine &Name = "") {
705 assert(IRB.GetInsertBlock()->getParent() == OwnerFn);
706
707 CallInst *Inst = IRB.CreateCall(Callee, Args, Name, FPMathTag: nullptr);
708 if (TrackInsertedCalls)
709 InsertedCalls.push_back(Elt: Inst);
710 return Inst;
711 }
712};
713
714/// AddressSanitizer: instrument the code in module to find memory bugs.
715struct AddressSanitizer {
716 AddressSanitizer(Module &M, const StackSafetyGlobalInfo *SSGI,
717 int InstrumentationWithCallsThreshold,
718 uint32_t MaxInlinePoisoningSize, bool CompileKernel = false,
719 bool Recover = false, bool UseAfterScope = false,
720 AsanDetectStackUseAfterReturnMode UseAfterReturn =
721 AsanDetectStackUseAfterReturnMode::Runtime)
722 : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
723 : CompileKernel),
724 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
725 UseAfterScope(UseAfterScope || ClUseAfterScope),
726 UseAfterReturn(ClUseAfterReturn.getNumOccurrences() ? ClUseAfterReturn
727 : UseAfterReturn),
728 SSGI(SSGI),
729 InstrumentationWithCallsThreshold(
730 ClInstrumentationWithCallsThreshold.getNumOccurrences() > 0
731 ? ClInstrumentationWithCallsThreshold
732 : InstrumentationWithCallsThreshold),
733 MaxInlinePoisoningSize(ClMaxInlinePoisoningSize.getNumOccurrences() > 0
734 ? ClMaxInlinePoisoningSize
735 : MaxInlinePoisoningSize) {
736 C = &(M.getContext());
737 DL = &M.getDataLayout();
738 LongSize = M.getDataLayout().getPointerSizeInBits();
739 IntptrTy = Type::getIntNTy(C&: *C, N: LongSize);
740 PtrTy = PointerType::getUnqual(C&: *C);
741 Int32Ty = Type::getInt32Ty(C&: *C);
742 TargetTriple = Triple(M.getTargetTriple());
743
744 Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan: this->CompileKernel);
745
746 assert(this->UseAfterReturn != AsanDetectStackUseAfterReturnMode::Invalid);
747 }
748
749 TypeSize getAllocaSizeInBytes(const AllocaInst &AI) const {
750 return *AI.getAllocationSize(DL: AI.getDataLayout());
751 }
752
753 /// Check if we want (and can) handle this alloca.
754 bool isInterestingAlloca(const AllocaInst &AI);
755
756 bool ignoreAccess(Instruction *Inst, Value *Ptr);
757 void getInterestingMemoryOperands(
758 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting);
759
760 void instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
761 InterestingMemoryOperand &O, bool UseCalls,
762 const DataLayout &DL, RuntimeCallInserter &RTCI);
763 void instrumentPointerComparisonOrSubtraction(Instruction *I,
764 RuntimeCallInserter &RTCI);
765 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
766 Value *Addr, MaybeAlign Alignment,
767 uint32_t TypeStoreSize, bool IsWrite,
768 Value *SizeArgument, bool UseCalls, uint32_t Exp,
769 RuntimeCallInserter &RTCI);
770 Instruction *instrumentAMDGPUAddress(Instruction *OrigIns,
771 Instruction *InsertBefore, Value *Addr,
772 uint32_t TypeStoreSize, bool IsWrite,
773 Value *SizeArgument);
774 Instruction *genAMDGPUReportBlock(IRBuilder<> &IRB, Value *Cond,
775 bool Recover);
776 void instrumentUnusualSizeOrAlignment(Instruction *I,
777 Instruction *InsertBefore, Value *Addr,
778 TypeSize TypeStoreSize, bool IsWrite,
779 Value *SizeArgument, bool UseCalls,
780 uint32_t Exp,
781 RuntimeCallInserter &RTCI);
782 void instrumentMaskedLoadOrStore(AddressSanitizer *Pass, const DataLayout &DL,
783 Type *IntptrTy, Value *Mask, Value *EVL,
784 Value *Stride, Instruction *I, Value *Addr,
785 MaybeAlign Alignment, unsigned Granularity,
786 Type *OpType, bool IsWrite,
787 Value *SizeArgument, bool UseCalls,
788 uint32_t Exp, RuntimeCallInserter &RTCI);
789 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
790 Value *ShadowValue, uint32_t TypeStoreSize);
791 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
792 bool IsWrite, size_t AccessSizeIndex,
793 Value *SizeArgument, uint32_t Exp,
794 RuntimeCallInserter &RTCI);
795 void instrumentMemIntrinsic(MemIntrinsic *MI, RuntimeCallInserter &RTCI);
796 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
797 bool suppressInstrumentationSiteForDebug(int &Instrumented);
798 bool instrumentFunction(Function &F, const TargetLibraryInfo *TLI);
799 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
800 bool maybeInsertDynamicShadowAtFunctionEntry(Function &F);
801 void markEscapedLocalAllocas(Function &F);
802
803private:
804 friend struct FunctionStackPoisoner;
805
806 void initializeCallbacks(Module &M, const TargetLibraryInfo *TLI);
807
808 bool LooksLikeCodeInBug11395(Instruction *I);
809 bool GlobalIsLinkerInitialized(GlobalVariable *G);
810 bool isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, Value *Addr,
811 TypeSize TypeStoreSize) const;
812
813 /// Helper to cleanup per-function state.
814 struct FunctionStateRAII {
815 AddressSanitizer *Pass;
816
817 FunctionStateRAII(AddressSanitizer *Pass) : Pass(Pass) {
818 assert(Pass->ProcessedAllocas.empty() &&
819 "last pass forgot to clear cache");
820 assert(!Pass->LocalDynamicShadow);
821 }
822
823 ~FunctionStateRAII() {
824 Pass->LocalDynamicShadow = nullptr;
825 Pass->ProcessedAllocas.clear();
826 }
827 };
828
829 LLVMContext *C;
830 const DataLayout *DL;
831 Triple TargetTriple;
832 int LongSize;
833 bool CompileKernel;
834 bool Recover;
835 bool UseAfterScope;
836 AsanDetectStackUseAfterReturnMode UseAfterReturn;
837 Type *IntptrTy;
838 Type *Int32Ty;
839 PointerType *PtrTy;
840 ShadowMapping Mapping;
841 FunctionCallee AsanHandleNoReturnFunc;
842 FunctionCallee AsanPtrCmpFunction, AsanPtrSubFunction;
843 Constant *AsanShadowGlobal;
844
845 // These arrays is indexed by AccessIsWrite, Experiment and log2(AccessSize).
846 FunctionCallee AsanErrorCallback[2][2][kNumberOfAccessSizes];
847 FunctionCallee AsanMemoryAccessCallback[2][2][kNumberOfAccessSizes];
848
849 // These arrays is indexed by AccessIsWrite and Experiment.
850 FunctionCallee AsanErrorCallbackSized[2][2];
851 FunctionCallee AsanMemoryAccessCallbackSized[2][2];
852
853 FunctionCallee AsanMemmove, AsanMemcpy, AsanMemset;
854 Value *LocalDynamicShadow = nullptr;
855 const StackSafetyGlobalInfo *SSGI;
856 DenseMap<const AllocaInst *, bool> ProcessedAllocas;
857
858 FunctionCallee AMDGPUAddressShared;
859 FunctionCallee AMDGPUAddressPrivate;
860 int InstrumentationWithCallsThreshold;
861 uint32_t MaxInlinePoisoningSize;
862};
863
864class ModuleAddressSanitizer {
865public:
866 ModuleAddressSanitizer(Module &M, bool InsertVersionCheck,
867 bool CompileKernel = false, bool Recover = false,
868 bool UseGlobalsGC = true, bool UseOdrIndicator = true,
869 AsanDtorKind DestructorKind = AsanDtorKind::Global,
870 AsanCtorKind ConstructorKind = AsanCtorKind::Global)
871 : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
872 : CompileKernel),
873 InsertVersionCheck(ClInsertVersionCheck.getNumOccurrences() > 0
874 ? ClInsertVersionCheck
875 : InsertVersionCheck),
876 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
877 UseGlobalsGC(UseGlobalsGC && ClUseGlobalsGC && !this->CompileKernel),
878 // Enable aliases as they should have no downside with ODR indicators.
879 UsePrivateAlias(ClUsePrivateAlias.getNumOccurrences() > 0
880 ? ClUsePrivateAlias
881 : UseOdrIndicator),
882 UseOdrIndicator(ClUseOdrIndicator.getNumOccurrences() > 0
883 ? ClUseOdrIndicator
884 : UseOdrIndicator),
885 // Not a typo: ClWithComdat is almost completely pointless without
886 // ClUseGlobalsGC (because then it only works on modules without
887 // globals, which are rare); it is a prerequisite for ClUseGlobalsGC;
888 // and both suffer from gold PR19002 for which UseGlobalsGC constructor
889 // argument is designed as workaround. Therefore, disable both
890 // ClWithComdat and ClUseGlobalsGC unless the frontend says it's ok to
891 // do globals-gc.
892 UseCtorComdat(UseGlobalsGC && ClWithComdat && !this->CompileKernel),
893 DestructorKind(DestructorKind),
894 ConstructorKind(ClConstructorKind.getNumOccurrences() > 0
895 ? ClConstructorKind
896 : ConstructorKind) {
897 C = &(M.getContext());
898 int LongSize = M.getDataLayout().getPointerSizeInBits();
899 IntptrTy = Type::getIntNTy(C&: *C, N: LongSize);
900 PtrTy = PointerType::getUnqual(C&: *C);
901 TargetTriple = Triple(M.getTargetTriple());
902 Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan: this->CompileKernel);
903
904 if (ClOverrideDestructorKind != AsanDtorKind::Invalid)
905 this->DestructorKind = ClOverrideDestructorKind;
906 assert(this->DestructorKind != AsanDtorKind::Invalid);
907 }
908
909 bool instrumentModule(Module &);
910
911private:
912 void initializeCallbacks(Module &M);
913
914 void instrumentGlobals(IRBuilder<> &IRB, Module &M, bool *CtorComdat);
915 void InstrumentGlobalsCOFF(IRBuilder<> &IRB, Module &M,
916 ArrayRef<GlobalVariable *> ExtendedGlobals,
917 ArrayRef<Constant *> MetadataInitializers);
918 void instrumentGlobalsELF(IRBuilder<> &IRB, Module &M,
919 ArrayRef<GlobalVariable *> ExtendedGlobals,
920 ArrayRef<Constant *> MetadataInitializers,
921 const std::string &UniqueModuleId);
922 void InstrumentGlobalsMachO(IRBuilder<> &IRB, Module &M,
923 ArrayRef<GlobalVariable *> ExtendedGlobals,
924 ArrayRef<Constant *> MetadataInitializers);
925 void
926 InstrumentGlobalsWithMetadataArray(IRBuilder<> &IRB, Module &M,
927 ArrayRef<GlobalVariable *> ExtendedGlobals,
928 ArrayRef<Constant *> MetadataInitializers);
929
930 GlobalVariable *CreateMetadataGlobal(Module &M, Constant *Initializer,
931 StringRef OriginalName);
932 void SetComdatForGlobalMetadata(GlobalVariable *G, GlobalVariable *Metadata,
933 StringRef InternalSuffix);
934 Instruction *CreateAsanModuleDtor(Module &M);
935
936 const GlobalVariable *getExcludedAliasedGlobal(const GlobalAlias &GA) const;
937 bool shouldInstrumentGlobal(GlobalVariable *G) const;
938 bool ShouldUseMachOGlobalsSection() const;
939 StringRef getGlobalMetadataSection() const;
940 void poisonOneInitializer(Function &GlobalInit, GlobalValue *ModuleName);
941 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
942 uint64_t getMinRedzoneSizeForGlobal() const {
943 return getRedzoneSizeForScale(MappingScale: Mapping.Scale);
944 }
945 uint64_t getRedzoneSizeForGlobal(uint64_t SizeInBytes) const;
946 int GetAsanVersion(const Module &M) const;
947
948 bool CompileKernel;
949 bool InsertVersionCheck;
950 bool Recover;
951 bool UseGlobalsGC;
952 bool UsePrivateAlias;
953 bool UseOdrIndicator;
954 bool UseCtorComdat;
955 AsanDtorKind DestructorKind;
956 AsanCtorKind ConstructorKind;
957 Type *IntptrTy;
958 PointerType *PtrTy;
959 LLVMContext *C;
960 Triple TargetTriple;
961 ShadowMapping Mapping;
962 FunctionCallee AsanPoisonGlobals;
963 FunctionCallee AsanUnpoisonGlobals;
964 FunctionCallee AsanRegisterGlobals;
965 FunctionCallee AsanUnregisterGlobals;
966 FunctionCallee AsanRegisterImageGlobals;
967 FunctionCallee AsanUnregisterImageGlobals;
968 FunctionCallee AsanRegisterElfGlobals;
969 FunctionCallee AsanUnregisterElfGlobals;
970
971 Function *AsanCtorFunction = nullptr;
972 Function *AsanDtorFunction = nullptr;
973};
974
975// Stack poisoning does not play well with exception handling.
976// When an exception is thrown, we essentially bypass the code
977// that unpoisones the stack. This is why the run-time library has
978// to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
979// stack in the interceptor. This however does not work inside the
980// actual function which catches the exception. Most likely because the
981// compiler hoists the load of the shadow value somewhere too high.
982// This causes asan to report a non-existing bug on 453.povray.
983// It sounds like an LLVM bug.
984struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
985 Function &F;
986 AddressSanitizer &ASan;
987 RuntimeCallInserter &RTCI;
988 DIBuilder DIB;
989 LLVMContext *C;
990 Type *IntptrTy;
991 Type *IntptrPtrTy;
992 ShadowMapping Mapping;
993
994 SmallVector<AllocaInst *, 16> AllocaVec;
995 SmallVector<AllocaInst *, 16> StaticAllocasToMoveUp;
996 SmallVector<Instruction *, 8> RetVec;
997
998 FunctionCallee AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1],
999 AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1];
1000 FunctionCallee AsanSetShadowFunc[0x100] = {};
1001 FunctionCallee AsanPoisonStackMemoryFunc, AsanUnpoisonStackMemoryFunc;
1002 FunctionCallee AsanAllocaPoisonFunc, AsanAllocasUnpoisonFunc;
1003
1004 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
1005 struct AllocaPoisonCall {
1006 IntrinsicInst *InsBefore;
1007 AllocaInst *AI;
1008 uint64_t Size;
1009 bool DoPoison;
1010 };
1011 SmallVector<AllocaPoisonCall, 8> DynamicAllocaPoisonCallVec;
1012 SmallVector<AllocaPoisonCall, 8> StaticAllocaPoisonCallVec;
1013 bool HasUntracedLifetimeIntrinsic = false;
1014
1015 SmallVector<AllocaInst *, 1> DynamicAllocaVec;
1016 SmallVector<IntrinsicInst *, 1> StackRestoreVec;
1017 AllocaInst *DynamicAllocaLayout = nullptr;
1018 IntrinsicInst *LocalEscapeCall = nullptr;
1019
1020 bool HasInlineAsm = false;
1021 bool HasReturnsTwiceCall = false;
1022 bool PoisonStack;
1023
1024 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan,
1025 RuntimeCallInserter &RTCI)
1026 : F(F), ASan(ASan), RTCI(RTCI),
1027 DIB(*F.getParent(), /*AllowUnresolved*/ false), C(ASan.C),
1028 IntptrTy(ASan.IntptrTy), IntptrPtrTy(PointerType::get(ElementType: IntptrTy, AddressSpace: 0)),
1029 Mapping(ASan.Mapping),
1030 PoisonStack(ClStack &&
1031 !Triple(F.getParent()->getTargetTriple()).isAMDGPU()) {}
1032
1033 bool runOnFunction() {
1034 if (!PoisonStack)
1035 return false;
1036
1037 if (ClRedzoneByvalArgs)
1038 copyArgsPassedByValToAllocas();
1039
1040 // Collect alloca, ret, lifetime instructions etc.
1041 for (BasicBlock *BB : depth_first(G: &F.getEntryBlock())) visit(BB&: *BB);
1042
1043 if (AllocaVec.empty() && DynamicAllocaVec.empty()) return false;
1044
1045 initializeCallbacks(M&: *F.getParent());
1046
1047 if (HasUntracedLifetimeIntrinsic) {
1048 // If there are lifetime intrinsics which couldn't be traced back to an
1049 // alloca, we may not know exactly when a variable enters scope, and
1050 // therefore should "fail safe" by not poisoning them.
1051 StaticAllocaPoisonCallVec.clear();
1052 DynamicAllocaPoisonCallVec.clear();
1053 }
1054
1055 processDynamicAllocas();
1056 processStaticAllocas();
1057
1058 if (ClDebugStack) {
1059 LLVM_DEBUG(dbgs() << F);
1060 }
1061 return true;
1062 }
1063
1064 // Arguments marked with the "byval" attribute are implicitly copied without
1065 // using an alloca instruction. To produce redzones for those arguments, we
1066 // copy them a second time into memory allocated with an alloca instruction.
1067 void copyArgsPassedByValToAllocas();
1068
1069 // Finds all Alloca instructions and puts
1070 // poisoned red zones around all of them.
1071 // Then unpoison everything back before the function returns.
1072 void processStaticAllocas();
1073 void processDynamicAllocas();
1074
1075 void createDynamicAllocasInitStorage();
1076
1077 // ----------------------- Visitors.
1078 /// Collect all Ret instructions, or the musttail call instruction if it
1079 /// precedes the return instruction.
1080 void visitReturnInst(ReturnInst &RI) {
1081 if (CallInst *CI = RI.getParent()->getTerminatingMustTailCall())
1082 RetVec.push_back(Elt: CI);
1083 else
1084 RetVec.push_back(Elt: &RI);
1085 }
1086
1087 /// Collect all Resume instructions.
1088 void visitResumeInst(ResumeInst &RI) { RetVec.push_back(Elt: &RI); }
1089
1090 /// Collect all CatchReturnInst instructions.
1091 void visitCleanupReturnInst(CleanupReturnInst &CRI) { RetVec.push_back(Elt: &CRI); }
1092
1093 void unpoisonDynamicAllocasBeforeInst(Instruction *InstBefore,
1094 Value *SavedStack) {
1095 IRBuilder<> IRB(InstBefore);
1096 Value *DynamicAreaPtr = IRB.CreatePtrToInt(V: SavedStack, DestTy: IntptrTy);
1097 // When we insert _asan_allocas_unpoison before @llvm.stackrestore, we
1098 // need to adjust extracted SP to compute the address of the most recent
1099 // alloca. We have a special @llvm.get.dynamic.area.offset intrinsic for
1100 // this purpose.
1101 if (!isa<ReturnInst>(Val: InstBefore)) {
1102 Function *DynamicAreaOffsetFunc = Intrinsic::getDeclaration(
1103 M: InstBefore->getModule(), id: Intrinsic::get_dynamic_area_offset,
1104 Tys: {IntptrTy});
1105
1106 Value *DynamicAreaOffset = IRB.CreateCall(Callee: DynamicAreaOffsetFunc, Args: {});
1107
1108 DynamicAreaPtr = IRB.CreateAdd(LHS: IRB.CreatePtrToInt(V: SavedStack, DestTy: IntptrTy),
1109 RHS: DynamicAreaOffset);
1110 }
1111
1112 RTCI.createRuntimeCall(
1113 IRB, Callee: AsanAllocasUnpoisonFunc,
1114 Args: {IRB.CreateLoad(Ty: IntptrTy, Ptr: DynamicAllocaLayout), DynamicAreaPtr});
1115 }
1116
1117 // Unpoison dynamic allocas redzones.
1118 void unpoisonDynamicAllocas() {
1119 for (Instruction *Ret : RetVec)
1120 unpoisonDynamicAllocasBeforeInst(InstBefore: Ret, SavedStack: DynamicAllocaLayout);
1121
1122 for (Instruction *StackRestoreInst : StackRestoreVec)
1123 unpoisonDynamicAllocasBeforeInst(InstBefore: StackRestoreInst,
1124 SavedStack: StackRestoreInst->getOperand(i: 0));
1125 }
1126
1127 // Deploy and poison redzones around dynamic alloca call. To do this, we
1128 // should replace this call with another one with changed parameters and
1129 // replace all its uses with new address, so
1130 // addr = alloca type, old_size, align
1131 // is replaced by
1132 // new_size = (old_size + additional_size) * sizeof(type)
1133 // tmp = alloca i8, new_size, max(align, 32)
1134 // addr = tmp + 32 (first 32 bytes are for the left redzone).
1135 // Additional_size is added to make new memory allocation contain not only
1136 // requested memory, but also left, partial and right redzones.
1137 void handleDynamicAllocaCall(AllocaInst *AI);
1138
1139 /// Collect Alloca instructions we want (and can) handle.
1140 void visitAllocaInst(AllocaInst &AI) {
1141 // FIXME: Handle scalable vectors instead of ignoring them.
1142 const Type *AllocaType = AI.getAllocatedType();
1143 const auto *STy = dyn_cast<StructType>(Val: AllocaType);
1144 if (!ASan.isInterestingAlloca(AI) || isa<ScalableVectorType>(Val: AllocaType) ||
1145 (STy && STy->containsHomogeneousScalableVectorTypes())) {
1146 if (AI.isStaticAlloca()) {
1147 // Skip over allocas that are present *before* the first instrumented
1148 // alloca, we don't want to move those around.
1149 if (AllocaVec.empty())
1150 return;
1151
1152 StaticAllocasToMoveUp.push_back(Elt: &AI);
1153 }
1154 return;
1155 }
1156
1157 if (!AI.isStaticAlloca())
1158 DynamicAllocaVec.push_back(Elt: &AI);
1159 else
1160 AllocaVec.push_back(Elt: &AI);
1161 }
1162
1163 /// Collect lifetime intrinsic calls to check for use-after-scope
1164 /// errors.
1165 void visitIntrinsicInst(IntrinsicInst &II) {
1166 Intrinsic::ID ID = II.getIntrinsicID();
1167 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(Elt: &II);
1168 if (ID == Intrinsic::localescape) LocalEscapeCall = &II;
1169 if (!ASan.UseAfterScope)
1170 return;
1171 if (!II.isLifetimeStartOrEnd())
1172 return;
1173 // Found lifetime intrinsic, add ASan instrumentation if necessary.
1174 auto *Size = cast<ConstantInt>(Val: II.getArgOperand(i: 0));
1175 // If size argument is undefined, don't do anything.
1176 if (Size->isMinusOne()) return;
1177 // Check that size doesn't saturate uint64_t and can
1178 // be stored in IntptrTy.
1179 const uint64_t SizeValue = Size->getValue().getLimitedValue();
1180 if (SizeValue == ~0ULL ||
1181 !ConstantInt::isValueValidForType(Ty: IntptrTy, V: SizeValue))
1182 return;
1183 // Find alloca instruction that corresponds to llvm.lifetime argument.
1184 // Currently we can only handle lifetime markers pointing to the
1185 // beginning of the alloca.
1186 AllocaInst *AI = findAllocaForValue(V: II.getArgOperand(i: 1), OffsetZero: true);
1187 if (!AI) {
1188 HasUntracedLifetimeIntrinsic = true;
1189 return;
1190 }
1191 // We're interested only in allocas we can handle.
1192 if (!ASan.isInterestingAlloca(AI: *AI))
1193 return;
1194 bool DoPoison = (ID == Intrinsic::lifetime_end);
1195 AllocaPoisonCall APC = {.InsBefore: &II, .AI: AI, .Size: SizeValue, .DoPoison: DoPoison};
1196 if (AI->isStaticAlloca())
1197 StaticAllocaPoisonCallVec.push_back(Elt: APC);
1198 else if (ClInstrumentDynamicAllocas)
1199 DynamicAllocaPoisonCallVec.push_back(Elt: APC);
1200 }
1201
1202 void visitCallBase(CallBase &CB) {
1203 if (CallInst *CI = dyn_cast<CallInst>(Val: &CB)) {
1204 HasInlineAsm |= CI->isInlineAsm() && &CB != ASan.LocalDynamicShadow;
1205 HasReturnsTwiceCall |= CI->canReturnTwice();
1206 }
1207 }
1208
1209 // ---------------------- Helpers.
1210 void initializeCallbacks(Module &M);
1211
1212 // Copies bytes from ShadowBytes into shadow memory for indexes where
1213 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that
1214 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten.
1215 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1216 IRBuilder<> &IRB, Value *ShadowBase);
1217 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1218 size_t Begin, size_t End, IRBuilder<> &IRB,
1219 Value *ShadowBase);
1220 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
1221 ArrayRef<uint8_t> ShadowBytes, size_t Begin,
1222 size_t End, IRBuilder<> &IRB, Value *ShadowBase);
1223
1224 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
1225
1226 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L,
1227 bool Dynamic);
1228 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue,
1229 Instruction *ThenTerm, Value *ValueIfFalse);
1230};
1231
1232} // end anonymous namespace
1233
1234void AddressSanitizerPass::printPipeline(
1235 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1236 static_cast<PassInfoMixin<AddressSanitizerPass> *>(this)->printPipeline(
1237 OS, MapClassName2PassName);
1238 OS << '<';
1239 if (Options.CompileKernel)
1240 OS << "kernel";
1241 OS << '>';
1242}
1243
1244AddressSanitizerPass::AddressSanitizerPass(
1245 const AddressSanitizerOptions &Options, bool UseGlobalGC,
1246 bool UseOdrIndicator, AsanDtorKind DestructorKind,
1247 AsanCtorKind ConstructorKind)
1248 : Options(Options), UseGlobalGC(UseGlobalGC),
1249 UseOdrIndicator(UseOdrIndicator), DestructorKind(DestructorKind),
1250 ConstructorKind(ConstructorKind) {}
1251
1252PreservedAnalyses AddressSanitizerPass::run(Module &M,
1253 ModuleAnalysisManager &MAM) {
1254 ModuleAddressSanitizer ModuleSanitizer(
1255 M, Options.InsertVersionCheck, Options.CompileKernel, Options.Recover,
1256 UseGlobalGC, UseOdrIndicator, DestructorKind, ConstructorKind);
1257 bool Modified = false;
1258 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
1259 const StackSafetyGlobalInfo *const SSGI =
1260 ClUseStackSafety ? &MAM.getResult<StackSafetyGlobalAnalysis>(IR&: M) : nullptr;
1261 for (Function &F : M) {
1262 AddressSanitizer FunctionSanitizer(
1263 M, SSGI, Options.InstrumentationWithCallsThreshold,
1264 Options.MaxInlinePoisoningSize, Options.CompileKernel, Options.Recover,
1265 Options.UseAfterScope, Options.UseAfterReturn);
1266 const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: F);
1267 Modified |= FunctionSanitizer.instrumentFunction(F, TLI: &TLI);
1268 }
1269 Modified |= ModuleSanitizer.instrumentModule(M);
1270 if (!Modified)
1271 return PreservedAnalyses::all();
1272
1273 PreservedAnalyses PA = PreservedAnalyses::none();
1274 // GlobalsAA is considered stateless and does not get invalidated unless
1275 // explicitly invalidated; PreservedAnalyses::none() is not enough. Sanitizers
1276 // make changes that require GlobalsAA to be invalidated.
1277 PA.abandon<GlobalsAA>();
1278 return PA;
1279}
1280
1281static size_t TypeStoreSizeToSizeIndex(uint32_t TypeSize) {
1282 size_t Res = llvm::countr_zero(Val: TypeSize / 8);
1283 assert(Res < kNumberOfAccessSizes);
1284 return Res;
1285}
1286
1287/// Check if \p G has been created by a trusted compiler pass.
1288static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) {
1289 // Do not instrument @llvm.global_ctors, @llvm.used, etc.
1290 if (G->getName().starts_with(Prefix: "llvm.") ||
1291 // Do not instrument gcov counter arrays.
1292 G->getName().starts_with(Prefix: "__llvm_gcov_ctr") ||
1293 // Do not instrument rtti proxy symbols for function sanitizer.
1294 G->getName().starts_with(Prefix: "__llvm_rtti_proxy"))
1295 return true;
1296
1297 // Do not instrument asan globals.
1298 if (G->getName().starts_with(Prefix: kAsanGenPrefix) ||
1299 G->getName().starts_with(Prefix: kSanCovGenPrefix) ||
1300 G->getName().starts_with(Prefix: kODRGenPrefix))
1301 return true;
1302
1303 return false;
1304}
1305
1306static bool isUnsupportedAMDGPUAddrspace(Value *Addr) {
1307 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1308 unsigned int AddrSpace = PtrTy->getPointerAddressSpace();
1309 if (AddrSpace == 3 || AddrSpace == 5)
1310 return true;
1311 return false;
1312}
1313
1314Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
1315 // Shadow >> scale
1316 Shadow = IRB.CreateLShr(LHS: Shadow, RHS: Mapping.Scale);
1317 if (Mapping.Offset == 0) return Shadow;
1318 // (Shadow >> scale) | offset
1319 Value *ShadowBase;
1320 if (LocalDynamicShadow)
1321 ShadowBase = LocalDynamicShadow;
1322 else
1323 ShadowBase = ConstantInt::get(Ty: IntptrTy, V: Mapping.Offset);
1324 if (Mapping.OrShadowOffset)
1325 return IRB.CreateOr(LHS: Shadow, RHS: ShadowBase);
1326 else
1327 return IRB.CreateAdd(LHS: Shadow, RHS: ShadowBase);
1328}
1329
1330// Instrument memset/memmove/memcpy
1331void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI,
1332 RuntimeCallInserter &RTCI) {
1333 InstrumentationIRBuilder IRB(MI);
1334 if (isa<MemTransferInst>(Val: MI)) {
1335 RTCI.createRuntimeCall(
1336 IRB, Callee: isa<MemMoveInst>(Val: MI) ? AsanMemmove : AsanMemcpy,
1337 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1338 IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 1), DestTy: PtrTy),
1339 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1340 } else if (isa<MemSetInst>(Val: MI)) {
1341 RTCI.createRuntimeCall(
1342 IRB, Callee: AsanMemset,
1343 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1344 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 1), DestTy: IRB.getInt32Ty(), isSigned: false),
1345 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1346 }
1347 MI->eraseFromParent();
1348}
1349
1350/// Check if we want (and can) handle this alloca.
1351bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1352 auto PreviouslySeenAllocaInfo = ProcessedAllocas.find(Val: &AI);
1353
1354 if (PreviouslySeenAllocaInfo != ProcessedAllocas.end())
1355 return PreviouslySeenAllocaInfo->getSecond();
1356
1357 bool IsInteresting =
1358 (AI.getAllocatedType()->isSized() &&
1359 // alloca() may be called with 0 size, ignore it.
1360 ((!AI.isStaticAlloca()) || !getAllocaSizeInBytes(AI).isZero()) &&
1361 // We are only interested in allocas not promotable to registers.
1362 // Promotable allocas are common under -O0.
1363 (!ClSkipPromotableAllocas || !isAllocaPromotable(AI: &AI)) &&
1364 // inalloca allocas are not treated as static, and we don't want
1365 // dynamic alloca instrumentation for them as well.
1366 !AI.isUsedWithInAlloca() &&
1367 // swifterror allocas are register promoted by ISel
1368 !AI.isSwiftError() &&
1369 // safe allocas are not interesting
1370 !(SSGI && SSGI->isSafe(AI)));
1371
1372 ProcessedAllocas[&AI] = IsInteresting;
1373 return IsInteresting;
1374}
1375
1376bool AddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) {
1377 // Instrument accesses from different address spaces only for AMDGPU.
1378 Type *PtrTy = cast<PointerType>(Val: Ptr->getType()->getScalarType());
1379 if (PtrTy->getPointerAddressSpace() != 0 &&
1380 !(TargetTriple.isAMDGPU() && !isUnsupportedAMDGPUAddrspace(Addr: Ptr)))
1381 return true;
1382
1383 // Ignore swifterror addresses.
1384 // swifterror memory addresses are mem2reg promoted by instruction
1385 // selection. As such they cannot have regular uses like an instrumentation
1386 // function and it makes no sense to track them as memory.
1387 if (Ptr->isSwiftError())
1388 return true;
1389
1390 // Treat memory accesses to promotable allocas as non-interesting since they
1391 // will not cause memory violations. This greatly speeds up the instrumented
1392 // executable at -O0.
1393 if (auto AI = dyn_cast_or_null<AllocaInst>(Val: Ptr))
1394 if (ClSkipPromotableAllocas && !isInterestingAlloca(AI: *AI))
1395 return true;
1396
1397 if (SSGI != nullptr && SSGI->stackAccessIsSafe(I: *Inst) &&
1398 findAllocaForValue(V: Ptr))
1399 return true;
1400
1401 return false;
1402}
1403
1404void AddressSanitizer::getInterestingMemoryOperands(
1405 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting) {
1406 // Do not instrument the load fetching the dynamic shadow address.
1407 if (LocalDynamicShadow == I)
1408 return;
1409
1410 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
1411 if (!ClInstrumentReads || ignoreAccess(Inst: I, Ptr: LI->getPointerOperand()))
1412 return;
1413 Interesting.emplace_back(Args&: I, Args: LI->getPointerOperandIndex(), Args: false,
1414 Args: LI->getType(), Args: LI->getAlign());
1415 } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
1416 if (!ClInstrumentWrites || ignoreAccess(Inst: I, Ptr: SI->getPointerOperand()))
1417 return;
1418 Interesting.emplace_back(Args&: I, Args: SI->getPointerOperandIndex(), Args: true,
1419 Args: SI->getValueOperand()->getType(), Args: SI->getAlign());
1420 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
1421 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: RMW->getPointerOperand()))
1422 return;
1423 Interesting.emplace_back(Args&: I, Args: RMW->getPointerOperandIndex(), Args: true,
1424 Args: RMW->getValOperand()->getType(), Args: std::nullopt);
1425 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
1426 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: XCHG->getPointerOperand()))
1427 return;
1428 Interesting.emplace_back(Args&: I, Args: XCHG->getPointerOperandIndex(), Args: true,
1429 Args: XCHG->getCompareOperand()->getType(),
1430 Args: std::nullopt);
1431 } else if (auto CI = dyn_cast<CallInst>(Val: I)) {
1432 switch (CI->getIntrinsicID()) {
1433 case Intrinsic::masked_load:
1434 case Intrinsic::masked_store:
1435 case Intrinsic::masked_gather:
1436 case Intrinsic::masked_scatter: {
1437 bool IsWrite = CI->getType()->isVoidTy();
1438 // Masked store has an initial operand for the value.
1439 unsigned OpOffset = IsWrite ? 1 : 0;
1440 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1441 return;
1442
1443 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1444 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1445 return;
1446 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1447 MaybeAlign Alignment = Align(1);
1448 // Otherwise no alignment guarantees. We probably got Undef.
1449 if (auto *Op = dyn_cast<ConstantInt>(Val: CI->getOperand(i_nocapture: 1 + OpOffset)))
1450 Alignment = Op->getMaybeAlignValue();
1451 Value *Mask = CI->getOperand(i_nocapture: 2 + OpOffset);
1452 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: Mask);
1453 break;
1454 }
1455 case Intrinsic::masked_expandload:
1456 case Intrinsic::masked_compressstore: {
1457 bool IsWrite = CI->getIntrinsicID() == Intrinsic::masked_compressstore;
1458 unsigned OpOffset = IsWrite ? 1 : 0;
1459 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1460 return;
1461 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1462 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1463 return;
1464 MaybeAlign Alignment = BasePtr->getPointerAlignment(DL: *DL);
1465 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1466
1467 IRBuilder IB(I);
1468 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1469 // Use the popcount of Mask as the effective vector length.
1470 Type *ExtTy = VectorType::get(ElementType: IntptrTy, Other: cast<VectorType>(Val: Ty));
1471 Value *ExtMask = IB.CreateZExt(V: Mask, DestTy: ExtTy);
1472 Value *EVL = IB.CreateAddReduce(Src: ExtMask);
1473 Value *TrueMask = ConstantInt::get(Ty: Mask->getType(), V: 1);
1474 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: TrueMask,
1475 Args&: EVL);
1476 break;
1477 }
1478 case Intrinsic::vp_load:
1479 case Intrinsic::vp_store:
1480 case Intrinsic::experimental_vp_strided_load:
1481 case Intrinsic::experimental_vp_strided_store: {
1482 auto *VPI = cast<VPIntrinsic>(Val: CI);
1483 unsigned IID = CI->getIntrinsicID();
1484 bool IsWrite = CI->getType()->isVoidTy();
1485 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1486 return;
1487 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1488 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1489 MaybeAlign Alignment = VPI->getOperand(i_nocapture: PtrOpNo)->getPointerAlignment(DL: *DL);
1490 Value *Stride = nullptr;
1491 if (IID == Intrinsic::experimental_vp_strided_store ||
1492 IID == Intrinsic::experimental_vp_strided_load) {
1493 Stride = VPI->getOperand(i_nocapture: PtrOpNo + 1);
1494 // Use the pointer alignment as the element alignment if the stride is a
1495 // mutiple of the pointer alignment. Otherwise, the element alignment
1496 // should be Align(1).
1497 unsigned PointerAlign = Alignment.valueOrOne().value();
1498 if (!isa<ConstantInt>(Val: Stride) ||
1499 cast<ConstantInt>(Val: Stride)->getZExtValue() % PointerAlign != 0)
1500 Alignment = Align(1);
1501 }
1502 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1503 Args: VPI->getMaskParam(), Args: VPI->getVectorLengthParam(),
1504 Args&: Stride);
1505 break;
1506 }
1507 case Intrinsic::vp_gather:
1508 case Intrinsic::vp_scatter: {
1509 auto *VPI = cast<VPIntrinsic>(Val: CI);
1510 unsigned IID = CI->getIntrinsicID();
1511 bool IsWrite = IID == Intrinsic::vp_scatter;
1512 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1513 return;
1514 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1515 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1516 MaybeAlign Alignment = VPI->getPointerAlignment();
1517 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1518 Args: VPI->getMaskParam(),
1519 Args: VPI->getVectorLengthParam());
1520 break;
1521 }
1522 default:
1523 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) {
1524 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1525 ignoreAccess(Inst: I, Ptr: CI->getArgOperand(i: ArgNo)))
1526 continue;
1527 Type *Ty = CI->getParamByValType(ArgNo);
1528 Interesting.emplace_back(Args&: I, Args&: ArgNo, Args: false, Args&: Ty, Args: Align(1));
1529 }
1530 }
1531 }
1532}
1533
1534static bool isPointerOperand(Value *V) {
1535 return V->getType()->isPointerTy() || isa<PtrToIntInst>(Val: V);
1536}
1537
1538// This is a rough heuristic; it may cause both false positives and
1539// false negatives. The proper implementation requires cooperation with
1540// the frontend.
1541static bool isInterestingPointerComparison(Instruction *I) {
1542 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: I)) {
1543 if (!Cmp->isRelational())
1544 return false;
1545 } else {
1546 return false;
1547 }
1548 return isPointerOperand(V: I->getOperand(i: 0)) &&
1549 isPointerOperand(V: I->getOperand(i: 1));
1550}
1551
1552// This is a rough heuristic; it may cause both false positives and
1553// false negatives. The proper implementation requires cooperation with
1554// the frontend.
1555static bool isInterestingPointerSubtraction(Instruction *I) {
1556 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) {
1557 if (BO->getOpcode() != Instruction::Sub)
1558 return false;
1559 } else {
1560 return false;
1561 }
1562 return isPointerOperand(V: I->getOperand(i: 0)) &&
1563 isPointerOperand(V: I->getOperand(i: 1));
1564}
1565
1566bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1567 // If a global variable does not have dynamic initialization we don't
1568 // have to instrument it. However, if a global does not have initializer
1569 // at all, we assume it has dynamic initializer (in other TU).
1570 if (!G->hasInitializer())
1571 return false;
1572
1573 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().IsDynInit)
1574 return false;
1575
1576 return true;
1577}
1578
1579void AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1580 Instruction *I, RuntimeCallInserter &RTCI) {
1581 IRBuilder<> IRB(I);
1582 FunctionCallee F = isa<ICmpInst>(Val: I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1583 Value *Param[2] = {I->getOperand(i: 0), I->getOperand(i: 1)};
1584 for (Value *&i : Param) {
1585 if (i->getType()->isPointerTy())
1586 i = IRB.CreatePointerCast(V: i, DestTy: IntptrTy);
1587 }
1588 RTCI.createRuntimeCall(IRB, Callee: F, Args: Param);
1589}
1590
1591static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1592 Instruction *InsertBefore, Value *Addr,
1593 MaybeAlign Alignment, unsigned Granularity,
1594 TypeSize TypeStoreSize, bool IsWrite,
1595 Value *SizeArgument, bool UseCalls,
1596 uint32_t Exp, RuntimeCallInserter &RTCI) {
1597 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1598 // if the data is properly aligned.
1599 if (!TypeStoreSize.isScalable()) {
1600 const auto FixedSize = TypeStoreSize.getFixedValue();
1601 switch (FixedSize) {
1602 case 8:
1603 case 16:
1604 case 32:
1605 case 64:
1606 case 128:
1607 if (!Alignment || *Alignment >= Granularity ||
1608 *Alignment >= FixedSize / 8)
1609 return Pass->instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment,
1610 TypeStoreSize: FixedSize, IsWrite, SizeArgument: nullptr, UseCalls,
1611 Exp, RTCI);
1612 }
1613 }
1614 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeStoreSize,
1615 IsWrite, SizeArgument: nullptr, UseCalls, Exp, RTCI);
1616}
1617
1618void AddressSanitizer::instrumentMaskedLoadOrStore(
1619 AddressSanitizer *Pass, const DataLayout &DL, Type *IntptrTy, Value *Mask,
1620 Value *EVL, Value *Stride, Instruction *I, Value *Addr,
1621 MaybeAlign Alignment, unsigned Granularity, Type *OpType, bool IsWrite,
1622 Value *SizeArgument, bool UseCalls, uint32_t Exp,
1623 RuntimeCallInserter &RTCI) {
1624 auto *VTy = cast<VectorType>(Val: OpType);
1625 TypeSize ElemTypeSize = DL.getTypeStoreSizeInBits(Ty: VTy->getScalarType());
1626 auto Zero = ConstantInt::get(Ty: IntptrTy, V: 0);
1627
1628 IRBuilder IB(I);
1629 Instruction *LoopInsertBefore = I;
1630 if (EVL) {
1631 // The end argument of SplitBlockAndInsertForLane is assumed bigger
1632 // than zero, so we should check whether EVL is zero here.
1633 Type *EVLType = EVL->getType();
1634 Value *IsEVLZero = IB.CreateICmpNE(LHS: EVL, RHS: ConstantInt::get(Ty: EVLType, V: 0));
1635 LoopInsertBefore = SplitBlockAndInsertIfThen(Cond: IsEVLZero, SplitBefore: I, Unreachable: false);
1636 IB.SetInsertPoint(LoopInsertBefore);
1637 // Cast EVL to IntptrTy.
1638 EVL = IB.CreateZExtOrTrunc(V: EVL, DestTy: IntptrTy);
1639 // To avoid undefined behavior for extracting with out of range index, use
1640 // the minimum of evl and element count as trip count.
1641 Value *EC = IB.CreateElementCount(DstType: IntptrTy, EC: VTy->getElementCount());
1642 EVL = IB.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: EVL, RHS: EC);
1643 } else {
1644 EVL = IB.CreateElementCount(DstType: IntptrTy, EC: VTy->getElementCount());
1645 }
1646
1647 // Cast Stride to IntptrTy.
1648 if (Stride)
1649 Stride = IB.CreateZExtOrTrunc(V: Stride, DestTy: IntptrTy);
1650
1651 SplitBlockAndInsertForEachLane(End: EVL, InsertBefore: LoopInsertBefore,
1652 Func: [&](IRBuilderBase &IRB, Value *Index) {
1653 Value *MaskElem = IRB.CreateExtractElement(Vec: Mask, Idx: Index);
1654 if (auto *MaskElemC = dyn_cast<ConstantInt>(Val: MaskElem)) {
1655 if (MaskElemC->isZero())
1656 // No check
1657 return;
1658 // Unconditional check
1659 } else {
1660 // Conditional check
1661 Instruction *ThenTerm = SplitBlockAndInsertIfThen(
1662 Cond: MaskElem, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false);
1663 IRB.SetInsertPoint(ThenTerm);
1664 }
1665
1666 Value *InstrumentedAddress;
1667 if (isa<VectorType>(Val: Addr->getType())) {
1668 assert(
1669 cast<VectorType>(Addr->getType())->getElementType()->isPointerTy() &&
1670 "Expected vector of pointer.");
1671 InstrumentedAddress = IRB.CreateExtractElement(Vec: Addr, Idx: Index);
1672 } else if (Stride) {
1673 Index = IRB.CreateMul(LHS: Index, RHS: Stride);
1674 InstrumentedAddress = IRB.CreatePtrAdd(Ptr: Addr, Offset: Index);
1675 } else {
1676 InstrumentedAddress = IRB.CreateGEP(Ty: VTy, Ptr: Addr, IdxList: {Zero, Index});
1677 }
1678 doInstrumentAddress(Pass, I, InsertBefore: &*IRB.GetInsertPoint(), Addr: InstrumentedAddress,
1679 Alignment, Granularity, TypeStoreSize: ElemTypeSize, IsWrite,
1680 SizeArgument, UseCalls, Exp, RTCI);
1681 });
1682}
1683
1684void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1685 InterestingMemoryOperand &O, bool UseCalls,
1686 const DataLayout &DL,
1687 RuntimeCallInserter &RTCI) {
1688 Value *Addr = O.getPtr();
1689
1690 // Optimization experiments.
1691 // The experiments can be used to evaluate potential optimizations that remove
1692 // instrumentation (assess false negatives). Instead of completely removing
1693 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1694 // experiments that want to remove instrumentation of this instruction).
1695 // If Exp is non-zero, this pass will emit special calls into runtime
1696 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1697 // make runtime terminate the program in a special way (with a different
1698 // exit status). Then you run the new compiler on a buggy corpus, collect
1699 // the special terminations (ideally, you don't see them at all -- no false
1700 // negatives) and make the decision on the optimization.
1701 uint32_t Exp = ClForceExperiment;
1702
1703 if (ClOpt && ClOptGlobals) {
1704 // If initialization order checking is disabled, a simple access to a
1705 // dynamically initialized global is always valid.
1706 GlobalVariable *G = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: Addr));
1707 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1708 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1709 NumOptimizedAccessesToGlobalVar++;
1710 return;
1711 }
1712 }
1713
1714 if (ClOpt && ClOptStack) {
1715 // A direct inbounds access to a stack variable is always valid.
1716 if (isa<AllocaInst>(Val: getUnderlyingObject(V: Addr)) &&
1717 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1718 NumOptimizedAccessesToStackVar++;
1719 return;
1720 }
1721 }
1722
1723 if (O.IsWrite)
1724 NumInstrumentedWrites++;
1725 else
1726 NumInstrumentedReads++;
1727
1728 unsigned Granularity = 1 << Mapping.Scale;
1729 if (O.MaybeMask) {
1730 instrumentMaskedLoadOrStore(Pass: this, DL, IntptrTy, Mask: O.MaybeMask, EVL: O.MaybeEVL,
1731 Stride: O.MaybeStride, I: O.getInsn(), Addr, Alignment: O.Alignment,
1732 Granularity, OpType: O.OpType, IsWrite: O.IsWrite, SizeArgument: nullptr,
1733 UseCalls, Exp, RTCI);
1734 } else {
1735 doInstrumentAddress(Pass: this, I: O.getInsn(), InsertBefore: O.getInsn(), Addr, Alignment: O.Alignment,
1736 Granularity, TypeStoreSize: O.TypeStoreSize, IsWrite: O.IsWrite, SizeArgument: nullptr,
1737 UseCalls, Exp, RTCI);
1738 }
1739}
1740
1741Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1742 Value *Addr, bool IsWrite,
1743 size_t AccessSizeIndex,
1744 Value *SizeArgument,
1745 uint32_t Exp,
1746 RuntimeCallInserter &RTCI) {
1747 InstrumentationIRBuilder IRB(InsertBefore);
1748 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp);
1749 CallInst *Call = nullptr;
1750 if (SizeArgument) {
1751 if (Exp == 0)
1752 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][0],
1753 Args: {Addr, SizeArgument});
1754 else
1755 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][1],
1756 Args: {Addr, SizeArgument, ExpVal});
1757 } else {
1758 if (Exp == 0)
1759 Call = RTCI.createRuntimeCall(
1760 IRB, Callee: AsanErrorCallback[IsWrite][0][AccessSizeIndex], Args: Addr);
1761 else
1762 Call = RTCI.createRuntimeCall(
1763 IRB, Callee: AsanErrorCallback[IsWrite][1][AccessSizeIndex], Args: {Addr, ExpVal});
1764 }
1765
1766 Call->setCannotMerge();
1767 return Call;
1768}
1769
1770Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1771 Value *ShadowValue,
1772 uint32_t TypeStoreSize) {
1773 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1774 // Addr & (Granularity - 1)
1775 Value *LastAccessedByte =
1776 IRB.CreateAnd(LHS: AddrLong, RHS: ConstantInt::get(Ty: IntptrTy, V: Granularity - 1));
1777 // (Addr & (Granularity - 1)) + size - 1
1778 if (TypeStoreSize / 8 > 1)
1779 LastAccessedByte = IRB.CreateAdd(
1780 LHS: LastAccessedByte, RHS: ConstantInt::get(Ty: IntptrTy, V: TypeStoreSize / 8 - 1));
1781 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1782 LastAccessedByte =
1783 IRB.CreateIntCast(V: LastAccessedByte, DestTy: ShadowValue->getType(), isSigned: false);
1784 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1785 return IRB.CreateICmpSGE(LHS: LastAccessedByte, RHS: ShadowValue);
1786}
1787
1788Instruction *AddressSanitizer::instrumentAMDGPUAddress(
1789 Instruction *OrigIns, Instruction *InsertBefore, Value *Addr,
1790 uint32_t TypeStoreSize, bool IsWrite, Value *SizeArgument) {
1791 // Do not instrument unsupported addrspaces.
1792 if (isUnsupportedAMDGPUAddrspace(Addr))
1793 return nullptr;
1794 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1795 // Follow host instrumentation for global and constant addresses.
1796 if (PtrTy->getPointerAddressSpace() != 0)
1797 return InsertBefore;
1798 // Instrument generic addresses in supported addressspaces.
1799 IRBuilder<> IRB(InsertBefore);
1800 Value *IsShared = IRB.CreateCall(Callee: AMDGPUAddressShared, Args: {Addr});
1801 Value *IsPrivate = IRB.CreateCall(Callee: AMDGPUAddressPrivate, Args: {Addr});
1802 Value *IsSharedOrPrivate = IRB.CreateOr(LHS: IsShared, RHS: IsPrivate);
1803 Value *Cmp = IRB.CreateNot(V: IsSharedOrPrivate);
1804 Value *AddrSpaceZeroLanding =
1805 SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false);
1806 InsertBefore = cast<Instruction>(Val: AddrSpaceZeroLanding);
1807 return InsertBefore;
1808}
1809
1810Instruction *AddressSanitizer::genAMDGPUReportBlock(IRBuilder<> &IRB,
1811 Value *Cond, bool Recover) {
1812 Module &M = *IRB.GetInsertBlock()->getModule();
1813 Value *ReportCond = Cond;
1814 if (!Recover) {
1815 auto Ballot = M.getOrInsertFunction(Name: kAMDGPUBallotName, RetTy: IRB.getInt64Ty(),
1816 Args: IRB.getInt1Ty());
1817 ReportCond = IRB.CreateIsNotNull(Arg: IRB.CreateCall(Callee: Ballot, Args: {Cond}));
1818 }
1819
1820 auto *Trm =
1821 SplitBlockAndInsertIfThen(Cond: ReportCond, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false,
1822 BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
1823 Trm->getParent()->setName("asan.report");
1824
1825 if (Recover)
1826 return Trm;
1827
1828 Trm = SplitBlockAndInsertIfThen(Cond, SplitBefore: Trm, Unreachable: false);
1829 IRB.SetInsertPoint(Trm);
1830 return IRB.CreateCall(
1831 Callee: M.getOrInsertFunction(Name: kAMDGPUUnreachableName, RetTy: IRB.getVoidTy()), Args: {});
1832}
1833
1834void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1835 Instruction *InsertBefore, Value *Addr,
1836 MaybeAlign Alignment,
1837 uint32_t TypeStoreSize, bool IsWrite,
1838 Value *SizeArgument, bool UseCalls,
1839 uint32_t Exp,
1840 RuntimeCallInserter &RTCI) {
1841 if (TargetTriple.isAMDGPU()) {
1842 InsertBefore = instrumentAMDGPUAddress(OrigIns, InsertBefore, Addr,
1843 TypeStoreSize, IsWrite, SizeArgument);
1844 if (!InsertBefore)
1845 return;
1846 }
1847
1848 InstrumentationIRBuilder IRB(InsertBefore);
1849 size_t AccessSizeIndex = TypeStoreSizeToSizeIndex(TypeSize: TypeStoreSize);
1850 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
1851
1852 if (UseCalls && ClOptimizeCallbacks) {
1853 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
1854 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1855 IRB.CreateCall(
1856 Callee: Intrinsic::getDeclaration(M, id: Intrinsic::asan_check_memaccess),
1857 Args: {IRB.CreatePointerCast(V: Addr, DestTy: PtrTy),
1858 ConstantInt::get(Ty: Int32Ty, V: AccessInfo.Packed)});
1859 return;
1860 }
1861
1862 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
1863 if (UseCalls) {
1864 if (Exp == 0)
1865 RTCI.createRuntimeCall(
1866 IRB, Callee: AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex], Args: AddrLong);
1867 else
1868 RTCI.createRuntimeCall(
1869 IRB, Callee: AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
1870 Args: {AddrLong, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
1871 return;
1872 }
1873
1874 Type *ShadowTy =
1875 IntegerType::get(C&: *C, NumBits: std::max(a: 8U, b: TypeStoreSize >> Mapping.Scale));
1876 Type *ShadowPtrTy = PointerType::get(ElementType: ShadowTy, AddressSpace: 0);
1877 Value *ShadowPtr = memToShadow(Shadow: AddrLong, IRB);
1878 const uint64_t ShadowAlign =
1879 std::max<uint64_t>(a: Alignment.valueOrOne().value() >> Mapping.Scale, b: 1);
1880 Value *ShadowValue = IRB.CreateAlignedLoad(
1881 Ty: ShadowTy, Ptr: IRB.CreateIntToPtr(V: ShadowPtr, DestTy: ShadowPtrTy), Align: Align(ShadowAlign));
1882
1883 Value *Cmp = IRB.CreateIsNotNull(Arg: ShadowValue);
1884 size_t Granularity = 1ULL << Mapping.Scale;
1885 Instruction *CrashTerm = nullptr;
1886
1887 bool GenSlowPath = (ClAlwaysSlowPath || (TypeStoreSize < 8 * Granularity));
1888
1889 if (TargetTriple.isAMDGCN()) {
1890 if (GenSlowPath) {
1891 auto *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
1892 Cmp = IRB.CreateAnd(LHS: Cmp, RHS: Cmp2);
1893 }
1894 CrashTerm = genAMDGPUReportBlock(IRB, Cond: Cmp, Recover);
1895 } else if (GenSlowPath) {
1896 // We use branch weights for the slow path check, to indicate that the slow
1897 // path is rarely taken. This seems to be the case for SPEC benchmarks.
1898 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
1899 Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false, BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
1900 assert(cast<BranchInst>(CheckTerm)->isUnconditional());
1901 BasicBlock *NextBB = CheckTerm->getSuccessor(Idx: 0);
1902 IRB.SetInsertPoint(CheckTerm);
1903 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
1904 if (Recover) {
1905 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp2, SplitBefore: CheckTerm, Unreachable: false);
1906 } else {
1907 BasicBlock *CrashBlock =
1908 BasicBlock::Create(Context&: *C, Name: "", Parent: NextBB->getParent(), InsertBefore: NextBB);
1909 CrashTerm = new UnreachableInst(*C, CrashBlock);
1910 BranchInst *NewTerm = BranchInst::Create(IfTrue: CrashBlock, IfFalse: NextBB, Cond: Cmp2);
1911 ReplaceInstWithInst(From: CheckTerm, To: NewTerm);
1912 }
1913 } else {
1914 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: !Recover);
1915 }
1916
1917 Instruction *Crash = generateCrashCode(
1918 InsertBefore: CrashTerm, Addr: AddrLong, IsWrite, AccessSizeIndex, SizeArgument, Exp, RTCI);
1919 if (OrigIns->getDebugLoc())
1920 Crash->setDebugLoc(OrigIns->getDebugLoc());
1921}
1922
1923// Instrument unusual size or unusual alignment.
1924// We can not do it with a single check, so we do 1-byte check for the first
1925// and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
1926// to report the actual access size.
1927void AddressSanitizer::instrumentUnusualSizeOrAlignment(
1928 Instruction *I, Instruction *InsertBefore, Value *Addr,
1929 TypeSize TypeStoreSize, bool IsWrite, Value *SizeArgument, bool UseCalls,
1930 uint32_t Exp, RuntimeCallInserter &RTCI) {
1931 InstrumentationIRBuilder IRB(InsertBefore);
1932 Value *NumBits = IRB.CreateTypeSize(DstType: IntptrTy, Size: TypeStoreSize);
1933 Value *Size = IRB.CreateLShr(LHS: NumBits, RHS: ConstantInt::get(Ty: IntptrTy, V: 3));
1934
1935 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
1936 if (UseCalls) {
1937 if (Exp == 0)
1938 RTCI.createRuntimeCall(IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][0],
1939 Args: {AddrLong, Size});
1940 else
1941 RTCI.createRuntimeCall(
1942 IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][1],
1943 Args: {AddrLong, Size, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
1944 } else {
1945 Value *SizeMinusOne = IRB.CreateSub(LHS: Size, RHS: ConstantInt::get(Ty: IntptrTy, V: 1));
1946 Value *LastByte = IRB.CreateIntToPtr(
1947 V: IRB.CreateAdd(LHS: AddrLong, RHS: SizeMinusOne),
1948 DestTy: Addr->getType());
1949 instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false, Exp,
1950 RTCI);
1951 instrumentAddress(OrigIns: I, InsertBefore, Addr: LastByte, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false,
1952 Exp, RTCI);
1953 }
1954}
1955
1956void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit,
1957 GlobalValue *ModuleName) {
1958 // Set up the arguments to our poison/unpoison functions.
1959 IRBuilder<> IRB(&GlobalInit.front(),
1960 GlobalInit.front().getFirstInsertionPt());
1961
1962 // Add a call to poison all external globals before the given function starts.
1963 Value *ModuleNameAddr = ConstantExpr::getPointerCast(C: ModuleName, Ty: IntptrTy);
1964 IRB.CreateCall(Callee: AsanPoisonGlobals, Args: ModuleNameAddr);
1965
1966 // Add calls to unpoison all globals before each return instruction.
1967 for (auto &BB : GlobalInit)
1968 if (ReturnInst *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator()))
1969 CallInst::Create(Func: AsanUnpoisonGlobals, NameStr: "", InsertBefore: RI->getIterator());
1970}
1971
1972void ModuleAddressSanitizer::createInitializerPoisonCalls(
1973 Module &M, GlobalValue *ModuleName) {
1974 GlobalVariable *GV = M.getGlobalVariable(Name: "llvm.global_ctors");
1975 if (!GV)
1976 return;
1977
1978 ConstantArray *CA = dyn_cast<ConstantArray>(Val: GV->getInitializer());
1979 if (!CA)
1980 return;
1981
1982 for (Use &OP : CA->operands()) {
1983 if (isa<ConstantAggregateZero>(Val: OP)) continue;
1984 ConstantStruct *CS = cast<ConstantStruct>(Val&: OP);
1985
1986 // Must have a function or null ptr.
1987 if (Function *F = dyn_cast<Function>(Val: CS->getOperand(i_nocapture: 1))) {
1988 if (F->getName() == kAsanModuleCtorName) continue;
1989 auto *Priority = cast<ConstantInt>(Val: CS->getOperand(i_nocapture: 0));
1990 // Don't instrument CTORs that will run before asan.module_ctor.
1991 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
1992 continue;
1993 poisonOneInitializer(GlobalInit&: *F, ModuleName);
1994 }
1995 }
1996}
1997
1998const GlobalVariable *
1999ModuleAddressSanitizer::getExcludedAliasedGlobal(const GlobalAlias &GA) const {
2000 // In case this function should be expanded to include rules that do not just
2001 // apply when CompileKernel is true, either guard all existing rules with an
2002 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
2003 // should also apply to user space.
2004 assert(CompileKernel && "Only expecting to be called when compiling kernel");
2005
2006 const Constant *C = GA.getAliasee();
2007
2008 // When compiling the kernel, globals that are aliased by symbols prefixed
2009 // by "__" are special and cannot be padded with a redzone.
2010 if (GA.getName().starts_with(Prefix: "__"))
2011 return dyn_cast<GlobalVariable>(Val: C->stripPointerCastsAndAliases());
2012
2013 return nullptr;
2014}
2015
2016bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
2017 Type *Ty = G->getValueType();
2018 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
2019
2020 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().NoAddress)
2021 return false;
2022 if (!Ty->isSized()) return false;
2023 if (!G->hasInitializer()) return false;
2024 // Globals in address space 1 and 4 are supported for AMDGPU.
2025 if (G->getAddressSpace() &&
2026 !(TargetTriple.isAMDGPU() && !isUnsupportedAMDGPUAddrspace(Addr: G)))
2027 return false;
2028 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
2029 // Two problems with thread-locals:
2030 // - The address of the main thread's copy can't be computed at link-time.
2031 // - Need to poison all copies, not just the main thread's one.
2032 if (G->isThreadLocal()) return false;
2033 // For now, just ignore this Global if the alignment is large.
2034 if (G->getAlign() && *G->getAlign() > getMinRedzoneSizeForGlobal()) return false;
2035
2036 // For non-COFF targets, only instrument globals known to be defined by this
2037 // TU.
2038 // FIXME: We can instrument comdat globals on ELF if we are using the
2039 // GC-friendly metadata scheme.
2040 if (!TargetTriple.isOSBinFormatCOFF()) {
2041 if (!G->hasExactDefinition() || G->hasComdat())
2042 return false;
2043 } else {
2044 // On COFF, don't instrument non-ODR linkages.
2045 if (G->isInterposable())
2046 return false;
2047 // If the global has AvailableExternally linkage, then it is not in this
2048 // module, which means it does not need to be instrumented.
2049 if (G->hasAvailableExternallyLinkage())
2050 return false;
2051 }
2052
2053 // If a comdat is present, it must have a selection kind that implies ODR
2054 // semantics: no duplicates, any, or exact match.
2055 if (Comdat *C = G->getComdat()) {
2056 switch (C->getSelectionKind()) {
2057 case Comdat::Any:
2058 case Comdat::ExactMatch:
2059 case Comdat::NoDeduplicate:
2060 break;
2061 case Comdat::Largest:
2062 case Comdat::SameSize:
2063 return false;
2064 }
2065 }
2066
2067 if (G->hasSection()) {
2068 // The kernel uses explicit sections for mostly special global variables
2069 // that we should not instrument. E.g. the kernel may rely on their layout
2070 // without redzones, or remove them at link time ("discard.*"), etc.
2071 if (CompileKernel)
2072 return false;
2073
2074 StringRef Section = G->getSection();
2075
2076 // Globals from llvm.metadata aren't emitted, do not instrument them.
2077 if (Section == "llvm.metadata") return false;
2078 // Do not instrument globals from special LLVM sections.
2079 if (Section.contains(Other: "__llvm") || Section.contains(Other: "__LLVM"))
2080 return false;
2081
2082 // Do not instrument function pointers to initialization and termination
2083 // routines: dynamic linker will not properly handle redzones.
2084 if (Section.starts_with(Prefix: ".preinit_array") ||
2085 Section.starts_with(Prefix: ".init_array") ||
2086 Section.starts_with(Prefix: ".fini_array")) {
2087 return false;
2088 }
2089
2090 // Do not instrument user-defined sections (with names resembling
2091 // valid C identifiers)
2092 if (TargetTriple.isOSBinFormatELF()) {
2093 if (llvm::all_of(Range&: Section,
2094 P: [](char c) { return llvm::isAlnum(C: c) || c == '_'; }))
2095 return false;
2096 }
2097
2098 // On COFF, if the section name contains '$', it is highly likely that the
2099 // user is using section sorting to create an array of globals similar to
2100 // the way initialization callbacks are registered in .init_array and
2101 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
2102 // to such globals is counterproductive, because the intent is that they
2103 // will form an array, and out-of-bounds accesses are expected.
2104 // See https://github.com/google/sanitizers/issues/305
2105 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
2106 if (TargetTriple.isOSBinFormatCOFF() && Section.contains(C: '$')) {
2107 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
2108 << *G << "\n");
2109 return false;
2110 }
2111
2112 if (TargetTriple.isOSBinFormatMachO()) {
2113 StringRef ParsedSegment, ParsedSection;
2114 unsigned TAA = 0, StubSize = 0;
2115 bool TAAParsed;
2116 cantFail(Err: MCSectionMachO::ParseSectionSpecifier(
2117 Spec: Section, Segment&: ParsedSegment, Section&: ParsedSection, TAA, TAAParsed, StubSize));
2118
2119 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
2120 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
2121 // them.
2122 if (ParsedSegment == "__OBJC" ||
2123 (ParsedSegment == "__DATA" && ParsedSection.starts_with(Prefix: "__objc_"))) {
2124 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
2125 return false;
2126 }
2127 // See https://github.com/google/sanitizers/issues/32
2128 // Constant CFString instances are compiled in the following way:
2129 // -- the string buffer is emitted into
2130 // __TEXT,__cstring,cstring_literals
2131 // -- the constant NSConstantString structure referencing that buffer
2132 // is placed into __DATA,__cfstring
2133 // Therefore there's no point in placing redzones into __DATA,__cfstring.
2134 // Moreover, it causes the linker to crash on OS X 10.7
2135 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
2136 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
2137 return false;
2138 }
2139 // The linker merges the contents of cstring_literals and removes the
2140 // trailing zeroes.
2141 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
2142 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
2143 return false;
2144 }
2145 }
2146 }
2147
2148 if (CompileKernel) {
2149 // Globals that prefixed by "__" are special and cannot be padded with a
2150 // redzone.
2151 if (G->getName().starts_with(Prefix: "__"))
2152 return false;
2153 }
2154
2155 return true;
2156}
2157
2158// On Mach-O platforms, we emit global metadata in a separate section of the
2159// binary in order to allow the linker to properly dead strip. This is only
2160// supported on recent versions of ld64.
2161bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
2162 if (!TargetTriple.isOSBinFormatMachO())
2163 return false;
2164
2165 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(Major: 10, Minor: 11))
2166 return true;
2167 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(Major: 9))
2168 return true;
2169 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(Major: 2))
2170 return true;
2171 if (TargetTriple.isDriverKit())
2172 return true;
2173 if (TargetTriple.isXROS())
2174 return true;
2175
2176 return false;
2177}
2178
2179StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
2180 switch (TargetTriple.getObjectFormat()) {
2181 case Triple::COFF: return ".ASAN$GL";
2182 case Triple::ELF: return "asan_globals";
2183 case Triple::MachO: return "__DATA,__asan_globals,regular";
2184 case Triple::Wasm:
2185 case Triple::GOFF:
2186 case Triple::SPIRV:
2187 case Triple::XCOFF:
2188 case Triple::DXContainer:
2189 report_fatal_error(
2190 reason: "ModuleAddressSanitizer not implemented for object file format");
2191 case Triple::UnknownObjectFormat:
2192 break;
2193 }
2194 llvm_unreachable("unsupported object format");
2195}
2196
2197void ModuleAddressSanitizer::initializeCallbacks(Module &M) {
2198 IRBuilder<> IRB(*C);
2199
2200 // Declare our poisoning and unpoisoning functions.
2201 AsanPoisonGlobals =
2202 M.getOrInsertFunction(Name: kAsanPoisonGlobalsName, RetTy: IRB.getVoidTy(), Args: IntptrTy);
2203 AsanUnpoisonGlobals =
2204 M.getOrInsertFunction(Name: kAsanUnpoisonGlobalsName, RetTy: IRB.getVoidTy());
2205
2206 // Declare functions that register/unregister globals.
2207 AsanRegisterGlobals = M.getOrInsertFunction(
2208 Name: kAsanRegisterGlobalsName, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
2209 AsanUnregisterGlobals = M.getOrInsertFunction(
2210 Name: kAsanUnregisterGlobalsName, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
2211
2212 // Declare the functions that find globals in a shared object and then invoke
2213 // the (un)register function on them.
2214 AsanRegisterImageGlobals = M.getOrInsertFunction(
2215 Name: kAsanRegisterImageGlobalsName, RetTy: IRB.getVoidTy(), Args: IntptrTy);
2216 AsanUnregisterImageGlobals = M.getOrInsertFunction(
2217 Name: kAsanUnregisterImageGlobalsName, RetTy: IRB.getVoidTy(), Args: IntptrTy);
2218
2219 AsanRegisterElfGlobals =
2220 M.getOrInsertFunction(Name: kAsanRegisterElfGlobalsName, RetTy: IRB.getVoidTy(),
2221 Args: IntptrTy, Args: IntptrTy, Args: IntptrTy);
2222 AsanUnregisterElfGlobals =
2223 M.getOrInsertFunction(Name: kAsanUnregisterElfGlobalsName, RetTy: IRB.getVoidTy(),
2224 Args: IntptrTy, Args: IntptrTy, Args: IntptrTy);
2225}
2226
2227// Put the metadata and the instrumented global in the same group. This ensures
2228// that the metadata is discarded if the instrumented global is discarded.
2229void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
2230 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
2231 Module &M = *G->getParent();
2232 Comdat *C = G->getComdat();
2233 if (!C) {
2234 if (!G->hasName()) {
2235 // If G is unnamed, it must be internal. Give it an artificial name
2236 // so we can put it in a comdat.
2237 assert(G->hasLocalLinkage());
2238 G->setName(Twine(kAsanGenPrefix) + "_anon_global");
2239 }
2240
2241 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
2242 std::string Name = std::string(G->getName());
2243 Name += InternalSuffix;
2244 C = M.getOrInsertComdat(Name);
2245 } else {
2246 C = M.getOrInsertComdat(Name: G->getName());
2247 }
2248
2249 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
2250 // linkage to internal linkage so that a symbol table entry is emitted. This
2251 // is necessary in order to create the comdat group.
2252 if (TargetTriple.isOSBinFormatCOFF()) {
2253 C->setSelectionKind(Comdat::NoDeduplicate);
2254 if (G->hasPrivateLinkage())
2255 G->setLinkage(GlobalValue::InternalLinkage);
2256 }
2257 G->setComdat(C);
2258 }
2259
2260 assert(G->hasComdat());
2261 Metadata->setComdat(G->getComdat());
2262}
2263
2264// Create a separate metadata global and put it in the appropriate ASan
2265// global registration section.
2266GlobalVariable *
2267ModuleAddressSanitizer::CreateMetadataGlobal(Module &M, Constant *Initializer,
2268 StringRef OriginalName) {
2269 auto Linkage = TargetTriple.isOSBinFormatMachO()
2270 ? GlobalVariable::InternalLinkage
2271 : GlobalVariable::PrivateLinkage;
2272 GlobalVariable *Metadata = new GlobalVariable(
2273 M, Initializer->getType(), false, Linkage, Initializer,
2274 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(Name: OriginalName));
2275 Metadata->setSection(getGlobalMetadataSection());
2276 // Place metadata in a large section for x86-64 ELF binaries to mitigate
2277 // relocation pressure.
2278 setGlobalVariableLargeSection(TargetTriple, GV&: *Metadata);
2279 return Metadata;
2280}
2281
2282Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor(Module &M) {
2283 AsanDtorFunction = Function::createWithDefaultAttr(
2284 Ty: FunctionType::get(Result: Type::getVoidTy(C&: *C), isVarArg: false),
2285 Linkage: GlobalValue::InternalLinkage, AddrSpace: 0, N: kAsanModuleDtorName, M: &M);
2286 AsanDtorFunction->addFnAttr(Kind: Attribute::NoUnwind);
2287 // Ensure Dtor cannot be discarded, even if in a comdat.
2288 appendToUsed(M, Values: {AsanDtorFunction});
2289 BasicBlock *AsanDtorBB = BasicBlock::Create(Context&: *C, Name: "", Parent: AsanDtorFunction);
2290
2291 return ReturnInst::Create(C&: *C, InsertAtEnd: AsanDtorBB);
2292}
2293
2294void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
2295 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2296 ArrayRef<Constant *> MetadataInitializers) {
2297 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2298 auto &DL = M.getDataLayout();
2299
2300 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2301 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2302 Constant *Initializer = MetadataInitializers[i];
2303 GlobalVariable *G = ExtendedGlobals[i];
2304 GlobalVariable *Metadata =
2305 CreateMetadataGlobal(M, Initializer, OriginalName: G->getName());
2306 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2307 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2308 MetadataGlobals[i] = Metadata;
2309
2310 // The MSVC linker always inserts padding when linking incrementally. We
2311 // cope with that by aligning each struct to its size, which must be a power
2312 // of two.
2313 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Ty: Initializer->getType());
2314 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2315 "global metadata will not be padded appropriately");
2316 Metadata->setAlignment(assumeAligned(Value: SizeOfGlobalStruct));
2317
2318 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: "");
2319 }
2320
2321 // Update llvm.compiler.used, adding the new metadata globals. This is
2322 // needed so that during LTO these variables stay alive.
2323 if (!MetadataGlobals.empty())
2324 appendToCompilerUsed(M, Values: MetadataGlobals);
2325}
2326
2327void ModuleAddressSanitizer::instrumentGlobalsELF(
2328 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2329 ArrayRef<Constant *> MetadataInitializers,
2330 const std::string &UniqueModuleId) {
2331 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2332
2333 // Putting globals in a comdat changes the semantic and potentially cause
2334 // false negative odr violations at link time. If odr indicators are used, we
2335 // keep the comdat sections, as link time odr violations will be dectected on
2336 // the odr indicator symbols.
2337 bool UseComdatForGlobalsGC = UseOdrIndicator && !UniqueModuleId.empty();
2338
2339 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2340 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2341 GlobalVariable *G = ExtendedGlobals[i];
2342 GlobalVariable *Metadata =
2343 CreateMetadataGlobal(M, Initializer: MetadataInitializers[i], OriginalName: G->getName());
2344 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2345 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2346 MetadataGlobals[i] = Metadata;
2347
2348 if (UseComdatForGlobalsGC)
2349 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: UniqueModuleId);
2350 }
2351
2352 // Update llvm.compiler.used, adding the new metadata globals. This is
2353 // needed so that during LTO these variables stay alive.
2354 if (!MetadataGlobals.empty())
2355 appendToCompilerUsed(M, Values: MetadataGlobals);
2356
2357 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2358 // to look up the loaded image that contains it. Second, we can store in it
2359 // whether registration has already occurred, to prevent duplicate
2360 // registration.
2361 //
2362 // Common linkage ensures that there is only one global per shared library.
2363 GlobalVariable *RegisteredFlag = new GlobalVariable(
2364 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2365 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2366 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2367
2368 // Create start and stop symbols.
2369 GlobalVariable *StartELFMetadata = new GlobalVariable(
2370 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2371 "__start_" + getGlobalMetadataSection());
2372 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2373 GlobalVariable *StopELFMetadata = new GlobalVariable(
2374 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2375 "__stop_" + getGlobalMetadataSection());
2376 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2377
2378 // Create a call to register the globals with the runtime.
2379 if (ConstructorKind == AsanCtorKind::Global)
2380 IRB.CreateCall(Callee: AsanRegisterElfGlobals,
2381 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2382 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2383 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2384
2385 // We also need to unregister globals at the end, e.g., when a shared library
2386 // gets closed.
2387 if (DestructorKind != AsanDtorKind::None && !MetadataGlobals.empty()) {
2388 IRBuilder<> IrbDtor(CreateAsanModuleDtor(M));
2389 IrbDtor.CreateCall(Callee: AsanUnregisterElfGlobals,
2390 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2391 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2392 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2393 }
2394}
2395
2396void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2397 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2398 ArrayRef<Constant *> MetadataInitializers) {
2399 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2400
2401 // On recent Mach-O platforms, use a structure which binds the liveness of
2402 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2403 // created to be added to llvm.compiler.used
2404 StructType *LivenessTy = StructType::get(elt1: IntptrTy, elts: IntptrTy);
2405 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2406
2407 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2408 Constant *Initializer = MetadataInitializers[i];
2409 GlobalVariable *G = ExtendedGlobals[i];
2410 GlobalVariable *Metadata =
2411 CreateMetadataGlobal(M, Initializer, OriginalName: G->getName());
2412
2413 // On recent Mach-O platforms, we emit the global metadata in a way that
2414 // allows the linker to properly strip dead globals.
2415 auto LivenessBinder =
2416 ConstantStruct::get(T: LivenessTy, Vs: Initializer->getAggregateElement(Elt: 0u),
2417 Vs: ConstantExpr::getPointerCast(C: Metadata, Ty: IntptrTy));
2418 GlobalVariable *Liveness = new GlobalVariable(
2419 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2420 Twine("__asan_binder_") + G->getName());
2421 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2422 LivenessGlobals[i] = Liveness;
2423 }
2424
2425 // Update llvm.compiler.used, adding the new liveness globals. This is
2426 // needed so that during LTO these variables stay alive. The alternative
2427 // would be to have the linker handling the LTO symbols, but libLTO
2428 // current API does not expose access to the section for each symbol.
2429 if (!LivenessGlobals.empty())
2430 appendToCompilerUsed(M, Values: LivenessGlobals);
2431
2432 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2433 // to look up the loaded image that contains it. Second, we can store in it
2434 // whether registration has already occurred, to prevent duplicate
2435 // registration.
2436 //
2437 // common linkage ensures that there is only one global per shared library.
2438 GlobalVariable *RegisteredFlag = new GlobalVariable(
2439 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2440 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2441 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2442
2443 if (ConstructorKind == AsanCtorKind::Global)
2444 IRB.CreateCall(Callee: AsanRegisterImageGlobals,
2445 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2446
2447 // We also need to unregister globals at the end, e.g., when a shared library
2448 // gets closed.
2449 if (DestructorKind != AsanDtorKind::None) {
2450 IRBuilder<> IrbDtor(CreateAsanModuleDtor(M));
2451 IrbDtor.CreateCall(Callee: AsanUnregisterImageGlobals,
2452 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2453 }
2454}
2455
2456void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2457 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2458 ArrayRef<Constant *> MetadataInitializers) {
2459 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2460 unsigned N = ExtendedGlobals.size();
2461 assert(N > 0);
2462
2463 // On platforms that don't have a custom metadata section, we emit an array
2464 // of global metadata structures.
2465 ArrayType *ArrayOfGlobalStructTy =
2466 ArrayType::get(ElementType: MetadataInitializers[0]->getType(), NumElements: N);
2467 auto AllGlobals = new GlobalVariable(
2468 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2469 ConstantArray::get(T: ArrayOfGlobalStructTy, V: MetadataInitializers), "");
2470 if (Mapping.Scale > 3)
2471 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2472
2473 if (ConstructorKind == AsanCtorKind::Global)
2474 IRB.CreateCall(Callee: AsanRegisterGlobals,
2475 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2476 ConstantInt::get(Ty: IntptrTy, V: N)});
2477
2478 // We also need to unregister globals at the end, e.g., when a shared library
2479 // gets closed.
2480 if (DestructorKind != AsanDtorKind::None) {
2481 IRBuilder<> IrbDtor(CreateAsanModuleDtor(M));
2482 IrbDtor.CreateCall(Callee: AsanUnregisterGlobals,
2483 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2484 ConstantInt::get(Ty: IntptrTy, V: N)});
2485 }
2486}
2487
2488// This function replaces all global variables with new variables that have
2489// trailing redzones. It also creates a function that poisons
2490// redzones and inserts this function into llvm.global_ctors.
2491// Sets *CtorComdat to true if the global registration code emitted into the
2492// asan constructor is comdat-compatible.
2493void ModuleAddressSanitizer::instrumentGlobals(IRBuilder<> &IRB, Module &M,
2494 bool *CtorComdat) {
2495 // Build set of globals that are aliased by some GA, where
2496 // getExcludedAliasedGlobal(GA) returns the relevant GlobalVariable.
2497 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2498 if (CompileKernel) {
2499 for (auto &GA : M.aliases()) {
2500 if (const GlobalVariable *GV = getExcludedAliasedGlobal(GA))
2501 AliasedGlobalExclusions.insert(Ptr: GV);
2502 }
2503 }
2504
2505 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2506 for (auto &G : M.globals()) {
2507 if (!AliasedGlobalExclusions.count(Ptr: &G) && shouldInstrumentGlobal(G: &G))
2508 GlobalsToChange.push_back(Elt: &G);
2509 }
2510
2511 size_t n = GlobalsToChange.size();
2512 auto &DL = M.getDataLayout();
2513
2514 // A global is described by a structure
2515 // size_t beg;
2516 // size_t size;
2517 // size_t size_with_redzone;
2518 // const char *name;
2519 // const char *module_name;
2520 // size_t has_dynamic_init;
2521 // size_t padding_for_windows_msvc_incremental_link;
2522 // size_t odr_indicator;
2523 // We initialize an array of such structures and pass it to a run-time call.
2524 StructType *GlobalStructTy =
2525 StructType::get(elt1: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy,
2526 elts: IntptrTy, elts: IntptrTy, elts: IntptrTy);
2527 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2528 SmallVector<Constant *, 16> Initializers(n);
2529
2530 bool HasDynamicallyInitializedGlobals = false;
2531
2532 // We shouldn't merge same module names, as this string serves as unique
2533 // module ID in runtime.
2534 GlobalVariable *ModuleName =
2535 n != 0
2536 ? createPrivateGlobalForString(M, Str: M.getModuleIdentifier(),
2537 /*AllowMerging*/ false, NamePrefix: kAsanGenPrefix)
2538 : nullptr;
2539
2540 for (size_t i = 0; i < n; i++) {
2541 GlobalVariable *G = GlobalsToChange[i];
2542
2543 GlobalValue::SanitizerMetadata MD;
2544 if (G->hasSanitizerMetadata())
2545 MD = G->getSanitizerMetadata();
2546
2547 // The runtime library tries demangling symbol names in the descriptor but
2548 // functionality like __cxa_demangle may be unavailable (e.g.
2549 // -static-libstdc++). So we demangle the symbol names here.
2550 std::string NameForGlobal = G->getName().str();
2551 GlobalVariable *Name =
2552 createPrivateGlobalForString(M, Str: llvm::demangle(MangledName: NameForGlobal),
2553 /*AllowMerging*/ true, NamePrefix: kAsanGenPrefix);
2554
2555 Type *Ty = G->getValueType();
2556 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2557 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2558 Type *RightRedZoneTy = ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: RightRedzoneSize);
2559
2560 StructType *NewTy = StructType::get(elt1: Ty, elts: RightRedZoneTy);
2561 Constant *NewInitializer = ConstantStruct::get(
2562 T: NewTy, Vs: G->getInitializer(), Vs: Constant::getNullValue(Ty: RightRedZoneTy));
2563
2564 // Create a new global variable with enough space for a redzone.
2565 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2566 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2567 Linkage = GlobalValue::InternalLinkage;
2568 GlobalVariable *NewGlobal = new GlobalVariable(
2569 M, NewTy, G->isConstant(), Linkage, NewInitializer, "", G,
2570 G->getThreadLocalMode(), G->getAddressSpace());
2571 NewGlobal->copyAttributesFrom(Src: G);
2572 NewGlobal->setComdat(G->getComdat());
2573 NewGlobal->setAlignment(Align(getMinRedzoneSizeForGlobal()));
2574 // Don't fold globals with redzones. ODR violation detector and redzone
2575 // poisoning implicitly creates a dependence on the global's address, so it
2576 // is no longer valid for it to be marked unnamed_addr.
2577 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2578
2579 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2580 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2581 G->isConstant()) {
2582 auto Seq = dyn_cast<ConstantDataSequential>(Val: G->getInitializer());
2583 if (Seq && Seq->isCString())
2584 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2585 }
2586
2587 // Transfer the debug info and type metadata. The payload starts at offset
2588 // zero so we can copy the metadata over as is.
2589 NewGlobal->copyMetadata(Src: G, Offset: 0);
2590
2591 Value *Indices2[2];
2592 Indices2[0] = IRB.getInt32(C: 0);
2593 Indices2[1] = IRB.getInt32(C: 0);
2594
2595 G->replaceAllUsesWith(
2596 V: ConstantExpr::getGetElementPtr(Ty: NewTy, C: NewGlobal, IdxList: Indices2, NW: true));
2597 NewGlobal->takeName(V: G);
2598 G->eraseFromParent();
2599 NewGlobals[i] = NewGlobal;
2600
2601 Constant *ODRIndicator = ConstantPointerNull::get(T: PtrTy);
2602 GlobalValue *InstrumentedGlobal = NewGlobal;
2603
2604 bool CanUsePrivateAliases =
2605 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2606 TargetTriple.isOSBinFormatWasm();
2607 if (CanUsePrivateAliases && UsePrivateAlias) {
2608 // Create local alias for NewGlobal to avoid crash on ODR between
2609 // instrumented and non-instrumented libraries.
2610 InstrumentedGlobal =
2611 GlobalAlias::create(Linkage: GlobalValue::PrivateLinkage, Name: "", Aliasee: NewGlobal);
2612 }
2613
2614 // ODR should not happen for local linkage.
2615 if (NewGlobal->hasLocalLinkage()) {
2616 ODRIndicator =
2617 ConstantExpr::getIntToPtr(C: ConstantInt::get(Ty: IntptrTy, V: -1), Ty: PtrTy);
2618 } else if (UseOdrIndicator) {
2619 // With local aliases, we need to provide another externally visible
2620 // symbol __odr_asan_XXX to detect ODR violation.
2621 auto *ODRIndicatorSym =
2622 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2623 Constant::getNullValue(Ty: IRB.getInt8Ty()),
2624 kODRGenPrefix + NameForGlobal, nullptr,
2625 NewGlobal->getThreadLocalMode());
2626
2627 // Set meaningful attributes for indicator symbol.
2628 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2629 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2630 ODRIndicatorSym->setAlignment(Align(1));
2631 ODRIndicator = ODRIndicatorSym;
2632 }
2633
2634 Constant *Initializer = ConstantStruct::get(
2635 T: GlobalStructTy,
2636 Vs: ConstantExpr::getPointerCast(C: InstrumentedGlobal, Ty: IntptrTy),
2637 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes),
2638 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes + RightRedzoneSize),
2639 Vs: ConstantExpr::getPointerCast(C: Name, Ty: IntptrTy),
2640 Vs: ConstantExpr::getPointerCast(C: ModuleName, Ty: IntptrTy),
2641 Vs: ConstantInt::get(Ty: IntptrTy, V: MD.IsDynInit),
2642 Vs: Constant::getNullValue(Ty: IntptrTy),
2643 Vs: ConstantExpr::getPointerCast(C: ODRIndicator, Ty: IntptrTy));
2644
2645 if (ClInitializers && MD.IsDynInit)
2646 HasDynamicallyInitializedGlobals = true;
2647
2648 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2649
2650 Initializers[i] = Initializer;
2651 }
2652
2653 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2654 // ConstantMerge'ing them.
2655 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2656 for (size_t i = 0; i < n; i++) {
2657 GlobalVariable *G = NewGlobals[i];
2658 if (G->getName().empty()) continue;
2659 GlobalsToAddToUsedList.push_back(Elt: G);
2660 }
2661 appendToCompilerUsed(M, Values: ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2662
2663 if (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) {
2664 // Use COMDAT and register globals even if n == 0 to ensure that (a) the
2665 // linkage unit will only have one module constructor, and (b) the register
2666 // function will be called. The module destructor is not created when n ==
2667 // 0.
2668 *CtorComdat = true;
2669 instrumentGlobalsELF(IRB, M, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers,
2670 UniqueModuleId: getUniqueModuleId(M: &M));
2671 } else if (n == 0) {
2672 // When UseGlobalsGC is false, COMDAT can still be used if n == 0, because
2673 // all compile units will have identical module constructor/destructor.
2674 *CtorComdat = TargetTriple.isOSBinFormatELF();
2675 } else {
2676 *CtorComdat = false;
2677 if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2678 InstrumentGlobalsCOFF(IRB, M, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2679 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2680 InstrumentGlobalsMachO(IRB, M, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2681 } else {
2682 InstrumentGlobalsWithMetadataArray(IRB, M, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2683 }
2684 }
2685
2686 // Create calls for poisoning before initializers run and unpoisoning after.
2687 if (HasDynamicallyInitializedGlobals)
2688 createInitializerPoisonCalls(M, ModuleName);
2689
2690 LLVM_DEBUG(dbgs() << M);
2691}
2692
2693uint64_t
2694ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2695 constexpr uint64_t kMaxRZ = 1 << 18;
2696 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2697
2698 uint64_t RZ = 0;
2699 if (SizeInBytes <= MinRZ / 2) {
2700 // Reduce redzone size for small size objects, e.g. int, char[1]. MinRZ is
2701 // at least 32 bytes, optimize when SizeInBytes is less than or equal to
2702 // half of MinRZ.
2703 RZ = MinRZ - SizeInBytes;
2704 } else {
2705 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2706 RZ = std::clamp(val: (SizeInBytes / MinRZ / 4) * MinRZ, lo: MinRZ, hi: kMaxRZ);
2707
2708 // Round up to multiple of MinRZ.
2709 if (SizeInBytes % MinRZ)
2710 RZ += MinRZ - (SizeInBytes % MinRZ);
2711 }
2712
2713 assert((RZ + SizeInBytes) % MinRZ == 0);
2714
2715 return RZ;
2716}
2717
2718int ModuleAddressSanitizer::GetAsanVersion(const Module &M) const {
2719 int LongSize = M.getDataLayout().getPointerSizeInBits();
2720 bool isAndroid = Triple(M.getTargetTriple()).isAndroid();
2721 int Version = 8;
2722 // 32-bit Android is one version ahead because of the switch to dynamic
2723 // shadow.
2724 Version += (LongSize == 32 && isAndroid);
2725 return Version;
2726}
2727
2728bool ModuleAddressSanitizer::instrumentModule(Module &M) {
2729 initializeCallbacks(M);
2730
2731 // Create a module constructor. A destructor is created lazily because not all
2732 // platforms, and not all modules need it.
2733 if (ConstructorKind == AsanCtorKind::Global) {
2734 if (CompileKernel) {
2735 // The kernel always builds with its own runtime, and therefore does not
2736 // need the init and version check calls.
2737 AsanCtorFunction = createSanitizerCtor(M, CtorName: kAsanModuleCtorName);
2738 } else {
2739 std::string AsanVersion = std::to_string(val: GetAsanVersion(M));
2740 std::string VersionCheckName =
2741 InsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2742 std::tie(args&: AsanCtorFunction, args: std::ignore) =
2743 createSanitizerCtorAndInitFunctions(M, CtorName: kAsanModuleCtorName,
2744 InitName: kAsanInitName, /*InitArgTypes=*/{},
2745 /*InitArgs=*/{}, VersionCheckName);
2746 }
2747 }
2748
2749 bool CtorComdat = true;
2750 if (ClGlobals) {
2751 assert(AsanCtorFunction || ConstructorKind == AsanCtorKind::None);
2752 if (AsanCtorFunction) {
2753 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2754 instrumentGlobals(IRB, M, CtorComdat: &CtorComdat);
2755 } else {
2756 IRBuilder<> IRB(*C);
2757 instrumentGlobals(IRB, M, CtorComdat: &CtorComdat);
2758 }
2759 }
2760
2761 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2762
2763 // Put the constructor and destructor in comdat if both
2764 // (1) global instrumentation is not TU-specific
2765 // (2) target is ELF.
2766 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2767 if (AsanCtorFunction) {
2768 AsanCtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleCtorName));
2769 appendToGlobalCtors(M, F: AsanCtorFunction, Priority, Data: AsanCtorFunction);
2770 }
2771 if (AsanDtorFunction) {
2772 AsanDtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleDtorName));
2773 appendToGlobalDtors(M, F: AsanDtorFunction, Priority, Data: AsanDtorFunction);
2774 }
2775 } else {
2776 if (AsanCtorFunction)
2777 appendToGlobalCtors(M, F: AsanCtorFunction, Priority);
2778 if (AsanDtorFunction)
2779 appendToGlobalDtors(M, F: AsanDtorFunction, Priority);
2780 }
2781
2782 return true;
2783}
2784
2785void AddressSanitizer::initializeCallbacks(Module &M, const TargetLibraryInfo *TLI) {
2786 IRBuilder<> IRB(*C);
2787 // Create __asan_report* callbacks.
2788 // IsWrite, TypeSize and Exp are encoded in the function name.
2789 for (int Exp = 0; Exp < 2; Exp++) {
2790 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2791 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2792 const std::string ExpStr = Exp ? "exp_" : "";
2793 const std::string EndingStr = Recover ? "_noabort" : "";
2794
2795 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2796 SmallVector<Type *, 2> Args1{1, IntptrTy};
2797 AttributeList AL2;
2798 AttributeList AL1;
2799 if (Exp) {
2800 Type *ExpType = Type::getInt32Ty(C&: *C);
2801 Args2.push_back(Elt: ExpType);
2802 Args1.push_back(Elt: ExpType);
2803 if (auto AK = TLI->getExtAttrForI32Param(Signed: false)) {
2804 AL2 = AL2.addParamAttribute(C&: *C, ArgNo: 2, Kind: AK);
2805 AL1 = AL1.addParamAttribute(C&: *C, ArgNo: 1, Kind: AK);
2806 }
2807 }
2808 AsanErrorCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2809 Name: kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2810 T: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), AttributeList: AL2);
2811
2812 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2813 Name: ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2814 T: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), AttributeList: AL2);
2815
2816 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2817 AccessSizeIndex++) {
2818 const std::string Suffix = TypeStr + itostr(X: 1ULL << AccessSizeIndex);
2819 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2820 M.getOrInsertFunction(
2821 Name: kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2822 T: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), AttributeList: AL1);
2823
2824 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2825 M.getOrInsertFunction(
2826 Name: ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2827 T: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), AttributeList: AL1);
2828 }
2829 }
2830 }
2831
2832 const std::string MemIntrinCallbackPrefix =
2833 (CompileKernel && !ClKasanMemIntrinCallbackPrefix)
2834 ? std::string("")
2835 : ClMemoryAccessCallbackPrefix;
2836 AsanMemmove = M.getOrInsertFunction(Name: MemIntrinCallbackPrefix + "memmove",
2837 RetTy: PtrTy, Args: PtrTy, Args: PtrTy, Args: IntptrTy);
2838 AsanMemcpy = M.getOrInsertFunction(Name: MemIntrinCallbackPrefix + "memcpy", RetTy: PtrTy,
2839 Args: PtrTy, Args: PtrTy, Args: IntptrTy);
2840 AsanMemset = M.getOrInsertFunction(Name: MemIntrinCallbackPrefix + "memset",
2841 AttributeList: TLI->getAttrList(C, ArgNos: {1}, /*Signed=*/false),
2842 RetTy: PtrTy, Args: PtrTy, Args: IRB.getInt32Ty(), Args: IntptrTy);
2843
2844 AsanHandleNoReturnFunc =
2845 M.getOrInsertFunction(Name: kAsanHandleNoReturnName, RetTy: IRB.getVoidTy());
2846
2847 AsanPtrCmpFunction =
2848 M.getOrInsertFunction(Name: kAsanPtrCmp, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
2849 AsanPtrSubFunction =
2850 M.getOrInsertFunction(Name: kAsanPtrSub, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
2851 if (Mapping.InGlobal)
2852 AsanShadowGlobal = M.getOrInsertGlobal(Name: "__asan_shadow",
2853 Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: 0));
2854
2855 AMDGPUAddressShared =
2856 M.getOrInsertFunction(Name: kAMDGPUAddressSharedName, RetTy: IRB.getInt1Ty(), Args: PtrTy);
2857 AMDGPUAddressPrivate =
2858 M.getOrInsertFunction(Name: kAMDGPUAddressPrivateName, RetTy: IRB.getInt1Ty(), Args: PtrTy);
2859}
2860
2861bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
2862 // For each NSObject descendant having a +load method, this method is invoked
2863 // by the ObjC runtime before any of the static constructors is called.
2864 // Therefore we need to instrument such methods with a call to __asan_init
2865 // at the beginning in order to initialize our runtime before any access to
2866 // the shadow memory.
2867 // We cannot just ignore these methods, because they may call other
2868 // instrumented functions.
2869 if (F.getName().contains(Other: " load]")) {
2870 FunctionCallee AsanInitFunction =
2871 declareSanitizerInitFunction(M&: *F.getParent(), InitName: kAsanInitName, InitArgTypes: {});
2872 IRBuilder<> IRB(&F.front(), F.front().begin());
2873 IRB.CreateCall(Callee: AsanInitFunction, Args: {});
2874 return true;
2875 }
2876 return false;
2877}
2878
2879bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
2880 // Generate code only when dynamic addressing is needed.
2881 if (Mapping.Offset != kDynamicShadowSentinel)
2882 return false;
2883
2884 IRBuilder<> IRB(&F.front().front());
2885 if (Mapping.InGlobal) {
2886 if (ClWithIfuncSuppressRemat) {
2887 // An empty inline asm with input reg == output reg.
2888 // An opaque pointer-to-int cast, basically.
2889 InlineAsm *Asm = InlineAsm::get(
2890 Ty: FunctionType::get(Result: IntptrTy, Params: {AsanShadowGlobal->getType()}, isVarArg: false),
2891 AsmString: StringRef(""), Constraints: StringRef("=r,0"),
2892 /*hasSideEffects=*/false);
2893 LocalDynamicShadow =
2894 IRB.CreateCall(Callee: Asm, Args: {AsanShadowGlobal}, Name: ".asan.shadow");
2895 } else {
2896 LocalDynamicShadow =
2897 IRB.CreatePointerCast(V: AsanShadowGlobal, DestTy: IntptrTy, Name: ".asan.shadow");
2898 }
2899 } else {
2900 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
2901 Name: kAsanShadowMemoryDynamicAddress, Ty: IntptrTy);
2902 LocalDynamicShadow = IRB.CreateLoad(Ty: IntptrTy, Ptr: GlobalDynamicAddress);
2903 }
2904 return true;
2905}
2906
2907void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
2908 // Find the one possible call to llvm.localescape and pre-mark allocas passed
2909 // to it as uninteresting. This assumes we haven't started processing allocas
2910 // yet. This check is done up front because iterating the use list in
2911 // isInterestingAlloca would be algorithmically slower.
2912 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
2913
2914 // Try to get the declaration of llvm.localescape. If it's not in the module,
2915 // we can exit early.
2916 if (!F.getParent()->getFunction(Name: "llvm.localescape")) return;
2917
2918 // Look for a call to llvm.localescape call in the entry block. It can't be in
2919 // any other block.
2920 for (Instruction &I : F.getEntryBlock()) {
2921 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &I);
2922 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
2923 // We found a call. Mark all the allocas passed in as uninteresting.
2924 for (Value *Arg : II->args()) {
2925 AllocaInst *AI = dyn_cast<AllocaInst>(Val: Arg->stripPointerCasts());
2926 assert(AI && AI->isStaticAlloca() &&
2927 "non-static alloca arg to localescape");
2928 ProcessedAllocas[AI] = false;
2929 }
2930 break;
2931 }
2932 }
2933}
2934
2935bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
2936 bool ShouldInstrument =
2937 ClDebugMin < 0 || ClDebugMax < 0 ||
2938 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
2939 Instrumented++;
2940 return !ShouldInstrument;
2941}
2942
2943bool AddressSanitizer::instrumentFunction(Function &F,
2944 const TargetLibraryInfo *TLI) {
2945 if (F.empty())
2946 return false;
2947 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
2948 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName()) return false;
2949 if (F.getName().starts_with(Prefix: "__asan_")) return false;
2950 if (F.isPresplitCoroutine())
2951 return false;
2952
2953 bool FunctionModified = false;
2954
2955 // If needed, insert __asan_init before checking for SanitizeAddress attr.
2956 // This function needs to be called even if the function body is not
2957 // instrumented.
2958 if (maybeInsertAsanInitAtFunctionEntry(F))
2959 FunctionModified = true;
2960
2961 // Leave if the function doesn't need instrumentation.
2962 if (!F.hasFnAttribute(Kind: Attribute::SanitizeAddress)) return FunctionModified;
2963
2964 if (F.hasFnAttribute(Kind: Attribute::DisableSanitizerInstrumentation))
2965 return FunctionModified;
2966
2967 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
2968
2969 initializeCallbacks(M&: *F.getParent(), TLI);
2970
2971 FunctionStateRAII CleanupObj(this);
2972
2973 RuntimeCallInserter RTCI(F);
2974
2975 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
2976
2977 // We can't instrument allocas used with llvm.localescape. Only static allocas
2978 // can be passed to that intrinsic.
2979 markEscapedLocalAllocas(F);
2980
2981 // We want to instrument every address only once per basic block (unless there
2982 // are calls between uses).
2983 SmallPtrSet<Value *, 16> TempsToInstrument;
2984 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
2985 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
2986 SmallVector<Instruction *, 8> NoReturnCalls;
2987 SmallVector<BasicBlock *, 16> AllBlocks;
2988 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
2989
2990 // Fill the set of memory operations to instrument.
2991 for (auto &BB : F) {
2992 AllBlocks.push_back(Elt: &BB);
2993 TempsToInstrument.clear();
2994 int NumInsnsPerBB = 0;
2995 for (auto &Inst : BB) {
2996 if (LooksLikeCodeInBug11395(I: &Inst)) return false;
2997 // Skip instructions inserted by another instrumentation.
2998 if (Inst.hasMetadata(KindID: LLVMContext::MD_nosanitize))
2999 continue;
3000 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
3001 getInterestingMemoryOperands(I: &Inst, Interesting&: InterestingOperands);
3002
3003 if (!InterestingOperands.empty()) {
3004 for (auto &Operand : InterestingOperands) {
3005 if (ClOpt && ClOptSameTemp) {
3006 Value *Ptr = Operand.getPtr();
3007 // If we have a mask, skip instrumentation if we've already
3008 // instrumented the full object. But don't add to TempsToInstrument
3009 // because we might get another load/store with a different mask.
3010 if (Operand.MaybeMask) {
3011 if (TempsToInstrument.count(Ptr))
3012 continue; // We've seen this (whole) temp in the current BB.
3013 } else {
3014 if (!TempsToInstrument.insert(Ptr).second)
3015 continue; // We've seen this temp in the current BB.
3016 }
3017 }
3018 OperandsToInstrument.push_back(Elt: Operand);
3019 NumInsnsPerBB++;
3020 }
3021 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
3022 isInterestingPointerComparison(I: &Inst)) ||
3023 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
3024 isInterestingPointerSubtraction(I: &Inst))) {
3025 PointerComparisonsOrSubtracts.push_back(Elt: &Inst);
3026 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: &Inst)) {
3027 // ok, take it.
3028 IntrinToInstrument.push_back(Elt: MI);
3029 NumInsnsPerBB++;
3030 } else {
3031 if (auto *CB = dyn_cast<CallBase>(Val: &Inst)) {
3032 // A call inside BB.
3033 TempsToInstrument.clear();
3034 if (CB->doesNotReturn())
3035 NoReturnCalls.push_back(Elt: CB);
3036 }
3037 if (CallInst *CI = dyn_cast<CallInst>(Val: &Inst))
3038 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
3039 }
3040 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
3041 }
3042 }
3043
3044 bool UseCalls = (InstrumentationWithCallsThreshold >= 0 &&
3045 OperandsToInstrument.size() + IntrinToInstrument.size() >
3046 (unsigned)InstrumentationWithCallsThreshold);
3047 const DataLayout &DL = F.getDataLayout();
3048 ObjectSizeOpts ObjSizeOpts;
3049 ObjSizeOpts.RoundToAlign = true;
3050 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext(), ObjSizeOpts);
3051
3052 // Instrument.
3053 int NumInstrumented = 0;
3054 for (auto &Operand : OperandsToInstrument) {
3055 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3056 instrumentMop(ObjSizeVis, O&: Operand, UseCalls,
3057 DL: F.getDataLayout(), RTCI);
3058 FunctionModified = true;
3059 }
3060 for (auto *Inst : IntrinToInstrument) {
3061 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3062 instrumentMemIntrinsic(MI: Inst, RTCI);
3063 FunctionModified = true;
3064 }
3065
3066 FunctionStackPoisoner FSP(F, *this, RTCI);
3067 bool ChangedStack = FSP.runOnFunction();
3068
3069 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
3070 // See e.g. https://github.com/google/sanitizers/issues/37
3071 for (auto *CI : NoReturnCalls) {
3072 IRBuilder<> IRB(CI);
3073 RTCI.createRuntimeCall(IRB, Callee: AsanHandleNoReturnFunc, Args: {});
3074 }
3075
3076 for (auto *Inst : PointerComparisonsOrSubtracts) {
3077 instrumentPointerComparisonOrSubtraction(I: Inst, RTCI);
3078 FunctionModified = true;
3079 }
3080
3081 if (ChangedStack || !NoReturnCalls.empty())
3082 FunctionModified = true;
3083
3084 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
3085 << F << "\n");
3086
3087 return FunctionModified;
3088}
3089
3090// Workaround for bug 11395: we don't want to instrument stack in functions
3091// with large assembly blobs (32-bit only), otherwise reg alloc may crash.
3092// FIXME: remove once the bug 11395 is fixed.
3093bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
3094 if (LongSize != 32) return false;
3095 CallInst *CI = dyn_cast<CallInst>(Val: I);
3096 if (!CI || !CI->isInlineAsm()) return false;
3097 if (CI->arg_size() <= 5)
3098 return false;
3099 // We have inline assembly with quite a few arguments.
3100 return true;
3101}
3102
3103void FunctionStackPoisoner::initializeCallbacks(Module &M) {
3104 IRBuilder<> IRB(*C);
3105 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always ||
3106 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3107 const char *MallocNameTemplate =
3108 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always
3109 ? kAsanStackMallocAlwaysNameTemplate
3110 : kAsanStackMallocNameTemplate;
3111 for (int Index = 0; Index <= kMaxAsanStackMallocSizeClass; Index++) {
3112 std::string Suffix = itostr(X: Index);
3113 AsanStackMallocFunc[Index] = M.getOrInsertFunction(
3114 Name: MallocNameTemplate + Suffix, RetTy: IntptrTy, Args: IntptrTy);
3115 AsanStackFreeFunc[Index] =
3116 M.getOrInsertFunction(Name: kAsanStackFreeNameTemplate + Suffix,
3117 RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
3118 }
3119 }
3120 if (ASan.UseAfterScope) {
3121 AsanPoisonStackMemoryFunc = M.getOrInsertFunction(
3122 Name: kAsanPoisonStackMemoryName, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
3123 AsanUnpoisonStackMemoryFunc = M.getOrInsertFunction(
3124 Name: kAsanUnpoisonStackMemoryName, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
3125 }
3126
3127 for (size_t Val : {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0xf1, 0xf2,
3128 0xf3, 0xf5, 0xf8}) {
3129 std::ostringstream Name;
3130 Name << kAsanSetShadowPrefix;
3131 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
3132 AsanSetShadowFunc[Val] =
3133 M.getOrInsertFunction(Name: Name.str(), RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
3134 }
3135
3136 AsanAllocaPoisonFunc = M.getOrInsertFunction(
3137 Name: kAsanAllocaPoison, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
3138 AsanAllocasUnpoisonFunc = M.getOrInsertFunction(
3139 Name: kAsanAllocasUnpoison, RetTy: IRB.getVoidTy(), Args: IntptrTy, Args: IntptrTy);
3140}
3141
3142void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
3143 ArrayRef<uint8_t> ShadowBytes,
3144 size_t Begin, size_t End,
3145 IRBuilder<> &IRB,
3146 Value *ShadowBase) {
3147 if (Begin >= End)
3148 return;
3149
3150 const size_t LargestStoreSizeInBytes =
3151 std::min<size_t>(a: sizeof(uint64_t), b: ASan.LongSize / 8);
3152
3153 const bool IsLittleEndian = F.getDataLayout().isLittleEndian();
3154
3155 // Poison given range in shadow using larges store size with out leading and
3156 // trailing zeros in ShadowMask. Zeros never change, so they need neither
3157 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
3158 // middle of a store.
3159 for (size_t i = Begin; i < End;) {
3160 if (!ShadowMask[i]) {
3161 assert(!ShadowBytes[i]);
3162 ++i;
3163 continue;
3164 }
3165
3166 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
3167 // Fit store size into the range.
3168 while (StoreSizeInBytes > End - i)
3169 StoreSizeInBytes /= 2;
3170
3171 // Minimize store size by trimming trailing zeros.
3172 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
3173 while (j <= StoreSizeInBytes / 2)
3174 StoreSizeInBytes /= 2;
3175 }
3176
3177 uint64_t Val = 0;
3178 for (size_t j = 0; j < StoreSizeInBytes; j++) {
3179 if (IsLittleEndian)
3180 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
3181 else
3182 Val = (Val << 8) | ShadowBytes[i + j];
3183 }
3184
3185 Value *Ptr = IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i));
3186 Value *Poison = IRB.getIntN(N: StoreSizeInBytes * 8, C: Val);
3187 IRB.CreateAlignedStore(
3188 Val: Poison, Ptr: IRB.CreateIntToPtr(V: Ptr, DestTy: PointerType::getUnqual(C&: Poison->getContext())),
3189 Align: Align(1));
3190
3191 i += StoreSizeInBytes;
3192 }
3193}
3194
3195void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3196 ArrayRef<uint8_t> ShadowBytes,
3197 IRBuilder<> &IRB, Value *ShadowBase) {
3198 copyToShadow(ShadowMask, ShadowBytes, Begin: 0, End: ShadowMask.size(), IRB, ShadowBase);
3199}
3200
3201void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3202 ArrayRef<uint8_t> ShadowBytes,
3203 size_t Begin, size_t End,
3204 IRBuilder<> &IRB, Value *ShadowBase) {
3205 assert(ShadowMask.size() == ShadowBytes.size());
3206 size_t Done = Begin;
3207 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
3208 if (!ShadowMask[i]) {
3209 assert(!ShadowBytes[i]);
3210 continue;
3211 }
3212 uint8_t Val = ShadowBytes[i];
3213 if (!AsanSetShadowFunc[Val])
3214 continue;
3215
3216 // Skip same values.
3217 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
3218 }
3219
3220 if (j - i >= ASan.MaxInlinePoisoningSize) {
3221 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End: i, IRB, ShadowBase);
3222 RTCI.createRuntimeCall(
3223 IRB, Callee: AsanSetShadowFunc[Val],
3224 Args: {IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i)),
3225 ConstantInt::get(Ty: IntptrTy, V: j - i)});
3226 Done = j;
3227 }
3228 }
3229
3230 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End, IRB, ShadowBase);
3231}
3232
3233// Fake stack allocator (asan_fake_stack.h) has 11 size classes
3234// for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
3235static int StackMallocSizeClass(uint64_t LocalStackSize) {
3236 assert(LocalStackSize <= kMaxStackMallocSize);
3237 uint64_t MaxSize = kMinStackMallocSize;
3238 for (int i = 0;; i++, MaxSize *= 2)
3239 if (LocalStackSize <= MaxSize) return i;
3240 llvm_unreachable("impossible LocalStackSize");
3241}
3242
3243void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
3244 Instruction *CopyInsertPoint = &F.front().front();
3245 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
3246 // Insert after the dynamic shadow location is determined
3247 CopyInsertPoint = CopyInsertPoint->getNextNode();
3248 assert(CopyInsertPoint);
3249 }
3250 IRBuilder<> IRB(CopyInsertPoint);
3251 const DataLayout &DL = F.getDataLayout();
3252 for (Argument &Arg : F.args()) {
3253 if (Arg.hasByValAttr()) {
3254 Type *Ty = Arg.getParamByValType();
3255 const Align Alignment =
3256 DL.getValueOrABITypeAlignment(Alignment: Arg.getParamAlign(), Ty);
3257
3258 AllocaInst *AI = IRB.CreateAlloca(
3259 Ty, ArraySize: nullptr,
3260 Name: (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
3261 ".byval");
3262 AI->setAlignment(Alignment);
3263 Arg.replaceAllUsesWith(V: AI);
3264
3265 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3266 IRB.CreateMemCpy(Dst: AI, DstAlign: Alignment, Src: &Arg, SrcAlign: Alignment, Size: AllocSize);
3267 }
3268 }
3269}
3270
3271PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
3272 Value *ValueIfTrue,
3273 Instruction *ThenTerm,
3274 Value *ValueIfFalse) {
3275 PHINode *PHI = IRB.CreatePHI(Ty: IntptrTy, NumReservedValues: 2);
3276 BasicBlock *CondBlock = cast<Instruction>(Val: Cond)->getParent();
3277 PHI->addIncoming(V: ValueIfFalse, BB: CondBlock);
3278 BasicBlock *ThenBlock = ThenTerm->getParent();
3279 PHI->addIncoming(V: ValueIfTrue, BB: ThenBlock);
3280 return PHI;
3281}
3282
3283Value *FunctionStackPoisoner::createAllocaForLayout(
3284 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
3285 AllocaInst *Alloca;
3286 if (Dynamic) {
3287 Alloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(),
3288 ArraySize: ConstantInt::get(Ty: IRB.getInt64Ty(), V: L.FrameSize),
3289 Name: "MyAlloca");
3290 } else {
3291 Alloca = IRB.CreateAlloca(Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: L.FrameSize),
3292 ArraySize: nullptr, Name: "MyAlloca");
3293 assert(Alloca->isStaticAlloca());
3294 }
3295 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
3296 uint64_t FrameAlignment = std::max(a: L.FrameAlignment, b: uint64_t(ClRealignStack));
3297 Alloca->setAlignment(Align(FrameAlignment));
3298 return IRB.CreatePointerCast(V: Alloca, DestTy: IntptrTy);
3299}
3300
3301void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
3302 BasicBlock &FirstBB = *F.begin();
3303 IRBuilder<> IRB(dyn_cast<Instruction>(Val: FirstBB.begin()));
3304 DynamicAllocaLayout = IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr);
3305 IRB.CreateStore(Val: Constant::getNullValue(Ty: IntptrTy), Ptr: DynamicAllocaLayout);
3306 DynamicAllocaLayout->setAlignment(Align(32));
3307}
3308
3309void FunctionStackPoisoner::processDynamicAllocas() {
3310 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
3311 assert(DynamicAllocaPoisonCallVec.empty());
3312 return;
3313 }
3314
3315 // Insert poison calls for lifetime intrinsics for dynamic allocas.
3316 for (const auto &APC : DynamicAllocaPoisonCallVec) {
3317 assert(APC.InsBefore);
3318 assert(APC.AI);
3319 assert(ASan.isInterestingAlloca(*APC.AI));
3320 assert(!APC.AI->isStaticAlloca());
3321
3322 IRBuilder<> IRB(APC.InsBefore);
3323 poisonAlloca(V: APC.AI, Size: APC.Size, IRB, DoPoison: APC.DoPoison);
3324 // Dynamic allocas will be unpoisoned unconditionally below in
3325 // unpoisonDynamicAllocas.
3326 // Flag that we need unpoison static allocas.
3327 }
3328
3329 // Handle dynamic allocas.
3330 createDynamicAllocasInitStorage();
3331 for (auto &AI : DynamicAllocaVec)
3332 handleDynamicAllocaCall(AI);
3333 unpoisonDynamicAllocas();
3334}
3335
3336/// Collect instructions in the entry block after \p InsBefore which initialize
3337/// permanent storage for a function argument. These instructions must remain in
3338/// the entry block so that uninitialized values do not appear in backtraces. An
3339/// added benefit is that this conserves spill slots. This does not move stores
3340/// before instrumented / "interesting" allocas.
3341static void findStoresToUninstrumentedArgAllocas(
3342 AddressSanitizer &ASan, Instruction &InsBefore,
3343 SmallVectorImpl<Instruction *> &InitInsts) {
3344 Instruction *Start = InsBefore.getNextNonDebugInstruction();
3345 for (Instruction *It = Start; It; It = It->getNextNonDebugInstruction()) {
3346 // Argument initialization looks like:
3347 // 1) store <Argument>, <Alloca> OR
3348 // 2) <CastArgument> = cast <Argument> to ...
3349 // store <CastArgument> to <Alloca>
3350 // Do not consider any other kind of instruction.
3351 //
3352 // Note: This covers all known cases, but may not be exhaustive. An
3353 // alternative to pattern-matching stores is to DFS over all Argument uses:
3354 // this might be more general, but is probably much more complicated.
3355 if (isa<AllocaInst>(Val: It) || isa<CastInst>(Val: It))
3356 continue;
3357 if (auto *Store = dyn_cast<StoreInst>(Val: It)) {
3358 // The store destination must be an alloca that isn't interesting for
3359 // ASan to instrument. These are moved up before InsBefore, and they're
3360 // not interesting because allocas for arguments can be mem2reg'd.
3361 auto *Alloca = dyn_cast<AllocaInst>(Val: Store->getPointerOperand());
3362 if (!Alloca || ASan.isInterestingAlloca(AI: *Alloca))
3363 continue;
3364
3365 Value *Val = Store->getValueOperand();
3366 bool IsDirectArgInit = isa<Argument>(Val);
3367 bool IsArgInitViaCast =
3368 isa<CastInst>(Val) &&
3369 isa<Argument>(Val: cast<CastInst>(Val)->getOperand(i_nocapture: 0)) &&
3370 // Check that the cast appears directly before the store. Otherwise
3371 // moving the cast before InsBefore may break the IR.
3372 Val == It->getPrevNonDebugInstruction();
3373 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3374 if (!IsArgInit)
3375 continue;
3376
3377 if (IsArgInitViaCast)
3378 InitInsts.push_back(Elt: cast<Instruction>(Val));
3379 InitInsts.push_back(Elt: Store);
3380 continue;
3381 }
3382
3383 // Do not reorder past unknown instructions: argument initialization should
3384 // only involve casts and stores.
3385 return;
3386 }
3387}
3388
3389void FunctionStackPoisoner::processStaticAllocas() {
3390 if (AllocaVec.empty()) {
3391 assert(StaticAllocaPoisonCallVec.empty());
3392 return;
3393 }
3394
3395 int StackMallocIdx = -1;
3396 DebugLoc EntryDebugLocation;
3397 if (auto SP = F.getSubprogram())
3398 EntryDebugLocation =
3399 DILocation::get(Context&: SP->getContext(), Line: SP->getScopeLine(), Column: 0, Scope: SP);
3400
3401 Instruction *InsBefore = AllocaVec[0];
3402 IRBuilder<> IRB(InsBefore);
3403
3404 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3405 // debug info is broken, because only entry-block allocas are treated as
3406 // regular stack slots.
3407 auto InsBeforeB = InsBefore->getParent();
3408 assert(InsBeforeB == &F.getEntryBlock());
3409 for (auto *AI : StaticAllocasToMoveUp)
3410 if (AI->getParent() == InsBeforeB)
3411 AI->moveBefore(MovePos: InsBefore);
3412
3413 // Move stores of arguments into entry-block allocas as well. This prevents
3414 // extra stack slots from being generated (to house the argument values until
3415 // they can be stored into the allocas). This also prevents uninitialized
3416 // values from being shown in backtraces.
3417 SmallVector<Instruction *, 8> ArgInitInsts;
3418 findStoresToUninstrumentedArgAllocas(ASan, InsBefore&: *InsBefore, InitInsts&: ArgInitInsts);
3419 for (Instruction *ArgInitInst : ArgInitInsts)
3420 ArgInitInst->moveBefore(MovePos: InsBefore);
3421
3422 // If we have a call to llvm.localescape, keep it in the entry block.
3423 if (LocalEscapeCall) LocalEscapeCall->moveBefore(MovePos: InsBefore);
3424
3425 SmallVector<ASanStackVariableDescription, 16> SVD;
3426 SVD.reserve(N: AllocaVec.size());
3427 for (AllocaInst *AI : AllocaVec) {
3428 ASanStackVariableDescription D = {.Name: AI->getName().data(),
3429 .Size: ASan.getAllocaSizeInBytes(AI: *AI),
3430 .LifetimeSize: 0,
3431 .Alignment: AI->getAlign().value(),
3432 .AI: AI,
3433 .Offset: 0,
3434 .Line: 0};
3435 SVD.push_back(Elt: D);
3436 }
3437
3438 // Minimal header size (left redzone) is 4 pointers,
3439 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3440 uint64_t Granularity = 1ULL << Mapping.Scale;
3441 uint64_t MinHeaderSize = std::max(a: (uint64_t)ASan.LongSize / 2, b: Granularity);
3442 const ASanStackFrameLayout &L =
3443 ComputeASanStackFrameLayout(Vars&: SVD, Granularity, MinHeaderSize);
3444
3445 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3446 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3447 for (auto &Desc : SVD)
3448 AllocaToSVDMap[Desc.AI] = &Desc;
3449
3450 // Update SVD with information from lifetime intrinsics.
3451 for (const auto &APC : StaticAllocaPoisonCallVec) {
3452 assert(APC.InsBefore);
3453 assert(APC.AI);
3454 assert(ASan.isInterestingAlloca(*APC.AI));
3455 assert(APC.AI->isStaticAlloca());
3456
3457 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3458 Desc.LifetimeSize = Desc.Size;
3459 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3460 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3461 if (LifetimeLoc->getFile() == FnLoc->getFile())
3462 if (unsigned Line = LifetimeLoc->getLine())
3463 Desc.Line = std::min(a: Desc.Line ? Desc.Line : Line, b: Line);
3464 }
3465 }
3466 }
3467
3468 auto DescriptionString = ComputeASanStackFrameDescription(Vars: SVD);
3469 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3470 uint64_t LocalStackSize = L.FrameSize;
3471 bool DoStackMalloc =
3472 ASan.UseAfterReturn != AsanDetectStackUseAfterReturnMode::Never &&
3473 !ASan.CompileKernel && LocalStackSize <= kMaxStackMallocSize;
3474 bool DoDynamicAlloca = ClDynamicAllocaStack;
3475 // Don't do dynamic alloca or stack malloc if:
3476 // 1) There is inline asm: too often it makes assumptions on which registers
3477 // are available.
3478 // 2) There is a returns_twice call (typically setjmp), which is
3479 // optimization-hostile, and doesn't play well with introduced indirect
3480 // register-relative calculation of local variable addresses.
3481 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3482 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3483
3484 Value *StaticAlloca =
3485 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, Dynamic: false);
3486
3487 Value *FakeStack;
3488 Value *LocalStackBase;
3489 Value *LocalStackBaseAlloca;
3490 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3491
3492 if (DoStackMalloc) {
3493 LocalStackBaseAlloca =
3494 IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr, Name: "asan_local_stack_base");
3495 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3496 // void *FakeStack = __asan_option_detect_stack_use_after_return
3497 // ? __asan_stack_malloc_N(LocalStackSize)
3498 // : nullptr;
3499 // void *LocalStackBase = (FakeStack) ? FakeStack :
3500 // alloca(LocalStackSize);
3501 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3502 Name: kAsanOptionDetectUseAfterReturn, Ty: IRB.getInt32Ty());
3503 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3504 LHS: IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: OptionDetectUseAfterReturn),
3505 RHS: Constant::getNullValue(Ty: IRB.getInt32Ty()));
3506 Instruction *Term =
3507 SplitBlockAndInsertIfThen(Cond: UseAfterReturnIsEnabled, SplitBefore: InsBefore, Unreachable: false);
3508 IRBuilder<> IRBIf(Term);
3509 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3510 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3511 Value *FakeStackValue =
3512 RTCI.createRuntimeCall(IRB&: IRBIf, Callee: AsanStackMallocFunc[StackMallocIdx],
3513 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3514 IRB.SetInsertPoint(InsBefore);
3515 FakeStack = createPHI(IRB, Cond: UseAfterReturnIsEnabled, ValueIfTrue: FakeStackValue, ThenTerm: Term,
3516 ValueIfFalse: ConstantInt::get(Ty: IntptrTy, V: 0));
3517 } else {
3518 // assert(ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode:Always)
3519 // void *FakeStack = __asan_stack_malloc_N(LocalStackSize);
3520 // void *LocalStackBase = (FakeStack) ? FakeStack :
3521 // alloca(LocalStackSize);
3522 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3523 FakeStack =
3524 RTCI.createRuntimeCall(IRB, Callee: AsanStackMallocFunc[StackMallocIdx],
3525 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3526 }
3527 Value *NoFakeStack =
3528 IRB.CreateICmpEQ(LHS: FakeStack, RHS: Constant::getNullValue(Ty: IntptrTy));
3529 Instruction *Term =
3530 SplitBlockAndInsertIfThen(Cond: NoFakeStack, SplitBefore: InsBefore, Unreachable: false);
3531 IRBuilder<> IRBIf(Term);
3532 Value *AllocaValue =
3533 DoDynamicAlloca ? createAllocaForLayout(IRB&: IRBIf, L, Dynamic: true) : StaticAlloca;
3534
3535 IRB.SetInsertPoint(InsBefore);
3536 LocalStackBase = createPHI(IRB, Cond: NoFakeStack, ValueIfTrue: AllocaValue, ThenTerm: Term, ValueIfFalse: FakeStack);
3537 IRB.CreateStore(Val: LocalStackBase, Ptr: LocalStackBaseAlloca);
3538 DIExprFlags |= DIExpression::DerefBefore;
3539 } else {
3540 // void *FakeStack = nullptr;
3541 // void *LocalStackBase = alloca(LocalStackSize);
3542 FakeStack = ConstantInt::get(Ty: IntptrTy, V: 0);
3543 LocalStackBase =
3544 DoDynamicAlloca ? createAllocaForLayout(IRB, L, Dynamic: true) : StaticAlloca;
3545 LocalStackBaseAlloca = LocalStackBase;
3546 }
3547
3548 // It shouldn't matter whether we pass an `alloca` or a `ptrtoint` as the
3549 // dbg.declare address opereand, but passing a `ptrtoint` seems to confuse
3550 // later passes and can result in dropped variable coverage in debug info.
3551 Value *LocalStackBaseAllocaPtr =
3552 isa<PtrToIntInst>(Val: LocalStackBaseAlloca)
3553 ? cast<PtrToIntInst>(Val: LocalStackBaseAlloca)->getPointerOperand()
3554 : LocalStackBaseAlloca;
3555 assert(isa<AllocaInst>(LocalStackBaseAllocaPtr) &&
3556 "Variable descriptions relative to ASan stack base will be dropped");
3557
3558 // Replace Alloca instructions with base+offset.
3559 for (const auto &Desc : SVD) {
3560 AllocaInst *AI = Desc.AI;
3561 replaceDbgDeclare(Address: AI, NewAddress: LocalStackBaseAllocaPtr, Builder&: DIB, DIExprFlags,
3562 Offset: Desc.Offset);
3563 Value *NewAllocaPtr = IRB.CreateIntToPtr(
3564 V: IRB.CreateAdd(LHS: LocalStackBase, RHS: ConstantInt::get(Ty: IntptrTy, V: Desc.Offset)),
3565 DestTy: AI->getType());
3566 AI->replaceAllUsesWith(V: NewAllocaPtr);
3567 }
3568
3569 // The left-most redzone has enough space for at least 4 pointers.
3570 // Write the Magic value to redzone[0].
3571 Value *BasePlus0 = IRB.CreateIntToPtr(V: LocalStackBase, DestTy: IntptrPtrTy);
3572 IRB.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kCurrentStackFrameMagic),
3573 Ptr: BasePlus0);
3574 // Write the frame description constant to redzone[1].
3575 Value *BasePlus1 = IRB.CreateIntToPtr(
3576 V: IRB.CreateAdd(LHS: LocalStackBase,
3577 RHS: ConstantInt::get(Ty: IntptrTy, V: ASan.LongSize / 8)),
3578 DestTy: IntptrPtrTy);
3579 GlobalVariable *StackDescriptionGlobal =
3580 createPrivateGlobalForString(M&: *F.getParent(), Str: DescriptionString,
3581 /*AllowMerging*/ true, NamePrefix: kAsanGenPrefix);
3582 Value *Description = IRB.CreatePointerCast(V: StackDescriptionGlobal, DestTy: IntptrTy);
3583 IRB.CreateStore(Val: Description, Ptr: BasePlus1);
3584 // Write the PC to redzone[2].
3585 Value *BasePlus2 = IRB.CreateIntToPtr(
3586 V: IRB.CreateAdd(LHS: LocalStackBase,
3587 RHS: ConstantInt::get(Ty: IntptrTy, V: 2 * ASan.LongSize / 8)),
3588 DestTy: IntptrPtrTy);
3589 IRB.CreateStore(Val: IRB.CreatePointerCast(V: &F, DestTy: IntptrTy), Ptr: BasePlus2);
3590
3591 const auto &ShadowAfterScope = GetShadowBytesAfterScope(Vars: SVD, Layout: L);
3592
3593 // Poison the stack red zones at the entry.
3594 Value *ShadowBase = ASan.memToShadow(Shadow: LocalStackBase, IRB);
3595 // As mask we must use most poisoned case: red zones and after scope.
3596 // As bytes we can use either the same or just red zones only.
3597 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowAfterScope, IRB, ShadowBase);
3598
3599 if (!StaticAllocaPoisonCallVec.empty()) {
3600 const auto &ShadowInScope = GetShadowBytes(Vars: SVD, Layout: L);
3601
3602 // Poison static allocas near lifetime intrinsics.
3603 for (const auto &APC : StaticAllocaPoisonCallVec) {
3604 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3605 assert(Desc.Offset % L.Granularity == 0);
3606 size_t Begin = Desc.Offset / L.Granularity;
3607 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3608
3609 IRBuilder<> IRB(APC.InsBefore);
3610 copyToShadow(ShadowMask: ShadowAfterScope,
3611 ShadowBytes: APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3612 IRB, ShadowBase);
3613 }
3614 }
3615
3616 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3617 SmallVector<uint8_t, 64> ShadowAfterReturn;
3618
3619 // (Un)poison the stack before all ret instructions.
3620 for (Instruction *Ret : RetVec) {
3621 IRBuilder<> IRBRet(Ret);
3622 // Mark the current frame as retired.
3623 IRBRet.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kRetiredStackFrameMagic),
3624 Ptr: BasePlus0);
3625 if (DoStackMalloc) {
3626 assert(StackMallocIdx >= 0);
3627 // if FakeStack != 0 // LocalStackBase == FakeStack
3628 // // In use-after-return mode, poison the whole stack frame.
3629 // if StackMallocIdx <= 4
3630 // // For small sizes inline the whole thing:
3631 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3632 // **SavedFlagPtr(FakeStack) = 0
3633 // else
3634 // __asan_stack_free_N(FakeStack, LocalStackSize)
3635 // else
3636 // <This is not a fake stack; unpoison the redzones>
3637 Value *Cmp =
3638 IRBRet.CreateICmpNE(LHS: FakeStack, RHS: Constant::getNullValue(Ty: IntptrTy));
3639 Instruction *ThenTerm, *ElseTerm;
3640 SplitBlockAndInsertIfThenElse(Cond: Cmp, SplitBefore: Ret, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm);
3641
3642 IRBuilder<> IRBPoison(ThenTerm);
3643 if (ASan.MaxInlinePoisoningSize != 0 && StackMallocIdx <= 4) {
3644 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3645 ShadowAfterReturn.resize(N: ClassSize / L.Granularity,
3646 NV: kAsanStackUseAfterReturnMagic);
3647 copyToShadow(ShadowMask: ShadowAfterReturn, ShadowBytes: ShadowAfterReturn, IRB&: IRBPoison,
3648 ShadowBase);
3649 Value *SavedFlagPtrPtr = IRBPoison.CreateAdd(
3650 LHS: FakeStack,
3651 RHS: ConstantInt::get(Ty: IntptrTy, V: ClassSize - ASan.LongSize / 8));
3652 Value *SavedFlagPtr = IRBPoison.CreateLoad(
3653 Ty: IntptrTy, Ptr: IRBPoison.CreateIntToPtr(V: SavedFlagPtrPtr, DestTy: IntptrPtrTy));
3654 IRBPoison.CreateStore(
3655 Val: Constant::getNullValue(Ty: IRBPoison.getInt8Ty()),
3656 Ptr: IRBPoison.CreateIntToPtr(V: SavedFlagPtr, DestTy: IRBPoison.getPtrTy()));
3657 } else {
3658 // For larger frames call __asan_stack_free_*.
3659 RTCI.createRuntimeCall(
3660 IRB&: IRBPoison, Callee: AsanStackFreeFunc[StackMallocIdx],
3661 Args: {FakeStack, ConstantInt::get(Ty: IntptrTy, V: LocalStackSize)});
3662 }
3663
3664 IRBuilder<> IRBElse(ElseTerm);
3665 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBElse, ShadowBase);
3666 } else {
3667 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBRet, ShadowBase);
3668 }
3669 }
3670
3671 // We are done. Remove the old unused alloca instructions.
3672 for (auto *AI : AllocaVec)
3673 AI->eraseFromParent();
3674}
3675
3676void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3677 IRBuilder<> &IRB, bool DoPoison) {
3678 // For now just insert the call to ASan runtime.
3679 Value *AddrArg = IRB.CreatePointerCast(V, DestTy: IntptrTy);
3680 Value *SizeArg = ConstantInt::get(Ty: IntptrTy, V: Size);
3681 RTCI.createRuntimeCall(
3682 IRB, Callee: DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3683 Args: {AddrArg, SizeArg});
3684}
3685
3686// Handling llvm.lifetime intrinsics for a given %alloca:
3687// (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3688// (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3689// invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3690// could be poisoned by previous llvm.lifetime.end instruction, as the
3691// variable may go in and out of scope several times, e.g. in loops).
3692// (3) if we poisoned at least one %alloca in a function,
3693// unpoison the whole stack frame at function exit.
3694void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3695 IRBuilder<> IRB(AI);
3696
3697 const Align Alignment = std::max(a: Align(kAllocaRzSize), b: AI->getAlign());
3698 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3699
3700 Value *Zero = Constant::getNullValue(Ty: IntptrTy);
3701 Value *AllocaRzSize = ConstantInt::get(Ty: IntptrTy, V: kAllocaRzSize);
3702 Value *AllocaRzMask = ConstantInt::get(Ty: IntptrTy, V: AllocaRedzoneMask);
3703
3704 // Since we need to extend alloca with additional memory to locate
3705 // redzones, and OldSize is number of allocated blocks with
3706 // ElementSize size, get allocated memory size in bytes by
3707 // OldSize * ElementSize.
3708 const unsigned ElementSize =
3709 F.getDataLayout().getTypeAllocSize(Ty: AI->getAllocatedType());
3710 Value *OldSize =
3711 IRB.CreateMul(LHS: IRB.CreateIntCast(V: AI->getArraySize(), DestTy: IntptrTy, isSigned: false),
3712 RHS: ConstantInt::get(Ty: IntptrTy, V: ElementSize));
3713
3714 // PartialSize = OldSize % 32
3715 Value *PartialSize = IRB.CreateAnd(LHS: OldSize, RHS: AllocaRzMask);
3716
3717 // Misalign = kAllocaRzSize - PartialSize;
3718 Value *Misalign = IRB.CreateSub(LHS: AllocaRzSize, RHS: PartialSize);
3719
3720 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3721 Value *Cond = IRB.CreateICmpNE(LHS: Misalign, RHS: AllocaRzSize);
3722 Value *PartialPadding = IRB.CreateSelect(C: Cond, True: Misalign, False: Zero);
3723
3724 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3725 // Alignment is added to locate left redzone, PartialPadding for possible
3726 // partial redzone and kAllocaRzSize for right redzone respectively.
3727 Value *AdditionalChunkSize = IRB.CreateAdd(
3728 LHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value() + kAllocaRzSize),
3729 RHS: PartialPadding);
3730
3731 Value *NewSize = IRB.CreateAdd(LHS: OldSize, RHS: AdditionalChunkSize);
3732
3733 // Insert new alloca with new NewSize and Alignment params.
3734 AllocaInst *NewAlloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(), ArraySize: NewSize);
3735 NewAlloca->setAlignment(Alignment);
3736
3737 // NewAddress = Address + Alignment
3738 Value *NewAddress =
3739 IRB.CreateAdd(LHS: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy),
3740 RHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value()));
3741
3742 // Insert __asan_alloca_poison call for new created alloca.
3743 RTCI.createRuntimeCall(IRB, Callee: AsanAllocaPoisonFunc, Args: {NewAddress, OldSize});
3744
3745 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3746 // for unpoisoning stuff.
3747 IRB.CreateStore(Val: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy), Ptr: DynamicAllocaLayout);
3748
3749 Value *NewAddressPtr = IRB.CreateIntToPtr(V: NewAddress, DestTy: AI->getType());
3750
3751 // Replace all uses of AddessReturnedByAlloca with NewAddressPtr.
3752 AI->replaceAllUsesWith(V: NewAddressPtr);
3753
3754 // We are done. Erase old alloca from parent.
3755 AI->eraseFromParent();
3756}
3757
3758// isSafeAccess returns true if Addr is always inbounds with respect to its
3759// base object. For example, it is a field access or an array access with
3760// constant inbounds index.
3761bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3762 Value *Addr, TypeSize TypeStoreSize) const {
3763 if (TypeStoreSize.isScalable())
3764 // TODO: We can use vscale_range to convert a scalable value to an
3765 // upper bound on the access size.
3766 return false;
3767
3768 SizeOffsetAPInt SizeOffset = ObjSizeVis.compute(V: Addr);
3769 if (!SizeOffset.bothKnown())
3770 return false;
3771
3772 uint64_t Size = SizeOffset.Size.getZExtValue();
3773 int64_t Offset = SizeOffset.Offset.getSExtValue();
3774
3775 // Three checks are required to ensure safety:
3776 // . Offset >= 0 (since the offset is given from the base ptr)
3777 // . Size >= Offset (unsigned)
3778 // . Size - Offset >= NeededSize (unsigned)
3779 return Offset >= 0 && Size >= uint64_t(Offset) &&
3780 Size - uint64_t(Offset) >= TypeStoreSize / 8;
3781}
3782