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 void 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 const Type *AllocaType = AI.getAllocatedType();
1232 const auto *STy = dyn_cast<StructType>(Val: AllocaType);
1233 if (!ASan.isInterestingAlloca(AI) || isa<ScalableVectorType>(Val: AllocaType) ||
1234 (STy && STy->containsHomogeneousScalableVectorTypes())) {
1235 if (AI.isStaticAlloca()) {
1236 // Skip over allocas that are present *before* the first instrumented
1237 // alloca, we don't want to move those around.
1238 if (AllocaVec.empty())
1239 return;
1240
1241 StaticAllocasToMoveUp.push_back(Elt: &AI);
1242 }
1243 return;
1244 }
1245
1246 if (!AI.isStaticAlloca())
1247 DynamicAllocaVec.push_back(Elt: &AI);
1248 else
1249 AllocaVec.push_back(Elt: &AI);
1250 }
1251
1252 /// Collect lifetime intrinsic calls to check for use-after-scope
1253 /// errors.
1254 void visitIntrinsicInst(IntrinsicInst &II) {
1255 Intrinsic::ID ID = II.getIntrinsicID();
1256 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(Elt: &II);
1257 if (ID == Intrinsic::localescape) LocalEscapeCall = &II;
1258 if (!ASan.UseAfterScope)
1259 return;
1260 if (!II.isLifetimeStartOrEnd())
1261 return;
1262 // Find alloca instruction that corresponds to llvm.lifetime argument.
1263 AllocaInst *AI = dyn_cast<AllocaInst>(Val: II.getArgOperand(i: 0));
1264 // We're interested only in allocas we can handle.
1265 if (!AI || !ASan.isInterestingAlloca(AI: *AI))
1266 return;
1267
1268 std::optional<TypeSize> Size = AI->getAllocationSize(DL: AI->getDataLayout());
1269 // Check that size is known and can be stored in IntptrTy.
1270 // TODO: Add support for scalable vectors if possible.
1271 if (!Size || Size->isScalable() ||
1272 !ConstantInt::isValueValidForType(Ty: IntptrTy, V: *Size))
1273 return;
1274
1275 bool DoPoison = (ID == Intrinsic::lifetime_end);
1276 AllocaPoisonCall APC = {.InsBefore: &II, .AI: AI, .Size: *Size, .DoPoison: DoPoison};
1277 if (AI->isStaticAlloca())
1278 StaticAllocaPoisonCallVec.push_back(Elt: APC);
1279 else if (ClInstrumentDynamicAllocas)
1280 DynamicAllocaPoisonCallVec.push_back(Elt: APC);
1281 }
1282
1283 void visitCallBase(CallBase &CB) {
1284 if (CallInst *CI = dyn_cast<CallInst>(Val: &CB)) {
1285 HasInlineAsm |= CI->isInlineAsm() && &CB != ASan.LocalDynamicShadow;
1286 HasReturnsTwiceCall |= CI->canReturnTwice();
1287 }
1288 }
1289
1290 // ---------------------- Helpers.
1291 void initializeCallbacks(Module &M);
1292
1293 // Copies bytes from ShadowBytes into shadow memory for indexes where
1294 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that
1295 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten.
1296 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1297 IRBuilder<> &IRB, Value *ShadowBase);
1298 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1299 size_t Begin, size_t End, IRBuilder<> &IRB,
1300 Value *ShadowBase);
1301 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
1302 ArrayRef<uint8_t> ShadowBytes, size_t Begin,
1303 size_t End, IRBuilder<> &IRB, Value *ShadowBase);
1304
1305 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
1306
1307 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L,
1308 bool Dynamic);
1309 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue,
1310 Instruction *ThenTerm, Value *ValueIfFalse);
1311};
1312
1313} // end anonymous namespace
1314
1315void AddressSanitizerPass::printPipeline(
1316 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1317 static_cast<PassInfoMixin<AddressSanitizerPass> *>(this)->printPipeline(
1318 OS, MapClassName2PassName);
1319 OS << '<';
1320 if (Options.CompileKernel)
1321 OS << "kernel;";
1322 if (Options.UseAfterScope)
1323 OS << "use-after-scope";
1324 OS << '>';
1325}
1326
1327AddressSanitizerPass::AddressSanitizerPass(
1328 const AddressSanitizerOptions &Options, bool UseGlobalGC,
1329 bool UseOdrIndicator, AsanDtorKind DestructorKind,
1330 AsanCtorKind ConstructorKind)
1331 : Options(Options), UseGlobalGC(UseGlobalGC),
1332 UseOdrIndicator(UseOdrIndicator), DestructorKind(DestructorKind),
1333 ConstructorKind(ConstructorKind) {}
1334
1335PreservedAnalyses AddressSanitizerPass::run(Module &M,
1336 ModuleAnalysisManager &MAM) {
1337 // Return early if nosanitize_address module flag is present for the module.
1338 // This implies that asan pass has already run before.
1339 if (checkIfAlreadyInstrumented(M, Flag: "nosanitize_address"))
1340 return PreservedAnalyses::all();
1341
1342 ModuleAddressSanitizer ModuleSanitizer(
1343 M, Options.InsertVersionCheck, Options.CompileKernel, Options.Recover,
1344 UseGlobalGC, UseOdrIndicator, DestructorKind, ConstructorKind);
1345 bool Modified = false;
1346 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
1347 const StackSafetyGlobalInfo *const SSGI =
1348 ClUseStackSafety ? &MAM.getResult<StackSafetyGlobalAnalysis>(IR&: M) : nullptr;
1349 for (Function &F : M) {
1350 if (F.empty())
1351 continue;
1352 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage)
1353 continue;
1354 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName())
1355 continue;
1356 if (F.getName().starts_with(Prefix: "__asan_"))
1357 continue;
1358 if (F.isPresplitCoroutine())
1359 continue;
1360 AddressSanitizer FunctionSanitizer(
1361 M, SSGI, Options.InstrumentationWithCallsThreshold,
1362 Options.MaxInlinePoisoningSize, Options.CompileKernel, Options.Recover,
1363 Options.UseAfterScope, Options.UseAfterReturn);
1364 const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: F);
1365 const TargetTransformInfo &TTI = FAM.getResult<TargetIRAnalysis>(IR&: F);
1366 Modified |= FunctionSanitizer.instrumentFunction(F, TLI: &TLI, TTI: &TTI);
1367 }
1368 Modified |= ModuleSanitizer.instrumentModule();
1369 if (!Modified)
1370 return PreservedAnalyses::all();
1371
1372 PreservedAnalyses PA = PreservedAnalyses::none();
1373 // GlobalsAA is considered stateless and does not get invalidated unless
1374 // explicitly invalidated; PreservedAnalyses::none() is not enough. Sanitizers
1375 // make changes that require GlobalsAA to be invalidated.
1376 PA.abandon<GlobalsAA>();
1377 return PA;
1378}
1379
1380static size_t TypeStoreSizeToSizeIndex(uint32_t TypeSize) {
1381 size_t Res = llvm::countr_zero(Val: TypeSize / 8);
1382 assert(Res < kNumberOfAccessSizes);
1383 return Res;
1384}
1385
1386/// Check if \p G has been created by a trusted compiler pass.
1387static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) {
1388 // Do not instrument @llvm.global_ctors, @llvm.used, etc.
1389 if (G->getName().starts_with(Prefix: "llvm.") ||
1390 // Do not instrument gcov counter arrays.
1391 G->getName().starts_with(Prefix: "__llvm_gcov_ctr") ||
1392 // Do not instrument rtti proxy symbols for function sanitizer.
1393 G->getName().starts_with(Prefix: "__llvm_rtti_proxy"))
1394 return true;
1395
1396 // Do not instrument asan globals.
1397 if (G->getName().starts_with(Prefix: kAsanGenPrefix) ||
1398 G->getName().starts_with(Prefix: kSanCovGenPrefix) ||
1399 G->getName().starts_with(Prefix: kODRGenPrefix))
1400 return true;
1401
1402 return false;
1403}
1404
1405static bool isUnsupportedAMDGPUAddrspace(Value *Addr) {
1406 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1407 unsigned int AddrSpace = PtrTy->getPointerAddressSpace();
1408 // Globals in address space 1 and 4 are supported for AMDGPU.
1409 if (AddrSpace == 3 || AddrSpace == 5)
1410 return true;
1411 return false;
1412}
1413
1414static bool isSupportedAddrspace(const Triple &TargetTriple, Value *Addr) {
1415 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1416 unsigned int AddrSpace = PtrTy->getPointerAddressSpace();
1417
1418 if (!SrcAddrSpaces.empty())
1419 return SrcAddrSpaces.count(V: AddrSpace);
1420
1421 if (TargetTriple.isAMDGPU())
1422 return !isUnsupportedAMDGPUAddrspace(Addr);
1423
1424 return AddrSpace == 0;
1425}
1426
1427Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
1428 // Shadow >> scale
1429 Shadow = IRB.CreateLShr(LHS: Shadow, RHS: Mapping.Scale);
1430 if (Mapping.Offset == 0) return Shadow;
1431 // (Shadow >> scale) | offset
1432 Value *ShadowBase;
1433 if (LocalDynamicShadow)
1434 ShadowBase = LocalDynamicShadow;
1435 else
1436 ShadowBase = ConstantInt::get(Ty: IntptrTy, V: Mapping.Offset);
1437 if (Mapping.OrShadowOffset)
1438 return IRB.CreateOr(LHS: Shadow, RHS: ShadowBase);
1439 else
1440 return IRB.CreateAdd(LHS: Shadow, RHS: ShadowBase);
1441}
1442
1443// Instrument memset/memmove/memcpy
1444void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI,
1445 RuntimeCallInserter &RTCI) {
1446 InstrumentationIRBuilder IRB(MI);
1447 if (isa<MemTransferInst>(Val: MI)) {
1448 RTCI.createRuntimeCall(
1449 IRB, Callee: isa<MemMoveInst>(Val: MI) ? AsanMemmove : AsanMemcpy,
1450 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1451 IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 1), DestTy: PtrTy),
1452 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1453 } else if (isa<MemSetInst>(Val: MI)) {
1454 RTCI.createRuntimeCall(
1455 IRB, Callee: AsanMemset,
1456 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1457 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 1), DestTy: IRB.getInt32Ty(), isSigned: false),
1458 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1459 }
1460 MI->eraseFromParent();
1461}
1462
1463/// Check if we want (and can) handle this alloca.
1464bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1465 auto [It, Inserted] = ProcessedAllocas.try_emplace(Key: &AI);
1466
1467 if (!Inserted)
1468 return It->getSecond();
1469
1470 bool IsInteresting =
1471 (AI.getAllocatedType()->isSized() &&
1472 // alloca() may be called with 0 size, ignore it.
1473 ((!AI.isStaticAlloca()) || !getAllocaSizeInBytes(AI).isZero()) &&
1474 // We are only interested in allocas not promotable to registers.
1475 // Promotable allocas are common under -O0.
1476 (!ClSkipPromotableAllocas || !isAllocaPromotable(AI: &AI)) &&
1477 // inalloca allocas are not treated as static, and we don't want
1478 // dynamic alloca instrumentation for them as well.
1479 !AI.isUsedWithInAlloca() &&
1480 // swifterror allocas are register promoted by ISel
1481 !AI.isSwiftError() &&
1482 // safe allocas are not interesting
1483 !(SSGI && SSGI->isSafe(AI)));
1484
1485 It->second = IsInteresting;
1486 return IsInteresting;
1487}
1488
1489bool AddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) {
1490 // Check whether the target supports sanitizing the address space
1491 // of the pointer.
1492 if (!isSupportedAddrspace(TargetTriple, Addr: Ptr))
1493 return true;
1494
1495 // Ignore swifterror addresses.
1496 // swifterror memory addresses are mem2reg promoted by instruction
1497 // selection. As such they cannot have regular uses like an instrumentation
1498 // function and it makes no sense to track them as memory.
1499 if (Ptr->isSwiftError())
1500 return true;
1501
1502 // Treat memory accesses to promotable allocas as non-interesting since they
1503 // will not cause memory violations. This greatly speeds up the instrumented
1504 // executable at -O0.
1505 if (auto AI = dyn_cast_or_null<AllocaInst>(Val: Ptr))
1506 if (ClSkipPromotableAllocas && !isInterestingAlloca(AI: *AI))
1507 return true;
1508
1509 if (SSGI != nullptr && SSGI->stackAccessIsSafe(I: *Inst) &&
1510 findAllocaForValue(V: Ptr))
1511 return true;
1512
1513 return false;
1514}
1515
1516void AddressSanitizer::getInterestingMemoryOperands(
1517 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting,
1518 const TargetTransformInfo *TTI) {
1519 // Do not instrument the load fetching the dynamic shadow address.
1520 if (LocalDynamicShadow == I)
1521 return;
1522
1523 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
1524 if (!ClInstrumentReads || ignoreAccess(Inst: I, Ptr: LI->getPointerOperand()))
1525 return;
1526 Interesting.emplace_back(Args&: I, Args: LI->getPointerOperandIndex(), Args: false,
1527 Args: LI->getType(), Args: LI->getAlign());
1528 } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
1529 if (!ClInstrumentWrites || ignoreAccess(Inst: I, Ptr: SI->getPointerOperand()))
1530 return;
1531 Interesting.emplace_back(Args&: I, Args: SI->getPointerOperandIndex(), Args: true,
1532 Args: SI->getValueOperand()->getType(), Args: SI->getAlign());
1533 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
1534 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: RMW->getPointerOperand()))
1535 return;
1536 Interesting.emplace_back(Args&: I, Args: RMW->getPointerOperandIndex(), Args: true,
1537 Args: RMW->getValOperand()->getType(), Args: std::nullopt);
1538 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
1539 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: XCHG->getPointerOperand()))
1540 return;
1541 Interesting.emplace_back(Args&: I, Args: XCHG->getPointerOperandIndex(), Args: true,
1542 Args: XCHG->getCompareOperand()->getType(),
1543 Args: std::nullopt);
1544 } else if (auto CI = dyn_cast<CallInst>(Val: I)) {
1545 switch (CI->getIntrinsicID()) {
1546 case Intrinsic::masked_load:
1547 case Intrinsic::masked_store:
1548 case Intrinsic::masked_gather:
1549 case Intrinsic::masked_scatter: {
1550 bool IsWrite = CI->getType()->isVoidTy();
1551 // Masked store has an initial operand for the value.
1552 unsigned OpOffset = IsWrite ? 1 : 0;
1553 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1554 return;
1555
1556 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1557 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1558 return;
1559 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1560 MaybeAlign Alignment = CI->getParamAlign(ArgNo: 0);
1561 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1562 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: Mask);
1563 break;
1564 }
1565 case Intrinsic::masked_expandload:
1566 case Intrinsic::masked_compressstore: {
1567 bool IsWrite = CI->getIntrinsicID() == Intrinsic::masked_compressstore;
1568 unsigned OpOffset = IsWrite ? 1 : 0;
1569 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1570 return;
1571 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1572 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1573 return;
1574 MaybeAlign Alignment = BasePtr->getPointerAlignment(DL: *DL);
1575 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1576
1577 IRBuilder IB(I);
1578 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1579 // Use the popcount of Mask as the effective vector length.
1580 Type *ExtTy = VectorType::get(ElementType: IntptrTy, Other: cast<VectorType>(Val: Ty));
1581 Value *ExtMask = IB.CreateZExt(V: Mask, DestTy: ExtTy);
1582 Value *EVL = IB.CreateAddReduce(Src: ExtMask);
1583 Value *TrueMask = ConstantInt::get(Ty: Mask->getType(), V: 1);
1584 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: TrueMask,
1585 Args&: EVL);
1586 break;
1587 }
1588 case Intrinsic::vp_load:
1589 case Intrinsic::vp_store:
1590 case Intrinsic::experimental_vp_strided_load:
1591 case Intrinsic::experimental_vp_strided_store: {
1592 auto *VPI = cast<VPIntrinsic>(Val: CI);
1593 unsigned IID = CI->getIntrinsicID();
1594 bool IsWrite = CI->getType()->isVoidTy();
1595 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1596 return;
1597 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1598 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1599 MaybeAlign Alignment = VPI->getOperand(i_nocapture: PtrOpNo)->getPointerAlignment(DL: *DL);
1600 Value *Stride = nullptr;
1601 if (IID == Intrinsic::experimental_vp_strided_store ||
1602 IID == Intrinsic::experimental_vp_strided_load) {
1603 Stride = VPI->getOperand(i_nocapture: PtrOpNo + 1);
1604 // Use the pointer alignment as the element alignment if the stride is a
1605 // multiple of the pointer alignment. Otherwise, the element alignment
1606 // should be Align(1).
1607 unsigned PointerAlign = Alignment.valueOrOne().value();
1608 if (!isa<ConstantInt>(Val: Stride) ||
1609 cast<ConstantInt>(Val: Stride)->getZExtValue() % PointerAlign != 0)
1610 Alignment = Align(1);
1611 }
1612 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1613 Args: VPI->getMaskParam(), Args: VPI->getVectorLengthParam(),
1614 Args&: Stride);
1615 break;
1616 }
1617 case Intrinsic::vp_gather:
1618 case Intrinsic::vp_scatter: {
1619 auto *VPI = cast<VPIntrinsic>(Val: CI);
1620 unsigned IID = CI->getIntrinsicID();
1621 bool IsWrite = IID == Intrinsic::vp_scatter;
1622 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1623 return;
1624 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1625 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1626 MaybeAlign Alignment = VPI->getPointerAlignment();
1627 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1628 Args: VPI->getMaskParam(),
1629 Args: VPI->getVectorLengthParam());
1630 break;
1631 }
1632 default:
1633 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
1634 MemIntrinsicInfo IntrInfo;
1635 if (TTI->getTgtMemIntrinsic(Inst: II, Info&: IntrInfo))
1636 Interesting = IntrInfo.InterestingOperands;
1637 return;
1638 }
1639 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) {
1640 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1641 ignoreAccess(Inst: I, Ptr: CI->getArgOperand(i: ArgNo)))
1642 continue;
1643 Type *Ty = CI->getParamByValType(ArgNo);
1644 Interesting.emplace_back(Args&: I, Args&: ArgNo, Args: false, Args&: Ty, Args: Align(1));
1645 }
1646 }
1647 }
1648}
1649
1650static bool isPointerOperand(Value *V) {
1651 return V->getType()->isPointerTy() || isa<PtrToIntInst>(Val: V);
1652}
1653
1654// This is a rough heuristic; it may cause both false positives and
1655// false negatives. The proper implementation requires cooperation with
1656// the frontend.
1657static bool isInterestingPointerComparison(Instruction *I) {
1658 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: I)) {
1659 if (!Cmp->isRelational())
1660 return false;
1661 } else {
1662 return false;
1663 }
1664 return isPointerOperand(V: I->getOperand(i: 0)) &&
1665 isPointerOperand(V: I->getOperand(i: 1));
1666}
1667
1668// This is a rough heuristic; it may cause both false positives and
1669// false negatives. The proper implementation requires cooperation with
1670// the frontend.
1671static bool isInterestingPointerSubtraction(Instruction *I) {
1672 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) {
1673 if (BO->getOpcode() != Instruction::Sub)
1674 return false;
1675 } else {
1676 return false;
1677 }
1678 return isPointerOperand(V: I->getOperand(i: 0)) &&
1679 isPointerOperand(V: I->getOperand(i: 1));
1680}
1681
1682bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1683 // If a global variable does not have dynamic initialization we don't
1684 // have to instrument it. However, if a global does not have initializer
1685 // at all, we assume it has dynamic initializer (in other TU).
1686 if (!G->hasInitializer())
1687 return false;
1688
1689 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().IsDynInit)
1690 return false;
1691
1692 return true;
1693}
1694
1695void AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1696 Instruction *I, RuntimeCallInserter &RTCI) {
1697 IRBuilder<> IRB(I);
1698 FunctionCallee F = isa<ICmpInst>(Val: I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1699 Value *Param[2] = {I->getOperand(i: 0), I->getOperand(i: 1)};
1700 for (Value *&i : Param) {
1701 if (i->getType()->isPointerTy())
1702 i = IRB.CreatePointerCast(V: i, DestTy: IntptrTy);
1703 }
1704 RTCI.createRuntimeCall(IRB, Callee: F, Args: Param);
1705}
1706
1707static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1708 Instruction *InsertBefore, Value *Addr,
1709 MaybeAlign Alignment, unsigned Granularity,
1710 TypeSize TypeStoreSize, bool IsWrite,
1711 Value *SizeArgument, bool UseCalls,
1712 uint32_t Exp, RuntimeCallInserter &RTCI) {
1713 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1714 // if the data is properly aligned.
1715 if (!TypeStoreSize.isScalable()) {
1716 const auto FixedSize = TypeStoreSize.getFixedValue();
1717 switch (FixedSize) {
1718 case 8:
1719 case 16:
1720 case 32:
1721 case 64:
1722 case 128:
1723 if (!Alignment || *Alignment >= Granularity ||
1724 *Alignment >= FixedSize / 8)
1725 return Pass->instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment,
1726 TypeStoreSize: FixedSize, IsWrite, SizeArgument: nullptr, UseCalls,
1727 Exp, RTCI);
1728 }
1729 }
1730 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeStoreSize,
1731 IsWrite, SizeArgument: nullptr, UseCalls, Exp, RTCI);
1732}
1733
1734void AddressSanitizer::instrumentMaskedLoadOrStore(
1735 AddressSanitizer *Pass, const DataLayout &DL, Type *IntptrTy, Value *Mask,
1736 Value *EVL, Value *Stride, Instruction *I, Value *Addr,
1737 MaybeAlign Alignment, unsigned Granularity, Type *OpType, bool IsWrite,
1738 Value *SizeArgument, bool UseCalls, uint32_t Exp,
1739 RuntimeCallInserter &RTCI) {
1740 auto *VTy = cast<VectorType>(Val: OpType);
1741 TypeSize ElemTypeSize = DL.getTypeStoreSizeInBits(Ty: VTy->getScalarType());
1742 auto Zero = ConstantInt::get(Ty: IntptrTy, V: 0);
1743
1744 IRBuilder IB(I);
1745 Instruction *LoopInsertBefore = I;
1746 if (EVL) {
1747 // The end argument of SplitBlockAndInsertForLane is assumed bigger
1748 // than zero, so we should check whether EVL is zero here.
1749 Type *EVLType = EVL->getType();
1750 Value *IsEVLZero = IB.CreateICmpNE(LHS: EVL, RHS: ConstantInt::get(Ty: EVLType, V: 0));
1751 LoopInsertBefore = SplitBlockAndInsertIfThen(Cond: IsEVLZero, SplitBefore: I, Unreachable: false);
1752 IB.SetInsertPoint(LoopInsertBefore);
1753 // Cast EVL to IntptrTy.
1754 EVL = IB.CreateZExtOrTrunc(V: EVL, DestTy: IntptrTy);
1755 // To avoid undefined behavior for extracting with out of range index, use
1756 // the minimum of evl and element count as trip count.
1757 Value *EC = IB.CreateElementCount(Ty: IntptrTy, EC: VTy->getElementCount());
1758 EVL = IB.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: EVL, RHS: EC);
1759 } else {
1760 EVL = IB.CreateElementCount(Ty: IntptrTy, EC: VTy->getElementCount());
1761 }
1762
1763 // Cast Stride to IntptrTy.
1764 if (Stride)
1765 Stride = IB.CreateZExtOrTrunc(V: Stride, DestTy: IntptrTy);
1766
1767 SplitBlockAndInsertForEachLane(End: EVL, InsertBefore: LoopInsertBefore->getIterator(),
1768 Func: [&](IRBuilderBase &IRB, Value *Index) {
1769 Value *MaskElem = IRB.CreateExtractElement(Vec: Mask, Idx: Index);
1770 if (auto *MaskElemC = dyn_cast<ConstantInt>(Val: MaskElem)) {
1771 if (MaskElemC->isZero())
1772 // No check
1773 return;
1774 // Unconditional check
1775 } else {
1776 // Conditional check
1777 Instruction *ThenTerm = SplitBlockAndInsertIfThen(
1778 Cond: MaskElem, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false);
1779 IRB.SetInsertPoint(ThenTerm);
1780 }
1781
1782 Value *InstrumentedAddress;
1783 if (isa<VectorType>(Val: Addr->getType())) {
1784 assert(
1785 cast<VectorType>(Addr->getType())->getElementType()->isPointerTy() &&
1786 "Expected vector of pointer.");
1787 InstrumentedAddress = IRB.CreateExtractElement(Vec: Addr, Idx: Index);
1788 } else if (Stride) {
1789 Index = IRB.CreateMul(LHS: Index, RHS: Stride);
1790 InstrumentedAddress = IRB.CreatePtrAdd(Ptr: Addr, Offset: Index);
1791 } else {
1792 InstrumentedAddress = IRB.CreateGEP(Ty: VTy, Ptr: Addr, IdxList: {Zero, Index});
1793 }
1794 doInstrumentAddress(Pass, I, InsertBefore: &*IRB.GetInsertPoint(), Addr: InstrumentedAddress,
1795 Alignment, Granularity, TypeStoreSize: ElemTypeSize, IsWrite,
1796 SizeArgument, UseCalls, Exp, RTCI);
1797 });
1798}
1799
1800void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1801 InterestingMemoryOperand &O, bool UseCalls,
1802 const DataLayout &DL,
1803 RuntimeCallInserter &RTCI) {
1804 Value *Addr = O.getPtr();
1805
1806 // Optimization experiments.
1807 // The experiments can be used to evaluate potential optimizations that remove
1808 // instrumentation (assess false negatives). Instead of completely removing
1809 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1810 // experiments that want to remove instrumentation of this instruction).
1811 // If Exp is non-zero, this pass will emit special calls into runtime
1812 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1813 // make runtime terminate the program in a special way (with a different
1814 // exit status). Then you run the new compiler on a buggy corpus, collect
1815 // the special terminations (ideally, you don't see them at all -- no false
1816 // negatives) and make the decision on the optimization.
1817 uint32_t Exp = ClForceExperiment;
1818
1819 if (ClOpt && ClOptGlobals) {
1820 // If initialization order checking is disabled, a simple access to a
1821 // dynamically initialized global is always valid.
1822 GlobalVariable *G = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: Addr));
1823 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1824 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1825 NumOptimizedAccessesToGlobalVar++;
1826 return;
1827 }
1828 }
1829
1830 if (ClOpt && ClOptStack) {
1831 // A direct inbounds access to a stack variable is always valid.
1832 if (isa<AllocaInst>(Val: getUnderlyingObject(V: Addr)) &&
1833 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1834 NumOptimizedAccessesToStackVar++;
1835 return;
1836 }
1837 }
1838
1839 if (O.IsWrite)
1840 NumInstrumentedWrites++;
1841 else
1842 NumInstrumentedReads++;
1843
1844 if (O.MaybeByteOffset) {
1845 Type *Ty = Type::getInt8Ty(C&: *C);
1846 IRBuilder IB(O.getInsn());
1847
1848 Value *OffsetOp = O.MaybeByteOffset;
1849 if (TargetTriple.isRISCV()) {
1850 Type *OffsetTy = OffsetOp->getType();
1851 // RVV indexed loads/stores zero-extend offset operands which are narrower
1852 // than XLEN to XLEN.
1853 if (OffsetTy->getScalarType()->getIntegerBitWidth() <
1854 static_cast<unsigned>(LongSize)) {
1855 VectorType *OrigType = cast<VectorType>(Val: OffsetTy);
1856 Type *ExtendTy = VectorType::get(ElementType: IntptrTy, Other: OrigType);
1857 OffsetOp = IB.CreateZExt(V: OffsetOp, DestTy: ExtendTy);
1858 }
1859 }
1860 Addr = IB.CreateGEP(Ty, Ptr: Addr, IdxList: {OffsetOp});
1861 }
1862
1863 unsigned Granularity = 1 << Mapping.Scale;
1864 if (O.MaybeMask) {
1865 instrumentMaskedLoadOrStore(Pass: this, DL, IntptrTy, Mask: O.MaybeMask, EVL: O.MaybeEVL,
1866 Stride: O.MaybeStride, I: O.getInsn(), Addr, Alignment: O.Alignment,
1867 Granularity, OpType: O.OpType, IsWrite: O.IsWrite, SizeArgument: nullptr,
1868 UseCalls, Exp, RTCI);
1869 } else {
1870 doInstrumentAddress(Pass: this, I: O.getInsn(), InsertBefore: O.getInsn(), Addr, Alignment: O.Alignment,
1871 Granularity, TypeStoreSize: O.TypeStoreSize, IsWrite: O.IsWrite, SizeArgument: nullptr,
1872 UseCalls, Exp, RTCI);
1873 }
1874}
1875
1876Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1877 Value *Addr, bool IsWrite,
1878 size_t AccessSizeIndex,
1879 Value *SizeArgument,
1880 uint32_t Exp,
1881 RuntimeCallInserter &RTCI) {
1882 InstrumentationIRBuilder IRB(InsertBefore);
1883 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp);
1884 CallInst *Call = nullptr;
1885 if (SizeArgument) {
1886 if (Exp == 0)
1887 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][0],
1888 Args: {Addr, SizeArgument});
1889 else
1890 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][1],
1891 Args: {Addr, SizeArgument, ExpVal});
1892 } else {
1893 if (Exp == 0)
1894 Call = RTCI.createRuntimeCall(
1895 IRB, Callee: AsanErrorCallback[IsWrite][0][AccessSizeIndex], Args: Addr);
1896 else
1897 Call = RTCI.createRuntimeCall(
1898 IRB, Callee: AsanErrorCallback[IsWrite][1][AccessSizeIndex], Args: {Addr, ExpVal});
1899 }
1900
1901 Call->setCannotMerge();
1902 return Call;
1903}
1904
1905Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1906 Value *ShadowValue,
1907 uint32_t TypeStoreSize) {
1908 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1909 // Addr & (Granularity - 1)
1910 Value *LastAccessedByte =
1911 IRB.CreateAnd(LHS: AddrLong, RHS: ConstantInt::get(Ty: IntptrTy, V: Granularity - 1));
1912 // (Addr & (Granularity - 1)) + size - 1
1913 if (TypeStoreSize / 8 > 1)
1914 LastAccessedByte = IRB.CreateAdd(
1915 LHS: LastAccessedByte, RHS: ConstantInt::get(Ty: IntptrTy, V: TypeStoreSize / 8 - 1));
1916 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1917 LastAccessedByte =
1918 IRB.CreateIntCast(V: LastAccessedByte, DestTy: ShadowValue->getType(), isSigned: false);
1919 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1920 return IRB.CreateICmpSGE(LHS: LastAccessedByte, RHS: ShadowValue);
1921}
1922
1923Instruction *AddressSanitizer::instrumentAMDGPUAddress(
1924 Instruction *OrigIns, Instruction *InsertBefore, Value *Addr,
1925 uint32_t TypeStoreSize, bool IsWrite, Value *SizeArgument) {
1926 // Do not instrument unsupported addrspaces.
1927 if (isUnsupportedAMDGPUAddrspace(Addr))
1928 return nullptr;
1929 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1930 // Follow host instrumentation for global and constant addresses.
1931 if (PtrTy->getPointerAddressSpace() != 0)
1932 return InsertBefore;
1933 // Instrument generic addresses in supported addressspaces.
1934 IRBuilder<> IRB(InsertBefore);
1935 Value *IsShared = IRB.CreateCall(Callee: AMDGPUAddressShared, Args: {Addr});
1936 Value *IsPrivate = IRB.CreateCall(Callee: AMDGPUAddressPrivate, Args: {Addr});
1937 Value *IsSharedOrPrivate = IRB.CreateOr(LHS: IsShared, RHS: IsPrivate);
1938 Value *Cmp = IRB.CreateNot(V: IsSharedOrPrivate);
1939 Value *AddrSpaceZeroLanding =
1940 SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false);
1941 InsertBefore = cast<Instruction>(Val: AddrSpaceZeroLanding);
1942 return InsertBefore;
1943}
1944
1945Instruction *AddressSanitizer::genAMDGPUReportBlock(IRBuilder<> &IRB,
1946 Value *Cond, bool Recover) {
1947 Value *ReportCond = Cond;
1948 if (!Recover) {
1949 auto Ballot = Inserter.insertFunction(Name: kAMDGPUBallotName, Args: IRB.getInt64Ty(),
1950 Args: IRB.getInt1Ty());
1951 ReportCond = IRB.CreateIsNotNull(Arg: IRB.CreateCall(Callee: Ballot, Args: {Cond}));
1952 }
1953
1954 auto *Trm =
1955 SplitBlockAndInsertIfThen(Cond: ReportCond, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false,
1956 BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
1957 Trm->getParent()->setName("asan.report");
1958
1959 if (Recover)
1960 return Trm;
1961
1962 Trm = SplitBlockAndInsertIfThen(Cond, SplitBefore: Trm, Unreachable: false);
1963 IRB.SetInsertPoint(Trm);
1964 return IRB.CreateCall(
1965 Callee: Inserter.insertFunction(Name: kAMDGPUUnreachableName, Args: IRB.getVoidTy()), Args: {});
1966}
1967
1968void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1969 Instruction *InsertBefore, Value *Addr,
1970 MaybeAlign Alignment,
1971 uint32_t TypeStoreSize, bool IsWrite,
1972 Value *SizeArgument, bool UseCalls,
1973 uint32_t Exp,
1974 RuntimeCallInserter &RTCI) {
1975 if (TargetTriple.isAMDGPU()) {
1976 InsertBefore = instrumentAMDGPUAddress(OrigIns, InsertBefore, Addr,
1977 TypeStoreSize, IsWrite, SizeArgument);
1978 if (!InsertBefore)
1979 return;
1980 }
1981
1982 InstrumentationIRBuilder IRB(InsertBefore);
1983 size_t AccessSizeIndex = TypeStoreSizeToSizeIndex(TypeSize: TypeStoreSize);
1984
1985 if (UseCalls && ClOptimizeCallbacks) {
1986 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
1987 IRB.CreateIntrinsic(ID: Intrinsic::asan_check_memaccess, OverloadTypes: {},
1988 Args: {IRB.CreatePointerCast(V: Addr, DestTy: PtrTy),
1989 ConstantInt::get(Ty: Int32Ty, V: AccessInfo.Packed)});
1990 return;
1991 }
1992
1993 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
1994 if (UseCalls) {
1995 if (Exp == 0)
1996 RTCI.createRuntimeCall(
1997 IRB, Callee: AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex], Args: AddrLong);
1998 else
1999 RTCI.createRuntimeCall(
2000 IRB, Callee: AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
2001 Args: {AddrLong, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
2002 return;
2003 }
2004
2005 Type *ShadowTy =
2006 IntegerType::get(C&: *C, NumBits: std::max(a: 8U, b: TypeStoreSize >> Mapping.Scale));
2007 Type *ShadowPtrTy = PointerType::get(C&: *C, AddressSpace: ClShadowAddrSpace);
2008 Value *ShadowPtr = memToShadow(Shadow: AddrLong, IRB);
2009 const uint64_t ShadowAlign =
2010 std::max<uint64_t>(a: Alignment.valueOrOne().value() >> Mapping.Scale, b: 1);
2011 Value *ShadowValue = IRB.CreateAlignedLoad(
2012 Ty: ShadowTy, Ptr: IRB.CreateIntToPtr(V: ShadowPtr, DestTy: ShadowPtrTy), Align: Align(ShadowAlign));
2013
2014 Value *Cmp = IRB.CreateIsNotNull(Arg: ShadowValue);
2015 size_t Granularity = 1ULL << Mapping.Scale;
2016 Instruction *CrashTerm = nullptr;
2017
2018 bool GenSlowPath = (ClAlwaysSlowPath || (TypeStoreSize < 8 * Granularity));
2019
2020 if (TargetTriple.isAMDGCN()) {
2021 if (GenSlowPath) {
2022 auto *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
2023 Cmp = IRB.CreateAnd(LHS: Cmp, RHS: Cmp2);
2024 }
2025 CrashTerm = genAMDGPUReportBlock(IRB, Cond: Cmp, Recover);
2026 } else if (GenSlowPath) {
2027 // We use branch weights for the slow path check, to indicate that the slow
2028 // path is rarely taken. This seems to be the case for SPEC benchmarks.
2029 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
2030 Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false, BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
2031 BasicBlock *NextBB = cast<UncondBrInst>(Val: CheckTerm)->getSuccessor();
2032 IRB.SetInsertPoint(CheckTerm);
2033 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
2034 if (Recover) {
2035 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp2, SplitBefore: CheckTerm, Unreachable: false);
2036 } else {
2037 BasicBlock *CrashBlock =
2038 BasicBlock::Create(Context&: *C, Name: "", Parent: NextBB->getParent(), InsertBefore: NextBB);
2039 CrashTerm = new UnreachableInst(*C, CrashBlock);
2040 CondBrInst *NewTerm = CondBrInst::Create(Cond: Cmp2, IfTrue: CrashBlock, IfFalse: NextBB);
2041 ReplaceInstWithInst(From: CheckTerm, To: NewTerm);
2042 }
2043 } else {
2044 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: !Recover);
2045 }
2046
2047 Instruction *Crash = generateCrashCode(
2048 InsertBefore: CrashTerm, Addr: AddrLong, IsWrite, AccessSizeIndex, SizeArgument, Exp, RTCI);
2049 if (OrigIns->getDebugLoc())
2050 Crash->setDebugLoc(OrigIns->getDebugLoc());
2051}
2052
2053// Instrument unusual size or unusual alignment.
2054// We can not do it with a single check, so we do 1-byte check for the first
2055// and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
2056// to report the actual access size.
2057void AddressSanitizer::instrumentUnusualSizeOrAlignment(
2058 Instruction *I, Instruction *InsertBefore, Value *Addr,
2059 TypeSize TypeStoreSize, bool IsWrite, Value *SizeArgument, bool UseCalls,
2060 uint32_t Exp, RuntimeCallInserter &RTCI) {
2061 InstrumentationIRBuilder IRB(InsertBefore);
2062 Value *NumBits = IRB.CreateTypeSize(Ty: IntptrTy, Size: TypeStoreSize);
2063 Value *Size = IRB.CreateLShr(LHS: NumBits, RHS: ConstantInt::get(Ty: IntptrTy, V: 3));
2064
2065 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
2066 if (UseCalls) {
2067 if (Exp == 0)
2068 RTCI.createRuntimeCall(IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][0],
2069 Args: {AddrLong, Size});
2070 else
2071 RTCI.createRuntimeCall(
2072 IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][1],
2073 Args: {AddrLong, Size, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
2074 } else {
2075 Value *SizeMinusOne = IRB.CreateSub(LHS: Size, RHS: ConstantInt::get(Ty: IntptrTy, V: 1));
2076 Value *LastByte = IRB.CreateIntToPtr(
2077 V: IRB.CreateAdd(LHS: AddrLong, RHS: SizeMinusOne),
2078 DestTy: Addr->getType());
2079 instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false, Exp,
2080 RTCI);
2081 instrumentAddress(OrigIns: I, InsertBefore, Addr: LastByte, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false,
2082 Exp, RTCI);
2083 }
2084}
2085
2086void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit) {
2087 // Set up the arguments to our poison/unpoison functions.
2088 IRBuilder<> IRB(&GlobalInit.front(),
2089 GlobalInit.front().getFirstInsertionPt());
2090
2091 // Add a call to poison all external globals before the given function starts.
2092 Value *ModuleNameAddr =
2093 ConstantExpr::getPointerCast(C: getOrCreateModuleName(), Ty: IntptrTy);
2094 IRB.CreateCall(Callee: AsanPoisonGlobals, Args: ModuleNameAddr);
2095
2096 // Add calls to unpoison all globals before each return instruction.
2097 for (auto &BB : GlobalInit)
2098 if (ReturnInst *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator()))
2099 CallInst::Create(Func: AsanUnpoisonGlobals, NameStr: "", InsertBefore: RI->getIterator());
2100}
2101
2102void ModuleAddressSanitizer::createInitializerPoisonCalls() {
2103 GlobalVariable *GV = M.getGlobalVariable(Name: "llvm.global_ctors");
2104 if (!GV)
2105 return;
2106
2107 ConstantArray *CA = dyn_cast<ConstantArray>(Val: GV->getInitializer());
2108 if (!CA)
2109 return;
2110
2111 for (Use &OP : CA->operands()) {
2112 if (isa<ConstantAggregateZero>(Val: OP)) continue;
2113 ConstantStruct *CS = cast<ConstantStruct>(Val&: OP);
2114
2115 // Must have a function or null ptr.
2116 if (Function *F = dyn_cast<Function>(Val: CS->getOperand(i_nocapture: 1))) {
2117 if (F->getName() == kAsanModuleCtorName) continue;
2118 auto *Priority = cast<ConstantInt>(Val: CS->getOperand(i_nocapture: 0));
2119 // Don't instrument CTORs that will run before asan.module_ctor.
2120 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
2121 continue;
2122 poisonOneInitializer(GlobalInit&: *F);
2123 }
2124 }
2125}
2126
2127const GlobalVariable *
2128ModuleAddressSanitizer::getExcludedAliasedGlobal(const GlobalAlias &GA) const {
2129 // In case this function should be expanded to include rules that do not just
2130 // apply when CompileKernel is true, either guard all existing rules with an
2131 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
2132 // should also apply to user space.
2133 assert(CompileKernel && "Only expecting to be called when compiling kernel");
2134
2135 const Constant *C = GA.getAliasee();
2136
2137 // When compiling the kernel, globals that are aliased by symbols prefixed
2138 // by "__" are special and cannot be padded with a redzone.
2139 if (GA.getName().starts_with(Prefix: "__"))
2140 return dyn_cast<GlobalVariable>(Val: C->stripPointerCastsAndAliases());
2141
2142 return nullptr;
2143}
2144
2145bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
2146 Type *Ty = G->getValueType();
2147 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
2148
2149 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().NoAddress)
2150 return false;
2151 if (!Ty->isSized()) return false;
2152 if (!G->hasInitializer()) return false;
2153 if (!isSupportedAddrspace(TargetTriple, Addr: G))
2154 return false;
2155 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
2156 // Two problems with thread-locals:
2157 // - The address of the main thread's copy can't be computed at link-time.
2158 // - Need to poison all copies, not just the main thread's one.
2159 if (G->isThreadLocal()) return false;
2160 // For now, just ignore this Global if the alignment is large.
2161 if (G->getAlign() && *G->getAlign() > getMinRedzoneSizeForGlobal()) return false;
2162
2163 // For non-COFF targets, only instrument globals known to be defined by this
2164 // TU.
2165 // FIXME: We can instrument comdat globals on ELF if we are using the
2166 // GC-friendly metadata scheme.
2167 if (!TargetTriple.isOSBinFormatCOFF()) {
2168 if (!G->hasExactDefinition() || G->hasComdat())
2169 return false;
2170 } else {
2171 // On COFF, don't instrument non-ODR linkages.
2172 if (G->isInterposable())
2173 return false;
2174 // If the global has AvailableExternally linkage, then it is not in this
2175 // module, which means it does not need to be instrumented.
2176 if (G->hasAvailableExternallyLinkage())
2177 return false;
2178 }
2179
2180 // If a comdat is present, it must have a selection kind that implies ODR
2181 // semantics: no duplicates, any, or exact match.
2182 if (Comdat *C = G->getComdat()) {
2183 switch (C->getSelectionKind()) {
2184 case Comdat::Any:
2185 case Comdat::ExactMatch:
2186 case Comdat::NoDeduplicate:
2187 break;
2188 case Comdat::Largest:
2189 case Comdat::SameSize:
2190 return false;
2191 }
2192 }
2193
2194 if (G->hasSection()) {
2195 // The kernel uses explicit sections for mostly special global variables
2196 // that we should not instrument. E.g. the kernel may rely on their layout
2197 // without redzones, or remove them at link time ("discard.*"), etc.
2198 if (CompileKernel)
2199 return false;
2200
2201 StringRef Section = G->getSection();
2202
2203 // Globals from llvm.metadata aren't emitted, do not instrument them.
2204 if (Section == "llvm.metadata") return false;
2205 // Do not instrument globals from special LLVM sections.
2206 if (Section.contains(Other: "__llvm") || Section.contains(Other: "__LLVM"))
2207 return false;
2208
2209 // Do not instrument function pointers to initialization and termination
2210 // routines: dynamic linker will not properly handle redzones.
2211 if (Section.starts_with(Prefix: ".preinit_array") ||
2212 Section.starts_with(Prefix: ".init_array") ||
2213 Section.starts_with(Prefix: ".fini_array")) {
2214 return false;
2215 }
2216
2217 // Do not instrument user-defined sections (with names resembling
2218 // valid C identifiers)
2219 if (TargetTriple.isOSBinFormatELF()) {
2220 if (llvm::all_of(Range&: Section,
2221 P: [](char c) { return llvm::isAlnum(C: c) || c == '_'; }))
2222 return false;
2223 }
2224
2225 // On COFF, if the section name contains '$', it is highly likely that the
2226 // user is using section sorting to create an array of globals similar to
2227 // the way initialization callbacks are registered in .init_array and
2228 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
2229 // to such globals is counterproductive, because the intent is that they
2230 // will form an array, and out-of-bounds accesses are expected.
2231 // See https://github.com/google/sanitizers/issues/305
2232 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
2233 if (TargetTriple.isOSBinFormatCOFF() && Section.contains(C: '$')) {
2234 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
2235 << *G << "\n");
2236 return false;
2237 }
2238
2239 if (TargetTriple.isOSBinFormatMachO()) {
2240 StringRef ParsedSegment, ParsedSection;
2241 unsigned TAA = 0, StubSize = 0;
2242 bool TAAParsed;
2243 cantFail(Err: MCSectionMachO::ParseSectionSpecifier(
2244 Spec: Section, Segment&: ParsedSegment, Section&: ParsedSection, TAA, TAAParsed, StubSize));
2245
2246 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
2247 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
2248 // them.
2249 if (ParsedSegment == "__OBJC" ||
2250 (ParsedSegment == "__DATA" && ParsedSection.starts_with(Prefix: "__objc_"))) {
2251 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
2252 return false;
2253 }
2254 // See https://github.com/google/sanitizers/issues/32
2255 // Constant CFString instances are compiled in the following way:
2256 // -- the string buffer is emitted into
2257 // __TEXT,__cstring,cstring_literals
2258 // -- the constant NSConstantString structure referencing that buffer
2259 // is placed into __DATA,__cfstring
2260 // Therefore there's no point in placing redzones into __DATA,__cfstring.
2261 // Moreover, it causes the linker to crash on OS X 10.7
2262 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
2263 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
2264 return false;
2265 }
2266 // The linker merges the contents of cstring_literals and removes the
2267 // trailing zeroes.
2268 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
2269 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
2270 return false;
2271 }
2272 }
2273 }
2274
2275 if (CompileKernel) {
2276 // Globals that prefixed by "__" are special and cannot be padded with a
2277 // redzone.
2278 if (G->getName().starts_with(Prefix: "__"))
2279 return false;
2280 }
2281
2282 return true;
2283}
2284
2285// On Mach-O platforms, we emit global metadata in a separate section of the
2286// binary in order to allow the linker to properly dead strip. This is only
2287// supported on recent versions of ld64.
2288bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
2289 if (!TargetTriple.isOSBinFormatMachO())
2290 return false;
2291
2292 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(Major: 10, Minor: 11))
2293 return true;
2294 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(Major: 9))
2295 return true;
2296 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(Major: 2))
2297 return true;
2298 if (TargetTriple.isDriverKit())
2299 return true;
2300 if (TargetTriple.isXROS())
2301 return true;
2302
2303 return false;
2304}
2305
2306StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
2307 switch (TargetTriple.getObjectFormat()) {
2308 case Triple::COFF: return ".ASAN$GL";
2309 case Triple::ELF: return "asan_globals";
2310 case Triple::MachO: return "__DATA,__asan_globals,regular";
2311 case Triple::Wasm:
2312 case Triple::GOFF:
2313 case Triple::SPIRV:
2314 case Triple::XCOFF:
2315 case Triple::DXContainer:
2316 report_fatal_error(
2317 reason: "ModuleAddressSanitizer not implemented for object file format");
2318 case Triple::UnknownObjectFormat:
2319 break;
2320 }
2321 llvm_unreachable("unsupported object format");
2322}
2323
2324void ModuleAddressSanitizer::initializeCallbacks() {
2325 IRBuilder<> IRB(*C);
2326
2327 // Declare our poisoning and unpoisoning functions.
2328 AsanPoisonGlobals = Inserter.insertFunction(Name: kAsanPoisonGlobalsName,
2329 Args: IRB.getVoidTy(), Args&: IntptrTy);
2330 AsanUnpoisonGlobals =
2331 Inserter.insertFunction(Name: kAsanUnpoisonGlobalsName, Args: IRB.getVoidTy());
2332
2333 // Declare functions that register/unregister globals.
2334 AsanRegisterGlobals = Inserter.insertFunction(
2335 Name: kAsanRegisterGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2336 AsanUnregisterGlobals = Inserter.insertFunction(
2337 Name: kAsanUnregisterGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2338
2339 // Declare the functions that find globals in a shared object and then invoke
2340 // the (un)register function on them.
2341 AsanRegisterImageGlobals = Inserter.insertFunction(
2342 Name: kAsanRegisterImageGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy);
2343 AsanUnregisterImageGlobals = Inserter.insertFunction(
2344 Name: kAsanUnregisterImageGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy);
2345
2346 AsanRegisterElfGlobals =
2347 Inserter.insertFunction(Name: kAsanRegisterElfGlobalsName, Args: IRB.getVoidTy(),
2348 Args&: IntptrTy, Args&: IntptrTy, Args&: IntptrTy);
2349 AsanUnregisterElfGlobals =
2350 Inserter.insertFunction(Name: kAsanUnregisterElfGlobalsName, Args: IRB.getVoidTy(),
2351 Args&: IntptrTy, Args&: IntptrTy, Args&: IntptrTy);
2352}
2353
2354// Put the metadata and the instrumented global in the same group. This ensures
2355// that the metadata is discarded if the instrumented global is discarded.
2356void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
2357 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
2358 Module &M = *G->getParent();
2359 Comdat *C = G->getComdat();
2360 if (!C) {
2361 if (!G->hasName()) {
2362 // If G is unnamed, it must be internal. Give it an artificial name
2363 // so we can put it in a comdat.
2364 assert(G->hasLocalLinkage());
2365 G->setName(genName(suffix: "anon_global"));
2366 }
2367
2368 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
2369 std::string Name = std::string(G->getName());
2370 Name += InternalSuffix;
2371 C = M.getOrInsertComdat(Name);
2372 } else {
2373 C = M.getOrInsertComdat(Name: G->getName());
2374 }
2375
2376 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
2377 // linkage to internal linkage so that a symbol table entry is emitted. This
2378 // is necessary in order to create the comdat group.
2379 if (TargetTriple.isOSBinFormatCOFF()) {
2380 C->setSelectionKind(Comdat::NoDeduplicate);
2381 if (G->hasPrivateLinkage())
2382 G->setLinkage(GlobalValue::InternalLinkage);
2383 }
2384 G->setComdat(C);
2385 }
2386
2387 assert(G->hasComdat());
2388 Metadata->setComdat(G->getComdat());
2389}
2390
2391// Create a separate metadata global and put it in the appropriate ASan
2392// global registration section.
2393GlobalVariable *
2394ModuleAddressSanitizer::CreateMetadataGlobal(Constant *Initializer,
2395 StringRef OriginalName) {
2396 auto Linkage = TargetTriple.isOSBinFormatMachO()
2397 ? GlobalVariable::InternalLinkage
2398 : GlobalVariable::PrivateLinkage;
2399 GlobalVariable *Metadata = new GlobalVariable(
2400 M, Initializer->getType(), false, Linkage, Initializer,
2401 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(Name: OriginalName));
2402 Metadata->setSection(getGlobalMetadataSection());
2403 // Place metadata in a large section for x86-64 ELF binaries to mitigate
2404 // relocation pressure.
2405 setGlobalVariableLargeSection(TargetTriple, GV&: *Metadata);
2406 return Metadata;
2407}
2408
2409Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor() {
2410 AsanDtorFunction = Function::createWithDefaultAttr(
2411 Ty: FunctionType::get(Result: Type::getVoidTy(C&: *C), isVarArg: false),
2412 Linkage: GlobalValue::InternalLinkage, AddrSpace: 0, N: kAsanModuleDtorName, M: &M);
2413 AsanDtorFunction->addFnAttr(Kind: Attribute::NoUnwind);
2414 // Ensure Dtor cannot be discarded, even if in a comdat.
2415 appendToUsed(M, Values: {AsanDtorFunction});
2416 BasicBlock *AsanDtorBB = BasicBlock::Create(Context&: *C, Name: "", Parent: AsanDtorFunction);
2417
2418 return ReturnInst::Create(C&: *C, InsertAtEnd: AsanDtorBB);
2419}
2420
2421void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
2422 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2423 ArrayRef<Constant *> MetadataInitializers) {
2424 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2425 auto &DL = M.getDataLayout();
2426
2427 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2428 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2429 Constant *Initializer = MetadataInitializers[i];
2430 GlobalVariable *G = ExtendedGlobals[i];
2431 GlobalVariable *Metadata = CreateMetadataGlobal(Initializer, OriginalName: G->getName());
2432 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2433 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2434 MetadataGlobals[i] = Metadata;
2435
2436 // The MSVC linker always inserts padding when linking incrementally. We
2437 // cope with that by aligning each struct to its size, which must be a power
2438 // of two.
2439 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Ty: Initializer->getType());
2440 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2441 "global metadata will not be padded appropriately");
2442 Metadata->setAlignment(assumeAligned(Value: SizeOfGlobalStruct));
2443
2444 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: "");
2445 }
2446
2447 // Update llvm.compiler.used, adding the new metadata globals. This is
2448 // needed so that during LTO these variables stay alive.
2449 if (!MetadataGlobals.empty())
2450 appendToCompilerUsed(M, Values: MetadataGlobals);
2451}
2452
2453void ModuleAddressSanitizer::instrumentGlobalsELF(
2454 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2455 ArrayRef<Constant *> MetadataInitializers,
2456 const std::string &UniqueModuleId) {
2457 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2458
2459 // Putting globals in a comdat changes the semantic and potentially cause
2460 // false negative odr violations at link time. If odr indicators are used, we
2461 // keep the comdat sections, as link time odr violations will be detected on
2462 // the odr indicator symbols.
2463 bool UseComdatForGlobalsGC = UseOdrIndicator && !UniqueModuleId.empty();
2464
2465 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2466 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2467 GlobalVariable *G = ExtendedGlobals[i];
2468 GlobalVariable *Metadata =
2469 CreateMetadataGlobal(Initializer: MetadataInitializers[i], OriginalName: G->getName());
2470 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2471 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2472 MetadataGlobals[i] = Metadata;
2473
2474 if (UseComdatForGlobalsGC)
2475 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: UniqueModuleId);
2476 }
2477
2478 // Update llvm.compiler.used, adding the new metadata globals. This is
2479 // needed so that during LTO these variables stay alive.
2480 if (!MetadataGlobals.empty())
2481 appendToCompilerUsed(M, Values: MetadataGlobals);
2482
2483 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2484 // to look up the loaded image that contains it. Second, we can store in it
2485 // whether registration has already occurred, to prevent duplicate
2486 // registration.
2487 //
2488 // Common linkage ensures that there is only one global per shared library.
2489 GlobalVariable *RegisteredFlag = new GlobalVariable(
2490 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2491 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2492 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2493
2494 // Create start and stop symbols.
2495 GlobalVariable *StartELFMetadata = new GlobalVariable(
2496 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2497 "__start_" + getGlobalMetadataSection());
2498 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2499 GlobalVariable *StopELFMetadata = new GlobalVariable(
2500 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2501 "__stop_" + getGlobalMetadataSection());
2502 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2503
2504 // Create a call to register the globals with the runtime.
2505 if (ConstructorKind == AsanCtorKind::Global)
2506 IRB.CreateCall(Callee: AsanRegisterElfGlobals,
2507 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2508 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2509 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2510
2511 // We also need to unregister globals at the end, e.g., when a shared library
2512 // gets closed.
2513 if (DestructorKind != AsanDtorKind::None && !MetadataGlobals.empty()) {
2514 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2515 IrbDtor.CreateCall(Callee: AsanUnregisterElfGlobals,
2516 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2517 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2518 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2519 }
2520}
2521
2522void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2523 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2524 ArrayRef<Constant *> MetadataInitializers) {
2525 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2526
2527 // On recent Mach-O platforms, use a structure which binds the liveness of
2528 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2529 // created to be added to llvm.compiler.used
2530 StructType *LivenessTy = StructType::get(elt1: IntptrTy, elts: IntptrTy);
2531 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2532
2533 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2534 Constant *Initializer = MetadataInitializers[i];
2535 GlobalVariable *G = ExtendedGlobals[i];
2536 GlobalVariable *Metadata = CreateMetadataGlobal(Initializer, OriginalName: G->getName());
2537
2538 // On recent Mach-O platforms, we emit the global metadata in a way that
2539 // allows the linker to properly strip dead globals.
2540 auto LivenessBinder =
2541 ConstantStruct::get(T: LivenessTy, Vs: Initializer->getAggregateElement(Elt: 0u),
2542 Vs: ConstantExpr::getPointerCast(C: Metadata, Ty: IntptrTy));
2543 GlobalVariable *Liveness = new GlobalVariable(
2544 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2545 Twine("__asan_binder_") + G->getName());
2546 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2547 LivenessGlobals[i] = Liveness;
2548 }
2549
2550 // Update llvm.compiler.used, adding the new liveness globals. This is
2551 // needed so that during LTO these variables stay alive. The alternative
2552 // would be to have the linker handling the LTO symbols, but libLTO
2553 // current API does not expose access to the section for each symbol.
2554 if (!LivenessGlobals.empty())
2555 appendToCompilerUsed(M, Values: LivenessGlobals);
2556
2557 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2558 // to look up the loaded image that contains it. Second, we can store in it
2559 // whether registration has already occurred, to prevent duplicate
2560 // registration.
2561 //
2562 // common linkage ensures that there is only one global per shared library.
2563 GlobalVariable *RegisteredFlag = new GlobalVariable(
2564 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2565 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2566 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2567
2568 if (ConstructorKind == AsanCtorKind::Global)
2569 IRB.CreateCall(Callee: AsanRegisterImageGlobals,
2570 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2571
2572 // We also need to unregister globals at the end, e.g., when a shared library
2573 // gets closed.
2574 if (DestructorKind != AsanDtorKind::None) {
2575 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2576 IrbDtor.CreateCall(Callee: AsanUnregisterImageGlobals,
2577 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2578 }
2579}
2580
2581void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2582 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2583 ArrayRef<Constant *> MetadataInitializers) {
2584 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2585 unsigned N = ExtendedGlobals.size();
2586 assert(N > 0);
2587
2588 // On platforms that don't have a custom metadata section, we emit an array
2589 // of global metadata structures.
2590 ArrayType *ArrayOfGlobalStructTy =
2591 ArrayType::get(ElementType: MetadataInitializers[0]->getType(), NumElements: N);
2592 auto AllGlobals = new GlobalVariable(
2593 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2594 ConstantArray::get(T: ArrayOfGlobalStructTy, V: MetadataInitializers), "");
2595 if (Mapping.Scale > 3)
2596 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2597
2598 if (ConstructorKind == AsanCtorKind::Global)
2599 IRB.CreateCall(Callee: AsanRegisterGlobals,
2600 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2601 ConstantInt::get(Ty: IntptrTy, V: N)});
2602
2603 // We also need to unregister globals at the end, e.g., when a shared library
2604 // gets closed.
2605 if (DestructorKind != AsanDtorKind::None) {
2606 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2607 IrbDtor.CreateCall(Callee: AsanUnregisterGlobals,
2608 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2609 ConstantInt::get(Ty: IntptrTy, V: N)});
2610 }
2611}
2612
2613// This function replaces all global variables with new variables that have
2614// trailing redzones. It also creates a function that poisons
2615// redzones and inserts this function into llvm.global_ctors.
2616// Sets *CtorComdat to true if the global registration code emitted into the
2617// asan constructor is comdat-compatible.
2618void ModuleAddressSanitizer::instrumentGlobals(IRBuilder<> &IRB,
2619 bool *CtorComdat) {
2620 // Build set of globals that are aliased by some GA, where
2621 // getExcludedAliasedGlobal(GA) returns the relevant GlobalVariable.
2622 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2623 if (CompileKernel) {
2624 for (auto &GA : M.aliases()) {
2625 if (const GlobalVariable *GV = getExcludedAliasedGlobal(GA))
2626 AliasedGlobalExclusions.insert(Ptr: GV);
2627 }
2628 }
2629
2630 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2631 for (auto &G : M.globals()) {
2632 if (!AliasedGlobalExclusions.count(Ptr: &G) && shouldInstrumentGlobal(G: &G))
2633 GlobalsToChange.push_back(Elt: &G);
2634 }
2635
2636 size_t n = GlobalsToChange.size();
2637 auto &DL = M.getDataLayout();
2638
2639 // A global is described by a structure
2640 // size_t beg;
2641 // size_t size;
2642 // size_t size_with_redzone;
2643 // const char *name;
2644 // const char *module_name;
2645 // size_t has_dynamic_init;
2646 // size_t padding_for_windows_msvc_incremental_link;
2647 // size_t odr_indicator;
2648 // We initialize an array of such structures and pass it to a run-time call.
2649 StructType *GlobalStructTy =
2650 StructType::get(elt1: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy,
2651 elts: IntptrTy, elts: IntptrTy, elts: IntptrTy);
2652 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2653 SmallVector<Constant *, 16> Initializers(n);
2654
2655 for (size_t i = 0; i < n; i++) {
2656 GlobalVariable *G = GlobalsToChange[i];
2657
2658 GlobalValue::SanitizerMetadata MD;
2659 if (G->hasSanitizerMetadata())
2660 MD = G->getSanitizerMetadata();
2661
2662 // The runtime library tries demangling symbol names in the descriptor but
2663 // functionality like __cxa_demangle may be unavailable (e.g.
2664 // -static-libstdc++). So we demangle the symbol names here.
2665 std::string NameForGlobal = G->getName().str();
2666 GlobalVariable *Name =
2667 createPrivateGlobalForString(M, Str: llvm::demangle(MangledName: NameForGlobal),
2668 /*AllowMerging*/ true, NamePrefix: genName(suffix: "global"));
2669
2670 Type *Ty = G->getValueType();
2671 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2672 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2673 Type *RightRedZoneTy = ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: RightRedzoneSize);
2674
2675 StructType *NewTy = StructType::get(elt1: Ty, elts: RightRedZoneTy);
2676 Constant *NewInitializer = ConstantStruct::get(
2677 T: NewTy, Vs: G->getInitializer(), Vs: Constant::getNullValue(Ty: RightRedZoneTy));
2678
2679 // Create a new global variable with enough space for a redzone.
2680 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2681 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2682 Linkage = GlobalValue::InternalLinkage;
2683 GlobalVariable *NewGlobal = new GlobalVariable(
2684 M, NewTy, G->isConstant(), Linkage, NewInitializer, "", G,
2685 G->getThreadLocalMode(), G->getAddressSpace());
2686 NewGlobal->copyAttributesFrom(Src: G);
2687 NewGlobal->setComdat(G->getComdat());
2688 NewGlobal->setAlignment(Align(getMinRedzoneSizeForGlobal()));
2689 // Don't fold globals with redzones. ODR violation detector and redzone
2690 // poisoning implicitly creates a dependence on the global's address, so it
2691 // is no longer valid for it to be marked unnamed_addr.
2692 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2693
2694 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2695 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2696 G->isConstant()) {
2697 auto Seq = dyn_cast<ConstantDataSequential>(Val: G->getInitializer());
2698 if (Seq && Seq->isCString())
2699 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2700 }
2701
2702 // Transfer the debug info and type metadata. The payload starts at offset
2703 // zero so we can copy the metadata over as is.
2704 NewGlobal->copyMetadata(Src: G, Offset: 0);
2705
2706 G->replaceAllUsesWith(V: NewGlobal);
2707 NewGlobal->takeName(V: G);
2708 G->eraseFromParent();
2709 NewGlobals[i] = NewGlobal;
2710
2711 Constant *ODRIndicator = Constant::getNullValue(Ty: IntptrTy);
2712 GlobalValue *InstrumentedGlobal = NewGlobal;
2713
2714 bool CanUsePrivateAliases =
2715 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2716 TargetTriple.isOSBinFormatWasm();
2717 if (CanUsePrivateAliases && UsePrivateAlias) {
2718 // Create local alias for NewGlobal to avoid crash on ODR between
2719 // instrumented and non-instrumented libraries.
2720 InstrumentedGlobal =
2721 GlobalAlias::create(Linkage: GlobalValue::PrivateLinkage, Name: "", Aliasee: NewGlobal);
2722 }
2723
2724 // ODR should not happen for local linkage.
2725 if (NewGlobal->hasLocalLinkage()) {
2726 ODRIndicator = ConstantInt::getAllOnesValue(Ty: IntptrTy);
2727 } else if (UseOdrIndicator) {
2728 // With local aliases, we need to provide another externally visible
2729 // symbol __odr_asan_XXX to detect ODR violation.
2730 auto *ODRIndicatorSym =
2731 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2732 Constant::getNullValue(Ty: IRB.getInt8Ty()),
2733 kODRGenPrefix + NameForGlobal, nullptr,
2734 NewGlobal->getThreadLocalMode());
2735
2736 // Set meaningful attributes for indicator symbol.
2737 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2738 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2739 ODRIndicatorSym->setAlignment(Align(1));
2740 ODRIndicator = ConstantExpr::getPtrToInt(C: ODRIndicatorSym, Ty: IntptrTy);
2741 }
2742
2743 Constant *Initializer = ConstantStruct::get(
2744 T: GlobalStructTy,
2745 Vs: ConstantExpr::getPointerCast(C: InstrumentedGlobal, Ty: IntptrTy),
2746 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes),
2747 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes + RightRedzoneSize),
2748 Vs: ConstantExpr::getPointerCast(C: Name, Ty: IntptrTy),
2749 Vs: ConstantExpr::getPointerCast(C: getOrCreateModuleName(), Ty: IntptrTy),
2750 Vs: ConstantInt::get(Ty: IntptrTy, V: MD.IsDynInit),
2751 Vs: Constant::getNullValue(Ty: IntptrTy), Vs: ODRIndicator);
2752
2753 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2754
2755 Initializers[i] = Initializer;
2756 }
2757
2758 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2759 // ConstantMerge'ing them.
2760 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2761 for (size_t i = 0; i < n; i++) {
2762 GlobalVariable *G = NewGlobals[i];
2763 if (G->getName().empty()) continue;
2764 GlobalsToAddToUsedList.push_back(Elt: G);
2765 }
2766 appendToCompilerUsed(M, Values: ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2767
2768 if (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) {
2769 // Use COMDAT and register globals even if n == 0 to ensure that (a) the
2770 // linkage unit will only have one module constructor, and (b) the register
2771 // function will be called. The module destructor is not created when n ==
2772 // 0.
2773 *CtorComdat = true;
2774 instrumentGlobalsELF(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers, UniqueModuleId: getUniqueModuleId(M: &M));
2775 } else if (n == 0) {
2776 // When UseGlobalsGC is false, COMDAT can still be used if n == 0, because
2777 // all compile units will have identical module constructor/destructor.
2778 *CtorComdat = TargetTriple.isOSBinFormatELF();
2779 } else {
2780 *CtorComdat = false;
2781 if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2782 InstrumentGlobalsCOFF(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2783 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2784 InstrumentGlobalsMachO(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2785 } else {
2786 InstrumentGlobalsWithMetadataArray(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2787 }
2788 }
2789
2790 // Create calls for poisoning before initializers run and unpoisoning after.
2791 if (ClInitializers)
2792 createInitializerPoisonCalls();
2793
2794 LLVM_DEBUG(dbgs() << M);
2795}
2796
2797uint64_t
2798ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2799 constexpr uint64_t kMaxRZ = 1 << 18;
2800 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2801
2802 uint64_t RZ = 0;
2803 if (SizeInBytes <= MinRZ / 2) {
2804 // Reduce redzone size for small size objects, e.g. int, char[1]. MinRZ is
2805 // at least 32 bytes, optimize when SizeInBytes is less than or equal to
2806 // half of MinRZ.
2807 RZ = MinRZ - SizeInBytes;
2808 } else {
2809 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2810 RZ = std::clamp(val: (SizeInBytes / MinRZ / 4) * MinRZ, lo: MinRZ, hi: kMaxRZ);
2811
2812 // Round up to multiple of MinRZ.
2813 if (SizeInBytes % MinRZ)
2814 RZ += MinRZ - (SizeInBytes % MinRZ);
2815 }
2816
2817 assert((RZ + SizeInBytes) % MinRZ == 0);
2818
2819 return RZ;
2820}
2821
2822int ModuleAddressSanitizer::GetAsanVersion() const {
2823 int LongSize = M.getDataLayout().getPointerSizeInBits();
2824 bool isAndroid = M.getTargetTriple().isAndroid();
2825 int Version = 8;
2826 // 32-bit Android is one version ahead because of the switch to dynamic
2827 // shadow.
2828 Version += (LongSize == 32 && isAndroid);
2829 return Version;
2830}
2831
2832GlobalVariable *ModuleAddressSanitizer::getOrCreateModuleName() {
2833 if (!ModuleName) {
2834 // We shouldn't merge same module names, as this string serves as unique
2835 // module ID in runtime.
2836 ModuleName =
2837 createPrivateGlobalForString(M, Str: M.getModuleIdentifier(),
2838 /*AllowMerging*/ false, NamePrefix: genName(suffix: "module"));
2839 }
2840 return ModuleName;
2841}
2842
2843bool ModuleAddressSanitizer::instrumentModule() {
2844 initializeCallbacks();
2845
2846 for (Function &F : M)
2847 removeASanIncompatibleFnAttributes(F, /*ReadsArgMem=*/false);
2848
2849 // Create a module constructor. A destructor is created lazily because not all
2850 // platforms, and not all modules need it.
2851 if (ConstructorKind == AsanCtorKind::Global) {
2852 if (CompileKernel) {
2853 // The kernel always builds with its own runtime, and therefore does not
2854 // need the init and version check calls.
2855 AsanCtorFunction = createSanitizerCtor(M, CtorName: kAsanModuleCtorName);
2856 } else {
2857 std::string AsanVersion = std::to_string(val: GetAsanVersion());
2858 std::string VersionCheckName =
2859 InsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2860 std::tie(args&: AsanCtorFunction, args: std::ignore) =
2861 createSanitizerCtorAndInitFunctions(
2862 M, CtorName: kAsanModuleCtorName, InitName: kAsanInitName, /*InitArgTypes=*/{},
2863 /*InitArgs=*/{}, VersionCheckName);
2864 }
2865 }
2866
2867 bool CtorComdat = true;
2868 if (ClGlobals) {
2869 assert(AsanCtorFunction || ConstructorKind == AsanCtorKind::None);
2870 if (AsanCtorFunction) {
2871 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2872 instrumentGlobals(IRB, CtorComdat: &CtorComdat);
2873 } else {
2874 IRBuilder<> IRB(*C);
2875 instrumentGlobals(IRB, CtorComdat: &CtorComdat);
2876 }
2877 }
2878
2879 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2880
2881 // Put the constructor and destructor in comdat if both
2882 // (1) global instrumentation is not TU-specific
2883 // (2) target is ELF.
2884 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2885 if (AsanCtorFunction) {
2886 AsanCtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleCtorName));
2887 appendToGlobalCtors(M, F: AsanCtorFunction, Priority, Data: AsanCtorFunction);
2888 }
2889 if (AsanDtorFunction) {
2890 AsanDtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleDtorName));
2891 appendToGlobalDtors(M, F: AsanDtorFunction, Priority, Data: AsanDtorFunction);
2892 }
2893 } else {
2894 if (AsanCtorFunction)
2895 appendToGlobalCtors(M, F: AsanCtorFunction, Priority);
2896 if (AsanDtorFunction)
2897 appendToGlobalDtors(M, F: AsanDtorFunction, Priority);
2898 }
2899
2900 return true;
2901}
2902
2903void AddressSanitizer::initializeCallbacks(const TargetLibraryInfo *TLI) {
2904 IRBuilder<> IRB(*C);
2905 // Create __asan_report* callbacks.
2906 // IsWrite, TypeSize and Exp are encoded in the function name.
2907 for (int Exp = 0; Exp < 2; Exp++) {
2908 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2909 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2910 const std::string ExpStr = Exp ? "exp_" : "";
2911 const std::string EndingStr = Recover ? "_noabort" : "";
2912
2913 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2914 SmallVector<Type *, 2> Args1{1, IntptrTy};
2915 AttributeList AL2;
2916 AttributeList AL1;
2917 if (Exp) {
2918 Type *ExpType = Type::getInt32Ty(C&: *C);
2919 Args2.push_back(Elt: ExpType);
2920 Args1.push_back(Elt: ExpType);
2921 if (auto AK = TLI->getExtAttrForI32Param(Signed: false)) {
2922 AL2 = AL2.addParamAttribute(C&: *C, ArgNo: 2, Kind: AK);
2923 AL1 = AL1.addParamAttribute(C&: *C, ArgNo: 1, Kind: AK);
2924 }
2925 }
2926 AsanErrorCallbackSized[AccessIsWrite][Exp] = Inserter.insertFunction(
2927 Name: kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2928 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), Args&: AL2);
2929
2930 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] =
2931 Inserter.insertFunction(
2932 Name: ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2933 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), Args&: AL2);
2934
2935 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2936 AccessSizeIndex++) {
2937 const std::string Suffix = TypeStr + itostr(X: 1ULL << AccessSizeIndex);
2938 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2939 Inserter.insertFunction(
2940 Name: kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2941 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), Args&: AL1);
2942
2943 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2944 Inserter.insertFunction(
2945 Name: ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2946 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), Args&: AL1);
2947 }
2948 }
2949 }
2950
2951 const std::string MemIntrinCallbackPrefix =
2952 (CompileKernel && !ClKasanMemIntrinCallbackPrefix)
2953 ? std::string("")
2954 : ClMemoryAccessCallbackPrefix;
2955 AsanMemmove = Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memmove",
2956 Args&: PtrTy, Args&: PtrTy, Args&: PtrTy, Args&: IntptrTy);
2957 AsanMemcpy = Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memcpy",
2958 Args&: PtrTy, Args&: PtrTy, Args&: PtrTy, Args&: IntptrTy);
2959 AsanMemset =
2960 Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memset",
2961 Args: TLI->getAttrList(C, ArgNos: {1},
2962 /*Signed=*/false),
2963 Args&: PtrTy, Args&: PtrTy, Args: IRB.getInt32Ty(), Args&: IntptrTy);
2964
2965 AsanHandleNoReturnFunc =
2966 Inserter.insertFunction(Name: kAsanHandleNoReturnName, Args: IRB.getVoidTy());
2967
2968 AsanPtrCmpFunction =
2969 Inserter.insertFunction(Name: kAsanPtrCmp, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2970 AsanPtrSubFunction =
2971 Inserter.insertFunction(Name: kAsanPtrSub, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2972 if (Mapping.InGlobal)
2973 AsanShadowGlobal = M.getOrInsertGlobal(Name: "__asan_shadow",
2974 Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: 0));
2975
2976 AMDGPUAddressShared =
2977 Inserter.insertFunction(Name: kAMDGPUAddressSharedName, Args: IRB.getInt1Ty(), Args&: PtrTy);
2978 AMDGPUAddressPrivate = Inserter.insertFunction(Name: kAMDGPUAddressPrivateName,
2979 Args: IRB.getInt1Ty(), Args&: PtrTy);
2980}
2981
2982bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
2983 // For each NSObject descendant having a +load method, this method is invoked
2984 // by the ObjC runtime before any of the static constructors is called.
2985 // Therefore we need to instrument such methods with a call to __asan_init
2986 // at the beginning in order to initialize our runtime before any access to
2987 // the shadow memory.
2988 // We cannot just ignore these methods, because they may call other
2989 // instrumented functions.
2990 if (F.getName().contains(Other: " load]")) {
2991 FunctionCallee AsanInitFunction =
2992 declareSanitizerInitFunction(M&: *F.getParent(), InitName: kAsanInitName, InitArgTypes: {});
2993 IRBuilder<> IRB(&F.front(), F.front().begin());
2994 IRB.CreateCall(Callee: AsanInitFunction, Args: {});
2995 return true;
2996 }
2997 return false;
2998}
2999
3000bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
3001 // Generate code only when dynamic addressing is needed.
3002 if (Mapping.Offset != kDynamicShadowSentinel)
3003 return false;
3004
3005 IRBuilder<> IRB(&F.front().front());
3006 if (Mapping.InGlobal) {
3007 if (ClWithIfuncSuppressRemat) {
3008 // An empty inline asm with input reg == output reg.
3009 // An opaque pointer-to-int cast, basically.
3010 InlineAsm *Asm = InlineAsm::get(
3011 Ty: FunctionType::get(Result: IntptrTy, Params: {AsanShadowGlobal->getType()}, isVarArg: false),
3012 AsmString: StringRef(""), Constraints: StringRef("=r,0"),
3013 /*hasSideEffects=*/false);
3014 LocalDynamicShadow =
3015 IRB.CreateCall(Callee: Asm, Args: {AsanShadowGlobal}, Name: ".asan.shadow");
3016 } else {
3017 LocalDynamicShadow =
3018 IRB.CreatePointerCast(V: AsanShadowGlobal, DestTy: IntptrTy, Name: ".asan.shadow");
3019 }
3020 } else {
3021 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
3022 Name: kAsanShadowMemoryDynamicAddress, Ty: IntptrTy);
3023 LocalDynamicShadow = IRB.CreateLoad(Ty: IntptrTy, Ptr: GlobalDynamicAddress);
3024 }
3025 return true;
3026}
3027
3028void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
3029 // Find the one possible call to llvm.localescape and pre-mark allocas passed
3030 // to it as uninteresting. This assumes we haven't started processing allocas
3031 // yet. This check is done up front because iterating the use list in
3032 // isInterestingAlloca would be algorithmically slower.
3033 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
3034
3035 // Try to get the declaration of llvm.localescape. If it's not in the module,
3036 // we can exit early.
3037 if (!F.getParent()->getFunction(Name: "llvm.localescape")) return;
3038
3039 // Look for a call to llvm.localescape call in the entry block. It can't be in
3040 // any other block.
3041 for (Instruction &I : F.getEntryBlock()) {
3042 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &I);
3043 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
3044 // We found a call. Mark all the allocas passed in as uninteresting.
3045 for (Value *Arg : II->args()) {
3046 AllocaInst *AI = dyn_cast<AllocaInst>(Val: Arg->stripPointerCasts());
3047 assert(AI && AI->isStaticAlloca() &&
3048 "non-static alloca arg to localescape");
3049 ProcessedAllocas[AI] = false;
3050 }
3051 break;
3052 }
3053 }
3054}
3055// Mitigation for https://github.com/google/sanitizers/issues/749
3056// We don't instrument Windows catch-block parameters to avoid
3057// interfering with exception handling assumptions.
3058void AddressSanitizer::markCatchParametersAsUninteresting(Function &F) {
3059 for (BasicBlock &BB : F) {
3060 for (Instruction &I : BB) {
3061 if (auto *CatchPad = dyn_cast<CatchPadInst>(Val: &I)) {
3062 // Mark the parameters to a catch-block as uninteresting to avoid
3063 // instrumenting them.
3064 for (Value *Operand : CatchPad->arg_operands())
3065 if (auto *AI = dyn_cast<AllocaInst>(Val: Operand))
3066 ProcessedAllocas[AI] = false;
3067 }
3068 }
3069 }
3070}
3071
3072bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
3073 bool ShouldInstrument =
3074 ClDebugMin < 0 || ClDebugMax < 0 ||
3075 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
3076 Instrumented++;
3077 return !ShouldInstrument;
3078}
3079
3080bool AddressSanitizer::instrumentFunction(Function &F,
3081 const TargetLibraryInfo *TLI,
3082 const TargetTransformInfo *TTI) {
3083 bool FunctionModified = false;
3084
3085 // Do not apply any instrumentation for naked functions.
3086 if (F.hasFnAttribute(Kind: Attribute::Naked))
3087 return FunctionModified;
3088
3089 // If needed, insert __asan_init before checking for SanitizeAddress attr.
3090 // This function needs to be called even if the function body is not
3091 // instrumented.
3092 if (maybeInsertAsanInitAtFunctionEntry(F))
3093 FunctionModified = true;
3094
3095 // Leave if the function doesn't need instrumentation.
3096 if (!F.hasFnAttribute(Kind: Attribute::SanitizeAddress)) return FunctionModified;
3097
3098 if (F.hasFnAttribute(Kind: Attribute::DisableSanitizerInstrumentation))
3099 return FunctionModified;
3100
3101 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
3102
3103 initializeCallbacks(TLI);
3104
3105 FunctionStateRAII CleanupObj(this);
3106
3107 RuntimeCallInserter RTCI(F);
3108
3109 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
3110
3111 // We can't instrument allocas used with llvm.localescape. Only static allocas
3112 // can be passed to that intrinsic.
3113 markEscapedLocalAllocas(F);
3114
3115 if (TargetTriple.isOSWindows())
3116 markCatchParametersAsUninteresting(F);
3117
3118 // We want to instrument every address only once per basic block (unless there
3119 // are calls between uses).
3120 SmallPtrSet<Value *, 16> TempsToInstrument;
3121 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
3122 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
3123 SmallVector<Instruction *, 8> NoReturnCalls;
3124 SmallVector<BasicBlock *, 16> AllBlocks;
3125 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
3126
3127 // Fill the set of memory operations to instrument.
3128 for (auto &BB : F) {
3129 AllBlocks.push_back(Elt: &BB);
3130 TempsToInstrument.clear();
3131 int NumInsnsPerBB = 0;
3132 for (auto &Inst : BB) {
3133 if (LooksLikeCodeInBug11395(I: &Inst)) return false;
3134 // Skip instructions inserted by another instrumentation.
3135 if (Inst.hasMetadata(KindID: LLVMContext::MD_nosanitize))
3136 continue;
3137 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
3138 getInterestingMemoryOperands(I: &Inst, Interesting&: InterestingOperands, TTI);
3139
3140 if (!InterestingOperands.empty()) {
3141 for (auto &Operand : InterestingOperands) {
3142 if (ClOpt && ClOptSameTemp) {
3143 Value *Ptr = Operand.getPtr();
3144 // If we have a mask, skip instrumentation if we've already
3145 // instrumented the full object. But don't add to TempsToInstrument
3146 // because we might get another load/store with a different mask.
3147 if (Operand.MaybeMask) {
3148 if (TempsToInstrument.count(Ptr))
3149 continue; // We've seen this (whole) temp in the current BB.
3150 } else {
3151 if (!TempsToInstrument.insert(Ptr).second)
3152 continue; // We've seen this temp in the current BB.
3153 }
3154 }
3155 OperandsToInstrument.push_back(Elt: Operand);
3156 NumInsnsPerBB++;
3157 }
3158 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
3159 isInterestingPointerComparison(I: &Inst)) ||
3160 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
3161 isInterestingPointerSubtraction(I: &Inst))) {
3162 PointerComparisonsOrSubtracts.push_back(Elt: &Inst);
3163 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: &Inst)) {
3164 // ok, take it.
3165 IntrinToInstrument.push_back(Elt: MI);
3166 NumInsnsPerBB++;
3167 } else {
3168 if (auto *CB = dyn_cast<CallBase>(Val: &Inst)) {
3169 // A call inside BB.
3170 TempsToInstrument.clear();
3171 if (CB->doesNotReturn())
3172 NoReturnCalls.push_back(Elt: CB);
3173 }
3174 if (CallInst *CI = dyn_cast<CallInst>(Val: &Inst))
3175 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
3176 }
3177 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
3178 }
3179 }
3180
3181 bool UseCalls = (InstrumentationWithCallsThreshold >= 0 &&
3182 OperandsToInstrument.size() + IntrinToInstrument.size() >
3183 (unsigned)InstrumentationWithCallsThreshold);
3184 const DataLayout &DL = F.getDataLayout();
3185 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext());
3186
3187 // Instrument.
3188 int NumInstrumented = 0;
3189 for (auto &Operand : OperandsToInstrument) {
3190 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3191 instrumentMop(ObjSizeVis, O&: Operand, UseCalls,
3192 DL: F.getDataLayout(), RTCI);
3193 FunctionModified = true;
3194 }
3195 for (auto *Inst : IntrinToInstrument) {
3196 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3197 instrumentMemIntrinsic(MI: Inst, RTCI);
3198 FunctionModified = true;
3199 }
3200
3201 FunctionStackPoisoner FSP(F, *this, RTCI);
3202 bool ChangedStack = FSP.runOnFunction();
3203
3204 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
3205 // See e.g. https://github.com/google/sanitizers/issues/37
3206 for (auto *CI : NoReturnCalls) {
3207 IRBuilder<> IRB(CI);
3208 RTCI.createRuntimeCall(IRB, Callee: AsanHandleNoReturnFunc, Args: {});
3209 }
3210
3211 for (auto *Inst : PointerComparisonsOrSubtracts) {
3212 instrumentPointerComparisonOrSubtraction(I: Inst, RTCI);
3213 FunctionModified = true;
3214 }
3215
3216 if (ChangedStack || !NoReturnCalls.empty())
3217 FunctionModified = true;
3218
3219 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
3220 << F << "\n");
3221
3222 return FunctionModified;
3223}
3224
3225// Workaround for bug 11395: we don't want to instrument stack in functions
3226// with large assembly blobs (32-bit only), otherwise reg alloc may crash.
3227// FIXME: remove once the bug 11395 is fixed.
3228bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
3229 if (LongSize != 32) return false;
3230 CallInst *CI = dyn_cast<CallInst>(Val: I);
3231 if (!CI || !CI->isInlineAsm()) return false;
3232 if (CI->arg_size() <= 5)
3233 return false;
3234 // We have inline assembly with quite a few arguments.
3235 return true;
3236}
3237
3238void FunctionStackPoisoner::initializeCallbacks(Module &) {
3239 IRBuilder<> IRB(*C);
3240 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always ||
3241 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3242 const char *MallocNameTemplate =
3243 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always
3244 ? kAsanStackMallocAlwaysNameTemplate
3245 : kAsanStackMallocNameTemplate;
3246 for (int Index = 0; Index <= kMaxAsanStackMallocSizeClass; Index++) {
3247 std::string Suffix = itostr(X: Index);
3248 AsanStackMallocFunc[Index] = ASan.Inserter.insertFunction(
3249 Name: MallocNameTemplate + Suffix, Args&: IntptrTy, Args&: IntptrTy);
3250 AsanStackFreeFunc[Index] =
3251 ASan.Inserter.insertFunction(Name: kAsanStackFreeNameTemplate + Suffix,
3252 Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3253 }
3254 }
3255 if (ASan.UseAfterScope) {
3256 AsanPoisonStackMemoryFunc = ASan.Inserter.insertFunction(
3257 Name: kAsanPoisonStackMemoryName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3258 AsanUnpoisonStackMemoryFunc = ASan.Inserter.insertFunction(
3259 Name: kAsanUnpoisonStackMemoryName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3260 }
3261
3262 for (size_t Val : {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0xf1, 0xf2,
3263 0xf3, 0xf5, 0xf8}) {
3264 std::ostringstream Name;
3265 Name << kAsanSetShadowPrefix;
3266 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
3267 AsanSetShadowFunc[Val] = ASan.Inserter.insertFunction(
3268 Name: Name.str(), Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3269 }
3270
3271 AsanAllocaPoisonFunc = ASan.Inserter.insertFunction(
3272 Name: kAsanAllocaPoison, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3273 AsanAllocasUnpoisonFunc = ASan.Inserter.insertFunction(
3274 Name: kAsanAllocasUnpoison, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3275}
3276
3277void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
3278 ArrayRef<uint8_t> ShadowBytes,
3279 size_t Begin, size_t End,
3280 IRBuilder<> &IRB,
3281 Value *ShadowBase) {
3282 if (Begin >= End)
3283 return;
3284
3285 const size_t LargestStoreSizeInBytes =
3286 std::min<size_t>(a: sizeof(uint64_t), b: ASan.LongSize / 8);
3287
3288 const bool IsLittleEndian = F.getDataLayout().isLittleEndian();
3289
3290 // Poison given range in shadow using larges store size with out leading and
3291 // trailing zeros in ShadowMask. Zeros never change, so they need neither
3292 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
3293 // middle of a store.
3294 for (size_t i = Begin; i < End;) {
3295 if (!ShadowMask[i]) {
3296 assert(!ShadowBytes[i]);
3297 ++i;
3298 continue;
3299 }
3300
3301 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
3302 // Fit store size into the range.
3303 while (StoreSizeInBytes > End - i)
3304 StoreSizeInBytes /= 2;
3305
3306 // Minimize store size by trimming trailing zeros.
3307 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
3308 while (j <= StoreSizeInBytes / 2)
3309 StoreSizeInBytes /= 2;
3310 }
3311
3312 uint64_t Val = 0;
3313 for (size_t j = 0; j < StoreSizeInBytes; j++) {
3314 if (IsLittleEndian)
3315 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
3316 else
3317 Val = (Val << 8) | ShadowBytes[i + j];
3318 }
3319
3320 Value *Ptr = IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i));
3321 Value *Poison = IRB.getIntN(N: StoreSizeInBytes * 8, C: Val);
3322 IRB.CreateAlignedStore(
3323 Val: Poison, Ptr: IRB.CreateIntToPtr(V: Ptr, DestTy: PointerType::getUnqual(C&: Poison->getContext())),
3324 Align: Align(1));
3325
3326 i += StoreSizeInBytes;
3327 }
3328}
3329
3330void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3331 ArrayRef<uint8_t> ShadowBytes,
3332 IRBuilder<> &IRB, Value *ShadowBase) {
3333 copyToShadow(ShadowMask, ShadowBytes, Begin: 0, End: ShadowMask.size(), IRB, ShadowBase);
3334}
3335
3336void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3337 ArrayRef<uint8_t> ShadowBytes,
3338 size_t Begin, size_t End,
3339 IRBuilder<> &IRB, Value *ShadowBase) {
3340 assert(ShadowMask.size() == ShadowBytes.size());
3341 size_t Done = Begin;
3342 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
3343 if (!ShadowMask[i]) {
3344 assert(!ShadowBytes[i]);
3345 continue;
3346 }
3347 uint8_t Val = ShadowBytes[i];
3348 if (!AsanSetShadowFunc[Val])
3349 continue;
3350
3351 // Skip same values.
3352 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
3353 }
3354
3355 if (j - i >= ASan.MaxInlinePoisoningSize) {
3356 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End: i, IRB, ShadowBase);
3357 RTCI.createRuntimeCall(
3358 IRB, Callee: AsanSetShadowFunc[Val],
3359 Args: {IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i)),
3360 ConstantInt::get(Ty: IntptrTy, V: j - i)});
3361 Done = j;
3362 }
3363 }
3364
3365 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End, IRB, ShadowBase);
3366}
3367
3368// Fake stack allocator (asan_fake_stack.h) has 11 size classes
3369// for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
3370static int StackMallocSizeClass(uint64_t LocalStackSize) {
3371 assert(LocalStackSize <= kMaxStackMallocSize);
3372 uint64_t MaxSize = kMinStackMallocSize;
3373 for (int i = 0;; i++, MaxSize *= 2)
3374 if (LocalStackSize <= MaxSize) return i;
3375 llvm_unreachable("impossible LocalStackSize");
3376}
3377
3378void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
3379 Instruction *CopyInsertPoint = &F.front().front();
3380 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
3381 // Insert after the dynamic shadow location is determined
3382 CopyInsertPoint = CopyInsertPoint->getNextNode();
3383 assert(CopyInsertPoint);
3384 }
3385 IRBuilder<> IRB(CopyInsertPoint);
3386 const DataLayout &DL = F.getDataLayout();
3387 for (Argument &Arg : F.args()) {
3388 if (Arg.hasByValAttr()) {
3389 Type *Ty = Arg.getParamByValType();
3390 const Align Alignment =
3391 DL.getValueOrABITypeAlignment(Alignment: Arg.getParamAlign(), Ty);
3392
3393 AllocaInst *AI = IRB.CreateAlloca(
3394 Ty, ArraySize: nullptr,
3395 Name: (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
3396 ".byval");
3397 AI->setAlignment(Alignment);
3398 Arg.replaceAllUsesWith(V: AI);
3399
3400 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3401 IRB.CreateMemCpy(Dst: AI, DstAlign: Alignment, Src: &Arg, SrcAlign: Alignment, Size: AllocSize);
3402 }
3403 }
3404}
3405
3406PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
3407 Value *ValueIfTrue,
3408 Instruction *ThenTerm,
3409 Value *ValueIfFalse) {
3410 PHINode *PHI = IRB.CreatePHI(Ty: ValueIfTrue->getType(), NumReservedValues: 2);
3411 BasicBlock *CondBlock = cast<Instruction>(Val: Cond)->getParent();
3412 PHI->addIncoming(V: ValueIfFalse, BB: CondBlock);
3413 BasicBlock *ThenBlock = ThenTerm->getParent();
3414 PHI->addIncoming(V: ValueIfTrue, BB: ThenBlock);
3415 return PHI;
3416}
3417
3418Value *FunctionStackPoisoner::createAllocaForLayout(
3419 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
3420 AllocaInst *Alloca;
3421 if (Dynamic) {
3422 Alloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(),
3423 ArraySize: ConstantInt::get(Ty: IRB.getInt64Ty(), V: L.FrameSize),
3424 Name: "MyAlloca");
3425 } else {
3426 Alloca = IRB.CreateAlloca(Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: L.FrameSize),
3427 ArraySize: nullptr, Name: "MyAlloca");
3428 assert(Alloca->isStaticAlloca());
3429 }
3430 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
3431 uint64_t FrameAlignment = std::max(a: L.FrameAlignment, b: uint64_t(ClRealignStack));
3432 Alloca->setAlignment(Align(FrameAlignment));
3433 return Alloca;
3434}
3435
3436void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
3437 BasicBlock &FirstBB = *F.begin();
3438 IRBuilder<> IRB(dyn_cast<Instruction>(Val: FirstBB.begin()));
3439 DynamicAllocaLayout = IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr);
3440 IRB.CreateStore(Val: Constant::getNullValue(Ty: IntptrTy), Ptr: DynamicAllocaLayout);
3441 DynamicAllocaLayout->setAlignment(Align(32));
3442}
3443
3444void FunctionStackPoisoner::processDynamicAllocas() {
3445 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
3446 assert(DynamicAllocaPoisonCallVec.empty());
3447 return;
3448 }
3449
3450 // Insert poison calls for lifetime intrinsics for dynamic allocas.
3451 for (const auto &APC : DynamicAllocaPoisonCallVec) {
3452 assert(APC.InsBefore);
3453 assert(APC.AI);
3454 assert(ASan.isInterestingAlloca(*APC.AI));
3455 assert(!APC.AI->isStaticAlloca());
3456
3457 IRBuilder<> IRB(APC.InsBefore);
3458 poisonAlloca(V: APC.AI, Size: APC.Size, IRB, DoPoison: APC.DoPoison);
3459 // Dynamic allocas will be unpoisoned unconditionally below in
3460 // unpoisonDynamicAllocas.
3461 // Flag that we need unpoison static allocas.
3462 }
3463
3464 // Handle dynamic allocas.
3465 createDynamicAllocasInitStorage();
3466 for (auto &AI : DynamicAllocaVec)
3467 handleDynamicAllocaCall(AI);
3468 unpoisonDynamicAllocas();
3469}
3470
3471/// Collect instructions in the entry block after \p InsBefore which initialize
3472/// permanent storage for a function argument. These instructions must remain in
3473/// the entry block so that uninitialized values do not appear in backtraces. An
3474/// added benefit is that this conserves spill slots. This does not move stores
3475/// before instrumented / "interesting" allocas.
3476static void findStoresToUninstrumentedArgAllocas(
3477 AddressSanitizer &ASan, Instruction &InsBefore,
3478 SmallVectorImpl<Instruction *> &InitInsts) {
3479 Instruction *Start = InsBefore.getNextNode();
3480 for (Instruction *It = Start; It; It = It->getNextNode()) {
3481 // Argument initialization looks like:
3482 // 1) store <Argument>, <Alloca> OR
3483 // 2) <CastArgument> = cast <Argument> to ...
3484 // store <CastArgument> to <Alloca>
3485 // Do not consider any other kind of instruction.
3486 //
3487 // Note: This covers all known cases, but may not be exhaustive. An
3488 // alternative to pattern-matching stores is to DFS over all Argument uses:
3489 // this might be more general, but is probably much more complicated.
3490 if (isa<AllocaInst>(Val: It) || isa<CastInst>(Val: It))
3491 continue;
3492 if (auto *Store = dyn_cast<StoreInst>(Val: It)) {
3493 // The store destination must be an alloca that isn't interesting for
3494 // ASan to instrument. These are moved up before InsBefore, and they're
3495 // not interesting because allocas for arguments can be mem2reg'd.
3496 auto *Alloca = dyn_cast<AllocaInst>(Val: Store->getPointerOperand());
3497 if (!Alloca || ASan.isInterestingAlloca(AI: *Alloca))
3498 continue;
3499
3500 Value *Val = Store->getValueOperand();
3501 bool IsDirectArgInit = isa<Argument>(Val);
3502 bool IsArgInitViaCast =
3503 isa<CastInst>(Val) &&
3504 isa<Argument>(Val: cast<CastInst>(Val)->getOperand(i_nocapture: 0)) &&
3505 // Check that the cast appears directly before the store. Otherwise
3506 // moving the cast before InsBefore may break the IR.
3507 Val == It->getPrevNode();
3508 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3509 if (!IsArgInit)
3510 continue;
3511
3512 if (IsArgInitViaCast)
3513 InitInsts.push_back(Elt: cast<Instruction>(Val));
3514 InitInsts.push_back(Elt: Store);
3515 continue;
3516 }
3517
3518 // Do not reorder past unknown instructions: argument initialization should
3519 // only involve casts and stores.
3520 return;
3521 }
3522}
3523
3524static StringRef getAllocaName(AllocaInst *AI) {
3525 // Alloca could have been renamed for uniqueness. Its true name will have been
3526 // recorded as an annotation.
3527 if (AI->hasMetadata(KindID: LLVMContext::MD_annotation)) {
3528 MDTuple *AllocaAnnotations =
3529 cast<MDTuple>(Val: AI->getMetadata(KindID: LLVMContext::MD_annotation));
3530 for (auto &Annotation : AllocaAnnotations->operands()) {
3531 if (!isa<MDTuple>(Val: Annotation))
3532 continue;
3533 auto AnnotationTuple = cast<MDTuple>(Val: Annotation);
3534 for (unsigned Index = 0; Index < AnnotationTuple->getNumOperands();
3535 Index++) {
3536 // All annotations are strings
3537 auto MetadataString =
3538 cast<MDString>(Val: AnnotationTuple->getOperand(I: Index));
3539 if (MetadataString->getString() == "alloca_name_altered")
3540 return cast<MDString>(Val: AnnotationTuple->getOperand(I: Index + 1))
3541 ->getString();
3542 }
3543 }
3544 }
3545 return AI->getName();
3546}
3547
3548void FunctionStackPoisoner::processStaticAllocas() {
3549 if (AllocaVec.empty()) {
3550 assert(StaticAllocaPoisonCallVec.empty());
3551 return;
3552 }
3553
3554 int StackMallocIdx = -1;
3555 DebugLoc EntryDebugLocation;
3556 if (auto SP = F.getSubprogram())
3557 EntryDebugLocation =
3558 DILocation::get(Context&: SP->getContext(), Line: SP->getScopeLine(), Column: 0, Scope: SP);
3559
3560 Instruction *InsBefore = AllocaVec[0];
3561 IRBuilder<> IRB(InsBefore);
3562
3563 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3564 // debug info is broken, because only entry-block allocas are treated as
3565 // regular stack slots.
3566 auto InsBeforeB = InsBefore->getParent();
3567 assert(InsBeforeB == &F.getEntryBlock());
3568 for (auto *AI : StaticAllocasToMoveUp)
3569 if (AI->getParent() == InsBeforeB)
3570 AI->moveBefore(InsertPos: InsBefore->getIterator());
3571
3572 // Move stores of arguments into entry-block allocas as well. This prevents
3573 // extra stack slots from being generated (to house the argument values until
3574 // they can be stored into the allocas). This also prevents uninitialized
3575 // values from being shown in backtraces.
3576 SmallVector<Instruction *, 8> ArgInitInsts;
3577 findStoresToUninstrumentedArgAllocas(ASan, InsBefore&: *InsBefore, InitInsts&: ArgInitInsts);
3578 for (Instruction *ArgInitInst : ArgInitInsts)
3579 ArgInitInst->moveBefore(InsertPos: InsBefore->getIterator());
3580
3581 // If we have a call to llvm.localescape, keep it in the entry block.
3582 if (LocalEscapeCall)
3583 LocalEscapeCall->moveBefore(InsertPos: InsBefore->getIterator());
3584
3585 SmallVector<ASanStackVariableDescription, 16> SVD;
3586 SVD.reserve(N: AllocaVec.size());
3587 for (AllocaInst *AI : AllocaVec) {
3588 StringRef Name = getAllocaName(AI);
3589 ASanStackVariableDescription D = {.Name: Name.data(),
3590 .Size: ASan.getAllocaSizeInBytes(AI: *AI),
3591 .LifetimeSize: 0,
3592 .Alignment: AI->getAlign().value(),
3593 .AI: AI,
3594 .Offset: 0,
3595 .Line: 0};
3596 SVD.push_back(Elt: D);
3597 }
3598
3599 // Minimal header size (left redzone) is 4 pointers,
3600 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3601 uint64_t Granularity = 1ULL << Mapping.Scale;
3602 uint64_t MinHeaderSize = std::max(a: (uint64_t)ASan.LongSize / 2, b: Granularity);
3603 const ASanStackFrameLayout &L =
3604 ComputeASanStackFrameLayout(Vars&: SVD, Granularity, MinHeaderSize);
3605
3606 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3607 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3608 for (auto &Desc : SVD)
3609 AllocaToSVDMap[Desc.AI] = &Desc;
3610
3611 // Update SVD with information from lifetime intrinsics.
3612 for (const auto &APC : StaticAllocaPoisonCallVec) {
3613 assert(APC.InsBefore);
3614 assert(APC.AI);
3615 assert(ASan.isInterestingAlloca(*APC.AI));
3616 assert(APC.AI->isStaticAlloca());
3617
3618 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3619 Desc.LifetimeSize = Desc.Size;
3620 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3621 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3622 if (LifetimeLoc->getFile() == FnLoc->getFile())
3623 if (unsigned Line = LifetimeLoc->getLine())
3624 Desc.Line = std::min(a: Desc.Line ? Desc.Line : Line, b: Line);
3625 }
3626 }
3627 }
3628
3629 auto DescriptionString = ComputeASanStackFrameDescription(Vars: SVD);
3630 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3631 uint64_t LocalStackSize = L.FrameSize;
3632 bool DoStackMalloc =
3633 ASan.UseAfterReturn != AsanDetectStackUseAfterReturnMode::Never &&
3634 !ASan.CompileKernel && LocalStackSize <= kMaxStackMallocSize;
3635 bool DoDynamicAlloca = ClDynamicAllocaStack;
3636 // Don't do dynamic alloca or stack malloc if:
3637 // 1) There is inline asm: too often it makes assumptions on which registers
3638 // are available.
3639 // 2) There is a returns_twice call (typically setjmp), which is
3640 // optimization-hostile, and doesn't play well with introduced indirect
3641 // register-relative calculation of local variable addresses.
3642 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3643 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3644
3645 Type *PtrTy = F.getDataLayout().getAllocaPtrType(Ctx&: F.getContext());
3646 Value *StaticAlloca =
3647 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, Dynamic: false);
3648
3649 Value *FakeStackPtr;
3650 Value *FakeStackInt;
3651 Value *LocalStackBase;
3652 Value *LocalStackBaseAlloca;
3653 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3654
3655 if (DoStackMalloc) {
3656 LocalStackBaseAlloca =
3657 IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr, Name: "asan_local_stack_base");
3658 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3659 // void *FakeStack = __asan_option_detect_stack_use_after_return
3660 // ? __asan_stack_malloc_N(LocalStackSize)
3661 // : nullptr;
3662 // void *LocalStackBase = (FakeStack) ? FakeStack :
3663 // alloca(LocalStackSize);
3664 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3665 Name: kAsanOptionDetectUseAfterReturn, Ty: IRB.getInt32Ty());
3666 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3667 LHS: IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: OptionDetectUseAfterReturn),
3668 RHS: Constant::getNullValue(Ty: IRB.getInt32Ty()));
3669 Instruction *Term =
3670 SplitBlockAndInsertIfThen(Cond: UseAfterReturnIsEnabled, SplitBefore: InsBefore, Unreachable: false);
3671 IRBuilder<> IRBIf(Term);
3672 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3673 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3674 Value *FakeStackValue =
3675 RTCI.createRuntimeCall(IRB&: IRBIf, Callee: AsanStackMallocFunc[StackMallocIdx],
3676 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3677 IRB.SetInsertPoint(InsBefore);
3678 FakeStackInt = createPHI(IRB, Cond: UseAfterReturnIsEnabled, ValueIfTrue: FakeStackValue,
3679 ThenTerm: Term, ValueIfFalse: ConstantInt::get(Ty: IntptrTy, V: 0));
3680 } else {
3681 // assert(ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode:Always)
3682 // void *FakeStack = __asan_stack_malloc_N(LocalStackSize);
3683 // void *LocalStackBase = (FakeStack) ? FakeStack :
3684 // alloca(LocalStackSize);
3685 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3686 FakeStackInt =
3687 RTCI.createRuntimeCall(IRB, Callee: AsanStackMallocFunc[StackMallocIdx],
3688 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3689 }
3690 FakeStackPtr = IRB.CreateIntToPtr(V: FakeStackInt, DestTy: PtrTy);
3691 Value *NoFakeStack =
3692 IRB.CreateICmpEQ(LHS: FakeStackInt, RHS: Constant::getNullValue(Ty: IntptrTy));
3693 Instruction *Term =
3694 SplitBlockAndInsertIfThen(Cond: NoFakeStack, SplitBefore: InsBefore, Unreachable: false);
3695 IRBuilder<> IRBIf(Term);
3696 Value *AllocaValue =
3697 DoDynamicAlloca ? createAllocaForLayout(IRB&: IRBIf, L, Dynamic: true) : StaticAlloca;
3698
3699 IRB.SetInsertPoint(InsBefore);
3700 LocalStackBase =
3701 createPHI(IRB, Cond: NoFakeStack, ValueIfTrue: AllocaValue, ThenTerm: Term, ValueIfFalse: FakeStackPtr);
3702 IRB.CreateStore(Val: LocalStackBase, Ptr: LocalStackBaseAlloca);
3703 DIExprFlags |= DIExpression::DerefBefore;
3704 } else {
3705 // void *FakeStack = nullptr;
3706 // void *LocalStackBase = alloca(LocalStackSize);
3707 FakeStackInt = Constant::getNullValue(Ty: IntptrTy);
3708 FakeStackPtr = Constant::getNullValue(Ty: PtrTy);
3709 LocalStackBase =
3710 DoDynamicAlloca ? createAllocaForLayout(IRB, L, Dynamic: true) : StaticAlloca;
3711 LocalStackBaseAlloca = LocalStackBase;
3712 }
3713
3714 // Replace Alloca instructions with base+offset.
3715 SmallVector<Value *> NewAllocaPtrs;
3716 for (const auto &Desc : SVD) {
3717 AllocaInst *AI = Desc.AI;
3718 replaceDbgDeclare(Address: AI, NewAddress: LocalStackBaseAlloca, Builder&: DIB, DIExprFlags, Offset: Desc.Offset);
3719 Value *NewAllocaPtr = IRB.CreatePtrAdd(
3720 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: Desc.Offset));
3721 AI->replaceAllUsesWith(V: NewAllocaPtr);
3722 NewAllocaPtrs.push_back(Elt: NewAllocaPtr);
3723 }
3724
3725 // The left-most redzone has enough space for at least 4 pointers.
3726 // Write the Magic value to redzone[0].
3727 IRB.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kCurrentStackFrameMagic),
3728 Ptr: LocalStackBase);
3729 // Write the frame description constant to redzone[1].
3730 Value *BasePlus1 = IRB.CreatePtrAdd(
3731 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: ASan.LongSize / 8));
3732 GlobalVariable *StackDescriptionGlobal =
3733 createPrivateGlobalForString(M&: *F.getParent(), Str: DescriptionString,
3734 /*AllowMerging*/ true, NamePrefix: genName(suffix: "stack"));
3735 Value *Description = IRB.CreatePointerCast(V: StackDescriptionGlobal, DestTy: IntptrTy);
3736 IRB.CreateStore(Val: Description, Ptr: BasePlus1);
3737 // Write the PC to redzone[2].
3738 Value *BasePlus2 = IRB.CreatePtrAdd(
3739 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: 2 * ASan.LongSize / 8));
3740 IRB.CreateStore(Val: IRB.CreatePointerCast(V: &F, DestTy: IntptrTy), Ptr: BasePlus2);
3741
3742 const auto &ShadowAfterScope = GetShadowBytesAfterScope(Vars: SVD, Layout: L);
3743
3744 // Poison the stack red zones at the entry.
3745 Value *ShadowBase =
3746 ASan.memToShadow(Shadow: IRB.CreatePtrToInt(V: LocalStackBase, DestTy: IntptrTy), IRB);
3747 // As mask we must use most poisoned case: red zones and after scope.
3748 // As bytes we can use either the same or just red zones only.
3749 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowAfterScope, IRB, ShadowBase);
3750
3751 if (!StaticAllocaPoisonCallVec.empty()) {
3752 const auto &ShadowInScope = GetShadowBytes(Vars: SVD, Layout: L);
3753
3754 // Poison static allocas near lifetime intrinsics.
3755 for (const auto &APC : StaticAllocaPoisonCallVec) {
3756 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3757 assert(Desc.Offset % L.Granularity == 0);
3758 size_t Begin = Desc.Offset / L.Granularity;
3759 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3760
3761 IRBuilder<> IRB(APC.InsBefore);
3762 copyToShadow(ShadowMask: ShadowAfterScope,
3763 ShadowBytes: APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3764 IRB, ShadowBase);
3765 }
3766 }
3767
3768 // Remove lifetime markers now that these are no longer allocas.
3769 for (Value *NewAllocaPtr : NewAllocaPtrs) {
3770 for (User *U : make_early_inc_range(Range: NewAllocaPtr->users())) {
3771 auto *I = cast<Instruction>(Val: U);
3772 if (I->isLifetimeStartOrEnd())
3773 I->eraseFromParent();
3774 }
3775 }
3776
3777 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3778 SmallVector<uint8_t, 64> ShadowAfterReturn;
3779
3780 // (Un)poison the stack before all ret instructions.
3781 for (Instruction *Ret : RetVec) {
3782 IRBuilder<> IRBRet(Ret);
3783 // Mark the current frame as retired.
3784 IRBRet.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kRetiredStackFrameMagic),
3785 Ptr: LocalStackBase);
3786 if (DoStackMalloc) {
3787 assert(StackMallocIdx >= 0);
3788 // if FakeStack != 0 // LocalStackBase == FakeStack
3789 // // In use-after-return mode, poison the whole stack frame.
3790 // if StackMallocIdx <= 4
3791 // // For small sizes inline the whole thing:
3792 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3793 // **SavedFlagPtr(FakeStack) = 0
3794 // else
3795 // __asan_stack_free_N(FakeStack, LocalStackSize)
3796 // else
3797 // <This is not a fake stack; unpoison the redzones>
3798 Value *Cmp =
3799 IRBRet.CreateICmpNE(LHS: FakeStackInt, RHS: Constant::getNullValue(Ty: IntptrTy));
3800 Instruction *ThenTerm, *ElseTerm;
3801 SplitBlockAndInsertIfThenElse(Cond: Cmp, SplitBefore: Ret, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm);
3802
3803 IRBuilder<> IRBPoison(ThenTerm);
3804 if (ASan.MaxInlinePoisoningSize != 0 && StackMallocIdx <= 4) {
3805 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3806 ShadowAfterReturn.resize(N: ClassSize / L.Granularity,
3807 NV: kAsanStackUseAfterReturnMagic);
3808 copyToShadow(ShadowMask: ShadowAfterReturn, ShadowBytes: ShadowAfterReturn, IRB&: IRBPoison,
3809 ShadowBase);
3810 Value *SavedFlagPtrPtr = IRBPoison.CreatePtrAdd(
3811 Ptr: FakeStackPtr,
3812 Offset: ConstantInt::get(Ty: IntptrTy, V: ClassSize - ASan.LongSize / 8));
3813 Value *SavedFlagPtr = IRBPoison.CreateLoad(Ty: IntptrTy, Ptr: SavedFlagPtrPtr);
3814 IRBPoison.CreateStore(
3815 Val: Constant::getNullValue(Ty: IRBPoison.getInt8Ty()),
3816 Ptr: IRBPoison.CreateIntToPtr(V: SavedFlagPtr, DestTy: IRBPoison.getPtrTy()));
3817 } else {
3818 // For larger frames call __asan_stack_free_*.
3819 RTCI.createRuntimeCall(
3820 IRB&: IRBPoison, Callee: AsanStackFreeFunc[StackMallocIdx],
3821 Args: {FakeStackInt, ConstantInt::get(Ty: IntptrTy, V: LocalStackSize)});
3822 }
3823
3824 IRBuilder<> IRBElse(ElseTerm);
3825 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBElse, ShadowBase);
3826 } else {
3827 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBRet, ShadowBase);
3828 }
3829 }
3830
3831 // We are done. Remove the old unused alloca instructions.
3832 for (auto *AI : AllocaVec)
3833 AI->eraseFromParent();
3834}
3835
3836void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3837 IRBuilder<> &IRB, bool DoPoison) {
3838 // For now just insert the call to ASan runtime.
3839 Value *AddrArg = IRB.CreatePointerCast(V, DestTy: IntptrTy);
3840 Value *SizeArg = ConstantInt::get(Ty: IntptrTy, V: Size);
3841 RTCI.createRuntimeCall(
3842 IRB, Callee: DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3843 Args: {AddrArg, SizeArg});
3844}
3845
3846// Handling llvm.lifetime intrinsics for a given %alloca:
3847// (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3848// (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3849// invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3850// could be poisoned by previous llvm.lifetime.end instruction, as the
3851// variable may go in and out of scope several times, e.g. in loops).
3852// (3) if we poisoned at least one %alloca in a function,
3853// unpoison the whole stack frame at function exit.
3854void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3855 IRBuilder<> IRB(AI);
3856
3857 const Align Alignment = std::max(a: Align(kAllocaRzSize), b: AI->getAlign());
3858 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3859
3860 Value *Zero = Constant::getNullValue(Ty: IntptrTy);
3861 Value *AllocaRzSize = ConstantInt::get(Ty: IntptrTy, V: kAllocaRzSize);
3862 Value *AllocaRzMask = ConstantInt::get(Ty: IntptrTy, V: AllocaRedzoneMask);
3863
3864 // Since we need to extend alloca with additional memory to locate
3865 // redzones, and OldSize is number of allocated blocks with
3866 // ElementSize size, get allocated memory size in bytes by
3867 // OldSize * ElementSize.
3868 Value *OldSize = IRB.CreateAllocationSize(DestTy: IntptrTy, AI);
3869
3870 // PartialSize = OldSize % 32
3871 Value *PartialSize = IRB.CreateAnd(LHS: OldSize, RHS: AllocaRzMask);
3872
3873 // Misalign = kAllocaRzSize - PartialSize;
3874 Value *Misalign = IRB.CreateSub(LHS: AllocaRzSize, RHS: PartialSize);
3875
3876 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3877 Value *Cond = IRB.CreateICmpNE(LHS: Misalign, RHS: AllocaRzSize);
3878 Value *PartialPadding = IRB.CreateSelect(C: Cond, True: Misalign, False: Zero);
3879
3880 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3881 // Alignment is added to locate left redzone, PartialPadding for possible
3882 // partial redzone and kAllocaRzSize for right redzone respectively.
3883 Value *AdditionalChunkSize = IRB.CreateAdd(
3884 LHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value() + kAllocaRzSize),
3885 RHS: PartialPadding);
3886
3887 Value *NewSize = IRB.CreateAdd(LHS: OldSize, RHS: AdditionalChunkSize);
3888
3889 // Insert new alloca with new NewSize and Alignment params.
3890 AllocaInst *NewAlloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(), ArraySize: NewSize);
3891 NewAlloca->setAlignment(Alignment);
3892
3893 // NewAddress = Address + Alignment
3894 Value *NewAddress =
3895 IRB.CreateAdd(LHS: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy),
3896 RHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value()));
3897
3898 // Insert __asan_alloca_poison call for new created alloca.
3899 RTCI.createRuntimeCall(IRB, Callee: AsanAllocaPoisonFunc, Args: {NewAddress, OldSize});
3900
3901 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3902 // for unpoisoning stuff.
3903 IRB.CreateStore(Val: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy), Ptr: DynamicAllocaLayout);
3904
3905 Value *NewAddressPtr = IRB.CreateIntToPtr(V: NewAddress, DestTy: AI->getType());
3906
3907 // Remove lifetime markers now that this is no longer an alloca.
3908 for (User *U : make_early_inc_range(Range: AI->users())) {
3909 auto *I = cast<Instruction>(Val: U);
3910 if (I->isLifetimeStartOrEnd())
3911 I->eraseFromParent();
3912 }
3913
3914 // Replace all uses of AddressReturnedByAlloca with NewAddressPtr.
3915 AI->replaceAllUsesWith(V: NewAddressPtr);
3916
3917 // We are done. Erase old alloca from parent.
3918 AI->eraseFromParent();
3919}
3920
3921// isSafeAccess returns true if Addr is always inbounds with respect to its
3922// base object. For example, it is a field access or an array access with
3923// constant inbounds index.
3924bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3925 Value *Addr, TypeSize TypeStoreSize) const {
3926 if (TypeStoreSize.isScalable())
3927 // TODO: We can use vscale_range to convert a scalable value to an
3928 // upper bound on the access size.
3929 return false;
3930
3931 SizeOffsetAPInt SizeOffset = ObjSizeVis.compute(V: Addr);
3932 if (!SizeOffset.bothKnown())
3933 return false;
3934
3935 uint64_t Size = SizeOffset.Size.getZExtValue();
3936 int64_t Offset = SizeOffset.Offset.getSExtValue();
3937
3938 // Three checks are required to ensure safety:
3939 // . Offset >= 0 (since the offset is given from the base ptr)
3940 // . Size >= Offset (unsigned)
3941 // . Size - Offset >= NeededSize (unsigned)
3942 return Offset >= 0 && Size >= uint64_t(Offset) &&
3943 Size - uint64_t(Offset) >= TypeStoreSize / 8;
3944}
3945