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/SmallSet.h"
24#include "llvm/ADT/SmallVector.h"
25#include "llvm/ADT/Statistic.h"
26#include "llvm/ADT/StringExtras.h"
27#include "llvm/ADT/StringRef.h"
28#include "llvm/ADT/Twine.h"
29#include "llvm/Analysis/GlobalsModRef.h"
30#include "llvm/Analysis/MemoryBuiltins.h"
31#include "llvm/Analysis/StackSafetyAnalysis.h"
32#include "llvm/Analysis/TargetLibraryInfo.h"
33#include "llvm/Analysis/TargetTransformInfo.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/BinaryFormat/MachO.h"
36#include "llvm/Demangle/Demangle.h"
37#include "llvm/IR/Argument.h"
38#include "llvm/IR/Attributes.h"
39#include "llvm/IR/BasicBlock.h"
40#include "llvm/IR/Comdat.h"
41#include "llvm/IR/Constant.h"
42#include "llvm/IR/Constants.h"
43#include "llvm/IR/DIBuilder.h"
44#include "llvm/IR/DataLayout.h"
45#include "llvm/IR/DebugInfoMetadata.h"
46#include "llvm/IR/DebugLoc.h"
47#include "llvm/IR/DerivedTypes.h"
48#include "llvm/IR/EHPersonalities.h"
49#include "llvm/IR/Function.h"
50#include "llvm/IR/GlobalAlias.h"
51#include "llvm/IR/GlobalValue.h"
52#include "llvm/IR/GlobalVariable.h"
53#include "llvm/IR/IRBuilder.h"
54#include "llvm/IR/InlineAsm.h"
55#include "llvm/IR/InstVisitor.h"
56#include "llvm/IR/InstrTypes.h"
57#include "llvm/IR/Instruction.h"
58#include "llvm/IR/Instructions.h"
59#include "llvm/IR/IntrinsicInst.h"
60#include "llvm/IR/Intrinsics.h"
61#include "llvm/IR/LLVMContext.h"
62#include "llvm/IR/MDBuilder.h"
63#include "llvm/IR/Metadata.h"
64#include "llvm/IR/Module.h"
65#include "llvm/IR/Type.h"
66#include "llvm/IR/Use.h"
67#include "llvm/IR/Value.h"
68#include "llvm/MC/MCSectionMachO.h"
69#include "llvm/Support/Casting.h"
70#include "llvm/Support/CommandLine.h"
71#include "llvm/Support/Debug.h"
72#include "llvm/Support/ErrorHandling.h"
73#include "llvm/Support/MathExtras.h"
74#include "llvm/Support/ModRef.h"
75#include "llvm/Support/raw_ostream.h"
76#include "llvm/TargetParser/Triple.h"
77#include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h"
78#include "llvm/Transforms/Instrumentation/AddressSanitizerOptions.h"
79#include "llvm/Transforms/Utils/ASanStackFrameLayout.h"
80#include "llvm/Transforms/Utils/BasicBlockUtils.h"
81#include "llvm/Transforms/Utils/Instrumentation.h"
82#include "llvm/Transforms/Utils/Local.h"
83#include "llvm/Transforms/Utils/ModuleUtils.h"
84#include "llvm/Transforms/Utils/PromoteMemToReg.h"
85#include <algorithm>
86#include <cassert>
87#include <cstddef>
88#include <cstdint>
89#include <iomanip>
90#include <limits>
91#include <sstream>
92#include <string>
93#include <tuple>
94#include <utility>
95
96using namespace llvm;
97
98#define DEBUG_TYPE "asan"
99
100static const uint64_t kDefaultShadowScale = 3;
101static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
102static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
103static const uint64_t kDynamicShadowSentinel =
104 std::numeric_limits<uint64_t>::max();
105static const uint64_t kSmallX86_64ShadowOffsetBase = 0x7FFFFFFF; // < 2G.
106static const uint64_t kSmallX86_64ShadowOffsetAlignMask = ~0xFFFULL;
107static const uint64_t kLinuxKasan_ShadowOffset64 = 0xdffffc0000000000;
108static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 44;
109static const uint64_t kSystemZ_ShadowOffset64 = 1ULL << 52;
110static const uint64_t kMIPS_ShadowOffsetN32 = 1ULL << 29;
111static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa0000;
112static const uint64_t kMIPS64_ShadowOffset64 = 1ULL << 37;
113static const uint64_t kAArch64_ShadowOffset64 = 1ULL << 36;
114static const uint64_t kLoongArch64_ShadowOffset64 = 1ULL << 46;
115static const uint64_t kRISCV64_ShadowOffset64 = kDynamicShadowSentinel;
116static const uint64_t kFreeBSD_ShadowOffset32 = 1ULL << 30;
117static const uint64_t kFreeBSD_ShadowOffset64 = 1ULL << 46;
118static const uint64_t kFreeBSDAArch64_ShadowOffset64 = 1ULL << 47;
119static const uint64_t kFreeBSDKasan_ShadowOffset64 = 0xdffff7c000000000;
120static const uint64_t kNetBSD_ShadowOffset32 = 1ULL << 30;
121static const uint64_t kNetBSD_ShadowOffset64 = 1ULL << 46;
122static const uint64_t kNetBSDKasan_ShadowOffset64 = 0xdfff900000000000;
123static const uint64_t kPS_ShadowOffset64 = 1ULL << 40;
124static const uint64_t kWindowsShadowOffset32 = 3ULL << 28;
125static const uint64_t kWebAssemblyShadowOffset = 0;
126
127// The shadow memory space is dynamically allocated.
128static const uint64_t kWindowsShadowOffset64 = kDynamicShadowSentinel;
129
130static const size_t kMinStackMallocSize = 1 << 6; // 64B
131static const size_t kMaxStackMallocSize = 1 << 16; // 64K
132static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
133static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
134
135const char kAsanModuleCtorName[] = "asan.module_ctor";
136const char kAsanModuleDtorName[] = "asan.module_dtor";
137static const uint64_t kAsanCtorAndDtorPriority = 1;
138// On Emscripten, the system needs more than one priorities for constructors.
139static const uint64_t kAsanEmscriptenCtorAndDtorPriority = 50;
140const char kAsanReportErrorTemplate[] = "__asan_report_";
141const char kAsanRegisterGlobalsName[] = "__asan_register_globals";
142const char kAsanUnregisterGlobalsName[] = "__asan_unregister_globals";
143const char kAsanRegisterImageGlobalsName[] = "__asan_register_image_globals";
144const char kAsanUnregisterImageGlobalsName[] =
145 "__asan_unregister_image_globals";
146const char kAsanRegisterElfGlobalsName[] = "__asan_register_elf_globals";
147const char kAsanUnregisterElfGlobalsName[] = "__asan_unregister_elf_globals";
148const char kAsanPoisonGlobalsName[] = "__asan_before_dynamic_init";
149const char kAsanUnpoisonGlobalsName[] = "__asan_after_dynamic_init";
150const char kAsanInitName[] = "__asan_init";
151const char kAsanVersionCheckNamePrefix[] = "__asan_version_mismatch_check_v";
152const char kAsanPtrCmp[] = "__sanitizer_ptr_cmp";
153const char kAsanPtrSub[] = "__sanitizer_ptr_sub";
154const char kAsanHandleNoReturnName[] = "__asan_handle_no_return";
155static const int kMaxAsanStackMallocSizeClass = 10;
156const char kAsanStackMallocNameTemplate[] = "__asan_stack_malloc_";
157const char kAsanStackMallocAlwaysNameTemplate[] =
158 "__asan_stack_malloc_always_";
159const char kAsanStackFreeNameTemplate[] = "__asan_stack_free_";
160const char kAsanGenPrefix[] = "___asan_gen_";
161const char kODRGenPrefix[] = "__odr_asan_gen_";
162const char kSanCovGenPrefix[] = "__sancov_gen_";
163const char kAsanSetShadowPrefix[] = "__asan_set_shadow_";
164const char kAsanPoisonStackMemoryName[] = "__asan_poison_stack_memory";
165const char kAsanUnpoisonStackMemoryName[] = "__asan_unpoison_stack_memory";
166
167// ASan version script has __asan_* wildcard. Triple underscore prevents a
168// linker (gold) warning about attempting to export a local symbol.
169const char kAsanGlobalsRegisteredFlagName[] = "___asan_globals_registered";
170
171const char kAsanOptionDetectUseAfterReturn[] =
172 "__asan_option_detect_stack_use_after_return";
173
174const char kAsanShadowMemoryDynamicAddress[] =
175 "__asan_shadow_memory_dynamic_address";
176
177const char kAsanAllocaPoison[] = "__asan_alloca_poison";
178const char kAsanAllocasUnpoison[] = "__asan_allocas_unpoison";
179
180const char kAMDGPUAddressSharedName[] = "llvm.amdgcn.is.shared";
181const char kAMDGPUAddressPrivateName[] = "llvm.amdgcn.is.private";
182const char kAMDGPUBallotName[] = "llvm.amdgcn.ballot.i64";
183const char kAMDGPUUnreachableName[] = "llvm.amdgcn.unreachable";
184
185// Accesses sizes are powers of two: 1, 2, 4, 8, 16.
186static const size_t kNumberOfAccessSizes = 5;
187
188static const uint64_t kAllocaRzSize = 32;
189
190// ASanAccessInfo implementation constants.
191constexpr size_t kCompileKernelShift = 0;
192constexpr size_t kCompileKernelMask = 0x1;
193constexpr size_t kAccessSizeIndexShift = 1;
194constexpr size_t kAccessSizeIndexMask = 0xf;
195constexpr size_t kIsWriteShift = 5;
196constexpr size_t kIsWriteMask = 0x1;
197
198// Command-line flags.
199
200static cl::opt<bool> ClEnableKasan(
201 "asan-kernel", cl::desc("Enable KernelAddressSanitizer instrumentation"),
202 cl::Hidden, cl::init(Val: false));
203
204static cl::opt<bool> ClRecover(
205 "asan-recover",
206 cl::desc("Enable recovery mode (continue-after-error)."),
207 cl::Hidden, cl::init(Val: false));
208
209static cl::opt<bool> ClInsertVersionCheck(
210 "asan-guard-against-version-mismatch",
211 cl::desc("Guard against compiler/runtime version mismatch."), cl::Hidden,
212 cl::init(Val: true));
213
214// This flag may need to be replaced with -f[no-]asan-reads.
215static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
216 cl::desc("instrument read instructions"),
217 cl::Hidden, cl::init(Val: true));
218
219static cl::opt<bool> ClInstrumentWrites(
220 "asan-instrument-writes", cl::desc("instrument write instructions"),
221 cl::Hidden, cl::init(Val: true));
222
223static cl::opt<bool>
224 ClUseStackSafety("asan-use-stack-safety", cl::Hidden, cl::init(Val: true),
225 cl::Hidden, cl::desc("Use Stack Safety analysis results"),
226 cl::Optional);
227
228static cl::opt<bool> ClInstrumentAtomics(
229 "asan-instrument-atomics",
230 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden,
231 cl::init(Val: true));
232
233static cl::opt<bool>
234 ClInstrumentByval("asan-instrument-byval",
235 cl::desc("instrument byval call arguments"), cl::Hidden,
236 cl::init(Val: true));
237
238static cl::opt<bool> ClAlwaysSlowPath(
239 "asan-always-slow-path",
240 cl::desc("use instrumentation with slow path for all accesses"), cl::Hidden,
241 cl::init(Val: false));
242
243static cl::opt<bool> ClForceDynamicShadow(
244 "asan-force-dynamic-shadow",
245 cl::desc("Load shadow address into a local variable for each function"),
246 cl::Hidden, cl::init(Val: false));
247
248static cl::opt<bool>
249 ClWithIfunc("asan-with-ifunc",
250 cl::desc("Access dynamic shadow through an ifunc global on "
251 "platforms that support this"),
252 cl::Hidden, cl::init(Val: true));
253
254static cl::opt<int>
255 ClShadowAddrSpace("asan-shadow-addr-space",
256 cl::desc("Address space for pointers to the shadow map"),
257 cl::Hidden, cl::init(Val: 0));
258
259static cl::opt<bool> ClWithIfuncSuppressRemat(
260 "asan-with-ifunc-suppress-remat",
261 cl::desc("Suppress rematerialization of dynamic shadow address by passing "
262 "it through inline asm in prologue."),
263 cl::Hidden, cl::init(Val: true));
264
265// This flag limits the number of instructions to be instrumented
266// in any given BB. Normally, this should be set to unlimited (INT_MAX),
267// but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
268// set it to 10000.
269static cl::opt<int> ClMaxInsnsToInstrumentPerBB(
270 "asan-max-ins-per-bb", cl::init(Val: 10000),
271 cl::desc("maximal number of instructions to instrument in any given BB"),
272 cl::Hidden);
273
274// This flag may need to be replaced with -f[no]asan-stack.
275static cl::opt<bool> ClStack("asan-stack", cl::desc("Handle stack memory"),
276 cl::Hidden, cl::init(Val: true));
277static cl::opt<uint32_t> ClMaxInlinePoisoningSize(
278 "asan-max-inline-poisoning-size",
279 cl::desc(
280 "Inline shadow poisoning for blocks up to the given size in bytes."),
281 cl::Hidden, cl::init(Val: 64));
282
283static cl::opt<AsanDetectStackUseAfterReturnMode> ClUseAfterReturn(
284 "asan-use-after-return",
285 cl::desc("Sets the mode of detection for stack-use-after-return."),
286 cl::values(
287 clEnumValN(AsanDetectStackUseAfterReturnMode::Never, "never",
288 "Never detect stack use after return."),
289 clEnumValN(
290 AsanDetectStackUseAfterReturnMode::Runtime, "runtime",
291 "Detect stack use after return if "
292 "binary flag 'ASAN_OPTIONS=detect_stack_use_after_return' is set."),
293 clEnumValN(AsanDetectStackUseAfterReturnMode::Always, "always",
294 "Always detect stack use after return.")),
295 cl::Hidden, cl::init(Val: AsanDetectStackUseAfterReturnMode::Runtime));
296
297static cl::opt<bool> ClRedzoneByvalArgs("asan-redzone-byval-args",
298 cl::desc("Create redzones for byval "
299 "arguments (extra copy "
300 "required)"), cl::Hidden,
301 cl::init(Val: true));
302
303static cl::opt<bool> ClUseAfterScope("asan-use-after-scope",
304 cl::desc("Check stack-use-after-scope"),
305 cl::Hidden, cl::init(Val: false));
306
307// This flag may need to be replaced with -f[no]asan-globals.
308static cl::opt<bool> ClGlobals("asan-globals",
309 cl::desc("Handle global objects"), cl::Hidden,
310 cl::init(Val: true));
311
312static cl::opt<bool> ClInitializers("asan-initialization-order",
313 cl::desc("Handle C++ initializer order"),
314 cl::Hidden, cl::init(Val: true));
315
316static cl::opt<bool> ClInvalidPointerPairs(
317 "asan-detect-invalid-pointer-pair",
318 cl::desc("Instrument <, <=, >, >=, - with pointer operands"), cl::Hidden,
319 cl::init(Val: false));
320
321static cl::opt<bool> ClInvalidPointerCmp(
322 "asan-detect-invalid-pointer-cmp",
323 cl::desc("Instrument <, <=, >, >= with pointer operands"), cl::Hidden,
324 cl::init(Val: false));
325
326static cl::opt<bool> ClInvalidPointerSub(
327 "asan-detect-invalid-pointer-sub",
328 cl::desc("Instrument - operations with pointer operands"), cl::Hidden,
329 cl::init(Val: false));
330
331static cl::opt<unsigned> ClRealignStack(
332 "asan-realign-stack",
333 cl::desc("Realign stack to the value of this flag (power of two)"),
334 cl::Hidden, cl::init(Val: 32));
335
336static cl::opt<int> ClInstrumentationWithCallsThreshold(
337 "asan-instrumentation-with-call-threshold",
338 cl::desc("If the function being instrumented contains more than "
339 "this number of memory accesses, use callbacks instead of "
340 "inline checks (-1 means never use callbacks)."),
341 cl::Hidden, cl::init(Val: 7000));
342
343static cl::opt<std::string> ClMemoryAccessCallbackPrefix(
344 "asan-memory-access-callback-prefix",
345 cl::desc("Prefix for memory access callbacks"), cl::Hidden,
346 cl::init(Val: "__asan_"));
347
348static cl::opt<bool> ClKasanMemIntrinCallbackPrefix(
349 "asan-kernel-mem-intrinsic-prefix",
350 cl::desc("Use prefix for memory intrinsics in KASAN mode"), cl::Hidden,
351 cl::init(Val: false));
352
353static cl::opt<bool>
354 ClInstrumentDynamicAllocas("asan-instrument-dynamic-allocas",
355 cl::desc("instrument dynamic allocas"),
356 cl::Hidden, cl::init(Val: true));
357
358static cl::opt<bool> ClSkipPromotableAllocas(
359 "asan-skip-promotable-allocas",
360 cl::desc("Do not instrument promotable allocas"), cl::Hidden,
361 cl::init(Val: true));
362
363static cl::opt<AsanCtorKind> ClConstructorKind(
364 "asan-constructor-kind",
365 cl::desc("Sets the ASan constructor kind"),
366 cl::values(clEnumValN(AsanCtorKind::None, "none", "No constructors"),
367 clEnumValN(AsanCtorKind::Global, "global",
368 "Use global constructors")),
369 cl::init(Val: AsanCtorKind::Global), cl::Hidden);
370// These flags allow to change the shadow mapping.
371// The shadow mapping looks like
372// Shadow = (Mem >> scale) + offset
373
374static cl::opt<int> ClMappingScale("asan-mapping-scale",
375 cl::desc("scale of asan shadow mapping"),
376 cl::Hidden, cl::init(Val: 0));
377
378static cl::opt<uint64_t>
379 ClMappingOffset("asan-mapping-offset",
380 cl::desc("offset of asan shadow mapping [EXPERIMENTAL]"),
381 cl::Hidden, cl::init(Val: 0));
382
383// Optimization flags. Not user visible, used mostly for testing
384// and benchmarking the tool.
385
386static cl::opt<bool> ClOpt("asan-opt", cl::desc("Optimize instrumentation"),
387 cl::Hidden, cl::init(Val: true));
388
389static cl::opt<bool> ClOptimizeCallbacks("asan-optimize-callbacks",
390 cl::desc("Optimize callbacks"),
391 cl::Hidden, cl::init(Val: false));
392
393static cl::opt<bool> ClOptSameTemp(
394 "asan-opt-same-temp", cl::desc("Instrument the same temp just once"),
395 cl::Hidden, cl::init(Val: true));
396
397static cl::opt<bool> ClOptGlobals("asan-opt-globals",
398 cl::desc("Don't instrument scalar globals"),
399 cl::Hidden, cl::init(Val: true));
400
401static cl::opt<bool> ClOptStack(
402 "asan-opt-stack", cl::desc("Don't instrument scalar stack variables"),
403 cl::Hidden, cl::init(Val: false));
404
405static cl::opt<bool> ClDynamicAllocaStack(
406 "asan-stack-dynamic-alloca",
407 cl::desc("Use dynamic alloca to represent stack variables"), cl::Hidden,
408 cl::init(Val: true));
409
410static cl::opt<uint32_t> ClForceExperiment(
411 "asan-force-experiment",
412 cl::desc("Force optimization experiment (for testing)"), cl::Hidden,
413 cl::init(Val: 0));
414
415static cl::opt<bool>
416 ClUsePrivateAlias("asan-use-private-alias",
417 cl::desc("Use private aliases for global variables"),
418 cl::Hidden, cl::init(Val: true));
419
420static cl::opt<bool>
421 ClUseOdrIndicator("asan-use-odr-indicator",
422 cl::desc("Use odr indicators to improve ODR reporting"),
423 cl::Hidden, cl::init(Val: true));
424
425static cl::opt<bool>
426 ClUseGlobalsGC("asan-globals-live-support",
427 cl::desc("Use linker features to support dead "
428 "code stripping of globals"),
429 cl::Hidden, cl::init(Val: true));
430
431// This is on by default even though there is a bug in gold:
432// https://sourceware.org/bugzilla/show_bug.cgi?id=19002
433static cl::opt<bool>
434 ClWithComdat("asan-with-comdat",
435 cl::desc("Place ASan constructors in comdat sections"),
436 cl::Hidden, cl::init(Val: true));
437
438static cl::opt<AsanDtorKind> ClOverrideDestructorKind(
439 "asan-destructor-kind",
440 cl::desc("Sets the ASan destructor kind. The default is to use the value "
441 "provided to the pass constructor"),
442 cl::values(clEnumValN(AsanDtorKind::None, "none", "No destructors"),
443 clEnumValN(AsanDtorKind::Global, "global",
444 "Use global destructors")),
445 cl::init(Val: AsanDtorKind::Invalid), cl::Hidden);
446
447static SmallSet<unsigned, 8> SrcAddrSpaces;
448static cl::list<unsigned> ClAddrSpaces(
449 "asan-instrument-address-spaces",
450 cl::desc("Only instrument variables in the specified address spaces."),
451 cl::Hidden, cl::CommaSeparated, cl::callback(CB: [](const unsigned &AddrSpace) {
452 SrcAddrSpaces.insert(V: AddrSpace);
453 }));
454
455// Debug flags.
456
457static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
458 cl::init(Val: 0));
459
460static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
461 cl::Hidden, cl::init(Val: 0));
462
463static cl::opt<std::string> ClDebugFunc("asan-debug-func", cl::Hidden,
464 cl::desc("Debug func"));
465
466static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
467 cl::Hidden, cl::init(Val: -1));
468
469static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug max inst"),
470 cl::Hidden, cl::init(Val: -1));
471
472STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
473STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
474STATISTIC(NumOptimizedAccessesToGlobalVar,
475 "Number of optimized accesses to global vars");
476STATISTIC(NumOptimizedAccessesToStackVar,
477 "Number of optimized accesses to stack vars");
478
479namespace {
480
481/// This struct defines the shadow mapping using the rule:
482/// shadow = (mem >> Scale) ADD-or-OR Offset.
483/// If InGlobal is true, then
484/// extern char __asan_shadow[];
485/// shadow = (mem >> Scale) + &__asan_shadow
486struct ShadowMapping {
487 int Scale;
488 uint64_t Offset;
489 bool OrShadowOffset;
490 bool InGlobal;
491};
492
493} // end anonymous namespace
494
495static ShadowMapping getShadowMapping(const Triple &TargetTriple, int LongSize,
496 bool IsKasan) {
497 bool IsAndroid = TargetTriple.isAndroid();
498 bool IsIOS = TargetTriple.isiOS() || TargetTriple.isWatchOS() ||
499 TargetTriple.isDriverKit();
500 bool IsMacOS = TargetTriple.isMacOSX();
501 bool IsFreeBSD = TargetTriple.isOSFreeBSD();
502 bool IsNetBSD = TargetTriple.isOSNetBSD();
503 bool IsPS = TargetTriple.isPS();
504 bool IsLinux = TargetTriple.isOSLinux();
505 bool IsPPC64 = TargetTriple.getArch() == Triple::ppc64 ||
506 TargetTriple.getArch() == Triple::ppc64le;
507 bool IsSystemZ = TargetTriple.getArch() == Triple::systemz;
508 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64;
509 bool IsMIPSN32ABI = TargetTriple.isABIN32();
510 bool IsMIPS32 = TargetTriple.isMIPS32();
511 bool IsMIPS64 = TargetTriple.isMIPS64();
512 bool IsArmOrThumb = TargetTriple.isARM() || TargetTriple.isThumb();
513 bool IsAArch64 = TargetTriple.getArch() == Triple::aarch64 ||
514 TargetTriple.getArch() == Triple::aarch64_be;
515 bool IsLoongArch64 = TargetTriple.isLoongArch64();
516 bool IsRISCV64 = TargetTriple.getArch() == Triple::riscv64;
517 bool IsWindows = TargetTriple.isOSWindows();
518 bool IsFuchsia = TargetTriple.isOSFuchsia();
519 bool IsAMDGPU = TargetTriple.isAMDGPU();
520 bool IsHaiku = TargetTriple.isOSHaiku();
521 bool IsWasm = TargetTriple.isWasm();
522 bool IsBPF = TargetTriple.isBPF();
523
524 ShadowMapping Mapping;
525
526 Mapping.Scale = kDefaultShadowScale;
527 if (ClMappingScale.getNumOccurrences() > 0) {
528 Mapping.Scale = ClMappingScale;
529 }
530
531 if (LongSize == 32) {
532 if (IsAndroid)
533 Mapping.Offset = kDynamicShadowSentinel;
534 else if (IsMIPSN32ABI)
535 Mapping.Offset = kMIPS_ShadowOffsetN32;
536 else if (IsMIPS32)
537 Mapping.Offset = kMIPS32_ShadowOffset32;
538 else if (IsFreeBSD)
539 Mapping.Offset = kFreeBSD_ShadowOffset32;
540 else if (IsNetBSD)
541 Mapping.Offset = kNetBSD_ShadowOffset32;
542 else if (IsIOS)
543 Mapping.Offset = kDynamicShadowSentinel;
544 else if (IsWindows)
545 Mapping.Offset = kWindowsShadowOffset32;
546 else if (IsWasm)
547 Mapping.Offset = kWebAssemblyShadowOffset;
548 else
549 Mapping.Offset = kDefaultShadowOffset32;
550 } else { // LongSize == 64
551 // Fuchsia is always PIE, which means that the beginning of the address
552 // space is always available.
553 if (IsFuchsia) {
554 // kDynamicShadowSentinel tells instrumentation to use the dynamic shadow.
555 Mapping.Offset = kDynamicShadowSentinel;
556 } else if (IsPPC64)
557 Mapping.Offset = kPPC64_ShadowOffset64;
558 else if (IsSystemZ)
559 Mapping.Offset = kSystemZ_ShadowOffset64;
560 else if (IsFreeBSD && IsAArch64)
561 Mapping.Offset = kFreeBSDAArch64_ShadowOffset64;
562 else if (IsFreeBSD && !IsMIPS64) {
563 if (IsKasan)
564 Mapping.Offset = kFreeBSDKasan_ShadowOffset64;
565 else
566 Mapping.Offset = kFreeBSD_ShadowOffset64;
567 } else if (IsNetBSD) {
568 if (IsKasan)
569 Mapping.Offset = kNetBSDKasan_ShadowOffset64;
570 else
571 Mapping.Offset = kNetBSD_ShadowOffset64;
572 } else if (IsPS)
573 Mapping.Offset = kPS_ShadowOffset64;
574 else if (IsLinux && IsX86_64) {
575 if (IsKasan)
576 Mapping.Offset = kLinuxKasan_ShadowOffset64;
577 else
578 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
579 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
580 } else if (IsWindows && (IsX86_64 || IsAArch64)) {
581 Mapping.Offset = kWindowsShadowOffset64;
582 } else if (IsMIPS64)
583 Mapping.Offset = kMIPS64_ShadowOffset64;
584 else if (IsIOS)
585 Mapping.Offset = kDynamicShadowSentinel;
586 else if (IsMacOS && IsAArch64)
587 Mapping.Offset = kDynamicShadowSentinel;
588 else if (IsAArch64)
589 Mapping.Offset = kAArch64_ShadowOffset64;
590 else if (IsLoongArch64)
591 Mapping.Offset = kLoongArch64_ShadowOffset64;
592 else if (IsRISCV64)
593 Mapping.Offset = kRISCV64_ShadowOffset64;
594 else if (IsAMDGPU)
595 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
596 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
597 else if (IsHaiku && IsX86_64)
598 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
599 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
600 else if (IsBPF)
601 Mapping.Offset = kDynamicShadowSentinel;
602 else if (IsWasm)
603 Mapping.Offset = kWebAssemblyShadowOffset;
604 else
605 Mapping.Offset = kDefaultShadowOffset64;
606 }
607
608 if (ClForceDynamicShadow) {
609 Mapping.Offset = kDynamicShadowSentinel;
610 }
611
612 if (ClMappingOffset.getNumOccurrences() > 0) {
613 Mapping.Offset = ClMappingOffset;
614 }
615
616 // OR-ing shadow offset if more efficient (at least on x86) if the offset
617 // is a power of two, but on ppc64 and loongarch64 we have to use add since
618 // the shadow offset is not necessarily 1/8-th of the address space. On
619 // SystemZ, we could OR the constant in a single instruction, but it's more
620 // efficient to load it once and use indexed addressing.
621 Mapping.OrShadowOffset = !IsAArch64 && !IsPPC64 && !IsSystemZ && !IsPS &&
622 !IsRISCV64 && !IsLoongArch64 &&
623 !(Mapping.Offset & (Mapping.Offset - 1)) &&
624 Mapping.Offset != kDynamicShadowSentinel;
625 Mapping.InGlobal = ClWithIfunc && IsAndroid && IsArmOrThumb;
626
627 return Mapping;
628}
629
630void llvm::getAddressSanitizerParams(const Triple &TargetTriple, int LongSize,
631 bool IsKasan, uint64_t *ShadowBase,
632 int *MappingScale, bool *OrShadowOffset) {
633 auto Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan);
634 *ShadowBase = Mapping.Offset;
635 *MappingScale = Mapping.Scale;
636 *OrShadowOffset = Mapping.OrShadowOffset;
637}
638
639void llvm::removeASanIncompatibleFnAttributes(Function &F, bool ReadsArgMem) {
640 // Adding sanitizer checks invalidates previously inferred memory attributes.
641 //
642 // This is not only true for sanitized functions, because AttrInfer can
643 // infer those attributes on libc functions, which is not true if those
644 // are instrumented (Android) or intercepted.
645 //
646 // We might want to model ASan shadow memory more opaquely to get rid of
647 // this problem altogether, by hiding the shadow memory write in an
648 // intrinsic, essentially like in the AArch64StackTagging pass. But that's
649 // for another day.
650
651 bool Changed = false;
652 // We add memory(readwrite) to functions that don't already have that set and
653 // can access any non-inaccessible memory. Sanitizer instrumentation can
654 // read/write shadow memory, which is IRMemLocation::Other. Sanitizer
655 // instrumentation can instrument any memory accesses to non-inaccessible
656 // memory.
657 if (!F.getMemoryEffects()
658 .getWithoutLoc(Loc: IRMemLocation::InaccessibleMem)
659 .doesNotAccessMemory() &&
660 !isModAndRefSet(MRI: F.getMemoryEffects().getModRef(Loc: IRMemLocation::Other))) {
661 F.setMemoryEffects(F.getMemoryEffects() |
662 MemoryEffects::otherMemOnly(MR: ModRefInfo::ModRef));
663 Changed = true;
664 }
665 // HWASan reads from argument memory even for previously write-only accesses.
666 if (ReadsArgMem) {
667 if (F.getMemoryEffects().getModRef(Loc: IRMemLocation::ArgMem) ==
668 ModRefInfo::Mod) {
669 F.setMemoryEffects(F.getMemoryEffects() |
670 MemoryEffects::argMemOnly(MR: ModRefInfo::Ref));
671 Changed = true;
672 }
673 for (Argument &A : F.args()) {
674 if (A.hasAttribute(Kind: Attribute::WriteOnly)) {
675 A.removeAttr(Kind: Attribute::WriteOnly);
676 Changed = true;
677 }
678 }
679 }
680 if (Changed) {
681 // nobuiltin makes sure later passes don't restore assumptions about
682 // the function.
683 F.addFnAttr(Kind: Attribute::NoBuiltin);
684 }
685}
686
687ASanAccessInfo::ASanAccessInfo(int32_t Packed)
688 : Packed(Packed),
689 AccessSizeIndex((Packed >> kAccessSizeIndexShift) & kAccessSizeIndexMask),
690 IsWrite((Packed >> kIsWriteShift) & kIsWriteMask),
691 CompileKernel((Packed >> kCompileKernelShift) & kCompileKernelMask) {}
692
693ASanAccessInfo::ASanAccessInfo(bool IsWrite, bool CompileKernel,
694 uint8_t AccessSizeIndex)
695 : Packed((IsWrite << kIsWriteShift) +
696 (CompileKernel << kCompileKernelShift) +
697 (AccessSizeIndex << kAccessSizeIndexShift)),
698 AccessSizeIndex(AccessSizeIndex), IsWrite(IsWrite),
699 CompileKernel(CompileKernel) {}
700
701static uint64_t getRedzoneSizeForScale(int MappingScale) {
702 // Redzone used for stack and globals is at least 32 bytes.
703 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
704 return std::max(a: 32U, b: 1U << MappingScale);
705}
706
707static uint64_t GetCtorAndDtorPriority(Triple &TargetTriple) {
708 if (TargetTriple.isOSEmscripten())
709 return kAsanEmscriptenCtorAndDtorPriority;
710 else
711 return kAsanCtorAndDtorPriority;
712}
713
714static Twine genName(StringRef suffix) {
715 return Twine(kAsanGenPrefix) + suffix;
716}
717
718namespace {
719
720class AsanFunctionInserter {
721public:
722 AsanFunctionInserter(Module &M) : M(M) {}
723
724 template <typename... ArgTypes>
725 FunctionCallee insertFunction(StringRef Name, ArgTypes &&...Args) {
726 return M.getOrInsertFunction(Name, std::forward<ArgTypes>(Args)...);
727 }
728
729private:
730 Module &M;
731};
732
733} // end anonymous namespace
734
735namespace {
736/// Helper RAII class to post-process inserted asan runtime calls during a
737/// pass on a single Function. Upon end of scope, detects and applies the
738/// required funclet OpBundle.
739class RuntimeCallInserter {
740 Function *OwnerFn = nullptr;
741 bool TrackInsertedCalls = false;
742 SmallVector<CallInst *> InsertedCalls;
743
744public:
745 RuntimeCallInserter(Function &Fn) : OwnerFn(&Fn) {
746 if (Fn.hasPersonalityFn()) {
747 auto Personality = classifyEHPersonality(Pers: Fn.getPersonalityFn());
748 if (isScopedEHPersonality(Pers: Personality))
749 TrackInsertedCalls = true;
750 }
751 }
752
753 ~RuntimeCallInserter() {
754 if (InsertedCalls.empty())
755 return;
756 assert(TrackInsertedCalls && "Calls were wrongly tracked");
757
758 DenseMap<BasicBlock *, ColorVector> BlockColors = colorEHFunclets(F&: *OwnerFn);
759 for (CallInst *CI : InsertedCalls) {
760 BasicBlock *BB = CI->getParent();
761 assert(BB && "Instruction doesn't belong to a BasicBlock");
762 assert(BB->getParent() == OwnerFn &&
763 "Instruction doesn't belong to the expected Function!");
764
765 ColorVector &Colors = BlockColors[BB];
766 // funclet opbundles are only valid in monochromatic BBs.
767 // Note that unreachable BBs are seen as colorless by colorEHFunclets()
768 // and will be DCE'ed later.
769 if (Colors.empty())
770 continue;
771 if (Colors.size() != 1) {
772 OwnerFn->getContext().emitError(
773 ErrorStr: "Instruction's BasicBlock is not monochromatic");
774 continue;
775 }
776
777 BasicBlock *Color = Colors.front();
778 BasicBlock::iterator EHPadIt = Color->getFirstNonPHIIt();
779
780 if (EHPadIt != Color->end() && EHPadIt->isEHPad()) {
781 // Replace CI with a clone with an added funclet OperandBundle
782 OperandBundleDef OB("funclet", &*EHPadIt);
783 auto *NewCall = CallBase::addOperandBundle(CB: CI, ID: LLVMContext::OB_funclet,
784 OB, InsertPt: CI->getIterator());
785 NewCall->copyMetadata(SrcInst: *CI);
786 CI->replaceAllUsesWith(V: NewCall);
787 CI->eraseFromParent();
788 }
789 }
790 }
791
792 CallInst *createRuntimeCall(IRBuilder<> &IRB, FunctionCallee Callee,
793 ArrayRef<Value *> Args = {},
794 const Twine &Name = "") {
795 assert(IRB.GetInsertBlock()->getParent() == OwnerFn);
796
797 CallInst *Inst = IRB.CreateCall(Callee, Args, Name, FPMathTag: nullptr);
798 if (TrackInsertedCalls)
799 InsertedCalls.push_back(Elt: Inst);
800 return Inst;
801 }
802};
803
804/// AddressSanitizer: instrument the code in module to find memory bugs.
805struct AddressSanitizer {
806 AddressSanitizer(Module &M, const StackSafetyGlobalInfo *SSGI,
807 int InstrumentationWithCallsThreshold,
808 uint32_t MaxInlinePoisoningSize, bool CompileKernel = false,
809 bool Recover = false, bool UseAfterScope = false,
810 AsanDetectStackUseAfterReturnMode UseAfterReturn =
811 AsanDetectStackUseAfterReturnMode::Runtime)
812 : M(M), Inserter(M),
813 CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
814 : CompileKernel),
815 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
816 UseAfterScope(UseAfterScope || ClUseAfterScope),
817 UseAfterReturn(ClUseAfterReturn.getNumOccurrences() ? ClUseAfterReturn
818 : UseAfterReturn),
819 SSGI(SSGI),
820 InstrumentationWithCallsThreshold(
821 ClInstrumentationWithCallsThreshold.getNumOccurrences() > 0
822 ? ClInstrumentationWithCallsThreshold
823 : InstrumentationWithCallsThreshold),
824 MaxInlinePoisoningSize(ClMaxInlinePoisoningSize.getNumOccurrences() > 0
825 ? ClMaxInlinePoisoningSize
826 : MaxInlinePoisoningSize) {
827 C = &(M.getContext());
828 DL = &M.getDataLayout();
829 LongSize = M.getDataLayout().getPointerSizeInBits();
830 IntptrTy = Type::getIntNTy(C&: *C, N: LongSize);
831 PtrTy = PointerType::getUnqual(C&: *C);
832 Int32Ty = Type::getInt32Ty(C&: *C);
833 TargetTriple = M.getTargetTriple();
834
835 Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan: this->CompileKernel);
836
837 assert(this->UseAfterReturn != AsanDetectStackUseAfterReturnMode::Invalid);
838 }
839
840 TypeSize getAllocaSizeInBytes(const AllocaInst &AI) const {
841 return *AI.getAllocationSize(DL: AI.getDataLayout());
842 }
843
844 /// Check if we want (and can) handle this alloca.
845 bool isInterestingAlloca(const AllocaInst &AI);
846
847 bool ignoreAccess(Instruction *Inst, Value *Ptr);
848 void getInterestingMemoryOperands(
849 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting,
850 const TargetTransformInfo *TTI);
851
852 void instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
853 InterestingMemoryOperand &O, bool UseCalls,
854 const DataLayout &DL, RuntimeCallInserter &RTCI);
855 bool instrumentPointerComparisonOrSubtraction(Instruction *I,
856 RuntimeCallInserter &RTCI);
857 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
858 Value *Addr, MaybeAlign Alignment,
859 uint32_t TypeStoreSize, bool IsWrite,
860 Value *SizeArgument, bool UseCalls, uint32_t Exp,
861 RuntimeCallInserter &RTCI);
862 Instruction *instrumentAMDGPUAddress(Instruction *OrigIns,
863 Instruction *InsertBefore, Value *Addr,
864 uint32_t TypeStoreSize, bool IsWrite,
865 Value *SizeArgument);
866 Instruction *genAMDGPUReportBlock(IRBuilder<> &IRB, Value *Cond,
867 bool Recover);
868 void instrumentUnusualSizeOrAlignment(Instruction *I,
869 Instruction *InsertBefore, Value *Addr,
870 TypeSize TypeStoreSize, bool IsWrite,
871 Value *SizeArgument, bool UseCalls,
872 uint32_t Exp,
873 RuntimeCallInserter &RTCI);
874 void instrumentMaskedLoadOrStore(AddressSanitizer *Pass, const DataLayout &DL,
875 Type *IntptrTy, Value *Mask, Value *EVL,
876 Value *Stride, Instruction *I, Value *Addr,
877 MaybeAlign Alignment, unsigned Granularity,
878 Type *OpType, bool IsWrite,
879 Value *SizeArgument, bool UseCalls,
880 uint32_t Exp, RuntimeCallInserter &RTCI);
881 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
882 Value *ShadowValue, uint32_t TypeStoreSize);
883 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
884 bool IsWrite, size_t AccessSizeIndex,
885 Value *SizeArgument, uint32_t Exp,
886 RuntimeCallInserter &RTCI);
887 void instrumentMemIntrinsic(MemIntrinsic *MI, RuntimeCallInserter &RTCI);
888 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
889 bool suppressInstrumentationSiteForDebug(int &Instrumented);
890 bool instrumentFunction(Function &F, const TargetLibraryInfo *TLI,
891 const TargetTransformInfo *TTI);
892 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
893 bool maybeInsertDynamicShadowAtFunctionEntry(Function &F);
894 void markEscapedLocalAllocas(Function &F);
895 void markCatchParametersAsUninteresting(Function &F);
896
897private:
898 friend struct FunctionStackPoisoner;
899
900 void initializeCallbacks(const TargetLibraryInfo *TLI);
901
902 bool LooksLikeCodeInBug11395(Instruction *I);
903 bool GlobalIsLinkerInitialized(GlobalVariable *G);
904 bool isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, Value *Addr,
905 TypeSize TypeStoreSize) const;
906
907 /// Helper to cleanup per-function state.
908 struct FunctionStateRAII {
909 AddressSanitizer *Pass;
910
911 FunctionStateRAII(AddressSanitizer *Pass) : Pass(Pass) {
912 assert(Pass->ProcessedAllocas.empty() &&
913 "last pass forgot to clear cache");
914 assert(!Pass->LocalDynamicShadow);
915 }
916
917 ~FunctionStateRAII() {
918 Pass->LocalDynamicShadow = nullptr;
919 Pass->ProcessedAllocas.clear();
920 }
921 };
922
923 Module &M;
924 AsanFunctionInserter Inserter;
925 LLVMContext *C;
926 const DataLayout *DL;
927 Triple TargetTriple;
928 int LongSize;
929 bool CompileKernel;
930 bool Recover;
931 bool UseAfterScope;
932 AsanDetectStackUseAfterReturnMode UseAfterReturn;
933 Type *IntptrTy;
934 Type *Int32Ty;
935 PointerType *PtrTy;
936 ShadowMapping Mapping;
937 FunctionCallee AsanHandleNoReturnFunc;
938 FunctionCallee AsanPtrCmpFunction, AsanPtrSubFunction;
939 Constant *AsanShadowGlobal;
940
941 // These arrays is indexed by AccessIsWrite, Experiment and log2(AccessSize).
942 FunctionCallee AsanErrorCallback[2][2][kNumberOfAccessSizes];
943 FunctionCallee AsanMemoryAccessCallback[2][2][kNumberOfAccessSizes];
944
945 // These arrays is indexed by AccessIsWrite and Experiment.
946 FunctionCallee AsanErrorCallbackSized[2][2];
947 FunctionCallee AsanMemoryAccessCallbackSized[2][2];
948
949 FunctionCallee AsanMemmove, AsanMemcpy, AsanMemset;
950 Value *LocalDynamicShadow = nullptr;
951 const StackSafetyGlobalInfo *SSGI;
952 DenseMap<const AllocaInst *, bool> ProcessedAllocas;
953
954 FunctionCallee AMDGPUAddressShared;
955 FunctionCallee AMDGPUAddressPrivate;
956 int InstrumentationWithCallsThreshold;
957 uint32_t MaxInlinePoisoningSize;
958};
959
960class ModuleAddressSanitizer {
961public:
962 ModuleAddressSanitizer(Module &M, bool InsertVersionCheck,
963 bool CompileKernel = false, bool Recover = false,
964 bool UseGlobalsGC = true, bool UseOdrIndicator = true,
965 AsanDtorKind DestructorKind = AsanDtorKind::Global,
966 AsanCtorKind ConstructorKind = AsanCtorKind::Global)
967 : M(M), Inserter(M),
968 CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
969 : CompileKernel),
970 InsertVersionCheck(ClInsertVersionCheck.getNumOccurrences() > 0
971 ? ClInsertVersionCheck
972 : InsertVersionCheck),
973 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
974 UseGlobalsGC(UseGlobalsGC && ClUseGlobalsGC && !this->CompileKernel),
975 // Enable aliases as they should have no downside with ODR indicators.
976 UsePrivateAlias(ClUsePrivateAlias.getNumOccurrences() > 0
977 ? ClUsePrivateAlias
978 : UseOdrIndicator),
979 UseOdrIndicator(ClUseOdrIndicator.getNumOccurrences() > 0
980 ? ClUseOdrIndicator
981 : UseOdrIndicator),
982 // Not a typo: ClWithComdat is almost completely pointless without
983 // ClUseGlobalsGC (because then it only works on modules without
984 // globals, which are rare); it is a prerequisite for ClUseGlobalsGC;
985 // and both suffer from gold PR19002 for which UseGlobalsGC constructor
986 // argument is designed as workaround. Therefore, disable both
987 // ClWithComdat and ClUseGlobalsGC unless the frontend says it's ok to
988 // do globals-gc.
989 UseCtorComdat(UseGlobalsGC && ClWithComdat && !this->CompileKernel),
990 DestructorKind(DestructorKind),
991 ConstructorKind(ClConstructorKind.getNumOccurrences() > 0
992 ? ClConstructorKind
993 : ConstructorKind) {
994 C = &(M.getContext());
995 int LongSize = M.getDataLayout().getPointerSizeInBits();
996 IntptrTy = Type::getIntNTy(C&: *C, N: LongSize);
997 PtrTy = PointerType::getUnqual(C&: *C);
998 TargetTriple = M.getTargetTriple();
999 Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan: this->CompileKernel);
1000
1001 if (ClOverrideDestructorKind != AsanDtorKind::Invalid)
1002 this->DestructorKind = ClOverrideDestructorKind;
1003 assert(this->DestructorKind != AsanDtorKind::Invalid);
1004 }
1005
1006 bool instrumentModule();
1007
1008private:
1009 void initializeCallbacks();
1010
1011 void instrumentGlobals(IRBuilder<> &IRB, bool *CtorComdat);
1012 void InstrumentGlobalsCOFF(IRBuilder<> &IRB,
1013 ArrayRef<GlobalVariable *> ExtendedGlobals,
1014 ArrayRef<Constant *> MetadataInitializers);
1015 void instrumentGlobalsELF(IRBuilder<> &IRB,
1016 ArrayRef<GlobalVariable *> ExtendedGlobals,
1017 ArrayRef<Constant *> MetadataInitializers,
1018 const std::string &UniqueModuleId);
1019 void InstrumentGlobalsMachO(IRBuilder<> &IRB,
1020 ArrayRef<GlobalVariable *> ExtendedGlobals,
1021 ArrayRef<Constant *> MetadataInitializers);
1022 void
1023 InstrumentGlobalsWithMetadataArray(IRBuilder<> &IRB,
1024 ArrayRef<GlobalVariable *> ExtendedGlobals,
1025 ArrayRef<Constant *> MetadataInitializers);
1026
1027 GlobalVariable *CreateMetadataGlobal(Constant *Initializer,
1028 StringRef OriginalName);
1029 void SetComdatForGlobalMetadata(GlobalVariable *G, GlobalVariable *Metadata,
1030 StringRef InternalSuffix);
1031 Instruction *CreateAsanModuleDtor();
1032
1033 const GlobalVariable *getExcludedAliasedGlobal(const GlobalAlias &GA) const;
1034 bool shouldInstrumentGlobal(GlobalVariable *G) const;
1035 bool ShouldUseMachOGlobalsSection() const;
1036 StringRef getGlobalMetadataSection() const;
1037 void poisonOneInitializer(Function &GlobalInit);
1038 void createInitializerPoisonCalls();
1039 uint64_t getMinRedzoneSizeForGlobal() const {
1040 return getRedzoneSizeForScale(MappingScale: Mapping.Scale);
1041 }
1042 uint64_t getRedzoneSizeForGlobal(uint64_t SizeInBytes) const;
1043 int GetAsanVersion() const;
1044 GlobalVariable *getOrCreateModuleName();
1045
1046 Module &M;
1047 AsanFunctionInserter Inserter;
1048 bool CompileKernel;
1049 bool InsertVersionCheck;
1050 bool Recover;
1051 bool UseGlobalsGC;
1052 bool UsePrivateAlias;
1053 bool UseOdrIndicator;
1054 bool UseCtorComdat;
1055 AsanDtorKind DestructorKind;
1056 AsanCtorKind ConstructorKind;
1057 Type *IntptrTy;
1058 PointerType *PtrTy;
1059 LLVMContext *C;
1060 Triple TargetTriple;
1061 ShadowMapping Mapping;
1062 FunctionCallee AsanPoisonGlobals;
1063 FunctionCallee AsanUnpoisonGlobals;
1064 FunctionCallee AsanRegisterGlobals;
1065 FunctionCallee AsanUnregisterGlobals;
1066 FunctionCallee AsanRegisterImageGlobals;
1067 FunctionCallee AsanUnregisterImageGlobals;
1068 FunctionCallee AsanRegisterElfGlobals;
1069 FunctionCallee AsanUnregisterElfGlobals;
1070
1071 Function *AsanCtorFunction = nullptr;
1072 Function *AsanDtorFunction = nullptr;
1073 GlobalVariable *ModuleName = nullptr;
1074};
1075
1076// Stack poisoning does not play well with exception handling.
1077// When an exception is thrown, we essentially bypass the code
1078// that unpoisones the stack. This is why the run-time library has
1079// to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
1080// stack in the interceptor. This however does not work inside the
1081// actual function which catches the exception. Most likely because the
1082// compiler hoists the load of the shadow value somewhere too high.
1083// This causes asan to report a non-existing bug on 453.povray.
1084// It sounds like an LLVM bug.
1085struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
1086 Function &F;
1087 AddressSanitizer &ASan;
1088 RuntimeCallInserter &RTCI;
1089 DIBuilder DIB;
1090 LLVMContext *C;
1091 Type *IntptrTy;
1092 Type *IntptrPtrTy;
1093 ShadowMapping Mapping;
1094
1095 SmallVector<AllocaInst *, 16> AllocaVec;
1096 SmallVector<AllocaInst *, 16> StaticAllocasToMoveUp;
1097 SmallVector<Instruction *, 8> RetVec;
1098
1099 FunctionCallee AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1],
1100 AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1];
1101 FunctionCallee AsanSetShadowFunc[0x100] = {};
1102 FunctionCallee AsanPoisonStackMemoryFunc, AsanUnpoisonStackMemoryFunc;
1103 FunctionCallee AsanAllocaPoisonFunc, AsanAllocasUnpoisonFunc;
1104
1105 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
1106 struct AllocaPoisonCall {
1107 IntrinsicInst *InsBefore;
1108 AllocaInst *AI;
1109 uint64_t Size;
1110 bool DoPoison;
1111 };
1112 SmallVector<AllocaPoisonCall, 8> DynamicAllocaPoisonCallVec;
1113 SmallVector<AllocaPoisonCall, 8> StaticAllocaPoisonCallVec;
1114
1115 SmallVector<AllocaInst *, 1> DynamicAllocaVec;
1116 SmallVector<IntrinsicInst *, 1> StackRestoreVec;
1117 AllocaInst *DynamicAllocaLayout = nullptr;
1118 IntrinsicInst *LocalEscapeCall = nullptr;
1119
1120 bool HasInlineAsm = false;
1121 bool HasReturnsTwiceCall = false;
1122 bool PoisonStack;
1123
1124 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan,
1125 RuntimeCallInserter &RTCI)
1126 : F(F), ASan(ASan), RTCI(RTCI),
1127 DIB(*F.getParent(), /*AllowUnresolved*/ false), C(ASan.C),
1128 IntptrTy(ASan.IntptrTy),
1129 IntptrPtrTy(PointerType::get(C&: IntptrTy->getContext(), AddressSpace: 0)),
1130 Mapping(ASan.Mapping),
1131 PoisonStack(ClStack && !F.getParent()->getTargetTriple().isAMDGPU()) {}
1132
1133 bool runOnFunction() {
1134 if (!PoisonStack)
1135 return false;
1136
1137 if (ClRedzoneByvalArgs)
1138 copyArgsPassedByValToAllocas();
1139
1140 // Collect alloca, ret, lifetime instructions etc.
1141 for (BasicBlock *BB : depth_first(G: &F.getEntryBlock())) visit(BB&: *BB);
1142
1143 if (AllocaVec.empty() && DynamicAllocaVec.empty()) return false;
1144
1145 initializeCallbacks(M&: *F.getParent());
1146
1147 processDynamicAllocas();
1148 processStaticAllocas();
1149
1150 if (ClDebugStack) {
1151 LLVM_DEBUG(dbgs() << F);
1152 }
1153 return true;
1154 }
1155
1156 // Arguments marked with the "byval" attribute are implicitly copied without
1157 // using an alloca instruction. To produce redzones for those arguments, we
1158 // copy them a second time into memory allocated with an alloca instruction.
1159 void copyArgsPassedByValToAllocas();
1160
1161 // Finds all Alloca instructions and puts
1162 // poisoned red zones around all of them.
1163 // Then unpoison everything back before the function returns.
1164 void processStaticAllocas();
1165 void processDynamicAllocas();
1166
1167 void createDynamicAllocasInitStorage();
1168
1169 // ----------------------- Visitors.
1170 /// Collect all Ret instructions, or the musttail call instruction if it
1171 /// precedes the return instruction.
1172 void visitReturnInst(ReturnInst &RI) {
1173 if (CallInst *CI = RI.getParent()->getTerminatingMustTailCall())
1174 RetVec.push_back(Elt: CI);
1175 else
1176 RetVec.push_back(Elt: &RI);
1177 }
1178
1179 /// Collect all Resume instructions.
1180 void visitResumeInst(ResumeInst &RI) { RetVec.push_back(Elt: &RI); }
1181
1182 /// Collect all CatchReturnInst instructions.
1183 void visitCleanupReturnInst(CleanupReturnInst &CRI) { RetVec.push_back(Elt: &CRI); }
1184
1185 void unpoisonDynamicAllocasBeforeInst(Instruction *InstBefore,
1186 Value *SavedStack) {
1187 IRBuilder<> IRB(InstBefore);
1188 Value *DynamicAreaPtr = IRB.CreatePtrToInt(V: SavedStack, DestTy: IntptrTy);
1189 // When we insert _asan_allocas_unpoison before @llvm.stackrestore, we
1190 // need to adjust extracted SP to compute the address of the most recent
1191 // alloca. We have a special @llvm.get.dynamic.area.offset intrinsic for
1192 // this purpose.
1193 if (!isa<ReturnInst>(Val: InstBefore)) {
1194 Value *DynamicAreaOffset = IRB.CreateIntrinsic(
1195 ID: Intrinsic::get_dynamic_area_offset, OverloadTypes: {IntptrTy}, Args: {});
1196
1197 DynamicAreaPtr = IRB.CreateAdd(LHS: IRB.CreatePtrToInt(V: SavedStack, DestTy: IntptrTy),
1198 RHS: DynamicAreaOffset);
1199 }
1200
1201 RTCI.createRuntimeCall(
1202 IRB, Callee: AsanAllocasUnpoisonFunc,
1203 Args: {IRB.CreateLoad(Ty: IntptrTy, Ptr: DynamicAllocaLayout), DynamicAreaPtr});
1204 }
1205
1206 // Unpoison dynamic allocas redzones.
1207 void unpoisonDynamicAllocas() {
1208 for (Instruction *Ret : RetVec)
1209 unpoisonDynamicAllocasBeforeInst(InstBefore: Ret, SavedStack: DynamicAllocaLayout);
1210
1211 for (Instruction *StackRestoreInst : StackRestoreVec)
1212 unpoisonDynamicAllocasBeforeInst(InstBefore: StackRestoreInst,
1213 SavedStack: StackRestoreInst->getOperand(i: 0));
1214 }
1215
1216 // Deploy and poison redzones around dynamic alloca call. To do this, we
1217 // should replace this call with another one with changed parameters and
1218 // replace all its uses with new address, so
1219 // addr = alloca type, old_size, align
1220 // is replaced by
1221 // new_size = (old_size + additional_size) * sizeof(type)
1222 // tmp = alloca i8, new_size, max(align, 32)
1223 // addr = tmp + 32 (first 32 bytes are for the left redzone).
1224 // Additional_size is added to make new memory allocation contain not only
1225 // requested memory, but also left, partial and right redzones.
1226 void handleDynamicAllocaCall(AllocaInst *AI);
1227
1228 /// Collect Alloca instructions we want (and can) handle.
1229 void visitAllocaInst(AllocaInst &AI) {
1230 // FIXME: Handle scalable vectors instead of ignoring them.
1231 if (!ASan.isInterestingAlloca(AI) || AI.isScalable()) {
1232 if (AI.isStaticAlloca()) {
1233 // Skip over allocas that are present *before* the first instrumented
1234 // alloca, we don't want to move those around.
1235 if (AllocaVec.empty())
1236 return;
1237
1238 StaticAllocasToMoveUp.push_back(Elt: &AI);
1239 }
1240 return;
1241 }
1242
1243 if (!AI.isStaticAlloca())
1244 DynamicAllocaVec.push_back(Elt: &AI);
1245 else
1246 AllocaVec.push_back(Elt: &AI);
1247 }
1248
1249 /// Collect lifetime intrinsic calls to check for use-after-scope
1250 /// errors.
1251 void visitIntrinsicInst(IntrinsicInst &II) {
1252 Intrinsic::ID ID = II.getIntrinsicID();
1253 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(Elt: &II);
1254 if (ID == Intrinsic::localescape) LocalEscapeCall = &II;
1255 if (!ASan.UseAfterScope)
1256 return;
1257 if (!II.isLifetimeStartOrEnd())
1258 return;
1259 // Find alloca instruction that corresponds to llvm.lifetime argument.
1260 AllocaInst *AI = dyn_cast<AllocaInst>(Val: II.getArgOperand(i: 0));
1261 // We're interested only in allocas we can handle.
1262 if (!AI || !ASan.isInterestingAlloca(AI: *AI))
1263 return;
1264
1265 std::optional<TypeSize> Size = AI->getAllocationSize(DL: AI->getDataLayout());
1266 // Check that size is known and can be stored in IntptrTy.
1267 // TODO: Add support for scalable vectors if possible.
1268 if (!Size || Size->isScalable() ||
1269 !ConstantInt::isValueValidForType(Ty: IntptrTy, V: *Size))
1270 return;
1271
1272 bool DoPoison = (ID == Intrinsic::lifetime_end);
1273 AllocaPoisonCall APC = {.InsBefore: &II, .AI: AI, .Size: *Size, .DoPoison: DoPoison};
1274 if (AI->isStaticAlloca())
1275 StaticAllocaPoisonCallVec.push_back(Elt: APC);
1276 else if (ClInstrumentDynamicAllocas)
1277 DynamicAllocaPoisonCallVec.push_back(Elt: APC);
1278 }
1279
1280 void visitCallBase(CallBase &CB) {
1281 if (CallInst *CI = dyn_cast<CallInst>(Val: &CB)) {
1282 HasInlineAsm |= CI->isInlineAsm() && &CB != ASan.LocalDynamicShadow;
1283 HasReturnsTwiceCall |= CI->canReturnTwice();
1284 }
1285 }
1286
1287 // ---------------------- Helpers.
1288 void initializeCallbacks(Module &M);
1289
1290 // Copies bytes from ShadowBytes into shadow memory for indexes where
1291 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that
1292 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten.
1293 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1294 IRBuilder<> &IRB, Value *ShadowBase);
1295 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1296 size_t Begin, size_t End, IRBuilder<> &IRB,
1297 Value *ShadowBase);
1298 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
1299 ArrayRef<uint8_t> ShadowBytes, size_t Begin,
1300 size_t End, IRBuilder<> &IRB, Value *ShadowBase);
1301
1302 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
1303
1304 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L,
1305 bool Dynamic);
1306 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue,
1307 Instruction *ThenTerm, Value *ValueIfFalse);
1308};
1309
1310} // end anonymous namespace
1311
1312void AddressSanitizerPass::printPipeline(
1313 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1314 static_cast<PassInfoMixin<AddressSanitizerPass> *>(this)->printPipeline(
1315 OS, MapClassName2PassName);
1316 OS << '<';
1317 if (Options.CompileKernel)
1318 OS << "kernel;";
1319 if (Options.UseAfterScope)
1320 OS << "use-after-scope";
1321 OS << '>';
1322}
1323
1324AddressSanitizerPass::AddressSanitizerPass(
1325 const AddressSanitizerOptions &Options, bool UseGlobalGC,
1326 bool UseOdrIndicator, AsanDtorKind DestructorKind,
1327 AsanCtorKind ConstructorKind)
1328 : Options(Options), UseGlobalGC(UseGlobalGC),
1329 UseOdrIndicator(UseOdrIndicator), DestructorKind(DestructorKind),
1330 ConstructorKind(ConstructorKind) {}
1331
1332PreservedAnalyses AddressSanitizerPass::run(Module &M,
1333 ModuleAnalysisManager &MAM) {
1334 // Return early if nosanitize_address module flag is present for the module.
1335 // This implies that asan pass has already run before.
1336 if (checkIfAlreadyInstrumented(M, Flag: "nosanitize_address"))
1337 return PreservedAnalyses::all();
1338
1339 ModuleAddressSanitizer ModuleSanitizer(
1340 M, Options.InsertVersionCheck, Options.CompileKernel, Options.Recover,
1341 UseGlobalGC, UseOdrIndicator, DestructorKind, ConstructorKind);
1342 bool Modified = false;
1343 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
1344 const StackSafetyGlobalInfo *const SSGI =
1345 ClUseStackSafety ? &MAM.getResult<StackSafetyGlobalAnalysis>(IR&: M) : nullptr;
1346 for (Function &F : M) {
1347 if (F.empty())
1348 continue;
1349 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage)
1350 continue;
1351 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName())
1352 continue;
1353 if (F.getName().starts_with(Prefix: "__asan_"))
1354 continue;
1355 if (F.isPresplitCoroutine())
1356 continue;
1357 AddressSanitizer FunctionSanitizer(
1358 M, SSGI, Options.InstrumentationWithCallsThreshold,
1359 Options.MaxInlinePoisoningSize, Options.CompileKernel, Options.Recover,
1360 Options.UseAfterScope, Options.UseAfterReturn);
1361 const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: F);
1362 const TargetTransformInfo &TTI = FAM.getResult<TargetIRAnalysis>(IR&: F);
1363 Modified |= FunctionSanitizer.instrumentFunction(F, TLI: &TLI, TTI: &TTI);
1364 }
1365 Modified |= ModuleSanitizer.instrumentModule();
1366 if (!Modified)
1367 return PreservedAnalyses::all();
1368
1369 PreservedAnalyses PA = PreservedAnalyses::none();
1370 // GlobalsAA is considered stateless and does not get invalidated unless
1371 // explicitly invalidated; PreservedAnalyses::none() is not enough. Sanitizers
1372 // make changes that require GlobalsAA to be invalidated.
1373 PA.abandon<GlobalsAA>();
1374 return PA;
1375}
1376
1377static size_t TypeStoreSizeToSizeIndex(uint32_t TypeSize) {
1378 size_t Res = llvm::countr_zero(Val: TypeSize / 8);
1379 assert(Res < kNumberOfAccessSizes);
1380 return Res;
1381}
1382
1383/// Check if \p G has been created by a trusted compiler pass.
1384static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) {
1385 // Do not instrument @llvm.global_ctors, @llvm.used, etc.
1386 if (G->getName().starts_with(Prefix: "llvm.") ||
1387 // Do not instrument gcov counter arrays.
1388 G->getName().starts_with(Prefix: "__llvm_gcov_ctr") ||
1389 // Do not instrument rtti proxy symbols for function sanitizer.
1390 G->getName().starts_with(Prefix: "__llvm_rtti_proxy"))
1391 return true;
1392
1393 // Do not instrument asan globals.
1394 if (G->getName().starts_with(Prefix: kAsanGenPrefix) ||
1395 G->getName().starts_with(Prefix: kSanCovGenPrefix) ||
1396 G->getName().starts_with(Prefix: kODRGenPrefix))
1397 return true;
1398
1399 return false;
1400}
1401
1402static bool isUnsupportedAMDGPUAddrspace(Value *Addr) {
1403 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1404 unsigned int AddrSpace = PtrTy->getPointerAddressSpace();
1405 // Globals in address space 1 and 4 are supported for AMDGPU.
1406 if (AddrSpace == 3 || AddrSpace == 5)
1407 return true;
1408 return false;
1409}
1410
1411static bool isSupportedAddrspace(const Triple &TargetTriple, Value *Addr) {
1412 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1413 unsigned int AddrSpace = PtrTy->getPointerAddressSpace();
1414
1415 if (!SrcAddrSpaces.empty())
1416 return SrcAddrSpaces.count(V: AddrSpace);
1417
1418 if (TargetTriple.isAMDGPU())
1419 return !isUnsupportedAMDGPUAddrspace(Addr);
1420
1421 return AddrSpace == 0;
1422}
1423
1424Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
1425 if (TargetTriple.isOSDarwin() &&
1426 TargetTriple.getArch() == llvm::Triple::aarch64) {
1427 // Strip MTE-tag bits before translating to shadow address
1428 Shadow = IRB.CreateAnd(LHS: Shadow,
1429 RHS: ConstantInt::get(Ty: IntptrTy, V: ~(uint64_t(0x0f) << 56)));
1430 }
1431 // Shadow >> scale
1432 Shadow = IRB.CreateLShr(LHS: Shadow, RHS: Mapping.Scale);
1433 if (Mapping.Offset == 0) return Shadow;
1434 // (Shadow >> scale) | offset
1435 Value *ShadowBase;
1436 if (LocalDynamicShadow)
1437 ShadowBase = LocalDynamicShadow;
1438 else
1439 ShadowBase = ConstantInt::get(Ty: IntptrTy, V: Mapping.Offset);
1440 if (Mapping.OrShadowOffset)
1441 return IRB.CreateOr(LHS: Shadow, RHS: ShadowBase);
1442 else
1443 return IRB.CreateAdd(LHS: Shadow, RHS: ShadowBase);
1444}
1445
1446// Instrument memset/memmove/memcpy
1447void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI,
1448 RuntimeCallInserter &RTCI) {
1449 InstrumentationIRBuilder IRB(MI);
1450 if (isa<MemTransferInst>(Val: MI)) {
1451 RTCI.createRuntimeCall(
1452 IRB, Callee: isa<MemMoveInst>(Val: MI) ? AsanMemmove : AsanMemcpy,
1453 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1454 IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 1), DestTy: PtrTy),
1455 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1456 } else if (isa<MemSetInst>(Val: MI)) {
1457 RTCI.createRuntimeCall(
1458 IRB, Callee: AsanMemset,
1459 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1460 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 1), DestTy: IRB.getInt32Ty(), isSigned: false),
1461 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1462 }
1463 MI->eraseFromParent();
1464}
1465
1466/// Check if we want (and can) handle this alloca.
1467bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1468 auto [It, Inserted] = ProcessedAllocas.try_emplace(Key: &AI);
1469
1470 if (!Inserted)
1471 return It->getSecond();
1472
1473 bool IsInteresting = // alloca() may be called with 0 size, ignore it.
1474 (((!AI.isStaticAlloca()) || !getAllocaSizeInBytes(AI).isZero()) &&
1475 // We are only interested in allocas not promotable to registers.
1476 // Promotable allocas are common under -O0.
1477 (!ClSkipPromotableAllocas || !isAllocaPromotable(AI: &AI)) &&
1478 // inalloca allocas are not treated as static, and we don't want
1479 // dynamic alloca instrumentation for them as well.
1480 !AI.isUsedWithInAlloca() &&
1481 // swifterror allocas are register promoted by ISel
1482 !AI.isSwiftError() &&
1483 // safe allocas are not interesting
1484 !(SSGI && SSGI->isSafe(AI)));
1485
1486 It->second = IsInteresting;
1487 return IsInteresting;
1488}
1489
1490bool AddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) {
1491 // Check whether the target supports sanitizing the address space
1492 // of the pointer.
1493 if (!isSupportedAddrspace(TargetTriple, Addr: Ptr))
1494 return true;
1495
1496 // Ignore swifterror addresses.
1497 // swifterror memory addresses are mem2reg promoted by instruction
1498 // selection. As such they cannot have regular uses like an instrumentation
1499 // function and it makes no sense to track them as memory.
1500 if (Ptr->isSwiftError())
1501 return true;
1502
1503 // Treat memory accesses to promotable allocas as non-interesting since they
1504 // will not cause memory violations. This greatly speeds up the instrumented
1505 // executable at -O0.
1506 if (auto AI = dyn_cast_or_null<AllocaInst>(Val: Ptr))
1507 if (ClSkipPromotableAllocas && !isInterestingAlloca(AI: *AI))
1508 return true;
1509
1510 if (SSGI != nullptr && SSGI->stackAccessIsSafe(I: *Inst) &&
1511 findAllocaForValue(V: Ptr))
1512 return true;
1513
1514 return false;
1515}
1516
1517void AddressSanitizer::getInterestingMemoryOperands(
1518 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting,
1519 const TargetTransformInfo *TTI) {
1520 // Do not instrument the load fetching the dynamic shadow address.
1521 if (LocalDynamicShadow == I)
1522 return;
1523
1524 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
1525 if (!ClInstrumentReads || ignoreAccess(Inst: I, Ptr: LI->getPointerOperand()))
1526 return;
1527 Interesting.emplace_back(Args&: I, Args: LI->getPointerOperandIndex(), Args: false,
1528 Args: LI->getType(), Args: LI->getAlign());
1529 } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
1530 if (!ClInstrumentWrites || ignoreAccess(Inst: I, Ptr: SI->getPointerOperand()))
1531 return;
1532 Interesting.emplace_back(Args&: I, Args: SI->getPointerOperandIndex(), Args: true,
1533 Args: SI->getValueOperand()->getType(), Args: SI->getAlign());
1534 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
1535 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: RMW->getPointerOperand()))
1536 return;
1537 Interesting.emplace_back(Args&: I, Args: RMW->getPointerOperandIndex(), Args: true,
1538 Args: RMW->getValOperand()->getType(), Args: std::nullopt);
1539 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
1540 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: XCHG->getPointerOperand()))
1541 return;
1542 Interesting.emplace_back(Args&: I, Args: XCHG->getPointerOperandIndex(), Args: true,
1543 Args: XCHG->getCompareOperand()->getType(),
1544 Args: std::nullopt);
1545 } else if (auto CI = dyn_cast<CallInst>(Val: I)) {
1546 switch (CI->getIntrinsicID()) {
1547 case Intrinsic::masked_load:
1548 case Intrinsic::masked_store:
1549 case Intrinsic::masked_gather:
1550 case Intrinsic::masked_scatter: {
1551 bool IsWrite = CI->getType()->isVoidTy();
1552 // Masked store has an initial operand for the value.
1553 unsigned OpOffset = IsWrite ? 1 : 0;
1554 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1555 return;
1556
1557 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1558 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1559 return;
1560 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1561 MaybeAlign Alignment = CI->getParamAlign(ArgNo: 0);
1562 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1563 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: Mask);
1564 break;
1565 }
1566 case Intrinsic::masked_expandload:
1567 case Intrinsic::masked_compressstore: {
1568 bool IsWrite = CI->getIntrinsicID() == Intrinsic::masked_compressstore;
1569 unsigned OpOffset = IsWrite ? 1 : 0;
1570 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1571 return;
1572 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1573 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1574 return;
1575 MaybeAlign Alignment = BasePtr->getPointerAlignment(DL: *DL);
1576 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1577
1578 IRBuilder IB(I);
1579 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1580 // Use the popcount of Mask as the effective vector length.
1581 Type *ExtTy = VectorType::get(ElementType: IntptrTy, Other: cast<VectorType>(Val: Ty));
1582 Value *ExtMask = IB.CreateZExt(V: Mask, DestTy: ExtTy);
1583 Value *EVL = IB.CreateAddReduce(Src: ExtMask);
1584 Value *TrueMask = ConstantInt::get(Ty: Mask->getType(), V: 1);
1585 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: TrueMask,
1586 Args&: EVL);
1587 break;
1588 }
1589 case Intrinsic::vp_load:
1590 case Intrinsic::vp_store:
1591 case Intrinsic::experimental_vp_strided_load:
1592 case Intrinsic::experimental_vp_strided_store: {
1593 auto *VPI = cast<VPIntrinsic>(Val: CI);
1594 unsigned IID = CI->getIntrinsicID();
1595 bool IsWrite = CI->getType()->isVoidTy();
1596 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1597 return;
1598 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1599 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1600 MaybeAlign Alignment = VPI->getOperand(i_nocapture: PtrOpNo)->getPointerAlignment(DL: *DL);
1601 Value *Stride = nullptr;
1602 if (IID == Intrinsic::experimental_vp_strided_store ||
1603 IID == Intrinsic::experimental_vp_strided_load) {
1604 Stride = VPI->getOperand(i_nocapture: PtrOpNo + 1);
1605 // Use the pointer alignment as the element alignment if the stride is a
1606 // multiple of the pointer alignment. Otherwise, the element alignment
1607 // should be Align(1).
1608 unsigned PointerAlign = Alignment.valueOrOne().value();
1609 if (!isa<ConstantInt>(Val: Stride) ||
1610 cast<ConstantInt>(Val: Stride)->getZExtValue() % PointerAlign != 0)
1611 Alignment = Align(1);
1612 }
1613 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1614 Args: VPI->getMaskParam(), Args: VPI->getVectorLengthParam(),
1615 Args&: Stride);
1616 break;
1617 }
1618 case Intrinsic::vp_gather:
1619 case Intrinsic::vp_scatter: {
1620 auto *VPI = cast<VPIntrinsic>(Val: CI);
1621 unsigned IID = CI->getIntrinsicID();
1622 bool IsWrite = IID == Intrinsic::vp_scatter;
1623 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1624 return;
1625 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1626 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1627 MaybeAlign Alignment = VPI->getPointerAlignment();
1628 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1629 Args: VPI->getMaskParam(),
1630 Args: VPI->getVectorLengthParam());
1631 break;
1632 }
1633 default:
1634 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
1635 MemIntrinsicInfo IntrInfo;
1636 if (TTI->getTgtMemIntrinsic(Inst: II, Info&: IntrInfo))
1637 Interesting = IntrInfo.InterestingOperands;
1638 return;
1639 }
1640 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) {
1641 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1642 ignoreAccess(Inst: I, Ptr: CI->getArgOperand(i: ArgNo)))
1643 continue;
1644 Type *Ty = CI->getParamByValType(ArgNo);
1645 Interesting.emplace_back(Args&: I, Args&: ArgNo, Args: false, Args&: Ty, Args: Align(1));
1646 }
1647 }
1648 }
1649}
1650
1651static bool isPointerOperand(Value *V) {
1652 return V->getType()->isPointerTy() || isa<PtrToIntInst, PtrToAddrInst>(Val: V);
1653}
1654
1655// This is a rough heuristic; it may cause both false positives and
1656// false negatives. The proper implementation requires cooperation with
1657// the frontend.
1658static bool isInterestingPointerComparison(Instruction *I) {
1659 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: I)) {
1660 if (!Cmp->isRelational())
1661 return false;
1662 } else {
1663 return false;
1664 }
1665 return isPointerOperand(V: I->getOperand(i: 0)) &&
1666 isPointerOperand(V: I->getOperand(i: 1));
1667}
1668
1669// This is a rough heuristic; it may cause both false positives and
1670// false negatives. The proper implementation requires cooperation with
1671// the frontend.
1672static bool isInterestingPointerSubtraction(Instruction *I) {
1673 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) {
1674 if (BO->getOpcode() != Instruction::Sub)
1675 return false;
1676 } else {
1677 return false;
1678 }
1679 return isPointerOperand(V: I->getOperand(i: 0)) &&
1680 isPointerOperand(V: I->getOperand(i: 1));
1681}
1682
1683bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1684 // If a global variable does not have dynamic initialization we don't
1685 // have to instrument it. However, if a global does not have initializer
1686 // at all, we assume it has dynamic initializer (in other TU).
1687 if (!G->hasInitializer())
1688 return false;
1689
1690 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().IsDynInit)
1691 return false;
1692
1693 return true;
1694}
1695
1696bool AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1697 Instruction *I, RuntimeCallInserter &RTCI) {
1698 IRBuilder<> IRB(I);
1699 FunctionCallee F = isa<ICmpInst>(Val: I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1700 Value *Param[2] = {I->getOperand(i: 0), I->getOperand(i: 1)};
1701
1702 if (const auto *Ty = Param[0]->getType(); Ty->isVectorTy()) {
1703 const auto *VTy = dyn_cast<FixedVectorType>(Val: Ty);
1704 // TODO: Add support for scalable vectors if possible.
1705 if (!VTy)
1706 return false;
1707
1708 assert(Param[0]->getType() == Param[1]->getType() &&
1709 "invalid vector pointer pair instrumentation operands");
1710 for (unsigned Index = 0, NumElements = VTy->getNumElements();
1711 Index != NumElements; ++Index) {
1712 Value *ScalarParam[2] = {
1713 IRB.CreatePointerCast(
1714 V: IRB.CreateExtractElement(Vec: Param[0], Idx: IRB.getInt32(C: Index)),
1715 DestTy: IntptrTy),
1716 IRB.CreatePointerCast(
1717 V: IRB.CreateExtractElement(Vec: Param[1], Idx: IRB.getInt32(C: Index)),
1718 DestTy: IntptrTy)};
1719 RTCI.createRuntimeCall(IRB, Callee: F, Args: ScalarParam);
1720 }
1721 return true;
1722 }
1723
1724 for (Value *&P : Param)
1725 P = IRB.CreatePointerCast(V: P, DestTy: IntptrTy);
1726 RTCI.createRuntimeCall(IRB, Callee: F, Args: Param);
1727 return true;
1728}
1729
1730static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1731 Instruction *InsertBefore, Value *Addr,
1732 MaybeAlign Alignment, unsigned Granularity,
1733 TypeSize TypeStoreSize, bool IsWrite,
1734 Value *SizeArgument, bool UseCalls,
1735 uint32_t Exp, RuntimeCallInserter &RTCI) {
1736 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1737 // if the data is properly aligned.
1738 if (!TypeStoreSize.isScalable()) {
1739 const auto FixedSize = TypeStoreSize.getFixedValue();
1740 switch (FixedSize) {
1741 case 8:
1742 case 16:
1743 case 32:
1744 case 64:
1745 case 128:
1746 if (!Alignment || *Alignment >= Granularity ||
1747 *Alignment >= FixedSize / 8)
1748 return Pass->instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment,
1749 TypeStoreSize: FixedSize, IsWrite, SizeArgument: nullptr, UseCalls,
1750 Exp, RTCI);
1751 }
1752 }
1753 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeStoreSize,
1754 IsWrite, SizeArgument: nullptr, UseCalls, Exp, RTCI);
1755}
1756
1757void AddressSanitizer::instrumentMaskedLoadOrStore(
1758 AddressSanitizer *Pass, const DataLayout &DL, Type *IntptrTy, Value *Mask,
1759 Value *EVL, Value *Stride, Instruction *I, Value *Addr,
1760 MaybeAlign Alignment, unsigned Granularity, Type *OpType, bool IsWrite,
1761 Value *SizeArgument, bool UseCalls, uint32_t Exp,
1762 RuntimeCallInserter &RTCI) {
1763 auto *VTy = cast<VectorType>(Val: OpType);
1764 TypeSize ElemTypeSize = DL.getTypeStoreSizeInBits(Ty: VTy->getScalarType());
1765 auto Zero = ConstantInt::get(Ty: IntptrTy, V: 0);
1766
1767 IRBuilder IB(I);
1768 Instruction *LoopInsertBefore = I;
1769 if (EVL) {
1770 // The end argument of SplitBlockAndInsertForLane is assumed bigger
1771 // than zero, so we should check whether EVL is zero here.
1772 Type *EVLType = EVL->getType();
1773 Value *IsEVLZero = IB.CreateICmpNE(LHS: EVL, RHS: ConstantInt::get(Ty: EVLType, V: 0));
1774 LoopInsertBefore = SplitBlockAndInsertIfThen(Cond: IsEVLZero, SplitBefore: I, Unreachable: false);
1775 IB.SetInsertPoint(LoopInsertBefore);
1776 // Cast EVL to IntptrTy.
1777 EVL = IB.CreateZExtOrTrunc(V: EVL, DestTy: IntptrTy);
1778 // To avoid undefined behavior for extracting with out of range index, use
1779 // the minimum of evl and element count as trip count.
1780 Value *EC = IB.CreateElementCount(Ty: IntptrTy, EC: VTy->getElementCount());
1781 EVL = IB.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: EVL, RHS: EC);
1782 } else {
1783 EVL = IB.CreateElementCount(Ty: IntptrTy, EC: VTy->getElementCount());
1784 }
1785
1786 // Cast Stride to IntptrTy.
1787 if (Stride)
1788 Stride = IB.CreateZExtOrTrunc(V: Stride, DestTy: IntptrTy);
1789
1790 SplitBlockAndInsertForEachLane(End: EVL, InsertBefore: LoopInsertBefore->getIterator(),
1791 Func: [&](IRBuilderBase &IRB, Value *Index) {
1792 Value *MaskElem = IRB.CreateExtractElement(Vec: Mask, Idx: Index);
1793 if (auto *MaskElemC = dyn_cast<ConstantInt>(Val: MaskElem)) {
1794 if (MaskElemC->isZero())
1795 // No check
1796 return;
1797 // Unconditional check
1798 } else {
1799 // Conditional check
1800 Instruction *ThenTerm = SplitBlockAndInsertIfThen(
1801 Cond: MaskElem, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false);
1802 IRB.SetInsertPoint(ThenTerm);
1803 }
1804
1805 Value *InstrumentedAddress;
1806 if (isa<VectorType>(Val: Addr->getType())) {
1807 assert(
1808 cast<VectorType>(Addr->getType())->getElementType()->isPointerTy() &&
1809 "Expected vector of pointer.");
1810 InstrumentedAddress = IRB.CreateExtractElement(Vec: Addr, Idx: Index);
1811 } else if (Stride) {
1812 Index = IRB.CreateMul(LHS: Index, RHS: Stride);
1813 InstrumentedAddress = IRB.CreatePtrAdd(Ptr: Addr, Offset: Index);
1814 } else {
1815 InstrumentedAddress = IRB.CreateGEP(Ty: VTy, Ptr: Addr, IdxList: {Zero, Index});
1816 }
1817 doInstrumentAddress(Pass, I, InsertBefore: &*IRB.GetInsertPoint(), Addr: InstrumentedAddress,
1818 Alignment, Granularity, TypeStoreSize: ElemTypeSize, IsWrite,
1819 SizeArgument, UseCalls, Exp, RTCI);
1820 });
1821}
1822
1823void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1824 InterestingMemoryOperand &O, bool UseCalls,
1825 const DataLayout &DL,
1826 RuntimeCallInserter &RTCI) {
1827 Value *Addr = O.getPtr();
1828
1829 // Optimization experiments.
1830 // The experiments can be used to evaluate potential optimizations that remove
1831 // instrumentation (assess false negatives). Instead of completely removing
1832 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1833 // experiments that want to remove instrumentation of this instruction).
1834 // If Exp is non-zero, this pass will emit special calls into runtime
1835 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1836 // make runtime terminate the program in a special way (with a different
1837 // exit status). Then you run the new compiler on a buggy corpus, collect
1838 // the special terminations (ideally, you don't see them at all -- no false
1839 // negatives) and make the decision on the optimization.
1840 uint32_t Exp = ClForceExperiment;
1841
1842 if (ClOpt && ClOptGlobals) {
1843 // If initialization order checking is disabled, a simple access to a
1844 // dynamically initialized global is always valid.
1845 GlobalVariable *G = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: Addr));
1846 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1847 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1848 NumOptimizedAccessesToGlobalVar++;
1849 return;
1850 }
1851 }
1852
1853 if (ClOpt && ClOptStack) {
1854 // A direct inbounds access to a stack variable is always valid.
1855 if (isa<AllocaInst>(Val: getUnderlyingObject(V: Addr)) &&
1856 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1857 NumOptimizedAccessesToStackVar++;
1858 return;
1859 }
1860 }
1861
1862 if (O.IsWrite)
1863 NumInstrumentedWrites++;
1864 else
1865 NumInstrumentedReads++;
1866
1867 if (O.MaybeByteOffset) {
1868 Type *Ty = Type::getInt8Ty(C&: *C);
1869 IRBuilder IB(O.getInsn());
1870
1871 Value *OffsetOp = O.MaybeByteOffset;
1872 if (TargetTriple.isRISCV()) {
1873 Type *OffsetTy = OffsetOp->getType();
1874 // RVV indexed loads/stores zero-extend offset operands which are narrower
1875 // than XLEN to XLEN.
1876 if (OffsetTy->getScalarType()->getIntegerBitWidth() <
1877 static_cast<unsigned>(LongSize)) {
1878 VectorType *OrigType = cast<VectorType>(Val: OffsetTy);
1879 Type *ExtendTy = VectorType::get(ElementType: IntptrTy, Other: OrigType);
1880 OffsetOp = IB.CreateZExt(V: OffsetOp, DestTy: ExtendTy);
1881 }
1882 }
1883 Addr = IB.CreateGEP(Ty, Ptr: Addr, IdxList: {OffsetOp});
1884 }
1885
1886 unsigned Granularity = 1 << Mapping.Scale;
1887 if (O.MaybeMask) {
1888 instrumentMaskedLoadOrStore(Pass: this, DL, IntptrTy, Mask: O.MaybeMask, EVL: O.MaybeEVL,
1889 Stride: O.MaybeStride, I: O.getInsn(), Addr, Alignment: O.Alignment,
1890 Granularity, OpType: O.OpType, IsWrite: O.IsWrite, SizeArgument: nullptr,
1891 UseCalls, Exp, RTCI);
1892 } else {
1893 doInstrumentAddress(Pass: this, I: O.getInsn(), InsertBefore: O.getInsn(), Addr, Alignment: O.Alignment,
1894 Granularity, TypeStoreSize: O.TypeStoreSize, IsWrite: O.IsWrite, SizeArgument: nullptr,
1895 UseCalls, Exp, RTCI);
1896 }
1897}
1898
1899Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1900 Value *Addr, bool IsWrite,
1901 size_t AccessSizeIndex,
1902 Value *SizeArgument,
1903 uint32_t Exp,
1904 RuntimeCallInserter &RTCI) {
1905 InstrumentationIRBuilder IRB(InsertBefore);
1906 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp);
1907 CallInst *Call = nullptr;
1908 if (SizeArgument) {
1909 if (Exp == 0)
1910 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][0],
1911 Args: {Addr, SizeArgument});
1912 else
1913 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][1],
1914 Args: {Addr, SizeArgument, ExpVal});
1915 } else {
1916 if (Exp == 0)
1917 Call = RTCI.createRuntimeCall(
1918 IRB, Callee: AsanErrorCallback[IsWrite][0][AccessSizeIndex], Args: Addr);
1919 else
1920 Call = RTCI.createRuntimeCall(
1921 IRB, Callee: AsanErrorCallback[IsWrite][1][AccessSizeIndex], Args: {Addr, ExpVal});
1922 }
1923
1924 Call->setCannotMerge();
1925 return Call;
1926}
1927
1928Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1929 Value *ShadowValue,
1930 uint32_t TypeStoreSize) {
1931 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1932 // Addr & (Granularity - 1)
1933 Value *LastAccessedByte =
1934 IRB.CreateAnd(LHS: AddrLong, RHS: ConstantInt::get(Ty: IntptrTy, V: Granularity - 1));
1935 // (Addr & (Granularity - 1)) + size - 1
1936 if (TypeStoreSize / 8 > 1)
1937 LastAccessedByte = IRB.CreateAdd(
1938 LHS: LastAccessedByte, RHS: ConstantInt::get(Ty: IntptrTy, V: TypeStoreSize / 8 - 1));
1939 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1940 LastAccessedByte =
1941 IRB.CreateIntCast(V: LastAccessedByte, DestTy: ShadowValue->getType(), isSigned: false);
1942 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1943 return IRB.CreateICmpSGE(LHS: LastAccessedByte, RHS: ShadowValue);
1944}
1945
1946Instruction *AddressSanitizer::instrumentAMDGPUAddress(
1947 Instruction *OrigIns, Instruction *InsertBefore, Value *Addr,
1948 uint32_t TypeStoreSize, bool IsWrite, Value *SizeArgument) {
1949 // Do not instrument unsupported addrspaces.
1950 if (isUnsupportedAMDGPUAddrspace(Addr))
1951 return nullptr;
1952 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1953 // Follow host instrumentation for global and constant addresses.
1954 if (PtrTy->getPointerAddressSpace() != 0)
1955 return InsertBefore;
1956 // Instrument generic addresses in supported addressspaces.
1957 IRBuilder<> IRB(InsertBefore);
1958 Value *IsShared = IRB.CreateCall(Callee: AMDGPUAddressShared, Args: {Addr});
1959 Value *IsPrivate = IRB.CreateCall(Callee: AMDGPUAddressPrivate, Args: {Addr});
1960 Value *IsSharedOrPrivate = IRB.CreateOr(LHS: IsShared, RHS: IsPrivate);
1961 Value *Cmp = IRB.CreateNot(V: IsSharedOrPrivate);
1962 Value *AddrSpaceZeroLanding =
1963 SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false);
1964 InsertBefore = cast<Instruction>(Val: AddrSpaceZeroLanding);
1965 return InsertBefore;
1966}
1967
1968Instruction *AddressSanitizer::genAMDGPUReportBlock(IRBuilder<> &IRB,
1969 Value *Cond, bool Recover) {
1970 Value *ReportCond = Cond;
1971 if (!Recover) {
1972 auto Ballot = Inserter.insertFunction(Name: kAMDGPUBallotName, Args: IRB.getInt64Ty(),
1973 Args: IRB.getInt1Ty());
1974 ReportCond = IRB.CreateIsNotNull(Arg: IRB.CreateCall(Callee: Ballot, Args: {Cond}));
1975 }
1976
1977 auto *Trm =
1978 SplitBlockAndInsertIfThen(Cond: ReportCond, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false,
1979 BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
1980 Trm->getParent()->setName("asan.report");
1981
1982 if (Recover)
1983 return Trm;
1984
1985 Trm = SplitBlockAndInsertIfThen(Cond, SplitBefore: Trm, Unreachable: false);
1986 IRB.SetInsertPoint(Trm);
1987 return IRB.CreateCall(
1988 Callee: Inserter.insertFunction(Name: kAMDGPUUnreachableName, Args: IRB.getVoidTy()), Args: {});
1989}
1990
1991void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1992 Instruction *InsertBefore, Value *Addr,
1993 MaybeAlign Alignment,
1994 uint32_t TypeStoreSize, bool IsWrite,
1995 Value *SizeArgument, bool UseCalls,
1996 uint32_t Exp,
1997 RuntimeCallInserter &RTCI) {
1998 if (TargetTriple.isAMDGPU()) {
1999 InsertBefore = instrumentAMDGPUAddress(OrigIns, InsertBefore, Addr,
2000 TypeStoreSize, IsWrite, SizeArgument);
2001 if (!InsertBefore)
2002 return;
2003 }
2004
2005 InstrumentationIRBuilder IRB(InsertBefore);
2006 size_t AccessSizeIndex = TypeStoreSizeToSizeIndex(TypeSize: TypeStoreSize);
2007
2008 if (UseCalls && ClOptimizeCallbacks) {
2009 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
2010 IRB.CreateIntrinsic(ID: Intrinsic::asan_check_memaccess, OverloadTypes: {},
2011 Args: {IRB.CreatePointerCast(V: Addr, DestTy: PtrTy),
2012 ConstantInt::get(Ty: Int32Ty, V: AccessInfo.Packed)});
2013 return;
2014 }
2015
2016 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
2017 if (UseCalls) {
2018 if (Exp == 0)
2019 RTCI.createRuntimeCall(
2020 IRB, Callee: AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex], Args: AddrLong);
2021 else
2022 RTCI.createRuntimeCall(
2023 IRB, Callee: AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
2024 Args: {AddrLong, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
2025 return;
2026 }
2027
2028 Type *ShadowTy =
2029 IntegerType::get(C&: *C, NumBits: std::max(a: 8U, b: TypeStoreSize >> Mapping.Scale));
2030 Type *ShadowPtrTy = PointerType::get(C&: *C, AddressSpace: ClShadowAddrSpace);
2031 Value *ShadowPtr = memToShadow(Shadow: AddrLong, IRB);
2032 const uint64_t ShadowAlign =
2033 std::max<uint64_t>(a: Alignment.valueOrOne().value() >> Mapping.Scale, b: 1);
2034 Value *ShadowValue = IRB.CreateAlignedLoad(
2035 Ty: ShadowTy, Ptr: IRB.CreateIntToPtr(V: ShadowPtr, DestTy: ShadowPtrTy), Align: Align(ShadowAlign));
2036
2037 Value *Cmp = IRB.CreateIsNotNull(Arg: ShadowValue);
2038 size_t Granularity = 1ULL << Mapping.Scale;
2039 Instruction *CrashTerm = nullptr;
2040
2041 bool GenSlowPath = (ClAlwaysSlowPath || (TypeStoreSize < 8 * Granularity));
2042
2043 if (TargetTriple.isAMDGCN()) {
2044 if (GenSlowPath) {
2045 auto *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
2046 Cmp = IRB.CreateAnd(LHS: Cmp, RHS: Cmp2);
2047 }
2048 CrashTerm = genAMDGPUReportBlock(IRB, Cond: Cmp, Recover);
2049 } else if (GenSlowPath) {
2050 // We use branch weights for the slow path check, to indicate that the slow
2051 // path is rarely taken. This seems to be the case for SPEC benchmarks.
2052 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
2053 Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false, BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
2054 BasicBlock *NextBB = cast<UncondBrInst>(Val: CheckTerm)->getSuccessor();
2055 IRB.SetInsertPoint(CheckTerm);
2056 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
2057 if (Recover) {
2058 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp2, SplitBefore: CheckTerm, Unreachable: false);
2059 } else {
2060 BasicBlock *CrashBlock =
2061 BasicBlock::Create(Context&: *C, Name: "", Parent: NextBB->getParent(), InsertBefore: NextBB);
2062 CrashTerm = new UnreachableInst(*C, CrashBlock);
2063 CondBrInst *NewTerm = CondBrInst::Create(Cond: Cmp2, IfTrue: CrashBlock, IfFalse: NextBB);
2064 ReplaceInstWithInst(From: CheckTerm, To: NewTerm);
2065 }
2066 } else {
2067 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: !Recover);
2068 }
2069
2070 Instruction *Crash = generateCrashCode(
2071 InsertBefore: CrashTerm, Addr: AddrLong, IsWrite, AccessSizeIndex, SizeArgument, Exp, RTCI);
2072 if (OrigIns->getDebugLoc())
2073 Crash->setDebugLoc(OrigIns->getDebugLoc());
2074}
2075
2076// Instrument unusual size or unusual alignment.
2077// We can not do it with a single check, so we do 1-byte check for the first
2078// and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
2079// to report the actual access size.
2080void AddressSanitizer::instrumentUnusualSizeOrAlignment(
2081 Instruction *I, Instruction *InsertBefore, Value *Addr,
2082 TypeSize TypeStoreSize, bool IsWrite, Value *SizeArgument, bool UseCalls,
2083 uint32_t Exp, RuntimeCallInserter &RTCI) {
2084 InstrumentationIRBuilder IRB(InsertBefore);
2085 Value *NumBits = IRB.CreateTypeSize(Ty: IntptrTy, Size: TypeStoreSize);
2086 Value *Size = IRB.CreateLShr(LHS: NumBits, RHS: ConstantInt::get(Ty: IntptrTy, V: 3));
2087
2088 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
2089 if (UseCalls) {
2090 if (Exp == 0)
2091 RTCI.createRuntimeCall(IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][0],
2092 Args: {AddrLong, Size});
2093 else
2094 RTCI.createRuntimeCall(
2095 IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][1],
2096 Args: {AddrLong, Size, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
2097 } else {
2098 Value *SizeMinusOne = IRB.CreateSub(LHS: Size, RHS: ConstantInt::get(Ty: IntptrTy, V: 1));
2099 Value *LastByte = IRB.CreateIntToPtr(
2100 V: IRB.CreateAdd(LHS: AddrLong, RHS: SizeMinusOne),
2101 DestTy: Addr->getType());
2102 instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false, Exp,
2103 RTCI);
2104 instrumentAddress(OrigIns: I, InsertBefore, Addr: LastByte, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false,
2105 Exp, RTCI);
2106 }
2107}
2108
2109void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit) {
2110 // Set up the arguments to our poison/unpoison functions.
2111 IRBuilder<> IRB(&GlobalInit.front(),
2112 GlobalInit.front().getFirstInsertionPt());
2113
2114 // Add a call to poison all external globals before the given function starts.
2115 Value *ModuleNameAddr =
2116 ConstantExpr::getPointerCast(C: getOrCreateModuleName(), Ty: IntptrTy);
2117 CallInst *CallBefore = IRB.CreateCall(Callee: AsanPoisonGlobals, Args: ModuleNameAddr);
2118 if (DISubprogram *SP = GlobalInit.getSubprogram())
2119 CallBefore->setDebugLoc(
2120 DILocation::get(Context&: SP->getContext(), Line: SP->getScopeLine(), Column: 0, Scope: SP));
2121
2122 // Add calls to unpoison all globals before each return instruction.
2123 for (auto &BB : GlobalInit)
2124 if (ReturnInst *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator())) {
2125 CallInst *CallAfter =
2126 CallInst::Create(Func: AsanUnpoisonGlobals, NameStr: "", InsertBefore: RI->getIterator());
2127 if (RI->getDebugLoc())
2128 CallAfter->setDebugLoc(RI->getDebugLoc());
2129 else if (DISubprogram *SP = GlobalInit.getSubprogram())
2130 CallAfter->setDebugLoc(
2131 DILocation::get(Context&: SP->getContext(), Line: SP->getScopeLine(), Column: 0, Scope: SP));
2132 }
2133}
2134
2135void ModuleAddressSanitizer::createInitializerPoisonCalls() {
2136 GlobalVariable *GV = M.getGlobalVariable(Name: "llvm.global_ctors");
2137 if (!GV)
2138 return;
2139
2140 ConstantArray *CA = dyn_cast<ConstantArray>(Val: GV->getInitializer());
2141 if (!CA)
2142 return;
2143
2144 for (Use &OP : CA->operands()) {
2145 if (isa<ConstantAggregateZero>(Val: OP)) continue;
2146 ConstantStruct *CS = cast<ConstantStruct>(Val&: OP);
2147
2148 // Must have a function or null ptr.
2149 if (Function *F = dyn_cast<Function>(Val: CS->getOperand(i_nocapture: 1))) {
2150 if (F->getName() == kAsanModuleCtorName) continue;
2151 auto *Priority = cast<ConstantInt>(Val: CS->getOperand(i_nocapture: 0));
2152 // Don't instrument CTORs that will run before asan.module_ctor.
2153 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
2154 continue;
2155 poisonOneInitializer(GlobalInit&: *F);
2156 }
2157 }
2158}
2159
2160const GlobalVariable *
2161ModuleAddressSanitizer::getExcludedAliasedGlobal(const GlobalAlias &GA) const {
2162 // In case this function should be expanded to include rules that do not just
2163 // apply when CompileKernel is true, either guard all existing rules with an
2164 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
2165 // should also apply to user space.
2166 assert(CompileKernel && "Only expecting to be called when compiling kernel");
2167
2168 const Constant *C = GA.getAliasee();
2169
2170 // When compiling the kernel, globals that are aliased by symbols prefixed
2171 // by "__" are special and cannot be padded with a redzone.
2172 if (GA.getName().starts_with(Prefix: "__"))
2173 return dyn_cast<GlobalVariable>(Val: C->stripPointerCastsAndAliases());
2174
2175 return nullptr;
2176}
2177
2178bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
2179 Type *Ty = G->getValueType();
2180 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
2181
2182 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().NoAddress)
2183 return false;
2184 if (!Ty->isSized()) return false;
2185 if (!G->hasInitializer()) return false;
2186 if (!isSupportedAddrspace(TargetTriple, Addr: G))
2187 return false;
2188 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
2189 // Two problems with thread-locals:
2190 // - The address of the main thread's copy can't be computed at link-time.
2191 // - Need to poison all copies, not just the main thread's one.
2192 if (G->isThreadLocal()) return false;
2193 // For now, just ignore this Global if the alignment is large.
2194 if (G->getAlign() && *G->getAlign() > getMinRedzoneSizeForGlobal()) return false;
2195
2196 // For non-COFF targets, only instrument globals known to be defined by this
2197 // TU.
2198 // FIXME: We can instrument comdat globals on ELF if we are using the
2199 // GC-friendly metadata scheme.
2200 if (!TargetTriple.isOSBinFormatCOFF()) {
2201 if (!G->hasExactDefinition() || G->hasComdat())
2202 return false;
2203 } else {
2204 // On COFF, don't instrument non-ODR linkages.
2205 if (G->isInterposable())
2206 return false;
2207 // If the global has AvailableExternally linkage, then it is not in this
2208 // module, which means it does not need to be instrumented.
2209 if (G->hasAvailableExternallyLinkage())
2210 return false;
2211 }
2212
2213 // If a comdat is present, it must have a selection kind that implies ODR
2214 // semantics: no duplicates, any, or exact match.
2215 if (Comdat *C = G->getComdat()) {
2216 switch (C->getSelectionKind()) {
2217 case Comdat::Any:
2218 case Comdat::ExactMatch:
2219 case Comdat::NoDeduplicate:
2220 break;
2221 case Comdat::Largest:
2222 case Comdat::SameSize:
2223 return false;
2224 }
2225 }
2226
2227 if (G->hasSection()) {
2228 // The kernel uses explicit sections for mostly special global variables
2229 // that we should not instrument. E.g. the kernel may rely on their layout
2230 // without redzones, or remove them at link time ("discard.*"), etc.
2231 if (CompileKernel)
2232 return false;
2233
2234 StringRef Section = G->getSection();
2235
2236 // Globals from llvm.metadata aren't emitted, do not instrument them.
2237 if (Section == "llvm.metadata") return false;
2238 // Do not instrument globals from special LLVM sections.
2239 if (Section.contains(Other: "__llvm") || Section.contains(Other: "__LLVM"))
2240 return false;
2241
2242 // Do not instrument function pointers to initialization and termination
2243 // routines: dynamic linker will not properly handle redzones.
2244 if (Section.starts_with(Prefix: ".preinit_array") ||
2245 Section.starts_with(Prefix: ".init_array") ||
2246 Section.starts_with(Prefix: ".fini_array")) {
2247 return false;
2248 }
2249
2250 // Do not instrument user-defined sections (with names resembling
2251 // valid C identifiers)
2252 if (TargetTriple.isOSBinFormatELF()) {
2253 if (llvm::all_of(Range&: Section,
2254 P: [](char c) { return llvm::isAlnum(C: c) || c == '_'; }))
2255 return false;
2256 }
2257
2258 // On COFF, if the section name contains '$', it is highly likely that the
2259 // user is using section sorting to create an array of globals similar to
2260 // the way initialization callbacks are registered in .init_array and
2261 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
2262 // to such globals is counterproductive, because the intent is that they
2263 // will form an array, and out-of-bounds accesses are expected.
2264 // See https://github.com/google/sanitizers/issues/305
2265 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
2266 if (TargetTriple.isOSBinFormatCOFF() && Section.contains(C: '$')) {
2267 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
2268 << *G << "\n");
2269 return false;
2270 }
2271
2272 if (TargetTriple.isOSBinFormatMachO()) {
2273 StringRef ParsedSegment, ParsedSection;
2274 unsigned TAA = 0, StubSize = 0;
2275 bool TAAParsed;
2276 cantFail(Err: MCSectionMachO::ParseSectionSpecifier(
2277 Spec: Section, Segment&: ParsedSegment, Section&: ParsedSection, TAA, TAAParsed, StubSize));
2278
2279 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
2280 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
2281 // them.
2282 if (ParsedSegment == "__OBJC" ||
2283 (ParsedSegment == "__DATA" && ParsedSection.starts_with(Prefix: "__objc_"))) {
2284 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
2285 return false;
2286 }
2287 // See https://github.com/google/sanitizers/issues/32
2288 // Constant CFString instances are compiled in the following way:
2289 // -- the string buffer is emitted into
2290 // __TEXT,__cstring,cstring_literals
2291 // -- the constant NSConstantString structure referencing that buffer
2292 // is placed into __DATA,__cfstring
2293 // Therefore there's no point in placing redzones into __DATA,__cfstring.
2294 // Moreover, it causes the linker to crash on OS X 10.7
2295 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
2296 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
2297 return false;
2298 }
2299 // The linker merges the contents of cstring_literals and removes the
2300 // trailing zeroes.
2301 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
2302 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
2303 return false;
2304 }
2305 }
2306 }
2307
2308 if (CompileKernel) {
2309 // Globals that prefixed by "__" are special and cannot be padded with a
2310 // redzone.
2311 if (G->getName().starts_with(Prefix: "__"))
2312 return false;
2313 }
2314
2315 return true;
2316}
2317
2318// On Mach-O platforms, we emit global metadata in a separate section of the
2319// binary in order to allow the linker to properly dead strip. This is only
2320// supported on recent versions of ld64.
2321bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
2322 if (!TargetTriple.isOSBinFormatMachO())
2323 return false;
2324
2325 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(Major: 10, Minor: 11))
2326 return true;
2327 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(Major: 9))
2328 return true;
2329 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(Major: 2))
2330 return true;
2331 if (TargetTriple.isDriverKit())
2332 return true;
2333 if (TargetTriple.isXROS())
2334 return true;
2335
2336 return false;
2337}
2338
2339StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
2340 switch (TargetTriple.getObjectFormat()) {
2341 case Triple::COFF: return ".ASAN$GL";
2342 case Triple::ELF: return "asan_globals";
2343 case Triple::MachO: return "__DATA,__asan_globals,regular";
2344 case Triple::Wasm:
2345 case Triple::GOFF:
2346 case Triple::SPIRV:
2347 case Triple::XCOFF:
2348 case Triple::DXContainer:
2349 report_fatal_error(
2350 reason: "ModuleAddressSanitizer not implemented for object file format");
2351 case Triple::UnknownObjectFormat:
2352 break;
2353 }
2354 llvm_unreachable("unsupported object format");
2355}
2356
2357void ModuleAddressSanitizer::initializeCallbacks() {
2358 IRBuilder<> IRB(*C);
2359
2360 // Declare our poisoning and unpoisoning functions.
2361 AsanPoisonGlobals = Inserter.insertFunction(Name: kAsanPoisonGlobalsName,
2362 Args: IRB.getVoidTy(), Args&: IntptrTy);
2363 AsanUnpoisonGlobals =
2364 Inserter.insertFunction(Name: kAsanUnpoisonGlobalsName, Args: IRB.getVoidTy());
2365
2366 // Declare functions that register/unregister globals.
2367 AsanRegisterGlobals = Inserter.insertFunction(
2368 Name: kAsanRegisterGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2369 AsanUnregisterGlobals = Inserter.insertFunction(
2370 Name: kAsanUnregisterGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2371
2372 // Declare the functions that find globals in a shared object and then invoke
2373 // the (un)register function on them.
2374 AsanRegisterImageGlobals = Inserter.insertFunction(
2375 Name: kAsanRegisterImageGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy);
2376 AsanUnregisterImageGlobals = Inserter.insertFunction(
2377 Name: kAsanUnregisterImageGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy);
2378
2379 AsanRegisterElfGlobals =
2380 Inserter.insertFunction(Name: kAsanRegisterElfGlobalsName, Args: IRB.getVoidTy(),
2381 Args&: IntptrTy, Args&: IntptrTy, Args&: IntptrTy);
2382 AsanUnregisterElfGlobals =
2383 Inserter.insertFunction(Name: kAsanUnregisterElfGlobalsName, Args: IRB.getVoidTy(),
2384 Args&: IntptrTy, Args&: IntptrTy, Args&: IntptrTy);
2385}
2386
2387// Put the metadata and the instrumented global in the same group. This ensures
2388// that the metadata is discarded if the instrumented global is discarded.
2389void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
2390 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
2391 Module &M = *G->getParent();
2392 Comdat *C = G->getComdat();
2393 if (!C) {
2394 if (!G->hasName()) {
2395 // If G is unnamed, it must be internal. Give it an artificial name
2396 // so we can put it in a comdat.
2397 assert(G->hasLocalLinkage());
2398 G->setName(genName(suffix: "anon_global"));
2399 }
2400
2401 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
2402 std::string Name = std::string(G->getName());
2403 Name += InternalSuffix;
2404 C = M.getOrInsertComdat(Name);
2405 } else {
2406 C = M.getOrInsertComdat(Name: G->getName());
2407 }
2408
2409 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
2410 // linkage to internal linkage so that a symbol table entry is emitted. This
2411 // is necessary in order to create the comdat group.
2412 if (TargetTriple.isOSBinFormatCOFF()) {
2413 C->setSelectionKind(Comdat::NoDeduplicate);
2414 if (G->hasPrivateLinkage())
2415 G->setLinkage(GlobalValue::InternalLinkage);
2416 }
2417 G->setComdat(C);
2418 }
2419
2420 assert(G->hasComdat());
2421 Metadata->setComdat(G->getComdat());
2422}
2423
2424// Create a separate metadata global and put it in the appropriate ASan
2425// global registration section.
2426GlobalVariable *
2427ModuleAddressSanitizer::CreateMetadataGlobal(Constant *Initializer,
2428 StringRef OriginalName) {
2429 auto Linkage = TargetTriple.isOSBinFormatMachO()
2430 ? GlobalVariable::InternalLinkage
2431 : GlobalVariable::PrivateLinkage;
2432 GlobalVariable *Metadata = new GlobalVariable(
2433 M, Initializer->getType(), false, Linkage, Initializer,
2434 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(Name: OriginalName));
2435 Metadata->setSection(getGlobalMetadataSection());
2436 // Place metadata in a large section for x86-64 ELF binaries to mitigate
2437 // relocation pressure.
2438 setGlobalVariableLargeSection(TargetTriple, GV&: *Metadata);
2439 return Metadata;
2440}
2441
2442Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor() {
2443 AsanDtorFunction = Function::createWithDefaultAttr(
2444 Ty: FunctionType::get(Result: Type::getVoidTy(C&: *C), isVarArg: false),
2445 Linkage: GlobalValue::InternalLinkage, AddrSpace: 0, N: kAsanModuleDtorName, M: &M);
2446 AsanDtorFunction->addFnAttr(Kind: Attribute::NoUnwind);
2447 // Ensure Dtor cannot be discarded, even if in a comdat.
2448 appendToUsed(M, Values: {AsanDtorFunction});
2449 BasicBlock *AsanDtorBB = BasicBlock::Create(Context&: *C, Name: "", Parent: AsanDtorFunction);
2450
2451 return ReturnInst::Create(C&: *C, InsertAtEnd: AsanDtorBB);
2452}
2453
2454void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
2455 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2456 ArrayRef<Constant *> MetadataInitializers) {
2457 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2458 auto &DL = M.getDataLayout();
2459
2460 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2461 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2462 Constant *Initializer = MetadataInitializers[i];
2463 GlobalVariable *G = ExtendedGlobals[i];
2464 GlobalVariable *Metadata = CreateMetadataGlobal(Initializer, OriginalName: G->getName());
2465 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2466 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2467 MetadataGlobals[i] = Metadata;
2468
2469 // The MSVC linker always inserts padding when linking incrementally. We
2470 // cope with that by aligning each struct to its size, which must be a power
2471 // of two.
2472 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Ty: Initializer->getType());
2473 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2474 "global metadata will not be padded appropriately");
2475 Metadata->setAlignment(assumeAligned(Value: SizeOfGlobalStruct));
2476
2477 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: "");
2478 }
2479
2480 // Update llvm.compiler.used, adding the new metadata globals. This is
2481 // needed so that during LTO these variables stay alive.
2482 if (!MetadataGlobals.empty())
2483 appendToCompilerUsed(M, Values: MetadataGlobals);
2484}
2485
2486void ModuleAddressSanitizer::instrumentGlobalsELF(
2487 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2488 ArrayRef<Constant *> MetadataInitializers,
2489 const std::string &UniqueModuleId) {
2490 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2491
2492 // Putting globals in a comdat changes the semantic and potentially cause
2493 // false negative odr violations at link time. If odr indicators are used, we
2494 // keep the comdat sections, as link time odr violations will be detected on
2495 // the odr indicator symbols.
2496 bool UseComdatForGlobalsGC = UseOdrIndicator && !UniqueModuleId.empty();
2497
2498 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2499 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2500 GlobalVariable *G = ExtendedGlobals[i];
2501 GlobalVariable *Metadata =
2502 CreateMetadataGlobal(Initializer: MetadataInitializers[i], OriginalName: G->getName());
2503 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2504 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2505 MetadataGlobals[i] = Metadata;
2506
2507 if (UseComdatForGlobalsGC)
2508 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: UniqueModuleId);
2509 }
2510
2511 // Update llvm.compiler.used, adding the new metadata globals. This is
2512 // needed so that during LTO these variables stay alive.
2513 if (!MetadataGlobals.empty())
2514 appendToCompilerUsed(M, Values: MetadataGlobals);
2515
2516 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2517 // to look up the loaded image that contains it. Second, we can store in it
2518 // whether registration has already occurred, to prevent duplicate
2519 // registration.
2520 //
2521 // Common linkage ensures that there is only one global per shared library.
2522 GlobalVariable *RegisteredFlag = new GlobalVariable(
2523 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2524 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2525 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2526
2527 // Create start and stop symbols.
2528 GlobalVariable *StartELFMetadata = new GlobalVariable(
2529 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2530 "__start_" + getGlobalMetadataSection());
2531 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2532 GlobalVariable *StopELFMetadata = new GlobalVariable(
2533 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2534 "__stop_" + getGlobalMetadataSection());
2535 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2536
2537 // Create a call to register the globals with the runtime.
2538 if (ConstructorKind == AsanCtorKind::Global)
2539 IRB.CreateCall(Callee: AsanRegisterElfGlobals,
2540 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2541 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2542 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2543
2544 // We also need to unregister globals at the end, e.g., when a shared library
2545 // gets closed.
2546 if (DestructorKind != AsanDtorKind::None && !MetadataGlobals.empty()) {
2547 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2548 IrbDtor.CreateCall(Callee: AsanUnregisterElfGlobals,
2549 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2550 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2551 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2552 }
2553}
2554
2555void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2556 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2557 ArrayRef<Constant *> MetadataInitializers) {
2558 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2559
2560 // On recent Mach-O platforms, use a structure which binds the liveness of
2561 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2562 // created to be added to llvm.compiler.used
2563 StructType *LivenessTy = StructType::get(elt1: IntptrTy, elts: IntptrTy);
2564 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2565
2566 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2567 Constant *Initializer = MetadataInitializers[i];
2568 GlobalVariable *G = ExtendedGlobals[i];
2569 GlobalVariable *Metadata = CreateMetadataGlobal(Initializer, OriginalName: G->getName());
2570
2571 // On recent Mach-O platforms, we emit the global metadata in a way that
2572 // allows the linker to properly strip dead globals.
2573 auto LivenessBinder =
2574 ConstantStruct::get(T: LivenessTy, Vs: Initializer->getAggregateElement(Elt: 0u),
2575 Vs: ConstantExpr::getPointerCast(C: Metadata, Ty: IntptrTy));
2576 GlobalVariable *Liveness = new GlobalVariable(
2577 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2578 Twine("__asan_binder_") + G->getName());
2579 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2580 LivenessGlobals[i] = Liveness;
2581 }
2582
2583 // Update llvm.compiler.used, adding the new liveness globals. This is
2584 // needed so that during LTO these variables stay alive. The alternative
2585 // would be to have the linker handling the LTO symbols, but libLTO
2586 // current API does not expose access to the section for each symbol.
2587 if (!LivenessGlobals.empty())
2588 appendToCompilerUsed(M, Values: LivenessGlobals);
2589
2590 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2591 // to look up the loaded image that contains it. Second, we can store in it
2592 // whether registration has already occurred, to prevent duplicate
2593 // registration.
2594 //
2595 // common linkage ensures that there is only one global per shared library.
2596 GlobalVariable *RegisteredFlag = new GlobalVariable(
2597 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2598 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2599 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2600
2601 if (ConstructorKind == AsanCtorKind::Global)
2602 IRB.CreateCall(Callee: AsanRegisterImageGlobals,
2603 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2604
2605 // We also need to unregister globals at the end, e.g., when a shared library
2606 // gets closed.
2607 if (DestructorKind != AsanDtorKind::None) {
2608 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2609 IrbDtor.CreateCall(Callee: AsanUnregisterImageGlobals,
2610 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2611 }
2612}
2613
2614void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2615 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2616 ArrayRef<Constant *> MetadataInitializers) {
2617 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2618 unsigned N = ExtendedGlobals.size();
2619 assert(N > 0);
2620
2621 // On platforms that don't have a custom metadata section, we emit an array
2622 // of global metadata structures.
2623 ArrayType *ArrayOfGlobalStructTy =
2624 ArrayType::get(ElementType: MetadataInitializers[0]->getType(), NumElements: N);
2625 auto AllGlobals = new GlobalVariable(
2626 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2627 ConstantArray::get(T: ArrayOfGlobalStructTy, V: MetadataInitializers), "");
2628 if (Mapping.Scale > 3)
2629 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2630
2631 if (ConstructorKind == AsanCtorKind::Global)
2632 IRB.CreateCall(Callee: AsanRegisterGlobals,
2633 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2634 ConstantInt::get(Ty: IntptrTy, V: N)});
2635
2636 // We also need to unregister globals at the end, e.g., when a shared library
2637 // gets closed.
2638 if (DestructorKind != AsanDtorKind::None) {
2639 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2640 IrbDtor.CreateCall(Callee: AsanUnregisterGlobals,
2641 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2642 ConstantInt::get(Ty: IntptrTy, V: N)});
2643 }
2644}
2645
2646// This function replaces all global variables with new variables that have
2647// trailing redzones. It also creates a function that poisons
2648// redzones and inserts this function into llvm.global_ctors.
2649// Sets *CtorComdat to true if the global registration code emitted into the
2650// asan constructor is comdat-compatible.
2651void ModuleAddressSanitizer::instrumentGlobals(IRBuilder<> &IRB,
2652 bool *CtorComdat) {
2653 // Build set of globals that are aliased by some GA, where
2654 // getExcludedAliasedGlobal(GA) returns the relevant GlobalVariable.
2655 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2656 if (CompileKernel) {
2657 for (auto &GA : M.aliases()) {
2658 if (const GlobalVariable *GV = getExcludedAliasedGlobal(GA))
2659 AliasedGlobalExclusions.insert(Ptr: GV);
2660 }
2661 }
2662
2663 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2664 for (auto &G : M.globals()) {
2665 if (!AliasedGlobalExclusions.count(Ptr: &G) && shouldInstrumentGlobal(G: &G))
2666 GlobalsToChange.push_back(Elt: &G);
2667 }
2668
2669 size_t n = GlobalsToChange.size();
2670 auto &DL = M.getDataLayout();
2671
2672 // A global is described by a structure
2673 // size_t beg;
2674 // size_t size;
2675 // size_t size_with_redzone;
2676 // const char *name;
2677 // const char *module_name;
2678 // size_t has_dynamic_init;
2679 // size_t padding_for_windows_msvc_incremental_link;
2680 // size_t odr_indicator;
2681 // We initialize an array of such structures and pass it to a run-time call.
2682 StructType *GlobalStructTy =
2683 StructType::get(elt1: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy,
2684 elts: IntptrTy, elts: IntptrTy, elts: IntptrTy);
2685 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2686 SmallVector<Constant *, 16> Initializers(n);
2687
2688 for (size_t i = 0; i < n; i++) {
2689 GlobalVariable *G = GlobalsToChange[i];
2690
2691 GlobalValue::SanitizerMetadata MD;
2692 if (G->hasSanitizerMetadata())
2693 MD = G->getSanitizerMetadata();
2694
2695 // The runtime library tries demangling symbol names in the descriptor but
2696 // functionality like __cxa_demangle may be unavailable (e.g.
2697 // -static-libstdc++). So we demangle the symbol names here.
2698 std::string NameForGlobal = G->getName().str();
2699 GlobalVariable *Name =
2700 createPrivateGlobalForString(M, Str: llvm::demangle(MangledName: NameForGlobal),
2701 /*AllowMerging*/ true, NamePrefix: genName(suffix: "global"));
2702
2703 Type *Ty = G->getValueType();
2704 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2705 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2706 Type *RightRedZoneTy = ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: RightRedzoneSize);
2707
2708 StructType *NewTy = StructType::get(elt1: Ty, elts: RightRedZoneTy);
2709 Constant *NewInitializer = ConstantStruct::get(
2710 T: NewTy, Vs: G->getInitializer(), Vs: Constant::getNullValue(Ty: RightRedZoneTy));
2711
2712 // Create a new global variable with enough space for a redzone.
2713 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2714 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2715 Linkage = GlobalValue::InternalLinkage;
2716 GlobalVariable *NewGlobal = new GlobalVariable(
2717 M, NewTy, G->isConstant(), Linkage, NewInitializer, "", G,
2718 G->getThreadLocalMode(), G->getAddressSpace());
2719 NewGlobal->copyAttributesFrom(Src: G);
2720 NewGlobal->setComdat(G->getComdat());
2721 NewGlobal->setAlignment(Align(getMinRedzoneSizeForGlobal()));
2722 // Don't fold globals with redzones. ODR violation detector and redzone
2723 // poisoning implicitly creates a dependence on the global's address, so it
2724 // is no longer valid for it to be marked unnamed_addr.
2725 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2726
2727 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2728 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2729 G->isConstant()) {
2730 auto Seq = dyn_cast<ConstantDataSequential>(Val: G->getInitializer());
2731 if (Seq && Seq->isCString())
2732 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2733 }
2734
2735 // Transfer the debug info and type metadata. The payload starts at offset
2736 // zero so we can copy the metadata over as is.
2737 NewGlobal->copyMetadata(Src: G, Offset: 0);
2738
2739 G->replaceAllUsesWith(V: NewGlobal);
2740 NewGlobal->takeName(V: G);
2741 G->eraseFromParent();
2742 NewGlobals[i] = NewGlobal;
2743
2744 Constant *ODRIndicator = Constant::getNullValue(Ty: IntptrTy);
2745 GlobalValue *InstrumentedGlobal = NewGlobal;
2746
2747 bool CanUsePrivateAliases =
2748 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2749 TargetTriple.isOSBinFormatWasm();
2750 if (CanUsePrivateAliases && UsePrivateAlias) {
2751 // Create local alias for NewGlobal to avoid crash on ODR between
2752 // instrumented and non-instrumented libraries.
2753 InstrumentedGlobal =
2754 GlobalAlias::create(Linkage: GlobalValue::PrivateLinkage, Name: "", Aliasee: NewGlobal);
2755 }
2756
2757 // ODR should not happen for local linkage.
2758 if (NewGlobal->hasLocalLinkage()) {
2759 ODRIndicator = ConstantInt::getAllOnesValue(Ty: IntptrTy);
2760 } else if (UseOdrIndicator) {
2761 // With local aliases, we need to provide another externally visible
2762 // symbol __odr_asan_XXX to detect ODR violation.
2763 auto *ODRIndicatorSym =
2764 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2765 Constant::getNullValue(Ty: IRB.getInt8Ty()),
2766 kODRGenPrefix + NameForGlobal, nullptr,
2767 NewGlobal->getThreadLocalMode());
2768
2769 // Set meaningful attributes for indicator symbol.
2770 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2771 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2772 ODRIndicatorSym->setAlignment(Align(1));
2773 ODRIndicator = ConstantExpr::getPtrToInt(C: ODRIndicatorSym, Ty: IntptrTy);
2774 }
2775
2776 Constant *Initializer = ConstantStruct::get(
2777 T: GlobalStructTy,
2778 Vs: ConstantExpr::getPointerCast(C: InstrumentedGlobal, Ty: IntptrTy),
2779 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes),
2780 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes + RightRedzoneSize),
2781 Vs: ConstantExpr::getPointerCast(C: Name, Ty: IntptrTy),
2782 Vs: ConstantExpr::getPointerCast(C: getOrCreateModuleName(), Ty: IntptrTy),
2783 Vs: ConstantInt::get(Ty: IntptrTy, V: MD.IsDynInit),
2784 Vs: Constant::getNullValue(Ty: IntptrTy), Vs: ODRIndicator);
2785
2786 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2787
2788 Initializers[i] = Initializer;
2789 }
2790
2791 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2792 // ConstantMerge'ing them.
2793 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2794 for (size_t i = 0; i < n; i++) {
2795 GlobalVariable *G = NewGlobals[i];
2796 if (G->getName().empty()) continue;
2797 GlobalsToAddToUsedList.push_back(Elt: G);
2798 }
2799 appendToCompilerUsed(M, Values: ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2800
2801 if (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) {
2802 // Use COMDAT and register globals even if n == 0 to ensure that (a) the
2803 // linkage unit will only have one module constructor, and (b) the register
2804 // function will be called. The module destructor is not created when n ==
2805 // 0.
2806 *CtorComdat = true;
2807 instrumentGlobalsELF(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers, UniqueModuleId: getUniqueModuleId(M: &M));
2808 } else if (n == 0) {
2809 // When UseGlobalsGC is false, COMDAT can still be used if n == 0, because
2810 // all compile units will have identical module constructor/destructor.
2811 *CtorComdat = TargetTriple.isOSBinFormatELF();
2812 } else {
2813 *CtorComdat = false;
2814 if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2815 InstrumentGlobalsCOFF(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2816 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2817 InstrumentGlobalsMachO(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2818 } else {
2819 InstrumentGlobalsWithMetadataArray(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2820 }
2821 }
2822
2823 // Create calls for poisoning before initializers run and unpoisoning after.
2824 if (ClInitializers)
2825 createInitializerPoisonCalls();
2826
2827 LLVM_DEBUG(dbgs() << M);
2828}
2829
2830uint64_t
2831ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2832 constexpr uint64_t kMaxRZ = 1 << 18;
2833 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2834
2835 uint64_t RZ = 0;
2836 if (SizeInBytes <= MinRZ / 2) {
2837 // Reduce redzone size for small size objects, e.g. int, char[1]. MinRZ is
2838 // at least 32 bytes, optimize when SizeInBytes is less than or equal to
2839 // half of MinRZ.
2840 RZ = MinRZ - SizeInBytes;
2841 } else {
2842 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2843 RZ = std::clamp(val: (SizeInBytes / MinRZ / 4) * MinRZ, lo: MinRZ, hi: kMaxRZ);
2844
2845 // Round up to multiple of MinRZ.
2846 if (SizeInBytes % MinRZ)
2847 RZ += MinRZ - (SizeInBytes % MinRZ);
2848 }
2849
2850 assert((RZ + SizeInBytes) % MinRZ == 0);
2851
2852 return RZ;
2853}
2854
2855int ModuleAddressSanitizer::GetAsanVersion() const {
2856 int LongSize = M.getDataLayout().getPointerSizeInBits();
2857 bool isAndroid = M.getTargetTriple().isAndroid();
2858 int Version = 8;
2859 // 32-bit Android is one version ahead because of the switch to dynamic
2860 // shadow.
2861 Version += (LongSize == 32 && isAndroid);
2862 return Version;
2863}
2864
2865GlobalVariable *ModuleAddressSanitizer::getOrCreateModuleName() {
2866 if (!ModuleName) {
2867 // We shouldn't merge same module names, as this string serves as unique
2868 // module ID in runtime.
2869 ModuleName =
2870 createPrivateGlobalForString(M, Str: M.getModuleIdentifier(),
2871 /*AllowMerging*/ false, NamePrefix: genName(suffix: "module"));
2872 }
2873 return ModuleName;
2874}
2875
2876bool ModuleAddressSanitizer::instrumentModule() {
2877 initializeCallbacks();
2878
2879 for (Function &F : M)
2880 removeASanIncompatibleFnAttributes(F, /*ReadsArgMem=*/false);
2881
2882 // Create a module constructor. A destructor is created lazily because not all
2883 // platforms, and not all modules need it.
2884 if (ConstructorKind == AsanCtorKind::Global) {
2885 if (CompileKernel) {
2886 // The kernel always builds with its own runtime, and therefore does not
2887 // need the init and version check calls.
2888 AsanCtorFunction = createSanitizerCtor(M, CtorName: kAsanModuleCtorName);
2889 } else {
2890 std::string AsanVersion = std::to_string(val: GetAsanVersion());
2891 std::string VersionCheckName =
2892 InsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2893 std::tie(args&: AsanCtorFunction, args: std::ignore) =
2894 createSanitizerCtorAndInitFunctions(
2895 M, CtorName: kAsanModuleCtorName, InitName: kAsanInitName, /*InitArgTypes=*/{},
2896 /*InitArgs=*/{}, VersionCheckName);
2897 }
2898 }
2899
2900 bool CtorComdat = true;
2901 if (ClGlobals) {
2902 assert(AsanCtorFunction || ConstructorKind == AsanCtorKind::None);
2903 if (AsanCtorFunction) {
2904 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2905 instrumentGlobals(IRB, CtorComdat: &CtorComdat);
2906 } else {
2907 IRBuilder<> IRB(*C);
2908 instrumentGlobals(IRB, CtorComdat: &CtorComdat);
2909 }
2910 }
2911
2912 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2913
2914 // Put the constructor and destructor in comdat if both
2915 // (1) global instrumentation is not TU-specific
2916 // (2) target is ELF.
2917 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2918 if (AsanCtorFunction) {
2919 AsanCtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleCtorName));
2920 appendToGlobalCtors(M, F: AsanCtorFunction, Priority, Data: AsanCtorFunction);
2921 }
2922 if (AsanDtorFunction) {
2923 AsanDtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleDtorName));
2924 appendToGlobalDtors(M, F: AsanDtorFunction, Priority, Data: AsanDtorFunction);
2925 }
2926 } else {
2927 if (AsanCtorFunction)
2928 appendToGlobalCtors(M, F: AsanCtorFunction, Priority);
2929 if (AsanDtorFunction)
2930 appendToGlobalDtors(M, F: AsanDtorFunction, Priority);
2931 }
2932
2933 return true;
2934}
2935
2936void AddressSanitizer::initializeCallbacks(const TargetLibraryInfo *TLI) {
2937 IRBuilder<> IRB(*C);
2938 // Create __asan_report* callbacks.
2939 // IsWrite, TypeSize and Exp are encoded in the function name.
2940 for (int Exp = 0; Exp < 2; Exp++) {
2941 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2942 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2943 const std::string ExpStr = Exp ? "exp_" : "";
2944 const std::string EndingStr = Recover ? "_noabort" : "";
2945
2946 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2947 SmallVector<Type *, 2> Args1{1, IntptrTy};
2948 AttributeList AL2;
2949 AttributeList AL1;
2950 if (Exp) {
2951 Type *ExpType = Type::getInt32Ty(C&: *C);
2952 Args2.push_back(Elt: ExpType);
2953 Args1.push_back(Elt: ExpType);
2954 if (auto AK = TLI->getExtAttrForI32Param(Signed: false)) {
2955 AL2 = AL2.addParamAttribute(C&: *C, ArgNo: 2, Kind: AK);
2956 AL1 = AL1.addParamAttribute(C&: *C, ArgNo: 1, Kind: AK);
2957 }
2958 }
2959 AsanErrorCallbackSized[AccessIsWrite][Exp] = Inserter.insertFunction(
2960 Name: kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2961 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), Args&: AL2);
2962
2963 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] =
2964 Inserter.insertFunction(
2965 Name: ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2966 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), Args&: AL2);
2967
2968 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2969 AccessSizeIndex++) {
2970 const std::string Suffix = TypeStr + itostr(X: 1ULL << AccessSizeIndex);
2971 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2972 Inserter.insertFunction(
2973 Name: kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2974 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), Args&: AL1);
2975
2976 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2977 Inserter.insertFunction(
2978 Name: ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2979 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), Args&: AL1);
2980 }
2981 }
2982 }
2983
2984 const std::string MemIntrinCallbackPrefix =
2985 (CompileKernel && !ClKasanMemIntrinCallbackPrefix)
2986 ? std::string("")
2987 : ClMemoryAccessCallbackPrefix;
2988 AsanMemmove = Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memmove",
2989 Args&: PtrTy, Args&: PtrTy, Args&: PtrTy, Args&: IntptrTy);
2990 AsanMemcpy = Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memcpy",
2991 Args&: PtrTy, Args&: PtrTy, Args&: PtrTy, Args&: IntptrTy);
2992 AsanMemset =
2993 Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memset",
2994 Args: TLI->getAttrList(C, ArgNos: {1},
2995 /*Signed=*/false),
2996 Args&: PtrTy, Args&: PtrTy, Args: IRB.getInt32Ty(), Args&: IntptrTy);
2997
2998 AsanHandleNoReturnFunc =
2999 Inserter.insertFunction(Name: kAsanHandleNoReturnName, Args: IRB.getVoidTy());
3000
3001 AsanPtrCmpFunction =
3002 Inserter.insertFunction(Name: kAsanPtrCmp, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3003 AsanPtrSubFunction =
3004 Inserter.insertFunction(Name: kAsanPtrSub, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3005 if (Mapping.InGlobal)
3006 AsanShadowGlobal = M.getOrInsertGlobal(Name: "__asan_shadow",
3007 Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: 0));
3008
3009 AMDGPUAddressShared =
3010 Inserter.insertFunction(Name: kAMDGPUAddressSharedName, Args: IRB.getInt1Ty(), Args&: PtrTy);
3011 AMDGPUAddressPrivate = Inserter.insertFunction(Name: kAMDGPUAddressPrivateName,
3012 Args: IRB.getInt1Ty(), Args&: PtrTy);
3013}
3014
3015bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
3016 // For each NSObject descendant having a +load method, this method is invoked
3017 // by the ObjC runtime before any of the static constructors is called.
3018 // Therefore we need to instrument such methods with a call to __asan_init
3019 // at the beginning in order to initialize our runtime before any access to
3020 // the shadow memory.
3021 // We cannot just ignore these methods, because they may call other
3022 // instrumented functions.
3023 if (F.getName().contains(Other: " load]")) {
3024 FunctionCallee AsanInitFunction =
3025 declareSanitizerInitFunction(M&: *F.getParent(), InitName: kAsanInitName, InitArgTypes: {});
3026 IRBuilder<> IRB(&F.front(), F.front().begin());
3027 IRB.CreateCall(Callee: AsanInitFunction, Args: {});
3028 return true;
3029 }
3030 return false;
3031}
3032
3033bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
3034 // Generate code only when dynamic addressing is needed.
3035 if (Mapping.Offset != kDynamicShadowSentinel)
3036 return false;
3037
3038 IRBuilder<> IRB(&F.front().front());
3039 if (Mapping.InGlobal) {
3040 if (ClWithIfuncSuppressRemat) {
3041 // An empty inline asm with input reg == output reg.
3042 // An opaque pointer-to-int cast, basically.
3043 InlineAsm *Asm = InlineAsm::get(
3044 Ty: FunctionType::get(Result: IntptrTy, Params: {AsanShadowGlobal->getType()}, isVarArg: false),
3045 AsmString: StringRef(""), Constraints: StringRef("=r,0"),
3046 /*hasSideEffects=*/false);
3047 LocalDynamicShadow =
3048 IRB.CreateCall(Callee: Asm, Args: {AsanShadowGlobal}, Name: ".asan.shadow");
3049 } else {
3050 LocalDynamicShadow =
3051 IRB.CreatePointerCast(V: AsanShadowGlobal, DestTy: IntptrTy, Name: ".asan.shadow");
3052 }
3053 } else {
3054 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
3055 Name: kAsanShadowMemoryDynamicAddress, Ty: IntptrTy);
3056 LocalDynamicShadow = IRB.CreateLoad(Ty: IntptrTy, Ptr: GlobalDynamicAddress);
3057 }
3058 return true;
3059}
3060
3061void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
3062 // Find the one possible call to llvm.localescape and pre-mark allocas passed
3063 // to it as uninteresting. This assumes we haven't started processing allocas
3064 // yet. This check is done up front because iterating the use list in
3065 // isInterestingAlloca would be algorithmically slower.
3066 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
3067
3068 // Try to get the declaration of llvm.localescape. If it's not in the module,
3069 // we can exit early.
3070 if (!F.getParent()->getFunction(Name: "llvm.localescape")) return;
3071
3072 // Look for a call to llvm.localescape call in the entry block. It can't be in
3073 // any other block.
3074 for (Instruction &I : F.getEntryBlock()) {
3075 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &I);
3076 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
3077 // We found a call. Mark all the allocas passed in as uninteresting.
3078 for (Value *Arg : II->args()) {
3079 AllocaInst *AI = dyn_cast<AllocaInst>(Val: Arg->stripPointerCasts());
3080 assert(AI && AI->isStaticAlloca() &&
3081 "non-static alloca arg to localescape");
3082 ProcessedAllocas[AI] = false;
3083 }
3084 break;
3085 }
3086 }
3087}
3088// Mitigation for https://github.com/google/sanitizers/issues/749
3089// We don't instrument Windows catch-block parameters to avoid
3090// interfering with exception handling assumptions.
3091void AddressSanitizer::markCatchParametersAsUninteresting(Function &F) {
3092 for (BasicBlock &BB : F) {
3093 for (Instruction &I : BB) {
3094 if (auto *CatchPad = dyn_cast<CatchPadInst>(Val: &I)) {
3095 // Mark the parameters to a catch-block as uninteresting to avoid
3096 // instrumenting them.
3097 for (Value *Operand : CatchPad->arg_operands())
3098 if (auto *AI = dyn_cast<AllocaInst>(Val: Operand))
3099 ProcessedAllocas[AI] = false;
3100 }
3101 }
3102 }
3103}
3104
3105bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
3106 bool ShouldInstrument =
3107 ClDebugMin < 0 || ClDebugMax < 0 ||
3108 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
3109 Instrumented++;
3110 return !ShouldInstrument;
3111}
3112
3113bool AddressSanitizer::instrumentFunction(Function &F,
3114 const TargetLibraryInfo *TLI,
3115 const TargetTransformInfo *TTI) {
3116 bool FunctionModified = false;
3117
3118 // Do not apply any instrumentation for naked functions.
3119 if (F.hasFnAttribute(Kind: Attribute::Naked))
3120 return FunctionModified;
3121
3122 // If needed, insert __asan_init before checking for SanitizeAddress attr.
3123 // This function needs to be called even if the function body is not
3124 // instrumented.
3125 if (maybeInsertAsanInitAtFunctionEntry(F))
3126 FunctionModified = true;
3127
3128 // Leave if the function doesn't need instrumentation.
3129 if (!F.hasFnAttribute(Kind: Attribute::SanitizeAddress)) return FunctionModified;
3130
3131 if (F.hasFnAttribute(Kind: Attribute::DisableSanitizerInstrumentation))
3132 return FunctionModified;
3133
3134 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
3135
3136 initializeCallbacks(TLI);
3137
3138 FunctionStateRAII CleanupObj(this);
3139
3140 RuntimeCallInserter RTCI(F);
3141
3142 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
3143
3144 // We can't instrument allocas used with llvm.localescape. Only static allocas
3145 // can be passed to that intrinsic.
3146 markEscapedLocalAllocas(F);
3147
3148 if (TargetTriple.isOSWindows())
3149 markCatchParametersAsUninteresting(F);
3150
3151 // We want to instrument every address only once per basic block (unless there
3152 // are calls between uses).
3153 SmallPtrSet<Value *, 16> TempsToInstrument;
3154 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
3155 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
3156 SmallVector<Instruction *, 8> NoReturnCalls;
3157 SmallVector<BasicBlock *, 16> AllBlocks;
3158 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
3159
3160 // Fill the set of memory operations to instrument.
3161 for (auto &BB : F) {
3162 AllBlocks.push_back(Elt: &BB);
3163 TempsToInstrument.clear();
3164 int NumInsnsPerBB = 0;
3165 for (auto &Inst : BB) {
3166 if (LooksLikeCodeInBug11395(I: &Inst)) return false;
3167 // Skip instructions inserted by another instrumentation.
3168 if (Inst.hasMetadata(KindID: LLVMContext::MD_nosanitize))
3169 continue;
3170 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
3171 getInterestingMemoryOperands(I: &Inst, Interesting&: InterestingOperands, TTI);
3172
3173 if (!InterestingOperands.empty()) {
3174 for (auto &Operand : InterestingOperands) {
3175 if (ClOpt && ClOptSameTemp) {
3176 Value *Ptr = Operand.getPtr();
3177 // If we have a mask, skip instrumentation if we've already
3178 // instrumented the full object. But don't add to TempsToInstrument
3179 // because we might get another load/store with a different mask.
3180 if (Operand.MaybeMask) {
3181 if (TempsToInstrument.count(Ptr))
3182 continue; // We've seen this (whole) temp in the current BB.
3183 } else {
3184 if (!TempsToInstrument.insert(Ptr).second)
3185 continue; // We've seen this temp in the current BB.
3186 }
3187 }
3188 OperandsToInstrument.push_back(Elt: Operand);
3189 NumInsnsPerBB++;
3190 }
3191 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
3192 isInterestingPointerComparison(I: &Inst)) ||
3193 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
3194 isInterestingPointerSubtraction(I: &Inst))) {
3195 PointerComparisonsOrSubtracts.push_back(Elt: &Inst);
3196 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: &Inst)) {
3197 // ok, take it.
3198 IntrinToInstrument.push_back(Elt: MI);
3199 NumInsnsPerBB++;
3200 } else {
3201 if (auto *CB = dyn_cast<CallBase>(Val: &Inst)) {
3202 // A call inside BB.
3203 TempsToInstrument.clear();
3204 if (CB->doesNotReturn())
3205 NoReturnCalls.push_back(Elt: CB);
3206 }
3207 if (CallInst *CI = dyn_cast<CallInst>(Val: &Inst))
3208 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
3209 }
3210 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
3211 }
3212 }
3213
3214 bool UseCalls = (InstrumentationWithCallsThreshold >= 0 &&
3215 OperandsToInstrument.size() + IntrinToInstrument.size() >
3216 (unsigned)InstrumentationWithCallsThreshold);
3217 const DataLayout &DL = F.getDataLayout();
3218 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext());
3219
3220 // Instrument.
3221 int NumInstrumented = 0;
3222 for (auto &Operand : OperandsToInstrument) {
3223 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3224 instrumentMop(ObjSizeVis, O&: Operand, UseCalls,
3225 DL: F.getDataLayout(), RTCI);
3226 FunctionModified = true;
3227 }
3228 for (auto *Inst : IntrinToInstrument) {
3229 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3230 instrumentMemIntrinsic(MI: Inst, RTCI);
3231 FunctionModified = true;
3232 }
3233
3234 FunctionStackPoisoner FSP(F, *this, RTCI);
3235 bool ChangedStack = FSP.runOnFunction();
3236
3237 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
3238 // See e.g. https://github.com/google/sanitizers/issues/37
3239 for (auto *CI : NoReturnCalls) {
3240 IRBuilder<> IRB(CI);
3241 RTCI.createRuntimeCall(IRB, Callee: AsanHandleNoReturnFunc, Args: {});
3242 }
3243
3244 for (auto *Inst : PointerComparisonsOrSubtracts) {
3245 FunctionModified |= instrumentPointerComparisonOrSubtraction(I: Inst, RTCI);
3246 }
3247
3248 if (ChangedStack || !NoReturnCalls.empty())
3249 FunctionModified = true;
3250
3251 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
3252 << F << "\n");
3253
3254 return FunctionModified;
3255}
3256
3257// Workaround for bug 11395: we don't want to instrument stack in functions
3258// with large assembly blobs (32-bit only), otherwise reg alloc may crash.
3259// FIXME: remove once the bug 11395 is fixed.
3260bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
3261 if (LongSize != 32) return false;
3262 CallInst *CI = dyn_cast<CallInst>(Val: I);
3263 if (!CI || !CI->isInlineAsm()) return false;
3264 if (CI->arg_size() <= 5)
3265 return false;
3266 // We have inline assembly with quite a few arguments.
3267 return true;
3268}
3269
3270void FunctionStackPoisoner::initializeCallbacks(Module &) {
3271 IRBuilder<> IRB(*C);
3272 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always ||
3273 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3274 const char *MallocNameTemplate =
3275 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always
3276 ? kAsanStackMallocAlwaysNameTemplate
3277 : kAsanStackMallocNameTemplate;
3278 for (int Index = 0; Index <= kMaxAsanStackMallocSizeClass; Index++) {
3279 std::string Suffix = itostr(X: Index);
3280 AsanStackMallocFunc[Index] = ASan.Inserter.insertFunction(
3281 Name: MallocNameTemplate + Suffix, Args&: IntptrTy, Args&: IntptrTy);
3282 AsanStackFreeFunc[Index] =
3283 ASan.Inserter.insertFunction(Name: kAsanStackFreeNameTemplate + Suffix,
3284 Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3285 }
3286 }
3287 if (ASan.UseAfterScope) {
3288 AsanPoisonStackMemoryFunc = ASan.Inserter.insertFunction(
3289 Name: kAsanPoisonStackMemoryName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3290 AsanUnpoisonStackMemoryFunc = ASan.Inserter.insertFunction(
3291 Name: kAsanUnpoisonStackMemoryName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3292 }
3293
3294 for (size_t Val : {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0xf1, 0xf2,
3295 0xf3, 0xf5, 0xf8}) {
3296 std::ostringstream Name;
3297 Name << kAsanSetShadowPrefix;
3298 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
3299 AsanSetShadowFunc[Val] = ASan.Inserter.insertFunction(
3300 Name: Name.str(), Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3301 }
3302
3303 AsanAllocaPoisonFunc = ASan.Inserter.insertFunction(
3304 Name: kAsanAllocaPoison, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3305 AsanAllocasUnpoisonFunc = ASan.Inserter.insertFunction(
3306 Name: kAsanAllocasUnpoison, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3307}
3308
3309void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
3310 ArrayRef<uint8_t> ShadowBytes,
3311 size_t Begin, size_t End,
3312 IRBuilder<> &IRB,
3313 Value *ShadowBase) {
3314 if (Begin >= End)
3315 return;
3316
3317 const size_t LargestStoreSizeInBytes =
3318 std::min<size_t>(a: sizeof(uint64_t), b: ASan.LongSize / 8);
3319
3320 const bool IsLittleEndian = F.getDataLayout().isLittleEndian();
3321
3322 // Poison given range in shadow using larges store size with out leading and
3323 // trailing zeros in ShadowMask. Zeros never change, so they need neither
3324 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
3325 // middle of a store.
3326 for (size_t i = Begin; i < End;) {
3327 if (!ShadowMask[i]) {
3328 assert(!ShadowBytes[i]);
3329 ++i;
3330 continue;
3331 }
3332
3333 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
3334 // Fit store size into the range.
3335 while (StoreSizeInBytes > End - i)
3336 StoreSizeInBytes /= 2;
3337
3338 // Minimize store size by trimming trailing zeros.
3339 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
3340 while (j <= StoreSizeInBytes / 2)
3341 StoreSizeInBytes /= 2;
3342 }
3343
3344 uint64_t Val = 0;
3345 for (size_t j = 0; j < StoreSizeInBytes; j++) {
3346 if (IsLittleEndian)
3347 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
3348 else
3349 Val = (Val << 8) | ShadowBytes[i + j];
3350 }
3351
3352 Value *Ptr = IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i));
3353 Value *Poison = IRB.getIntN(N: StoreSizeInBytes * 8, C: Val);
3354 IRB.CreateAlignedStore(
3355 Val: Poison, Ptr: IRB.CreateIntToPtr(V: Ptr, DestTy: PointerType::getUnqual(C&: Poison->getContext())),
3356 Align: Align(1));
3357
3358 i += StoreSizeInBytes;
3359 }
3360}
3361
3362void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3363 ArrayRef<uint8_t> ShadowBytes,
3364 IRBuilder<> &IRB, Value *ShadowBase) {
3365 copyToShadow(ShadowMask, ShadowBytes, Begin: 0, End: ShadowMask.size(), IRB, ShadowBase);
3366}
3367
3368void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3369 ArrayRef<uint8_t> ShadowBytes,
3370 size_t Begin, size_t End,
3371 IRBuilder<> &IRB, Value *ShadowBase) {
3372 assert(ShadowMask.size() == ShadowBytes.size());
3373 size_t Done = Begin;
3374 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
3375 if (!ShadowMask[i]) {
3376 assert(!ShadowBytes[i]);
3377 continue;
3378 }
3379 uint8_t Val = ShadowBytes[i];
3380 if (!AsanSetShadowFunc[Val])
3381 continue;
3382
3383 // Skip same values.
3384 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
3385 }
3386
3387 if (j - i >= ASan.MaxInlinePoisoningSize) {
3388 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End: i, IRB, ShadowBase);
3389 RTCI.createRuntimeCall(
3390 IRB, Callee: AsanSetShadowFunc[Val],
3391 Args: {IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i)),
3392 ConstantInt::get(Ty: IntptrTy, V: j - i)});
3393 Done = j;
3394 }
3395 }
3396
3397 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End, IRB, ShadowBase);
3398}
3399
3400// Fake stack allocator (asan_fake_stack.h) has 11 size classes
3401// for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
3402static int StackMallocSizeClass(uint64_t LocalStackSize) {
3403 assert(LocalStackSize <= kMaxStackMallocSize);
3404 uint64_t MaxSize = kMinStackMallocSize;
3405 for (int i = 0;; i++, MaxSize *= 2)
3406 if (LocalStackSize <= MaxSize) return i;
3407 llvm_unreachable("impossible LocalStackSize");
3408}
3409
3410void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
3411 Instruction *CopyInsertPoint = &F.front().front();
3412 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
3413 // Insert after the dynamic shadow location is determined
3414 CopyInsertPoint = CopyInsertPoint->getNextNode();
3415 assert(CopyInsertPoint);
3416 }
3417 IRBuilder<> IRB(CopyInsertPoint);
3418 const DataLayout &DL = F.getDataLayout();
3419 for (Argument &Arg : F.args()) {
3420 if (Arg.hasByValAttr()) {
3421 Type *Ty = Arg.getParamByValType();
3422 const Align Alignment =
3423 DL.getValueOrABITypeAlignment(Alignment: Arg.getParamAlign(), Ty);
3424
3425 AllocaInst *AI = IRB.CreateAlloca(
3426 Ty, ArraySize: nullptr,
3427 Name: (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
3428 ".byval");
3429 AI->setAlignment(Alignment);
3430 Arg.replaceAllUsesWith(V: AI);
3431
3432 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3433 IRB.CreateMemCpy(Dst: AI, DstAlign: Alignment, Src: &Arg, SrcAlign: Alignment, Size: AllocSize);
3434 }
3435 }
3436}
3437
3438PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
3439 Value *ValueIfTrue,
3440 Instruction *ThenTerm,
3441 Value *ValueIfFalse) {
3442 PHINode *PHI = IRB.CreatePHI(Ty: ValueIfTrue->getType(), NumReservedValues: 2);
3443 BasicBlock *CondBlock = cast<Instruction>(Val: Cond)->getParent();
3444 PHI->addIncoming(V: ValueIfFalse, BB: CondBlock);
3445 BasicBlock *ThenBlock = ThenTerm->getParent();
3446 PHI->addIncoming(V: ValueIfTrue, BB: ThenBlock);
3447 return PHI;
3448}
3449
3450Value *FunctionStackPoisoner::createAllocaForLayout(
3451 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
3452 AllocaInst *Alloca;
3453 if (Dynamic) {
3454 Alloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(),
3455 ArraySize: ConstantInt::get(Ty: IRB.getInt64Ty(), V: L.FrameSize),
3456 Name: "MyAlloca");
3457 } else {
3458 Alloca = IRB.CreateAlloca(Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: L.FrameSize),
3459 ArraySize: nullptr, Name: "MyAlloca");
3460 assert(Alloca->isStaticAlloca());
3461 }
3462 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
3463 uint64_t FrameAlignment = std::max(a: L.FrameAlignment, b: uint64_t(ClRealignStack));
3464 Alloca->setAlignment(Align(FrameAlignment));
3465 return Alloca;
3466}
3467
3468void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
3469 BasicBlock &FirstBB = *F.begin();
3470 IRBuilder<> IRB(dyn_cast<Instruction>(Val: FirstBB.begin()));
3471 DynamicAllocaLayout = IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr);
3472 IRB.CreateStore(Val: Constant::getNullValue(Ty: IntptrTy), Ptr: DynamicAllocaLayout);
3473 DynamicAllocaLayout->setAlignment(Align(32));
3474}
3475
3476void FunctionStackPoisoner::processDynamicAllocas() {
3477 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
3478 assert(DynamicAllocaPoisonCallVec.empty());
3479 return;
3480 }
3481
3482 // Insert poison calls for lifetime intrinsics for dynamic allocas.
3483 for (const auto &APC : DynamicAllocaPoisonCallVec) {
3484 assert(APC.InsBefore);
3485 assert(APC.AI);
3486 assert(ASan.isInterestingAlloca(*APC.AI));
3487 assert(!APC.AI->isStaticAlloca());
3488
3489 IRBuilder<> IRB(APC.InsBefore);
3490 poisonAlloca(V: APC.AI, Size: APC.Size, IRB, DoPoison: APC.DoPoison);
3491 // Dynamic allocas will be unpoisoned unconditionally below in
3492 // unpoisonDynamicAllocas.
3493 // Flag that we need unpoison static allocas.
3494 }
3495
3496 // Handle dynamic allocas.
3497 createDynamicAllocasInitStorage();
3498 for (auto &AI : DynamicAllocaVec)
3499 handleDynamicAllocaCall(AI);
3500 unpoisonDynamicAllocas();
3501}
3502
3503/// Collect instructions in the entry block after \p InsBefore which initialize
3504/// permanent storage for a function argument. These instructions must remain in
3505/// the entry block so that uninitialized values do not appear in backtraces. An
3506/// added benefit is that this conserves spill slots. This does not move stores
3507/// before instrumented / "interesting" allocas.
3508static void findStoresToUninstrumentedArgAllocas(
3509 AddressSanitizer &ASan, Instruction &InsBefore,
3510 SmallVectorImpl<Instruction *> &InitInsts) {
3511 Instruction *Start = InsBefore.getNextNode();
3512 for (Instruction *It = Start; It; It = It->getNextNode()) {
3513 // Argument initialization looks like:
3514 // 1) store <Argument>, <Alloca> OR
3515 // 2) <CastArgument> = cast <Argument> to ...
3516 // store <CastArgument> to <Alloca>
3517 // Do not consider any other kind of instruction.
3518 //
3519 // Note: This covers all known cases, but may not be exhaustive. An
3520 // alternative to pattern-matching stores is to DFS over all Argument uses:
3521 // this might be more general, but is probably much more complicated.
3522 if (isa<AllocaInst>(Val: It) || isa<CastInst>(Val: It))
3523 continue;
3524 if (auto *Store = dyn_cast<StoreInst>(Val: It)) {
3525 // The store destination must be an alloca that isn't interesting for
3526 // ASan to instrument. These are moved up before InsBefore, and they're
3527 // not interesting because allocas for arguments can be mem2reg'd.
3528 auto *Alloca = dyn_cast<AllocaInst>(Val: Store->getPointerOperand());
3529 if (!Alloca || ASan.isInterestingAlloca(AI: *Alloca))
3530 continue;
3531
3532 Value *Val = Store->getValueOperand();
3533 bool IsDirectArgInit = isa<Argument>(Val);
3534 bool IsArgInitViaCast =
3535 isa<CastInst>(Val) &&
3536 isa<Argument>(Val: cast<CastInst>(Val)->getOperand(i_nocapture: 0)) &&
3537 // Check that the cast appears directly before the store. Otherwise
3538 // moving the cast before InsBefore may break the IR.
3539 Val == It->getPrevNode();
3540 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3541 if (!IsArgInit)
3542 continue;
3543
3544 if (IsArgInitViaCast)
3545 InitInsts.push_back(Elt: cast<Instruction>(Val));
3546 InitInsts.push_back(Elt: Store);
3547 continue;
3548 }
3549
3550 // Do not reorder past unknown instructions: argument initialization should
3551 // only involve casts and stores.
3552 return;
3553 }
3554}
3555
3556static StringRef getAllocaName(AllocaInst *AI) {
3557 // Alloca could have been renamed for uniqueness. Its true name will have been
3558 // recorded as an annotation.
3559 if (AI->hasMetadata(KindID: LLVMContext::MD_annotation)) {
3560 MDTuple *AllocaAnnotations =
3561 cast<MDTuple>(Val: AI->getMetadata(KindID: LLVMContext::MD_annotation));
3562 for (auto &Annotation : AllocaAnnotations->operands()) {
3563 if (!isa<MDTuple>(Val: Annotation))
3564 continue;
3565 auto AnnotationTuple = cast<MDTuple>(Val: Annotation);
3566 for (unsigned Index = 0; Index < AnnotationTuple->getNumOperands();
3567 Index++) {
3568 // All annotations are strings
3569 auto MetadataString =
3570 cast<MDString>(Val: AnnotationTuple->getOperand(I: Index));
3571 if (MetadataString->getString() == "alloca_name_altered")
3572 return cast<MDString>(Val: AnnotationTuple->getOperand(I: Index + 1))
3573 ->getString();
3574 }
3575 }
3576 }
3577 return AI->getName();
3578}
3579
3580void FunctionStackPoisoner::processStaticAllocas() {
3581 if (AllocaVec.empty()) {
3582 assert(StaticAllocaPoisonCallVec.empty());
3583 return;
3584 }
3585
3586 int StackMallocIdx = -1;
3587 DebugLoc EntryDebugLocation;
3588 if (auto SP = F.getSubprogram())
3589 EntryDebugLocation =
3590 DILocation::get(Context&: SP->getContext(), Line: SP->getScopeLine(), Column: 0, Scope: SP);
3591
3592 Instruction *InsBefore = AllocaVec[0];
3593 IRBuilder<> IRB(InsBefore);
3594
3595 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3596 // debug info is broken, because only entry-block allocas are treated as
3597 // regular stack slots.
3598 auto InsBeforeB = InsBefore->getParent();
3599 assert(InsBeforeB == &F.getEntryBlock());
3600 for (auto *AI : StaticAllocasToMoveUp)
3601 if (AI->getParent() == InsBeforeB)
3602 AI->moveBefore(InsertPos: InsBefore->getIterator());
3603
3604 // Move stores of arguments into entry-block allocas as well. This prevents
3605 // extra stack slots from being generated (to house the argument values until
3606 // they can be stored into the allocas). This also prevents uninitialized
3607 // values from being shown in backtraces.
3608 SmallVector<Instruction *, 8> ArgInitInsts;
3609 findStoresToUninstrumentedArgAllocas(ASan, InsBefore&: *InsBefore, InitInsts&: ArgInitInsts);
3610 for (Instruction *ArgInitInst : ArgInitInsts)
3611 ArgInitInst->moveBefore(InsertPos: InsBefore->getIterator());
3612
3613 // If we have a call to llvm.localescape, keep it in the entry block.
3614 if (LocalEscapeCall)
3615 LocalEscapeCall->moveBefore(InsertPos: InsBefore->getIterator());
3616
3617 SmallVector<ASanStackVariableDescription, 16> SVD;
3618 SVD.reserve(N: AllocaVec.size());
3619 for (AllocaInst *AI : AllocaVec) {
3620 StringRef Name = getAllocaName(AI);
3621 ASanStackVariableDescription D = {.Name: Name.data(),
3622 .Size: ASan.getAllocaSizeInBytes(AI: *AI),
3623 .LifetimeSize: 0,
3624 .Alignment: AI->getAlign().value(),
3625 .AI: AI,
3626 .Offset: 0,
3627 .Line: 0};
3628 SVD.push_back(Elt: D);
3629 }
3630
3631 // Minimal header size (left redzone) is 4 pointers,
3632 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3633 uint64_t Granularity = 1ULL << Mapping.Scale;
3634 uint64_t MinHeaderSize = std::max(a: (uint64_t)ASan.LongSize / 2, b: Granularity);
3635 const ASanStackFrameLayout &L =
3636 ComputeASanStackFrameLayout(Vars&: SVD, Granularity, MinHeaderSize);
3637
3638 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3639 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3640 for (auto &Desc : SVD)
3641 AllocaToSVDMap[Desc.AI] = &Desc;
3642
3643 // Update SVD with information from lifetime intrinsics.
3644 for (const auto &APC : StaticAllocaPoisonCallVec) {
3645 assert(APC.InsBefore);
3646 assert(APC.AI);
3647 assert(ASan.isInterestingAlloca(*APC.AI));
3648 assert(APC.AI->isStaticAlloca());
3649
3650 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3651 Desc.LifetimeSize = Desc.Size;
3652 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3653 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3654 if (LifetimeLoc->getFile() == FnLoc->getFile())
3655 if (unsigned Line = LifetimeLoc->getLine())
3656 Desc.Line = std::min(a: Desc.Line ? Desc.Line : Line, b: Line);
3657 }
3658 }
3659 }
3660
3661 auto DescriptionString = ComputeASanStackFrameDescription(Vars: SVD);
3662 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3663 uint64_t LocalStackSize = L.FrameSize;
3664 bool DoStackMalloc =
3665 ASan.UseAfterReturn != AsanDetectStackUseAfterReturnMode::Never &&
3666 !ASan.CompileKernel && LocalStackSize <= kMaxStackMallocSize;
3667 bool DoDynamicAlloca = ClDynamicAllocaStack;
3668 // Don't do dynamic alloca or stack malloc if:
3669 // 1) There is inline asm: too often it makes assumptions on which registers
3670 // are available.
3671 // 2) There is a returns_twice call (typically setjmp), which is
3672 // optimization-hostile, and doesn't play well with introduced indirect
3673 // register-relative calculation of local variable addresses.
3674 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3675 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3676
3677 Type *PtrTy = F.getDataLayout().getAllocaPtrType(Ctx&: F.getContext());
3678 Value *StaticAlloca =
3679 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, Dynamic: false);
3680
3681 Value *FakeStackPtr;
3682 Value *FakeStackInt;
3683 Value *LocalStackBase;
3684 Value *LocalStackBaseAlloca;
3685 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3686
3687 if (DoStackMalloc) {
3688 LocalStackBaseAlloca =
3689 IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr, Name: "asan_local_stack_base");
3690 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3691 // void *FakeStack = __asan_option_detect_stack_use_after_return
3692 // ? __asan_stack_malloc_N(LocalStackSize)
3693 // : nullptr;
3694 // void *LocalStackBase = (FakeStack) ? FakeStack :
3695 // alloca(LocalStackSize);
3696 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3697 Name: kAsanOptionDetectUseAfterReturn, Ty: IRB.getInt32Ty());
3698 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3699 LHS: IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: OptionDetectUseAfterReturn),
3700 RHS: Constant::getNullValue(Ty: IRB.getInt32Ty()));
3701 Instruction *Term =
3702 SplitBlockAndInsertIfThen(Cond: UseAfterReturnIsEnabled, SplitBefore: InsBefore, Unreachable: false);
3703 IRBuilder<> IRBIf(Term);
3704 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3705 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3706 Value *FakeStackValue =
3707 RTCI.createRuntimeCall(IRB&: IRBIf, Callee: AsanStackMallocFunc[StackMallocIdx],
3708 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3709 IRB.SetInsertPoint(InsBefore);
3710 FakeStackInt = createPHI(IRB, Cond: UseAfterReturnIsEnabled, ValueIfTrue: FakeStackValue,
3711 ThenTerm: Term, ValueIfFalse: ConstantInt::get(Ty: IntptrTy, V: 0));
3712 } else {
3713 // assert(ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode:Always)
3714 // void *FakeStack = __asan_stack_malloc_N(LocalStackSize);
3715 // void *LocalStackBase = (FakeStack) ? FakeStack :
3716 // alloca(LocalStackSize);
3717 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3718 FakeStackInt =
3719 RTCI.createRuntimeCall(IRB, Callee: AsanStackMallocFunc[StackMallocIdx],
3720 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3721 }
3722 FakeStackPtr = IRB.CreateIntToPtr(V: FakeStackInt, DestTy: PtrTy);
3723 Value *NoFakeStack =
3724 IRB.CreateICmpEQ(LHS: FakeStackInt, RHS: Constant::getNullValue(Ty: IntptrTy));
3725 Instruction *Term =
3726 SplitBlockAndInsertIfThen(Cond: NoFakeStack, SplitBefore: InsBefore, Unreachable: false);
3727 IRBuilder<> IRBIf(Term);
3728 Value *AllocaValue =
3729 DoDynamicAlloca ? createAllocaForLayout(IRB&: IRBIf, L, Dynamic: true) : StaticAlloca;
3730
3731 IRB.SetInsertPoint(InsBefore);
3732 LocalStackBase =
3733 createPHI(IRB, Cond: NoFakeStack, ValueIfTrue: AllocaValue, ThenTerm: Term, ValueIfFalse: FakeStackPtr);
3734 IRB.CreateStore(Val: LocalStackBase, Ptr: LocalStackBaseAlloca);
3735 DIExprFlags |= DIExpression::DerefBefore;
3736 } else {
3737 // void *FakeStack = nullptr;
3738 // void *LocalStackBase = alloca(LocalStackSize);
3739 FakeStackInt = Constant::getNullValue(Ty: IntptrTy);
3740 FakeStackPtr = Constant::getNullValue(Ty: PtrTy);
3741 LocalStackBase =
3742 DoDynamicAlloca ? createAllocaForLayout(IRB, L, Dynamic: true) : StaticAlloca;
3743 LocalStackBaseAlloca = LocalStackBase;
3744 }
3745
3746 // Replace Alloca instructions with base+offset.
3747 SmallVector<Value *> NewAllocaPtrs;
3748 for (const auto &Desc : SVD) {
3749 AllocaInst *AI = Desc.AI;
3750 replaceDbgDeclare(Address: AI, NewAddress: LocalStackBaseAlloca, Builder&: DIB, DIExprFlags, Offset: Desc.Offset);
3751 Value *NewAllocaPtr = IRB.CreatePtrAdd(
3752 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: Desc.Offset));
3753 if (NewAllocaPtr->getType() != AI->getType())
3754 NewAllocaPtr = IRB.CreateAddrSpaceCast(V: NewAllocaPtr, DestTy: AI->getType());
3755 AI->replaceAllUsesWith(V: NewAllocaPtr);
3756 NewAllocaPtrs.push_back(Elt: NewAllocaPtr);
3757 }
3758
3759 // The left-most redzone has enough space for at least 4 pointers.
3760 // Write the Magic value to redzone[0].
3761 IRB.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kCurrentStackFrameMagic),
3762 Ptr: LocalStackBase);
3763 // Write the frame description constant to redzone[1].
3764 Value *BasePlus1 = IRB.CreatePtrAdd(
3765 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: ASan.LongSize / 8));
3766 GlobalVariable *StackDescriptionGlobal =
3767 createPrivateGlobalForString(M&: *F.getParent(), Str: DescriptionString,
3768 /*AllowMerging*/ true, NamePrefix: genName(suffix: "stack"));
3769 Value *Description = IRB.CreatePointerCast(V: StackDescriptionGlobal, DestTy: IntptrTy);
3770 IRB.CreateStore(Val: Description, Ptr: BasePlus1);
3771 // Write the PC to redzone[2].
3772 Value *BasePlus2 = IRB.CreatePtrAdd(
3773 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: 2 * ASan.LongSize / 8));
3774 IRB.CreateStore(Val: IRB.CreatePointerCast(V: &F, DestTy: IntptrTy), Ptr: BasePlus2);
3775
3776 const auto &ShadowAfterScope = GetShadowBytesAfterScope(Vars: SVD, Layout: L);
3777
3778 // Poison the stack red zones at the entry.
3779 Value *ShadowBase =
3780 ASan.memToShadow(Shadow: IRB.CreatePtrToInt(V: LocalStackBase, DestTy: IntptrTy), IRB);
3781 // As mask we must use most poisoned case: red zones and after scope.
3782 // As bytes we can use either the same or just red zones only.
3783 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowAfterScope, IRB, ShadowBase);
3784
3785 if (!StaticAllocaPoisonCallVec.empty()) {
3786 const auto &ShadowInScope = GetShadowBytes(Vars: SVD, Layout: L);
3787
3788 // Poison static allocas near lifetime intrinsics.
3789 for (const auto &APC : StaticAllocaPoisonCallVec) {
3790 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3791 assert(Desc.Offset % L.Granularity == 0);
3792 size_t Begin = Desc.Offset / L.Granularity;
3793 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3794
3795 IRBuilder<> IRB(APC.InsBefore);
3796 copyToShadow(ShadowMask: ShadowAfterScope,
3797 ShadowBytes: APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3798 IRB, ShadowBase);
3799 }
3800 }
3801
3802 // Remove lifetime markers now that these are no longer allocas.
3803 for (Value *NewAllocaPtr : NewAllocaPtrs) {
3804 for (User *U : make_early_inc_range(Range: NewAllocaPtr->users())) {
3805 auto *I = cast<Instruction>(Val: U);
3806 if (I->isLifetimeStartOrEnd())
3807 I->eraseFromParent();
3808 }
3809 }
3810
3811 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3812 SmallVector<uint8_t, 64> ShadowAfterReturn;
3813
3814 // (Un)poison the stack before all ret instructions.
3815 for (Instruction *Ret : RetVec) {
3816 IRBuilder<> IRBRet(Ret);
3817 // Mark the current frame as retired.
3818 IRBRet.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kRetiredStackFrameMagic),
3819 Ptr: LocalStackBase);
3820 if (DoStackMalloc) {
3821 assert(StackMallocIdx >= 0);
3822 // if FakeStack != 0 // LocalStackBase == FakeStack
3823 // // In use-after-return mode, poison the whole stack frame.
3824 // if StackMallocIdx <= 4
3825 // // For small sizes inline the whole thing:
3826 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3827 // **SavedFlagPtr(FakeStack) = 0
3828 // else
3829 // __asan_stack_free_N(FakeStack, LocalStackSize)
3830 // else
3831 // <This is not a fake stack; unpoison the redzones>
3832 Value *Cmp =
3833 IRBRet.CreateICmpNE(LHS: FakeStackInt, RHS: Constant::getNullValue(Ty: IntptrTy));
3834 Instruction *ThenTerm, *ElseTerm;
3835 SplitBlockAndInsertIfThenElse(Cond: Cmp, SplitBefore: Ret, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm);
3836
3837 IRBuilder<> IRBPoison(ThenTerm);
3838 if (ASan.MaxInlinePoisoningSize != 0 && StackMallocIdx <= 4) {
3839 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3840 ShadowAfterReturn.resize(N: ClassSize / L.Granularity,
3841 NV: kAsanStackUseAfterReturnMagic);
3842 copyToShadow(ShadowMask: ShadowAfterReturn, ShadowBytes: ShadowAfterReturn, IRB&: IRBPoison,
3843 ShadowBase);
3844 Value *SavedFlagPtrPtr = IRBPoison.CreatePtrAdd(
3845 Ptr: FakeStackPtr,
3846 Offset: ConstantInt::get(Ty: IntptrTy, V: ClassSize - ASan.LongSize / 8));
3847 Value *SavedFlagPtr = IRBPoison.CreateLoad(Ty: IntptrTy, Ptr: SavedFlagPtrPtr);
3848 IRBPoison.CreateStore(
3849 Val: Constant::getNullValue(Ty: IRBPoison.getInt8Ty()),
3850 Ptr: IRBPoison.CreateIntToPtr(V: SavedFlagPtr, DestTy: IRBPoison.getPtrTy()));
3851 } else {
3852 // For larger frames call __asan_stack_free_*.
3853 RTCI.createRuntimeCall(
3854 IRB&: IRBPoison, Callee: AsanStackFreeFunc[StackMallocIdx],
3855 Args: {FakeStackInt, ConstantInt::get(Ty: IntptrTy, V: LocalStackSize)});
3856 }
3857
3858 IRBuilder<> IRBElse(ElseTerm);
3859 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBElse, ShadowBase);
3860 } else {
3861 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBRet, ShadowBase);
3862 }
3863 }
3864
3865 // We are done. Remove the old unused alloca instructions.
3866 for (auto *AI : AllocaVec)
3867 AI->eraseFromParent();
3868}
3869
3870void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3871 IRBuilder<> &IRB, bool DoPoison) {
3872 // For now just insert the call to ASan runtime.
3873 Value *AddrArg = IRB.CreatePointerCast(V, DestTy: IntptrTy);
3874 Value *SizeArg = ConstantInt::get(Ty: IntptrTy, V: Size);
3875 RTCI.createRuntimeCall(
3876 IRB, Callee: DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3877 Args: {AddrArg, SizeArg});
3878}
3879
3880// Handling llvm.lifetime intrinsics for a given %alloca:
3881// (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3882// (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3883// invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3884// could be poisoned by previous llvm.lifetime.end instruction, as the
3885// variable may go in and out of scope several times, e.g. in loops).
3886// (3) if we poisoned at least one %alloca in a function,
3887// unpoison the whole stack frame at function exit.
3888void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3889 IRBuilder<> IRB(AI);
3890
3891 const Align Alignment = std::max(a: Align(kAllocaRzSize), b: AI->getAlign());
3892 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3893
3894 Value *Zero = Constant::getNullValue(Ty: IntptrTy);
3895 Value *AllocaRzSize = ConstantInt::get(Ty: IntptrTy, V: kAllocaRzSize);
3896 Value *AllocaRzMask = ConstantInt::get(Ty: IntptrTy, V: AllocaRedzoneMask);
3897
3898 // Since we need to extend alloca with additional memory to locate
3899 // redzones, and OldSize is number of allocated blocks with
3900 // ElementSize size, get allocated memory size in bytes by
3901 // OldSize * ElementSize.
3902 Value *OldSize = IRB.CreateAllocationSize(DestTy: IntptrTy, AI);
3903
3904 // PartialSize = OldSize % 32
3905 Value *PartialSize = IRB.CreateAnd(LHS: OldSize, RHS: AllocaRzMask);
3906
3907 // Misalign = kAllocaRzSize - PartialSize;
3908 Value *Misalign = IRB.CreateSub(LHS: AllocaRzSize, RHS: PartialSize);
3909
3910 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3911 Value *Cond = IRB.CreateICmpNE(LHS: Misalign, RHS: AllocaRzSize);
3912 Value *PartialPadding = IRB.CreateSelect(C: Cond, True: Misalign, False: Zero);
3913
3914 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3915 // Alignment is added to locate left redzone, PartialPadding for possible
3916 // partial redzone and kAllocaRzSize for right redzone respectively.
3917 Value *AdditionalChunkSize = IRB.CreateAdd(
3918 LHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value() + kAllocaRzSize),
3919 RHS: PartialPadding);
3920
3921 Value *NewSize = IRB.CreateAdd(LHS: OldSize, RHS: AdditionalChunkSize);
3922
3923 // Insert new alloca with new NewSize and Alignment params.
3924 AllocaInst *NewAlloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(), ArraySize: NewSize);
3925 NewAlloca->setAlignment(Alignment);
3926
3927 // NewAddress = Address + Alignment
3928 Value *NewAddress =
3929 IRB.CreateAdd(LHS: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy),
3930 RHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value()));
3931
3932 // Insert __asan_alloca_poison call for new created alloca.
3933 RTCI.createRuntimeCall(IRB, Callee: AsanAllocaPoisonFunc, Args: {NewAddress, OldSize});
3934
3935 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3936 // for unpoisoning stuff.
3937 IRB.CreateStore(Val: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy), Ptr: DynamicAllocaLayout);
3938
3939 Value *NewAddressPtr = IRB.CreateIntToPtr(V: NewAddress, DestTy: AI->getType());
3940
3941 // Remove lifetime markers now that this is no longer an alloca.
3942 for (User *U : make_early_inc_range(Range: AI->users())) {
3943 auto *I = cast<Instruction>(Val: U);
3944 if (I->isLifetimeStartOrEnd())
3945 I->eraseFromParent();
3946 }
3947
3948 // Replace all uses of AddressReturnedByAlloca with NewAddressPtr.
3949 AI->replaceAllUsesWith(V: NewAddressPtr);
3950
3951 // We are done. Erase old alloca from parent.
3952 AI->eraseFromParent();
3953}
3954
3955// isSafeAccess returns true if Addr is always inbounds with respect to its
3956// base object. For example, it is a field access or an array access with
3957// constant inbounds index.
3958bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3959 Value *Addr, TypeSize TypeStoreSize) const {
3960 if (TypeStoreSize.isScalable())
3961 // TODO: We can use vscale_range to convert a scalable value to an
3962 // upper bound on the access size.
3963 return false;
3964
3965 SizeOffsetAPInt SizeOffset = ObjSizeVis.compute(V: Addr);
3966 if (!SizeOffset.bothKnown())
3967 return false;
3968
3969 uint64_t Size = SizeOffset.Size.getZExtValue();
3970 int64_t Offset = SizeOffset.Offset.getSExtValue();
3971
3972 // Three checks are required to ensure safety:
3973 // . Offset >= 0 (since the offset is given from the base ptr)
3974 // . Size >= Offset (unsigned)
3975 // . Size - Offset >= NeededSize (unsigned)
3976 return Offset >= 0 && Size >= uint64_t(Offset) &&
3977 Size - uint64_t(Offset) >= TypeStoreSize / 8;
3978}
3979