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 if (TargetTriple.isOSDarwin() &&
1429 TargetTriple.getArch() == llvm::Triple::aarch64) {
1430 // Strip MTE-tag bits before translating to shadow address
1431 Shadow = IRB.CreateAnd(LHS: Shadow,
1432 RHS: ConstantInt::get(Ty: IntptrTy, V: ~(uint64_t(0x0f) << 56)));
1433 }
1434 // Shadow >> scale
1435 Shadow = IRB.CreateLShr(LHS: Shadow, RHS: Mapping.Scale);
1436 if (Mapping.Offset == 0) return Shadow;
1437 // (Shadow >> scale) | offset
1438 Value *ShadowBase;
1439 if (LocalDynamicShadow)
1440 ShadowBase = LocalDynamicShadow;
1441 else
1442 ShadowBase = ConstantInt::get(Ty: IntptrTy, V: Mapping.Offset);
1443 if (Mapping.OrShadowOffset)
1444 return IRB.CreateOr(LHS: Shadow, RHS: ShadowBase);
1445 else
1446 return IRB.CreateAdd(LHS: Shadow, RHS: ShadowBase);
1447}
1448
1449// Instrument memset/memmove/memcpy
1450void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI,
1451 RuntimeCallInserter &RTCI) {
1452 InstrumentationIRBuilder IRB(MI);
1453 if (isa<MemTransferInst>(Val: MI)) {
1454 RTCI.createRuntimeCall(
1455 IRB, Callee: isa<MemMoveInst>(Val: MI) ? AsanMemmove : AsanMemcpy,
1456 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1457 IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 1), DestTy: PtrTy),
1458 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1459 } else if (isa<MemSetInst>(Val: MI)) {
1460 RTCI.createRuntimeCall(
1461 IRB, Callee: AsanMemset,
1462 Args: {IRB.CreateAddrSpaceCast(V: MI->getOperand(i_nocapture: 0), DestTy: PtrTy),
1463 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 1), DestTy: IRB.getInt32Ty(), isSigned: false),
1464 IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)});
1465 }
1466 MI->eraseFromParent();
1467}
1468
1469/// Check if we want (and can) handle this alloca.
1470bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1471 auto [It, Inserted] = ProcessedAllocas.try_emplace(Key: &AI);
1472
1473 if (!Inserted)
1474 return It->getSecond();
1475
1476 bool IsInteresting =
1477 (AI.getAllocatedType()->isSized() &&
1478 // alloca() may be called with 0 size, ignore it.
1479 ((!AI.isStaticAlloca()) || !getAllocaSizeInBytes(AI).isZero()) &&
1480 // We are only interested in allocas not promotable to registers.
1481 // Promotable allocas are common under -O0.
1482 (!ClSkipPromotableAllocas || !isAllocaPromotable(AI: &AI)) &&
1483 // inalloca allocas are not treated as static, and we don't want
1484 // dynamic alloca instrumentation for them as well.
1485 !AI.isUsedWithInAlloca() &&
1486 // swifterror allocas are register promoted by ISel
1487 !AI.isSwiftError() &&
1488 // safe allocas are not interesting
1489 !(SSGI && SSGI->isSafe(AI)));
1490
1491 It->second = IsInteresting;
1492 return IsInteresting;
1493}
1494
1495bool AddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) {
1496 // Check whether the target supports sanitizing the address space
1497 // of the pointer.
1498 if (!isSupportedAddrspace(TargetTriple, Addr: Ptr))
1499 return true;
1500
1501 // Ignore swifterror addresses.
1502 // swifterror memory addresses are mem2reg promoted by instruction
1503 // selection. As such they cannot have regular uses like an instrumentation
1504 // function and it makes no sense to track them as memory.
1505 if (Ptr->isSwiftError())
1506 return true;
1507
1508 // Treat memory accesses to promotable allocas as non-interesting since they
1509 // will not cause memory violations. This greatly speeds up the instrumented
1510 // executable at -O0.
1511 if (auto AI = dyn_cast_or_null<AllocaInst>(Val: Ptr))
1512 if (ClSkipPromotableAllocas && !isInterestingAlloca(AI: *AI))
1513 return true;
1514
1515 if (SSGI != nullptr && SSGI->stackAccessIsSafe(I: *Inst) &&
1516 findAllocaForValue(V: Ptr))
1517 return true;
1518
1519 return false;
1520}
1521
1522void AddressSanitizer::getInterestingMemoryOperands(
1523 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting,
1524 const TargetTransformInfo *TTI) {
1525 // Do not instrument the load fetching the dynamic shadow address.
1526 if (LocalDynamicShadow == I)
1527 return;
1528
1529 if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) {
1530 if (!ClInstrumentReads || ignoreAccess(Inst: I, Ptr: LI->getPointerOperand()))
1531 return;
1532 Interesting.emplace_back(Args&: I, Args: LI->getPointerOperandIndex(), Args: false,
1533 Args: LI->getType(), Args: LI->getAlign());
1534 } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) {
1535 if (!ClInstrumentWrites || ignoreAccess(Inst: I, Ptr: SI->getPointerOperand()))
1536 return;
1537 Interesting.emplace_back(Args&: I, Args: SI->getPointerOperandIndex(), Args: true,
1538 Args: SI->getValueOperand()->getType(), Args: SI->getAlign());
1539 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) {
1540 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: RMW->getPointerOperand()))
1541 return;
1542 Interesting.emplace_back(Args&: I, Args: RMW->getPointerOperandIndex(), Args: true,
1543 Args: RMW->getValOperand()->getType(), Args: std::nullopt);
1544 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(Val: I)) {
1545 if (!ClInstrumentAtomics || ignoreAccess(Inst: I, Ptr: XCHG->getPointerOperand()))
1546 return;
1547 Interesting.emplace_back(Args&: I, Args: XCHG->getPointerOperandIndex(), Args: true,
1548 Args: XCHG->getCompareOperand()->getType(),
1549 Args: std::nullopt);
1550 } else if (auto CI = dyn_cast<CallInst>(Val: I)) {
1551 switch (CI->getIntrinsicID()) {
1552 case Intrinsic::masked_load:
1553 case Intrinsic::masked_store:
1554 case Intrinsic::masked_gather:
1555 case Intrinsic::masked_scatter: {
1556 bool IsWrite = CI->getType()->isVoidTy();
1557 // Masked store has an initial operand for the value.
1558 unsigned OpOffset = IsWrite ? 1 : 0;
1559 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1560 return;
1561
1562 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1563 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1564 return;
1565 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1566 MaybeAlign Alignment = CI->getParamAlign(ArgNo: 0);
1567 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1568 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: Mask);
1569 break;
1570 }
1571 case Intrinsic::masked_expandload:
1572 case Intrinsic::masked_compressstore: {
1573 bool IsWrite = CI->getIntrinsicID() == Intrinsic::masked_compressstore;
1574 unsigned OpOffset = IsWrite ? 1 : 0;
1575 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1576 return;
1577 auto BasePtr = CI->getOperand(i_nocapture: OpOffset);
1578 if (ignoreAccess(Inst: I, Ptr: BasePtr))
1579 return;
1580 MaybeAlign Alignment = BasePtr->getPointerAlignment(DL: *DL);
1581 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1582
1583 IRBuilder IB(I);
1584 Value *Mask = CI->getOperand(i_nocapture: 1 + OpOffset);
1585 // Use the popcount of Mask as the effective vector length.
1586 Type *ExtTy = VectorType::get(ElementType: IntptrTy, Other: cast<VectorType>(Val: Ty));
1587 Value *ExtMask = IB.CreateZExt(V: Mask, DestTy: ExtTy);
1588 Value *EVL = IB.CreateAddReduce(Src: ExtMask);
1589 Value *TrueMask = ConstantInt::get(Ty: Mask->getType(), V: 1);
1590 Interesting.emplace_back(Args&: I, Args&: OpOffset, Args&: IsWrite, Args&: Ty, Args&: Alignment, Args&: TrueMask,
1591 Args&: EVL);
1592 break;
1593 }
1594 case Intrinsic::vp_load:
1595 case Intrinsic::vp_store:
1596 case Intrinsic::experimental_vp_strided_load:
1597 case Intrinsic::experimental_vp_strided_store: {
1598 auto *VPI = cast<VPIntrinsic>(Val: CI);
1599 unsigned IID = CI->getIntrinsicID();
1600 bool IsWrite = CI->getType()->isVoidTy();
1601 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1602 return;
1603 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1604 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1605 MaybeAlign Alignment = VPI->getOperand(i_nocapture: PtrOpNo)->getPointerAlignment(DL: *DL);
1606 Value *Stride = nullptr;
1607 if (IID == Intrinsic::experimental_vp_strided_store ||
1608 IID == Intrinsic::experimental_vp_strided_load) {
1609 Stride = VPI->getOperand(i_nocapture: PtrOpNo + 1);
1610 // Use the pointer alignment as the element alignment if the stride is a
1611 // multiple of the pointer alignment. Otherwise, the element alignment
1612 // should be Align(1).
1613 unsigned PointerAlign = Alignment.valueOrOne().value();
1614 if (!isa<ConstantInt>(Val: Stride) ||
1615 cast<ConstantInt>(Val: Stride)->getZExtValue() % PointerAlign != 0)
1616 Alignment = Align(1);
1617 }
1618 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1619 Args: VPI->getMaskParam(), Args: VPI->getVectorLengthParam(),
1620 Args&: Stride);
1621 break;
1622 }
1623 case Intrinsic::vp_gather:
1624 case Intrinsic::vp_scatter: {
1625 auto *VPI = cast<VPIntrinsic>(Val: CI);
1626 unsigned IID = CI->getIntrinsicID();
1627 bool IsWrite = IID == Intrinsic::vp_scatter;
1628 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1629 return;
1630 unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
1631 Type *Ty = IsWrite ? CI->getArgOperand(i: 0)->getType() : CI->getType();
1632 MaybeAlign Alignment = VPI->getPointerAlignment();
1633 Interesting.emplace_back(Args&: I, Args&: PtrOpNo, Args&: IsWrite, Args&: Ty, Args&: Alignment,
1634 Args: VPI->getMaskParam(),
1635 Args: VPI->getVectorLengthParam());
1636 break;
1637 }
1638 default:
1639 if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) {
1640 MemIntrinsicInfo IntrInfo;
1641 if (TTI->getTgtMemIntrinsic(Inst: II, Info&: IntrInfo))
1642 Interesting = IntrInfo.InterestingOperands;
1643 return;
1644 }
1645 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) {
1646 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1647 ignoreAccess(Inst: I, Ptr: CI->getArgOperand(i: ArgNo)))
1648 continue;
1649 Type *Ty = CI->getParamByValType(ArgNo);
1650 Interesting.emplace_back(Args&: I, Args&: ArgNo, Args: false, Args&: Ty, Args: Align(1));
1651 }
1652 }
1653 }
1654}
1655
1656static bool isPointerOperand(Value *V) {
1657 return V->getType()->isPointerTy() || isa<PtrToIntInst>(Val: V);
1658}
1659
1660// This is a rough heuristic; it may cause both false positives and
1661// false negatives. The proper implementation requires cooperation with
1662// the frontend.
1663static bool isInterestingPointerComparison(Instruction *I) {
1664 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: I)) {
1665 if (!Cmp->isRelational())
1666 return false;
1667 } else {
1668 return false;
1669 }
1670 return isPointerOperand(V: I->getOperand(i: 0)) &&
1671 isPointerOperand(V: I->getOperand(i: 1));
1672}
1673
1674// This is a rough heuristic; it may cause both false positives and
1675// false negatives. The proper implementation requires cooperation with
1676// the frontend.
1677static bool isInterestingPointerSubtraction(Instruction *I) {
1678 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) {
1679 if (BO->getOpcode() != Instruction::Sub)
1680 return false;
1681 } else {
1682 return false;
1683 }
1684 return isPointerOperand(V: I->getOperand(i: 0)) &&
1685 isPointerOperand(V: I->getOperand(i: 1));
1686}
1687
1688bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1689 // If a global variable does not have dynamic initialization we don't
1690 // have to instrument it. However, if a global does not have initializer
1691 // at all, we assume it has dynamic initializer (in other TU).
1692 if (!G->hasInitializer())
1693 return false;
1694
1695 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().IsDynInit)
1696 return false;
1697
1698 return true;
1699}
1700
1701void AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1702 Instruction *I, RuntimeCallInserter &RTCI) {
1703 IRBuilder<> IRB(I);
1704 FunctionCallee F = isa<ICmpInst>(Val: I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1705 Value *Param[2] = {I->getOperand(i: 0), I->getOperand(i: 1)};
1706 for (Value *&i : Param) {
1707 if (i->getType()->isPointerTy())
1708 i = IRB.CreatePointerCast(V: i, DestTy: IntptrTy);
1709 }
1710 RTCI.createRuntimeCall(IRB, Callee: F, Args: Param);
1711}
1712
1713static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1714 Instruction *InsertBefore, Value *Addr,
1715 MaybeAlign Alignment, unsigned Granularity,
1716 TypeSize TypeStoreSize, bool IsWrite,
1717 Value *SizeArgument, bool UseCalls,
1718 uint32_t Exp, RuntimeCallInserter &RTCI) {
1719 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1720 // if the data is properly aligned.
1721 if (!TypeStoreSize.isScalable()) {
1722 const auto FixedSize = TypeStoreSize.getFixedValue();
1723 switch (FixedSize) {
1724 case 8:
1725 case 16:
1726 case 32:
1727 case 64:
1728 case 128:
1729 if (!Alignment || *Alignment >= Granularity ||
1730 *Alignment >= FixedSize / 8)
1731 return Pass->instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment,
1732 TypeStoreSize: FixedSize, IsWrite, SizeArgument: nullptr, UseCalls,
1733 Exp, RTCI);
1734 }
1735 }
1736 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeStoreSize,
1737 IsWrite, SizeArgument: nullptr, UseCalls, Exp, RTCI);
1738}
1739
1740void AddressSanitizer::instrumentMaskedLoadOrStore(
1741 AddressSanitizer *Pass, const DataLayout &DL, Type *IntptrTy, Value *Mask,
1742 Value *EVL, Value *Stride, Instruction *I, Value *Addr,
1743 MaybeAlign Alignment, unsigned Granularity, Type *OpType, bool IsWrite,
1744 Value *SizeArgument, bool UseCalls, uint32_t Exp,
1745 RuntimeCallInserter &RTCI) {
1746 auto *VTy = cast<VectorType>(Val: OpType);
1747 TypeSize ElemTypeSize = DL.getTypeStoreSizeInBits(Ty: VTy->getScalarType());
1748 auto Zero = ConstantInt::get(Ty: IntptrTy, V: 0);
1749
1750 IRBuilder IB(I);
1751 Instruction *LoopInsertBefore = I;
1752 if (EVL) {
1753 // The end argument of SplitBlockAndInsertForLane is assumed bigger
1754 // than zero, so we should check whether EVL is zero here.
1755 Type *EVLType = EVL->getType();
1756 Value *IsEVLZero = IB.CreateICmpNE(LHS: EVL, RHS: ConstantInt::get(Ty: EVLType, V: 0));
1757 LoopInsertBefore = SplitBlockAndInsertIfThen(Cond: IsEVLZero, SplitBefore: I, Unreachable: false);
1758 IB.SetInsertPoint(LoopInsertBefore);
1759 // Cast EVL to IntptrTy.
1760 EVL = IB.CreateZExtOrTrunc(V: EVL, DestTy: IntptrTy);
1761 // To avoid undefined behavior for extracting with out of range index, use
1762 // the minimum of evl and element count as trip count.
1763 Value *EC = IB.CreateElementCount(Ty: IntptrTy, EC: VTy->getElementCount());
1764 EVL = IB.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: EVL, RHS: EC);
1765 } else {
1766 EVL = IB.CreateElementCount(Ty: IntptrTy, EC: VTy->getElementCount());
1767 }
1768
1769 // Cast Stride to IntptrTy.
1770 if (Stride)
1771 Stride = IB.CreateZExtOrTrunc(V: Stride, DestTy: IntptrTy);
1772
1773 SplitBlockAndInsertForEachLane(End: EVL, InsertBefore: LoopInsertBefore->getIterator(),
1774 Func: [&](IRBuilderBase &IRB, Value *Index) {
1775 Value *MaskElem = IRB.CreateExtractElement(Vec: Mask, Idx: Index);
1776 if (auto *MaskElemC = dyn_cast<ConstantInt>(Val: MaskElem)) {
1777 if (MaskElemC->isZero())
1778 // No check
1779 return;
1780 // Unconditional check
1781 } else {
1782 // Conditional check
1783 Instruction *ThenTerm = SplitBlockAndInsertIfThen(
1784 Cond: MaskElem, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false);
1785 IRB.SetInsertPoint(ThenTerm);
1786 }
1787
1788 Value *InstrumentedAddress;
1789 if (isa<VectorType>(Val: Addr->getType())) {
1790 assert(
1791 cast<VectorType>(Addr->getType())->getElementType()->isPointerTy() &&
1792 "Expected vector of pointer.");
1793 InstrumentedAddress = IRB.CreateExtractElement(Vec: Addr, Idx: Index);
1794 } else if (Stride) {
1795 Index = IRB.CreateMul(LHS: Index, RHS: Stride);
1796 InstrumentedAddress = IRB.CreatePtrAdd(Ptr: Addr, Offset: Index);
1797 } else {
1798 InstrumentedAddress = IRB.CreateGEP(Ty: VTy, Ptr: Addr, IdxList: {Zero, Index});
1799 }
1800 doInstrumentAddress(Pass, I, InsertBefore: &*IRB.GetInsertPoint(), Addr: InstrumentedAddress,
1801 Alignment, Granularity, TypeStoreSize: ElemTypeSize, IsWrite,
1802 SizeArgument, UseCalls, Exp, RTCI);
1803 });
1804}
1805
1806void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1807 InterestingMemoryOperand &O, bool UseCalls,
1808 const DataLayout &DL,
1809 RuntimeCallInserter &RTCI) {
1810 Value *Addr = O.getPtr();
1811
1812 // Optimization experiments.
1813 // The experiments can be used to evaluate potential optimizations that remove
1814 // instrumentation (assess false negatives). Instead of completely removing
1815 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1816 // experiments that want to remove instrumentation of this instruction).
1817 // If Exp is non-zero, this pass will emit special calls into runtime
1818 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1819 // make runtime terminate the program in a special way (with a different
1820 // exit status). Then you run the new compiler on a buggy corpus, collect
1821 // the special terminations (ideally, you don't see them at all -- no false
1822 // negatives) and make the decision on the optimization.
1823 uint32_t Exp = ClForceExperiment;
1824
1825 if (ClOpt && ClOptGlobals) {
1826 // If initialization order checking is disabled, a simple access to a
1827 // dynamically initialized global is always valid.
1828 GlobalVariable *G = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: Addr));
1829 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1830 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1831 NumOptimizedAccessesToGlobalVar++;
1832 return;
1833 }
1834 }
1835
1836 if (ClOpt && ClOptStack) {
1837 // A direct inbounds access to a stack variable is always valid.
1838 if (isa<AllocaInst>(Val: getUnderlyingObject(V: Addr)) &&
1839 isSafeAccess(ObjSizeVis, Addr, TypeStoreSize: O.TypeStoreSize)) {
1840 NumOptimizedAccessesToStackVar++;
1841 return;
1842 }
1843 }
1844
1845 if (O.IsWrite)
1846 NumInstrumentedWrites++;
1847 else
1848 NumInstrumentedReads++;
1849
1850 if (O.MaybeByteOffset) {
1851 Type *Ty = Type::getInt8Ty(C&: *C);
1852 IRBuilder IB(O.getInsn());
1853
1854 Value *OffsetOp = O.MaybeByteOffset;
1855 if (TargetTriple.isRISCV()) {
1856 Type *OffsetTy = OffsetOp->getType();
1857 // RVV indexed loads/stores zero-extend offset operands which are narrower
1858 // than XLEN to XLEN.
1859 if (OffsetTy->getScalarType()->getIntegerBitWidth() <
1860 static_cast<unsigned>(LongSize)) {
1861 VectorType *OrigType = cast<VectorType>(Val: OffsetTy);
1862 Type *ExtendTy = VectorType::get(ElementType: IntptrTy, Other: OrigType);
1863 OffsetOp = IB.CreateZExt(V: OffsetOp, DestTy: ExtendTy);
1864 }
1865 }
1866 Addr = IB.CreateGEP(Ty, Ptr: Addr, IdxList: {OffsetOp});
1867 }
1868
1869 unsigned Granularity = 1 << Mapping.Scale;
1870 if (O.MaybeMask) {
1871 instrumentMaskedLoadOrStore(Pass: this, DL, IntptrTy, Mask: O.MaybeMask, EVL: O.MaybeEVL,
1872 Stride: O.MaybeStride, I: O.getInsn(), Addr, Alignment: O.Alignment,
1873 Granularity, OpType: O.OpType, IsWrite: O.IsWrite, SizeArgument: nullptr,
1874 UseCalls, Exp, RTCI);
1875 } else {
1876 doInstrumentAddress(Pass: this, I: O.getInsn(), InsertBefore: O.getInsn(), Addr, Alignment: O.Alignment,
1877 Granularity, TypeStoreSize: O.TypeStoreSize, IsWrite: O.IsWrite, SizeArgument: nullptr,
1878 UseCalls, Exp, RTCI);
1879 }
1880}
1881
1882Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1883 Value *Addr, bool IsWrite,
1884 size_t AccessSizeIndex,
1885 Value *SizeArgument,
1886 uint32_t Exp,
1887 RuntimeCallInserter &RTCI) {
1888 InstrumentationIRBuilder IRB(InsertBefore);
1889 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp);
1890 CallInst *Call = nullptr;
1891 if (SizeArgument) {
1892 if (Exp == 0)
1893 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][0],
1894 Args: {Addr, SizeArgument});
1895 else
1896 Call = RTCI.createRuntimeCall(IRB, Callee: AsanErrorCallbackSized[IsWrite][1],
1897 Args: {Addr, SizeArgument, ExpVal});
1898 } else {
1899 if (Exp == 0)
1900 Call = RTCI.createRuntimeCall(
1901 IRB, Callee: AsanErrorCallback[IsWrite][0][AccessSizeIndex], Args: Addr);
1902 else
1903 Call = RTCI.createRuntimeCall(
1904 IRB, Callee: AsanErrorCallback[IsWrite][1][AccessSizeIndex], Args: {Addr, ExpVal});
1905 }
1906
1907 Call->setCannotMerge();
1908 return Call;
1909}
1910
1911Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1912 Value *ShadowValue,
1913 uint32_t TypeStoreSize) {
1914 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1915 // Addr & (Granularity - 1)
1916 Value *LastAccessedByte =
1917 IRB.CreateAnd(LHS: AddrLong, RHS: ConstantInt::get(Ty: IntptrTy, V: Granularity - 1));
1918 // (Addr & (Granularity - 1)) + size - 1
1919 if (TypeStoreSize / 8 > 1)
1920 LastAccessedByte = IRB.CreateAdd(
1921 LHS: LastAccessedByte, RHS: ConstantInt::get(Ty: IntptrTy, V: TypeStoreSize / 8 - 1));
1922 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1923 LastAccessedByte =
1924 IRB.CreateIntCast(V: LastAccessedByte, DestTy: ShadowValue->getType(), isSigned: false);
1925 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1926 return IRB.CreateICmpSGE(LHS: LastAccessedByte, RHS: ShadowValue);
1927}
1928
1929Instruction *AddressSanitizer::instrumentAMDGPUAddress(
1930 Instruction *OrigIns, Instruction *InsertBefore, Value *Addr,
1931 uint32_t TypeStoreSize, bool IsWrite, Value *SizeArgument) {
1932 // Do not instrument unsupported addrspaces.
1933 if (isUnsupportedAMDGPUAddrspace(Addr))
1934 return nullptr;
1935 Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType());
1936 // Follow host instrumentation for global and constant addresses.
1937 if (PtrTy->getPointerAddressSpace() != 0)
1938 return InsertBefore;
1939 // Instrument generic addresses in supported addressspaces.
1940 IRBuilder<> IRB(InsertBefore);
1941 Value *IsShared = IRB.CreateCall(Callee: AMDGPUAddressShared, Args: {Addr});
1942 Value *IsPrivate = IRB.CreateCall(Callee: AMDGPUAddressPrivate, Args: {Addr});
1943 Value *IsSharedOrPrivate = IRB.CreateOr(LHS: IsShared, RHS: IsPrivate);
1944 Value *Cmp = IRB.CreateNot(V: IsSharedOrPrivate);
1945 Value *AddrSpaceZeroLanding =
1946 SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false);
1947 InsertBefore = cast<Instruction>(Val: AddrSpaceZeroLanding);
1948 return InsertBefore;
1949}
1950
1951Instruction *AddressSanitizer::genAMDGPUReportBlock(IRBuilder<> &IRB,
1952 Value *Cond, bool Recover) {
1953 Value *ReportCond = Cond;
1954 if (!Recover) {
1955 auto Ballot = Inserter.insertFunction(Name: kAMDGPUBallotName, Args: IRB.getInt64Ty(),
1956 Args: IRB.getInt1Ty());
1957 ReportCond = IRB.CreateIsNotNull(Arg: IRB.CreateCall(Callee: Ballot, Args: {Cond}));
1958 }
1959
1960 auto *Trm =
1961 SplitBlockAndInsertIfThen(Cond: ReportCond, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false,
1962 BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
1963 Trm->getParent()->setName("asan.report");
1964
1965 if (Recover)
1966 return Trm;
1967
1968 Trm = SplitBlockAndInsertIfThen(Cond, SplitBefore: Trm, Unreachable: false);
1969 IRB.SetInsertPoint(Trm);
1970 return IRB.CreateCall(
1971 Callee: Inserter.insertFunction(Name: kAMDGPUUnreachableName, Args: IRB.getVoidTy()), Args: {});
1972}
1973
1974void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1975 Instruction *InsertBefore, Value *Addr,
1976 MaybeAlign Alignment,
1977 uint32_t TypeStoreSize, bool IsWrite,
1978 Value *SizeArgument, bool UseCalls,
1979 uint32_t Exp,
1980 RuntimeCallInserter &RTCI) {
1981 if (TargetTriple.isAMDGPU()) {
1982 InsertBefore = instrumentAMDGPUAddress(OrigIns, InsertBefore, Addr,
1983 TypeStoreSize, IsWrite, SizeArgument);
1984 if (!InsertBefore)
1985 return;
1986 }
1987
1988 InstrumentationIRBuilder IRB(InsertBefore);
1989 size_t AccessSizeIndex = TypeStoreSizeToSizeIndex(TypeSize: TypeStoreSize);
1990
1991 if (UseCalls && ClOptimizeCallbacks) {
1992 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
1993 IRB.CreateIntrinsic(ID: Intrinsic::asan_check_memaccess, OverloadTypes: {},
1994 Args: {IRB.CreatePointerCast(V: Addr, DestTy: PtrTy),
1995 ConstantInt::get(Ty: Int32Ty, V: AccessInfo.Packed)});
1996 return;
1997 }
1998
1999 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
2000 if (UseCalls) {
2001 if (Exp == 0)
2002 RTCI.createRuntimeCall(
2003 IRB, Callee: AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex], Args: AddrLong);
2004 else
2005 RTCI.createRuntimeCall(
2006 IRB, Callee: AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
2007 Args: {AddrLong, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
2008 return;
2009 }
2010
2011 Type *ShadowTy =
2012 IntegerType::get(C&: *C, NumBits: std::max(a: 8U, b: TypeStoreSize >> Mapping.Scale));
2013 Type *ShadowPtrTy = PointerType::get(C&: *C, AddressSpace: ClShadowAddrSpace);
2014 Value *ShadowPtr = memToShadow(Shadow: AddrLong, IRB);
2015 const uint64_t ShadowAlign =
2016 std::max<uint64_t>(a: Alignment.valueOrOne().value() >> Mapping.Scale, b: 1);
2017 Value *ShadowValue = IRB.CreateAlignedLoad(
2018 Ty: ShadowTy, Ptr: IRB.CreateIntToPtr(V: ShadowPtr, DestTy: ShadowPtrTy), Align: Align(ShadowAlign));
2019
2020 Value *Cmp = IRB.CreateIsNotNull(Arg: ShadowValue);
2021 size_t Granularity = 1ULL << Mapping.Scale;
2022 Instruction *CrashTerm = nullptr;
2023
2024 bool GenSlowPath = (ClAlwaysSlowPath || (TypeStoreSize < 8 * Granularity));
2025
2026 if (TargetTriple.isAMDGCN()) {
2027 if (GenSlowPath) {
2028 auto *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
2029 Cmp = IRB.CreateAnd(LHS: Cmp, RHS: Cmp2);
2030 }
2031 CrashTerm = genAMDGPUReportBlock(IRB, Cond: Cmp, Recover);
2032 } else if (GenSlowPath) {
2033 // We use branch weights for the slow path check, to indicate that the slow
2034 // path is rarely taken. This seems to be the case for SPEC benchmarks.
2035 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
2036 Cond: Cmp, SplitBefore: InsertBefore, Unreachable: false, BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights());
2037 BasicBlock *NextBB = cast<UncondBrInst>(Val: CheckTerm)->getSuccessor();
2038 IRB.SetInsertPoint(CheckTerm);
2039 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeStoreSize);
2040 if (Recover) {
2041 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp2, SplitBefore: CheckTerm, Unreachable: false);
2042 } else {
2043 BasicBlock *CrashBlock =
2044 BasicBlock::Create(Context&: *C, Name: "", Parent: NextBB->getParent(), InsertBefore: NextBB);
2045 CrashTerm = new UnreachableInst(*C, CrashBlock);
2046 CondBrInst *NewTerm = CondBrInst::Create(Cond: Cmp2, IfTrue: CrashBlock, IfFalse: NextBB);
2047 ReplaceInstWithInst(From: CheckTerm, To: NewTerm);
2048 }
2049 } else {
2050 CrashTerm = SplitBlockAndInsertIfThen(Cond: Cmp, SplitBefore: InsertBefore, Unreachable: !Recover);
2051 }
2052
2053 Instruction *Crash = generateCrashCode(
2054 InsertBefore: CrashTerm, Addr: AddrLong, IsWrite, AccessSizeIndex, SizeArgument, Exp, RTCI);
2055 if (OrigIns->getDebugLoc())
2056 Crash->setDebugLoc(OrigIns->getDebugLoc());
2057}
2058
2059// Instrument unusual size or unusual alignment.
2060// We can not do it with a single check, so we do 1-byte check for the first
2061// and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
2062// to report the actual access size.
2063void AddressSanitizer::instrumentUnusualSizeOrAlignment(
2064 Instruction *I, Instruction *InsertBefore, Value *Addr,
2065 TypeSize TypeStoreSize, bool IsWrite, Value *SizeArgument, bool UseCalls,
2066 uint32_t Exp, RuntimeCallInserter &RTCI) {
2067 InstrumentationIRBuilder IRB(InsertBefore);
2068 Value *NumBits = IRB.CreateTypeSize(Ty: IntptrTy, Size: TypeStoreSize);
2069 Value *Size = IRB.CreateLShr(LHS: NumBits, RHS: ConstantInt::get(Ty: IntptrTy, V: 3));
2070
2071 Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
2072 if (UseCalls) {
2073 if (Exp == 0)
2074 RTCI.createRuntimeCall(IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][0],
2075 Args: {AddrLong, Size});
2076 else
2077 RTCI.createRuntimeCall(
2078 IRB, Callee: AsanMemoryAccessCallbackSized[IsWrite][1],
2079 Args: {AddrLong, Size, ConstantInt::get(Ty: IRB.getInt32Ty(), V: Exp)});
2080 } else {
2081 Value *SizeMinusOne = IRB.CreateSub(LHS: Size, RHS: ConstantInt::get(Ty: IntptrTy, V: 1));
2082 Value *LastByte = IRB.CreateIntToPtr(
2083 V: IRB.CreateAdd(LHS: AddrLong, RHS: SizeMinusOne),
2084 DestTy: Addr->getType());
2085 instrumentAddress(OrigIns: I, InsertBefore, Addr, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false, Exp,
2086 RTCI);
2087 instrumentAddress(OrigIns: I, InsertBefore, Addr: LastByte, Alignment: {}, TypeStoreSize: 8, IsWrite, SizeArgument: Size, UseCalls: false,
2088 Exp, RTCI);
2089 }
2090}
2091
2092void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit) {
2093 // Set up the arguments to our poison/unpoison functions.
2094 IRBuilder<> IRB(&GlobalInit.front(),
2095 GlobalInit.front().getFirstInsertionPt());
2096
2097 // Add a call to poison all external globals before the given function starts.
2098 Value *ModuleNameAddr =
2099 ConstantExpr::getPointerCast(C: getOrCreateModuleName(), Ty: IntptrTy);
2100 IRB.CreateCall(Callee: AsanPoisonGlobals, Args: ModuleNameAddr);
2101
2102 // Add calls to unpoison all globals before each return instruction.
2103 for (auto &BB : GlobalInit)
2104 if (ReturnInst *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator()))
2105 CallInst::Create(Func: AsanUnpoisonGlobals, NameStr: "", InsertBefore: RI->getIterator());
2106}
2107
2108void ModuleAddressSanitizer::createInitializerPoisonCalls() {
2109 GlobalVariable *GV = M.getGlobalVariable(Name: "llvm.global_ctors");
2110 if (!GV)
2111 return;
2112
2113 ConstantArray *CA = dyn_cast<ConstantArray>(Val: GV->getInitializer());
2114 if (!CA)
2115 return;
2116
2117 for (Use &OP : CA->operands()) {
2118 if (isa<ConstantAggregateZero>(Val: OP)) continue;
2119 ConstantStruct *CS = cast<ConstantStruct>(Val&: OP);
2120
2121 // Must have a function or null ptr.
2122 if (Function *F = dyn_cast<Function>(Val: CS->getOperand(i_nocapture: 1))) {
2123 if (F->getName() == kAsanModuleCtorName) continue;
2124 auto *Priority = cast<ConstantInt>(Val: CS->getOperand(i_nocapture: 0));
2125 // Don't instrument CTORs that will run before asan.module_ctor.
2126 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
2127 continue;
2128 poisonOneInitializer(GlobalInit&: *F);
2129 }
2130 }
2131}
2132
2133const GlobalVariable *
2134ModuleAddressSanitizer::getExcludedAliasedGlobal(const GlobalAlias &GA) const {
2135 // In case this function should be expanded to include rules that do not just
2136 // apply when CompileKernel is true, either guard all existing rules with an
2137 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
2138 // should also apply to user space.
2139 assert(CompileKernel && "Only expecting to be called when compiling kernel");
2140
2141 const Constant *C = GA.getAliasee();
2142
2143 // When compiling the kernel, globals that are aliased by symbols prefixed
2144 // by "__" are special and cannot be padded with a redzone.
2145 if (GA.getName().starts_with(Prefix: "__"))
2146 return dyn_cast<GlobalVariable>(Val: C->stripPointerCastsAndAliases());
2147
2148 return nullptr;
2149}
2150
2151bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
2152 Type *Ty = G->getValueType();
2153 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
2154
2155 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().NoAddress)
2156 return false;
2157 if (!Ty->isSized()) return false;
2158 if (!G->hasInitializer()) return false;
2159 if (!isSupportedAddrspace(TargetTriple, Addr: G))
2160 return false;
2161 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
2162 // Two problems with thread-locals:
2163 // - The address of the main thread's copy can't be computed at link-time.
2164 // - Need to poison all copies, not just the main thread's one.
2165 if (G->isThreadLocal()) return false;
2166 // For now, just ignore this Global if the alignment is large.
2167 if (G->getAlign() && *G->getAlign() > getMinRedzoneSizeForGlobal()) return false;
2168
2169 // For non-COFF targets, only instrument globals known to be defined by this
2170 // TU.
2171 // FIXME: We can instrument comdat globals on ELF if we are using the
2172 // GC-friendly metadata scheme.
2173 if (!TargetTriple.isOSBinFormatCOFF()) {
2174 if (!G->hasExactDefinition() || G->hasComdat())
2175 return false;
2176 } else {
2177 // On COFF, don't instrument non-ODR linkages.
2178 if (G->isInterposable())
2179 return false;
2180 // If the global has AvailableExternally linkage, then it is not in this
2181 // module, which means it does not need to be instrumented.
2182 if (G->hasAvailableExternallyLinkage())
2183 return false;
2184 }
2185
2186 // If a comdat is present, it must have a selection kind that implies ODR
2187 // semantics: no duplicates, any, or exact match.
2188 if (Comdat *C = G->getComdat()) {
2189 switch (C->getSelectionKind()) {
2190 case Comdat::Any:
2191 case Comdat::ExactMatch:
2192 case Comdat::NoDeduplicate:
2193 break;
2194 case Comdat::Largest:
2195 case Comdat::SameSize:
2196 return false;
2197 }
2198 }
2199
2200 if (G->hasSection()) {
2201 // The kernel uses explicit sections for mostly special global variables
2202 // that we should not instrument. E.g. the kernel may rely on their layout
2203 // without redzones, or remove them at link time ("discard.*"), etc.
2204 if (CompileKernel)
2205 return false;
2206
2207 StringRef Section = G->getSection();
2208
2209 // Globals from llvm.metadata aren't emitted, do not instrument them.
2210 if (Section == "llvm.metadata") return false;
2211 // Do not instrument globals from special LLVM sections.
2212 if (Section.contains(Other: "__llvm") || Section.contains(Other: "__LLVM"))
2213 return false;
2214
2215 // Do not instrument function pointers to initialization and termination
2216 // routines: dynamic linker will not properly handle redzones.
2217 if (Section.starts_with(Prefix: ".preinit_array") ||
2218 Section.starts_with(Prefix: ".init_array") ||
2219 Section.starts_with(Prefix: ".fini_array")) {
2220 return false;
2221 }
2222
2223 // Do not instrument user-defined sections (with names resembling
2224 // valid C identifiers)
2225 if (TargetTriple.isOSBinFormatELF()) {
2226 if (llvm::all_of(Range&: Section,
2227 P: [](char c) { return llvm::isAlnum(C: c) || c == '_'; }))
2228 return false;
2229 }
2230
2231 // On COFF, if the section name contains '$', it is highly likely that the
2232 // user is using section sorting to create an array of globals similar to
2233 // the way initialization callbacks are registered in .init_array and
2234 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
2235 // to such globals is counterproductive, because the intent is that they
2236 // will form an array, and out-of-bounds accesses are expected.
2237 // See https://github.com/google/sanitizers/issues/305
2238 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
2239 if (TargetTriple.isOSBinFormatCOFF() && Section.contains(C: '$')) {
2240 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
2241 << *G << "\n");
2242 return false;
2243 }
2244
2245 if (TargetTriple.isOSBinFormatMachO()) {
2246 StringRef ParsedSegment, ParsedSection;
2247 unsigned TAA = 0, StubSize = 0;
2248 bool TAAParsed;
2249 cantFail(Err: MCSectionMachO::ParseSectionSpecifier(
2250 Spec: Section, Segment&: ParsedSegment, Section&: ParsedSection, TAA, TAAParsed, StubSize));
2251
2252 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
2253 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
2254 // them.
2255 if (ParsedSegment == "__OBJC" ||
2256 (ParsedSegment == "__DATA" && ParsedSection.starts_with(Prefix: "__objc_"))) {
2257 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
2258 return false;
2259 }
2260 // See https://github.com/google/sanitizers/issues/32
2261 // Constant CFString instances are compiled in the following way:
2262 // -- the string buffer is emitted into
2263 // __TEXT,__cstring,cstring_literals
2264 // -- the constant NSConstantString structure referencing that buffer
2265 // is placed into __DATA,__cfstring
2266 // Therefore there's no point in placing redzones into __DATA,__cfstring.
2267 // Moreover, it causes the linker to crash on OS X 10.7
2268 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
2269 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
2270 return false;
2271 }
2272 // The linker merges the contents of cstring_literals and removes the
2273 // trailing zeroes.
2274 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
2275 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
2276 return false;
2277 }
2278 }
2279 }
2280
2281 if (CompileKernel) {
2282 // Globals that prefixed by "__" are special and cannot be padded with a
2283 // redzone.
2284 if (G->getName().starts_with(Prefix: "__"))
2285 return false;
2286 }
2287
2288 return true;
2289}
2290
2291// On Mach-O platforms, we emit global metadata in a separate section of the
2292// binary in order to allow the linker to properly dead strip. This is only
2293// supported on recent versions of ld64.
2294bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
2295 if (!TargetTriple.isOSBinFormatMachO())
2296 return false;
2297
2298 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(Major: 10, Minor: 11))
2299 return true;
2300 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(Major: 9))
2301 return true;
2302 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(Major: 2))
2303 return true;
2304 if (TargetTriple.isDriverKit())
2305 return true;
2306 if (TargetTriple.isXROS())
2307 return true;
2308
2309 return false;
2310}
2311
2312StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
2313 switch (TargetTriple.getObjectFormat()) {
2314 case Triple::COFF: return ".ASAN$GL";
2315 case Triple::ELF: return "asan_globals";
2316 case Triple::MachO: return "__DATA,__asan_globals,regular";
2317 case Triple::Wasm:
2318 case Triple::GOFF:
2319 case Triple::SPIRV:
2320 case Triple::XCOFF:
2321 case Triple::DXContainer:
2322 report_fatal_error(
2323 reason: "ModuleAddressSanitizer not implemented for object file format");
2324 case Triple::UnknownObjectFormat:
2325 break;
2326 }
2327 llvm_unreachable("unsupported object format");
2328}
2329
2330void ModuleAddressSanitizer::initializeCallbacks() {
2331 IRBuilder<> IRB(*C);
2332
2333 // Declare our poisoning and unpoisoning functions.
2334 AsanPoisonGlobals = Inserter.insertFunction(Name: kAsanPoisonGlobalsName,
2335 Args: IRB.getVoidTy(), Args&: IntptrTy);
2336 AsanUnpoisonGlobals =
2337 Inserter.insertFunction(Name: kAsanUnpoisonGlobalsName, Args: IRB.getVoidTy());
2338
2339 // Declare functions that register/unregister globals.
2340 AsanRegisterGlobals = Inserter.insertFunction(
2341 Name: kAsanRegisterGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2342 AsanUnregisterGlobals = Inserter.insertFunction(
2343 Name: kAsanUnregisterGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2344
2345 // Declare the functions that find globals in a shared object and then invoke
2346 // the (un)register function on them.
2347 AsanRegisterImageGlobals = Inserter.insertFunction(
2348 Name: kAsanRegisterImageGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy);
2349 AsanUnregisterImageGlobals = Inserter.insertFunction(
2350 Name: kAsanUnregisterImageGlobalsName, Args: IRB.getVoidTy(), Args&: IntptrTy);
2351
2352 AsanRegisterElfGlobals =
2353 Inserter.insertFunction(Name: kAsanRegisterElfGlobalsName, Args: IRB.getVoidTy(),
2354 Args&: IntptrTy, Args&: IntptrTy, Args&: IntptrTy);
2355 AsanUnregisterElfGlobals =
2356 Inserter.insertFunction(Name: kAsanUnregisterElfGlobalsName, Args: IRB.getVoidTy(),
2357 Args&: IntptrTy, Args&: IntptrTy, Args&: IntptrTy);
2358}
2359
2360// Put the metadata and the instrumented global in the same group. This ensures
2361// that the metadata is discarded if the instrumented global is discarded.
2362void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
2363 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
2364 Module &M = *G->getParent();
2365 Comdat *C = G->getComdat();
2366 if (!C) {
2367 if (!G->hasName()) {
2368 // If G is unnamed, it must be internal. Give it an artificial name
2369 // so we can put it in a comdat.
2370 assert(G->hasLocalLinkage());
2371 G->setName(genName(suffix: "anon_global"));
2372 }
2373
2374 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
2375 std::string Name = std::string(G->getName());
2376 Name += InternalSuffix;
2377 C = M.getOrInsertComdat(Name);
2378 } else {
2379 C = M.getOrInsertComdat(Name: G->getName());
2380 }
2381
2382 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
2383 // linkage to internal linkage so that a symbol table entry is emitted. This
2384 // is necessary in order to create the comdat group.
2385 if (TargetTriple.isOSBinFormatCOFF()) {
2386 C->setSelectionKind(Comdat::NoDeduplicate);
2387 if (G->hasPrivateLinkage())
2388 G->setLinkage(GlobalValue::InternalLinkage);
2389 }
2390 G->setComdat(C);
2391 }
2392
2393 assert(G->hasComdat());
2394 Metadata->setComdat(G->getComdat());
2395}
2396
2397// Create a separate metadata global and put it in the appropriate ASan
2398// global registration section.
2399GlobalVariable *
2400ModuleAddressSanitizer::CreateMetadataGlobal(Constant *Initializer,
2401 StringRef OriginalName) {
2402 auto Linkage = TargetTriple.isOSBinFormatMachO()
2403 ? GlobalVariable::InternalLinkage
2404 : GlobalVariable::PrivateLinkage;
2405 GlobalVariable *Metadata = new GlobalVariable(
2406 M, Initializer->getType(), false, Linkage, Initializer,
2407 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(Name: OriginalName));
2408 Metadata->setSection(getGlobalMetadataSection());
2409 // Place metadata in a large section for x86-64 ELF binaries to mitigate
2410 // relocation pressure.
2411 setGlobalVariableLargeSection(TargetTriple, GV&: *Metadata);
2412 return Metadata;
2413}
2414
2415Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor() {
2416 AsanDtorFunction = Function::createWithDefaultAttr(
2417 Ty: FunctionType::get(Result: Type::getVoidTy(C&: *C), isVarArg: false),
2418 Linkage: GlobalValue::InternalLinkage, AddrSpace: 0, N: kAsanModuleDtorName, M: &M);
2419 AsanDtorFunction->addFnAttr(Kind: Attribute::NoUnwind);
2420 // Ensure Dtor cannot be discarded, even if in a comdat.
2421 appendToUsed(M, Values: {AsanDtorFunction});
2422 BasicBlock *AsanDtorBB = BasicBlock::Create(Context&: *C, Name: "", Parent: AsanDtorFunction);
2423
2424 return ReturnInst::Create(C&: *C, InsertAtEnd: AsanDtorBB);
2425}
2426
2427void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
2428 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2429 ArrayRef<Constant *> MetadataInitializers) {
2430 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2431 auto &DL = M.getDataLayout();
2432
2433 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2434 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2435 Constant *Initializer = MetadataInitializers[i];
2436 GlobalVariable *G = ExtendedGlobals[i];
2437 GlobalVariable *Metadata = CreateMetadataGlobal(Initializer, OriginalName: G->getName());
2438 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2439 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2440 MetadataGlobals[i] = Metadata;
2441
2442 // The MSVC linker always inserts padding when linking incrementally. We
2443 // cope with that by aligning each struct to its size, which must be a power
2444 // of two.
2445 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Ty: Initializer->getType());
2446 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2447 "global metadata will not be padded appropriately");
2448 Metadata->setAlignment(assumeAligned(Value: SizeOfGlobalStruct));
2449
2450 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: "");
2451 }
2452
2453 // Update llvm.compiler.used, adding the new metadata globals. This is
2454 // needed so that during LTO these variables stay alive.
2455 if (!MetadataGlobals.empty())
2456 appendToCompilerUsed(M, Values: MetadataGlobals);
2457}
2458
2459void ModuleAddressSanitizer::instrumentGlobalsELF(
2460 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2461 ArrayRef<Constant *> MetadataInitializers,
2462 const std::string &UniqueModuleId) {
2463 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2464
2465 // Putting globals in a comdat changes the semantic and potentially cause
2466 // false negative odr violations at link time. If odr indicators are used, we
2467 // keep the comdat sections, as link time odr violations will be detected on
2468 // the odr indicator symbols.
2469 bool UseComdatForGlobalsGC = UseOdrIndicator && !UniqueModuleId.empty();
2470
2471 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2472 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2473 GlobalVariable *G = ExtendedGlobals[i];
2474 GlobalVariable *Metadata =
2475 CreateMetadataGlobal(Initializer: MetadataInitializers[i], OriginalName: G->getName());
2476 MDNode *MD = MDNode::get(Context&: M.getContext(), MDs: ValueAsMetadata::get(V: G));
2477 Metadata->setMetadata(KindID: LLVMContext::MD_associated, Node: MD);
2478 MetadataGlobals[i] = Metadata;
2479
2480 if (UseComdatForGlobalsGC)
2481 SetComdatForGlobalMetadata(G, Metadata, InternalSuffix: UniqueModuleId);
2482 }
2483
2484 // Update llvm.compiler.used, adding the new metadata globals. This is
2485 // needed so that during LTO these variables stay alive.
2486 if (!MetadataGlobals.empty())
2487 appendToCompilerUsed(M, Values: MetadataGlobals);
2488
2489 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2490 // to look up the loaded image that contains it. Second, we can store in it
2491 // whether registration has already occurred, to prevent duplicate
2492 // registration.
2493 //
2494 // Common linkage ensures that there is only one global per shared library.
2495 GlobalVariable *RegisteredFlag = new GlobalVariable(
2496 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2497 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2498 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2499
2500 // Create start and stop symbols.
2501 GlobalVariable *StartELFMetadata = new GlobalVariable(
2502 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2503 "__start_" + getGlobalMetadataSection());
2504 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2505 GlobalVariable *StopELFMetadata = new GlobalVariable(
2506 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2507 "__stop_" + getGlobalMetadataSection());
2508 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2509
2510 // Create a call to register the globals with the runtime.
2511 if (ConstructorKind == AsanCtorKind::Global)
2512 IRB.CreateCall(Callee: AsanRegisterElfGlobals,
2513 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2514 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2515 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2516
2517 // We also need to unregister globals at the end, e.g., when a shared library
2518 // gets closed.
2519 if (DestructorKind != AsanDtorKind::None && !MetadataGlobals.empty()) {
2520 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2521 IrbDtor.CreateCall(Callee: AsanUnregisterElfGlobals,
2522 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy),
2523 IRB.CreatePointerCast(V: StartELFMetadata, DestTy: IntptrTy),
2524 IRB.CreatePointerCast(V: StopELFMetadata, DestTy: IntptrTy)});
2525 }
2526}
2527
2528void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2529 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2530 ArrayRef<Constant *> MetadataInitializers) {
2531 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2532
2533 // On recent Mach-O platforms, use a structure which binds the liveness of
2534 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2535 // created to be added to llvm.compiler.used
2536 StructType *LivenessTy = StructType::get(elt1: IntptrTy, elts: IntptrTy);
2537 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2538
2539 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2540 Constant *Initializer = MetadataInitializers[i];
2541 GlobalVariable *G = ExtendedGlobals[i];
2542 GlobalVariable *Metadata = CreateMetadataGlobal(Initializer, OriginalName: G->getName());
2543
2544 // On recent Mach-O platforms, we emit the global metadata in a way that
2545 // allows the linker to properly strip dead globals.
2546 auto LivenessBinder =
2547 ConstantStruct::get(T: LivenessTy, Vs: Initializer->getAggregateElement(Elt: 0u),
2548 Vs: ConstantExpr::getPointerCast(C: Metadata, Ty: IntptrTy));
2549 GlobalVariable *Liveness = new GlobalVariable(
2550 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2551 Twine("__asan_binder_") + G->getName());
2552 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2553 LivenessGlobals[i] = Liveness;
2554 }
2555
2556 // Update llvm.compiler.used, adding the new liveness globals. This is
2557 // needed so that during LTO these variables stay alive. The alternative
2558 // would be to have the linker handling the LTO symbols, but libLTO
2559 // current API does not expose access to the section for each symbol.
2560 if (!LivenessGlobals.empty())
2561 appendToCompilerUsed(M, Values: LivenessGlobals);
2562
2563 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2564 // to look up the loaded image that contains it. Second, we can store in it
2565 // whether registration has already occurred, to prevent duplicate
2566 // registration.
2567 //
2568 // common linkage ensures that there is only one global per shared library.
2569 GlobalVariable *RegisteredFlag = new GlobalVariable(
2570 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2571 ConstantInt::get(Ty: IntptrTy, V: 0), kAsanGlobalsRegisteredFlagName);
2572 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2573
2574 if (ConstructorKind == AsanCtorKind::Global)
2575 IRB.CreateCall(Callee: AsanRegisterImageGlobals,
2576 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2577
2578 // We also need to unregister globals at the end, e.g., when a shared library
2579 // gets closed.
2580 if (DestructorKind != AsanDtorKind::None) {
2581 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2582 IrbDtor.CreateCall(Callee: AsanUnregisterImageGlobals,
2583 Args: {IRB.CreatePointerCast(V: RegisteredFlag, DestTy: IntptrTy)});
2584 }
2585}
2586
2587void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2588 IRBuilder<> &IRB, ArrayRef<GlobalVariable *> ExtendedGlobals,
2589 ArrayRef<Constant *> MetadataInitializers) {
2590 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2591 unsigned N = ExtendedGlobals.size();
2592 assert(N > 0);
2593
2594 // On platforms that don't have a custom metadata section, we emit an array
2595 // of global metadata structures.
2596 ArrayType *ArrayOfGlobalStructTy =
2597 ArrayType::get(ElementType: MetadataInitializers[0]->getType(), NumElements: N);
2598 auto AllGlobals = new GlobalVariable(
2599 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2600 ConstantArray::get(T: ArrayOfGlobalStructTy, V: MetadataInitializers), "");
2601 if (Mapping.Scale > 3)
2602 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2603
2604 if (ConstructorKind == AsanCtorKind::Global)
2605 IRB.CreateCall(Callee: AsanRegisterGlobals,
2606 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2607 ConstantInt::get(Ty: IntptrTy, V: N)});
2608
2609 // We also need to unregister globals at the end, e.g., when a shared library
2610 // gets closed.
2611 if (DestructorKind != AsanDtorKind::None) {
2612 IRBuilder<> IrbDtor(CreateAsanModuleDtor());
2613 IrbDtor.CreateCall(Callee: AsanUnregisterGlobals,
2614 Args: {IRB.CreatePointerCast(V: AllGlobals, DestTy: IntptrTy),
2615 ConstantInt::get(Ty: IntptrTy, V: N)});
2616 }
2617}
2618
2619// This function replaces all global variables with new variables that have
2620// trailing redzones. It also creates a function that poisons
2621// redzones and inserts this function into llvm.global_ctors.
2622// Sets *CtorComdat to true if the global registration code emitted into the
2623// asan constructor is comdat-compatible.
2624void ModuleAddressSanitizer::instrumentGlobals(IRBuilder<> &IRB,
2625 bool *CtorComdat) {
2626 // Build set of globals that are aliased by some GA, where
2627 // getExcludedAliasedGlobal(GA) returns the relevant GlobalVariable.
2628 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2629 if (CompileKernel) {
2630 for (auto &GA : M.aliases()) {
2631 if (const GlobalVariable *GV = getExcludedAliasedGlobal(GA))
2632 AliasedGlobalExclusions.insert(Ptr: GV);
2633 }
2634 }
2635
2636 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2637 for (auto &G : M.globals()) {
2638 if (!AliasedGlobalExclusions.count(Ptr: &G) && shouldInstrumentGlobal(G: &G))
2639 GlobalsToChange.push_back(Elt: &G);
2640 }
2641
2642 size_t n = GlobalsToChange.size();
2643 auto &DL = M.getDataLayout();
2644
2645 // A global is described by a structure
2646 // size_t beg;
2647 // size_t size;
2648 // size_t size_with_redzone;
2649 // const char *name;
2650 // const char *module_name;
2651 // size_t has_dynamic_init;
2652 // size_t padding_for_windows_msvc_incremental_link;
2653 // size_t odr_indicator;
2654 // We initialize an array of such structures and pass it to a run-time call.
2655 StructType *GlobalStructTy =
2656 StructType::get(elt1: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy, elts: IntptrTy,
2657 elts: IntptrTy, elts: IntptrTy, elts: IntptrTy);
2658 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2659 SmallVector<Constant *, 16> Initializers(n);
2660
2661 for (size_t i = 0; i < n; i++) {
2662 GlobalVariable *G = GlobalsToChange[i];
2663
2664 GlobalValue::SanitizerMetadata MD;
2665 if (G->hasSanitizerMetadata())
2666 MD = G->getSanitizerMetadata();
2667
2668 // The runtime library tries demangling symbol names in the descriptor but
2669 // functionality like __cxa_demangle may be unavailable (e.g.
2670 // -static-libstdc++). So we demangle the symbol names here.
2671 std::string NameForGlobal = G->getName().str();
2672 GlobalVariable *Name =
2673 createPrivateGlobalForString(M, Str: llvm::demangle(MangledName: NameForGlobal),
2674 /*AllowMerging*/ true, NamePrefix: genName(suffix: "global"));
2675
2676 Type *Ty = G->getValueType();
2677 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2678 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2679 Type *RightRedZoneTy = ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: RightRedzoneSize);
2680
2681 StructType *NewTy = StructType::get(elt1: Ty, elts: RightRedZoneTy);
2682 Constant *NewInitializer = ConstantStruct::get(
2683 T: NewTy, Vs: G->getInitializer(), Vs: Constant::getNullValue(Ty: RightRedZoneTy));
2684
2685 // Create a new global variable with enough space for a redzone.
2686 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2687 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2688 Linkage = GlobalValue::InternalLinkage;
2689 GlobalVariable *NewGlobal = new GlobalVariable(
2690 M, NewTy, G->isConstant(), Linkage, NewInitializer, "", G,
2691 G->getThreadLocalMode(), G->getAddressSpace());
2692 NewGlobal->copyAttributesFrom(Src: G);
2693 NewGlobal->setComdat(G->getComdat());
2694 NewGlobal->setAlignment(Align(getMinRedzoneSizeForGlobal()));
2695 // Don't fold globals with redzones. ODR violation detector and redzone
2696 // poisoning implicitly creates a dependence on the global's address, so it
2697 // is no longer valid for it to be marked unnamed_addr.
2698 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2699
2700 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2701 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2702 G->isConstant()) {
2703 auto Seq = dyn_cast<ConstantDataSequential>(Val: G->getInitializer());
2704 if (Seq && Seq->isCString())
2705 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2706 }
2707
2708 // Transfer the debug info and type metadata. The payload starts at offset
2709 // zero so we can copy the metadata over as is.
2710 NewGlobal->copyMetadata(Src: G, Offset: 0);
2711
2712 G->replaceAllUsesWith(V: NewGlobal);
2713 NewGlobal->takeName(V: G);
2714 G->eraseFromParent();
2715 NewGlobals[i] = NewGlobal;
2716
2717 Constant *ODRIndicator = Constant::getNullValue(Ty: IntptrTy);
2718 GlobalValue *InstrumentedGlobal = NewGlobal;
2719
2720 bool CanUsePrivateAliases =
2721 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2722 TargetTriple.isOSBinFormatWasm();
2723 if (CanUsePrivateAliases && UsePrivateAlias) {
2724 // Create local alias for NewGlobal to avoid crash on ODR between
2725 // instrumented and non-instrumented libraries.
2726 InstrumentedGlobal =
2727 GlobalAlias::create(Linkage: GlobalValue::PrivateLinkage, Name: "", Aliasee: NewGlobal);
2728 }
2729
2730 // ODR should not happen for local linkage.
2731 if (NewGlobal->hasLocalLinkage()) {
2732 ODRIndicator = ConstantInt::getAllOnesValue(Ty: IntptrTy);
2733 } else if (UseOdrIndicator) {
2734 // With local aliases, we need to provide another externally visible
2735 // symbol __odr_asan_XXX to detect ODR violation.
2736 auto *ODRIndicatorSym =
2737 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2738 Constant::getNullValue(Ty: IRB.getInt8Ty()),
2739 kODRGenPrefix + NameForGlobal, nullptr,
2740 NewGlobal->getThreadLocalMode());
2741
2742 // Set meaningful attributes for indicator symbol.
2743 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2744 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2745 ODRIndicatorSym->setAlignment(Align(1));
2746 ODRIndicator = ConstantExpr::getPtrToInt(C: ODRIndicatorSym, Ty: IntptrTy);
2747 }
2748
2749 Constant *Initializer = ConstantStruct::get(
2750 T: GlobalStructTy,
2751 Vs: ConstantExpr::getPointerCast(C: InstrumentedGlobal, Ty: IntptrTy),
2752 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes),
2753 Vs: ConstantInt::get(Ty: IntptrTy, V: SizeInBytes + RightRedzoneSize),
2754 Vs: ConstantExpr::getPointerCast(C: Name, Ty: IntptrTy),
2755 Vs: ConstantExpr::getPointerCast(C: getOrCreateModuleName(), Ty: IntptrTy),
2756 Vs: ConstantInt::get(Ty: IntptrTy, V: MD.IsDynInit),
2757 Vs: Constant::getNullValue(Ty: IntptrTy), Vs: ODRIndicator);
2758
2759 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2760
2761 Initializers[i] = Initializer;
2762 }
2763
2764 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2765 // ConstantMerge'ing them.
2766 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2767 for (size_t i = 0; i < n; i++) {
2768 GlobalVariable *G = NewGlobals[i];
2769 if (G->getName().empty()) continue;
2770 GlobalsToAddToUsedList.push_back(Elt: G);
2771 }
2772 appendToCompilerUsed(M, Values: ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2773
2774 if (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) {
2775 // Use COMDAT and register globals even if n == 0 to ensure that (a) the
2776 // linkage unit will only have one module constructor, and (b) the register
2777 // function will be called. The module destructor is not created when n ==
2778 // 0.
2779 *CtorComdat = true;
2780 instrumentGlobalsELF(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers, UniqueModuleId: getUniqueModuleId(M: &M));
2781 } else if (n == 0) {
2782 // When UseGlobalsGC is false, COMDAT can still be used if n == 0, because
2783 // all compile units will have identical module constructor/destructor.
2784 *CtorComdat = TargetTriple.isOSBinFormatELF();
2785 } else {
2786 *CtorComdat = false;
2787 if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2788 InstrumentGlobalsCOFF(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2789 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2790 InstrumentGlobalsMachO(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2791 } else {
2792 InstrumentGlobalsWithMetadataArray(IRB, ExtendedGlobals: NewGlobals, MetadataInitializers: Initializers);
2793 }
2794 }
2795
2796 // Create calls for poisoning before initializers run and unpoisoning after.
2797 if (ClInitializers)
2798 createInitializerPoisonCalls();
2799
2800 LLVM_DEBUG(dbgs() << M);
2801}
2802
2803uint64_t
2804ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2805 constexpr uint64_t kMaxRZ = 1 << 18;
2806 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2807
2808 uint64_t RZ = 0;
2809 if (SizeInBytes <= MinRZ / 2) {
2810 // Reduce redzone size for small size objects, e.g. int, char[1]. MinRZ is
2811 // at least 32 bytes, optimize when SizeInBytes is less than or equal to
2812 // half of MinRZ.
2813 RZ = MinRZ - SizeInBytes;
2814 } else {
2815 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2816 RZ = std::clamp(val: (SizeInBytes / MinRZ / 4) * MinRZ, lo: MinRZ, hi: kMaxRZ);
2817
2818 // Round up to multiple of MinRZ.
2819 if (SizeInBytes % MinRZ)
2820 RZ += MinRZ - (SizeInBytes % MinRZ);
2821 }
2822
2823 assert((RZ + SizeInBytes) % MinRZ == 0);
2824
2825 return RZ;
2826}
2827
2828int ModuleAddressSanitizer::GetAsanVersion() const {
2829 int LongSize = M.getDataLayout().getPointerSizeInBits();
2830 bool isAndroid = M.getTargetTriple().isAndroid();
2831 int Version = 8;
2832 // 32-bit Android is one version ahead because of the switch to dynamic
2833 // shadow.
2834 Version += (LongSize == 32 && isAndroid);
2835 return Version;
2836}
2837
2838GlobalVariable *ModuleAddressSanitizer::getOrCreateModuleName() {
2839 if (!ModuleName) {
2840 // We shouldn't merge same module names, as this string serves as unique
2841 // module ID in runtime.
2842 ModuleName =
2843 createPrivateGlobalForString(M, Str: M.getModuleIdentifier(),
2844 /*AllowMerging*/ false, NamePrefix: genName(suffix: "module"));
2845 }
2846 return ModuleName;
2847}
2848
2849bool ModuleAddressSanitizer::instrumentModule() {
2850 initializeCallbacks();
2851
2852 for (Function &F : M)
2853 removeASanIncompatibleFnAttributes(F, /*ReadsArgMem=*/false);
2854
2855 // Create a module constructor. A destructor is created lazily because not all
2856 // platforms, and not all modules need it.
2857 if (ConstructorKind == AsanCtorKind::Global) {
2858 if (CompileKernel) {
2859 // The kernel always builds with its own runtime, and therefore does not
2860 // need the init and version check calls.
2861 AsanCtorFunction = createSanitizerCtor(M, CtorName: kAsanModuleCtorName);
2862 } else {
2863 std::string AsanVersion = std::to_string(val: GetAsanVersion());
2864 std::string VersionCheckName =
2865 InsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2866 std::tie(args&: AsanCtorFunction, args: std::ignore) =
2867 createSanitizerCtorAndInitFunctions(
2868 M, CtorName: kAsanModuleCtorName, InitName: kAsanInitName, /*InitArgTypes=*/{},
2869 /*InitArgs=*/{}, VersionCheckName);
2870 }
2871 }
2872
2873 bool CtorComdat = true;
2874 if (ClGlobals) {
2875 assert(AsanCtorFunction || ConstructorKind == AsanCtorKind::None);
2876 if (AsanCtorFunction) {
2877 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2878 instrumentGlobals(IRB, CtorComdat: &CtorComdat);
2879 } else {
2880 IRBuilder<> IRB(*C);
2881 instrumentGlobals(IRB, CtorComdat: &CtorComdat);
2882 }
2883 }
2884
2885 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2886
2887 // Put the constructor and destructor in comdat if both
2888 // (1) global instrumentation is not TU-specific
2889 // (2) target is ELF.
2890 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2891 if (AsanCtorFunction) {
2892 AsanCtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleCtorName));
2893 appendToGlobalCtors(M, F: AsanCtorFunction, Priority, Data: AsanCtorFunction);
2894 }
2895 if (AsanDtorFunction) {
2896 AsanDtorFunction->setComdat(M.getOrInsertComdat(Name: kAsanModuleDtorName));
2897 appendToGlobalDtors(M, F: AsanDtorFunction, Priority, Data: AsanDtorFunction);
2898 }
2899 } else {
2900 if (AsanCtorFunction)
2901 appendToGlobalCtors(M, F: AsanCtorFunction, Priority);
2902 if (AsanDtorFunction)
2903 appendToGlobalDtors(M, F: AsanDtorFunction, Priority);
2904 }
2905
2906 return true;
2907}
2908
2909void AddressSanitizer::initializeCallbacks(const TargetLibraryInfo *TLI) {
2910 IRBuilder<> IRB(*C);
2911 // Create __asan_report* callbacks.
2912 // IsWrite, TypeSize and Exp are encoded in the function name.
2913 for (int Exp = 0; Exp < 2; Exp++) {
2914 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2915 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2916 const std::string ExpStr = Exp ? "exp_" : "";
2917 const std::string EndingStr = Recover ? "_noabort" : "";
2918
2919 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2920 SmallVector<Type *, 2> Args1{1, IntptrTy};
2921 AttributeList AL2;
2922 AttributeList AL1;
2923 if (Exp) {
2924 Type *ExpType = Type::getInt32Ty(C&: *C);
2925 Args2.push_back(Elt: ExpType);
2926 Args1.push_back(Elt: ExpType);
2927 if (auto AK = TLI->getExtAttrForI32Param(Signed: false)) {
2928 AL2 = AL2.addParamAttribute(C&: *C, ArgNo: 2, Kind: AK);
2929 AL1 = AL1.addParamAttribute(C&: *C, ArgNo: 1, Kind: AK);
2930 }
2931 }
2932 AsanErrorCallbackSized[AccessIsWrite][Exp] = Inserter.insertFunction(
2933 Name: kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2934 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), Args&: AL2);
2935
2936 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] =
2937 Inserter.insertFunction(
2938 Name: ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2939 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args2, isVarArg: false), Args&: AL2);
2940
2941 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2942 AccessSizeIndex++) {
2943 const std::string Suffix = TypeStr + itostr(X: 1ULL << AccessSizeIndex);
2944 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2945 Inserter.insertFunction(
2946 Name: kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2947 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), Args&: AL1);
2948
2949 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2950 Inserter.insertFunction(
2951 Name: ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2952 Args: FunctionType::get(Result: IRB.getVoidTy(), Params: Args1, isVarArg: false), Args&: AL1);
2953 }
2954 }
2955 }
2956
2957 const std::string MemIntrinCallbackPrefix =
2958 (CompileKernel && !ClKasanMemIntrinCallbackPrefix)
2959 ? std::string("")
2960 : ClMemoryAccessCallbackPrefix;
2961 AsanMemmove = Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memmove",
2962 Args&: PtrTy, Args&: PtrTy, Args&: PtrTy, Args&: IntptrTy);
2963 AsanMemcpy = Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memcpy",
2964 Args&: PtrTy, Args&: PtrTy, Args&: PtrTy, Args&: IntptrTy);
2965 AsanMemset =
2966 Inserter.insertFunction(Name: MemIntrinCallbackPrefix + "memset",
2967 Args: TLI->getAttrList(C, ArgNos: {1},
2968 /*Signed=*/false),
2969 Args&: PtrTy, Args&: PtrTy, Args: IRB.getInt32Ty(), Args&: IntptrTy);
2970
2971 AsanHandleNoReturnFunc =
2972 Inserter.insertFunction(Name: kAsanHandleNoReturnName, Args: IRB.getVoidTy());
2973
2974 AsanPtrCmpFunction =
2975 Inserter.insertFunction(Name: kAsanPtrCmp, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2976 AsanPtrSubFunction =
2977 Inserter.insertFunction(Name: kAsanPtrSub, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
2978 if (Mapping.InGlobal)
2979 AsanShadowGlobal = M.getOrInsertGlobal(Name: "__asan_shadow",
2980 Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: 0));
2981
2982 AMDGPUAddressShared =
2983 Inserter.insertFunction(Name: kAMDGPUAddressSharedName, Args: IRB.getInt1Ty(), Args&: PtrTy);
2984 AMDGPUAddressPrivate = Inserter.insertFunction(Name: kAMDGPUAddressPrivateName,
2985 Args: IRB.getInt1Ty(), Args&: PtrTy);
2986}
2987
2988bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
2989 // For each NSObject descendant having a +load method, this method is invoked
2990 // by the ObjC runtime before any of the static constructors is called.
2991 // Therefore we need to instrument such methods with a call to __asan_init
2992 // at the beginning in order to initialize our runtime before any access to
2993 // the shadow memory.
2994 // We cannot just ignore these methods, because they may call other
2995 // instrumented functions.
2996 if (F.getName().contains(Other: " load]")) {
2997 FunctionCallee AsanInitFunction =
2998 declareSanitizerInitFunction(M&: *F.getParent(), InitName: kAsanInitName, InitArgTypes: {});
2999 IRBuilder<> IRB(&F.front(), F.front().begin());
3000 IRB.CreateCall(Callee: AsanInitFunction, Args: {});
3001 return true;
3002 }
3003 return false;
3004}
3005
3006bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
3007 // Generate code only when dynamic addressing is needed.
3008 if (Mapping.Offset != kDynamicShadowSentinel)
3009 return false;
3010
3011 IRBuilder<> IRB(&F.front().front());
3012 if (Mapping.InGlobal) {
3013 if (ClWithIfuncSuppressRemat) {
3014 // An empty inline asm with input reg == output reg.
3015 // An opaque pointer-to-int cast, basically.
3016 InlineAsm *Asm = InlineAsm::get(
3017 Ty: FunctionType::get(Result: IntptrTy, Params: {AsanShadowGlobal->getType()}, isVarArg: false),
3018 AsmString: StringRef(""), Constraints: StringRef("=r,0"),
3019 /*hasSideEffects=*/false);
3020 LocalDynamicShadow =
3021 IRB.CreateCall(Callee: Asm, Args: {AsanShadowGlobal}, Name: ".asan.shadow");
3022 } else {
3023 LocalDynamicShadow =
3024 IRB.CreatePointerCast(V: AsanShadowGlobal, DestTy: IntptrTy, Name: ".asan.shadow");
3025 }
3026 } else {
3027 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
3028 Name: kAsanShadowMemoryDynamicAddress, Ty: IntptrTy);
3029 LocalDynamicShadow = IRB.CreateLoad(Ty: IntptrTy, Ptr: GlobalDynamicAddress);
3030 }
3031 return true;
3032}
3033
3034void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
3035 // Find the one possible call to llvm.localescape and pre-mark allocas passed
3036 // to it as uninteresting. This assumes we haven't started processing allocas
3037 // yet. This check is done up front because iterating the use list in
3038 // isInterestingAlloca would be algorithmically slower.
3039 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
3040
3041 // Try to get the declaration of llvm.localescape. If it's not in the module,
3042 // we can exit early.
3043 if (!F.getParent()->getFunction(Name: "llvm.localescape")) return;
3044
3045 // Look for a call to llvm.localescape call in the entry block. It can't be in
3046 // any other block.
3047 for (Instruction &I : F.getEntryBlock()) {
3048 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: &I);
3049 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
3050 // We found a call. Mark all the allocas passed in as uninteresting.
3051 for (Value *Arg : II->args()) {
3052 AllocaInst *AI = dyn_cast<AllocaInst>(Val: Arg->stripPointerCasts());
3053 assert(AI && AI->isStaticAlloca() &&
3054 "non-static alloca arg to localescape");
3055 ProcessedAllocas[AI] = false;
3056 }
3057 break;
3058 }
3059 }
3060}
3061// Mitigation for https://github.com/google/sanitizers/issues/749
3062// We don't instrument Windows catch-block parameters to avoid
3063// interfering with exception handling assumptions.
3064void AddressSanitizer::markCatchParametersAsUninteresting(Function &F) {
3065 for (BasicBlock &BB : F) {
3066 for (Instruction &I : BB) {
3067 if (auto *CatchPad = dyn_cast<CatchPadInst>(Val: &I)) {
3068 // Mark the parameters to a catch-block as uninteresting to avoid
3069 // instrumenting them.
3070 for (Value *Operand : CatchPad->arg_operands())
3071 if (auto *AI = dyn_cast<AllocaInst>(Val: Operand))
3072 ProcessedAllocas[AI] = false;
3073 }
3074 }
3075 }
3076}
3077
3078bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
3079 bool ShouldInstrument =
3080 ClDebugMin < 0 || ClDebugMax < 0 ||
3081 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
3082 Instrumented++;
3083 return !ShouldInstrument;
3084}
3085
3086bool AddressSanitizer::instrumentFunction(Function &F,
3087 const TargetLibraryInfo *TLI,
3088 const TargetTransformInfo *TTI) {
3089 bool FunctionModified = false;
3090
3091 // Do not apply any instrumentation for naked functions.
3092 if (F.hasFnAttribute(Kind: Attribute::Naked))
3093 return FunctionModified;
3094
3095 // If needed, insert __asan_init before checking for SanitizeAddress attr.
3096 // This function needs to be called even if the function body is not
3097 // instrumented.
3098 if (maybeInsertAsanInitAtFunctionEntry(F))
3099 FunctionModified = true;
3100
3101 // Leave if the function doesn't need instrumentation.
3102 if (!F.hasFnAttribute(Kind: Attribute::SanitizeAddress)) return FunctionModified;
3103
3104 if (F.hasFnAttribute(Kind: Attribute::DisableSanitizerInstrumentation))
3105 return FunctionModified;
3106
3107 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
3108
3109 initializeCallbacks(TLI);
3110
3111 FunctionStateRAII CleanupObj(this);
3112
3113 RuntimeCallInserter RTCI(F);
3114
3115 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
3116
3117 // We can't instrument allocas used with llvm.localescape. Only static allocas
3118 // can be passed to that intrinsic.
3119 markEscapedLocalAllocas(F);
3120
3121 if (TargetTriple.isOSWindows())
3122 markCatchParametersAsUninteresting(F);
3123
3124 // We want to instrument every address only once per basic block (unless there
3125 // are calls between uses).
3126 SmallPtrSet<Value *, 16> TempsToInstrument;
3127 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
3128 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
3129 SmallVector<Instruction *, 8> NoReturnCalls;
3130 SmallVector<BasicBlock *, 16> AllBlocks;
3131 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
3132
3133 // Fill the set of memory operations to instrument.
3134 for (auto &BB : F) {
3135 AllBlocks.push_back(Elt: &BB);
3136 TempsToInstrument.clear();
3137 int NumInsnsPerBB = 0;
3138 for (auto &Inst : BB) {
3139 if (LooksLikeCodeInBug11395(I: &Inst)) return false;
3140 // Skip instructions inserted by another instrumentation.
3141 if (Inst.hasMetadata(KindID: LLVMContext::MD_nosanitize))
3142 continue;
3143 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
3144 getInterestingMemoryOperands(I: &Inst, Interesting&: InterestingOperands, TTI);
3145
3146 if (!InterestingOperands.empty()) {
3147 for (auto &Operand : InterestingOperands) {
3148 if (ClOpt && ClOptSameTemp) {
3149 Value *Ptr = Operand.getPtr();
3150 // If we have a mask, skip instrumentation if we've already
3151 // instrumented the full object. But don't add to TempsToInstrument
3152 // because we might get another load/store with a different mask.
3153 if (Operand.MaybeMask) {
3154 if (TempsToInstrument.count(Ptr))
3155 continue; // We've seen this (whole) temp in the current BB.
3156 } else {
3157 if (!TempsToInstrument.insert(Ptr).second)
3158 continue; // We've seen this temp in the current BB.
3159 }
3160 }
3161 OperandsToInstrument.push_back(Elt: Operand);
3162 NumInsnsPerBB++;
3163 }
3164 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
3165 isInterestingPointerComparison(I: &Inst)) ||
3166 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
3167 isInterestingPointerSubtraction(I: &Inst))) {
3168 PointerComparisonsOrSubtracts.push_back(Elt: &Inst);
3169 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: &Inst)) {
3170 // ok, take it.
3171 IntrinToInstrument.push_back(Elt: MI);
3172 NumInsnsPerBB++;
3173 } else {
3174 if (auto *CB = dyn_cast<CallBase>(Val: &Inst)) {
3175 // A call inside BB.
3176 TempsToInstrument.clear();
3177 if (CB->doesNotReturn())
3178 NoReturnCalls.push_back(Elt: CB);
3179 }
3180 if (CallInst *CI = dyn_cast<CallInst>(Val: &Inst))
3181 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
3182 }
3183 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
3184 }
3185 }
3186
3187 bool UseCalls = (InstrumentationWithCallsThreshold >= 0 &&
3188 OperandsToInstrument.size() + IntrinToInstrument.size() >
3189 (unsigned)InstrumentationWithCallsThreshold);
3190 const DataLayout &DL = F.getDataLayout();
3191 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext());
3192
3193 // Instrument.
3194 int NumInstrumented = 0;
3195 for (auto &Operand : OperandsToInstrument) {
3196 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3197 instrumentMop(ObjSizeVis, O&: Operand, UseCalls,
3198 DL: F.getDataLayout(), RTCI);
3199 FunctionModified = true;
3200 }
3201 for (auto *Inst : IntrinToInstrument) {
3202 if (!suppressInstrumentationSiteForDebug(Instrumented&: NumInstrumented))
3203 instrumentMemIntrinsic(MI: Inst, RTCI);
3204 FunctionModified = true;
3205 }
3206
3207 FunctionStackPoisoner FSP(F, *this, RTCI);
3208 bool ChangedStack = FSP.runOnFunction();
3209
3210 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
3211 // See e.g. https://github.com/google/sanitizers/issues/37
3212 for (auto *CI : NoReturnCalls) {
3213 IRBuilder<> IRB(CI);
3214 RTCI.createRuntimeCall(IRB, Callee: AsanHandleNoReturnFunc, Args: {});
3215 }
3216
3217 for (auto *Inst : PointerComparisonsOrSubtracts) {
3218 instrumentPointerComparisonOrSubtraction(I: Inst, RTCI);
3219 FunctionModified = true;
3220 }
3221
3222 if (ChangedStack || !NoReturnCalls.empty())
3223 FunctionModified = true;
3224
3225 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
3226 << F << "\n");
3227
3228 return FunctionModified;
3229}
3230
3231// Workaround for bug 11395: we don't want to instrument stack in functions
3232// with large assembly blobs (32-bit only), otherwise reg alloc may crash.
3233// FIXME: remove once the bug 11395 is fixed.
3234bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
3235 if (LongSize != 32) return false;
3236 CallInst *CI = dyn_cast<CallInst>(Val: I);
3237 if (!CI || !CI->isInlineAsm()) return false;
3238 if (CI->arg_size() <= 5)
3239 return false;
3240 // We have inline assembly with quite a few arguments.
3241 return true;
3242}
3243
3244void FunctionStackPoisoner::initializeCallbacks(Module &) {
3245 IRBuilder<> IRB(*C);
3246 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always ||
3247 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3248 const char *MallocNameTemplate =
3249 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always
3250 ? kAsanStackMallocAlwaysNameTemplate
3251 : kAsanStackMallocNameTemplate;
3252 for (int Index = 0; Index <= kMaxAsanStackMallocSizeClass; Index++) {
3253 std::string Suffix = itostr(X: Index);
3254 AsanStackMallocFunc[Index] = ASan.Inserter.insertFunction(
3255 Name: MallocNameTemplate + Suffix, Args&: IntptrTy, Args&: IntptrTy);
3256 AsanStackFreeFunc[Index] =
3257 ASan.Inserter.insertFunction(Name: kAsanStackFreeNameTemplate + Suffix,
3258 Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3259 }
3260 }
3261 if (ASan.UseAfterScope) {
3262 AsanPoisonStackMemoryFunc = ASan.Inserter.insertFunction(
3263 Name: kAsanPoisonStackMemoryName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3264 AsanUnpoisonStackMemoryFunc = ASan.Inserter.insertFunction(
3265 Name: kAsanUnpoisonStackMemoryName, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3266 }
3267
3268 for (size_t Val : {0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0xf1, 0xf2,
3269 0xf3, 0xf5, 0xf8}) {
3270 std::ostringstream Name;
3271 Name << kAsanSetShadowPrefix;
3272 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
3273 AsanSetShadowFunc[Val] = ASan.Inserter.insertFunction(
3274 Name: Name.str(), Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3275 }
3276
3277 AsanAllocaPoisonFunc = ASan.Inserter.insertFunction(
3278 Name: kAsanAllocaPoison, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3279 AsanAllocasUnpoisonFunc = ASan.Inserter.insertFunction(
3280 Name: kAsanAllocasUnpoison, Args: IRB.getVoidTy(), Args&: IntptrTy, Args&: IntptrTy);
3281}
3282
3283void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
3284 ArrayRef<uint8_t> ShadowBytes,
3285 size_t Begin, size_t End,
3286 IRBuilder<> &IRB,
3287 Value *ShadowBase) {
3288 if (Begin >= End)
3289 return;
3290
3291 const size_t LargestStoreSizeInBytes =
3292 std::min<size_t>(a: sizeof(uint64_t), b: ASan.LongSize / 8);
3293
3294 const bool IsLittleEndian = F.getDataLayout().isLittleEndian();
3295
3296 // Poison given range in shadow using larges store size with out leading and
3297 // trailing zeros in ShadowMask. Zeros never change, so they need neither
3298 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
3299 // middle of a store.
3300 for (size_t i = Begin; i < End;) {
3301 if (!ShadowMask[i]) {
3302 assert(!ShadowBytes[i]);
3303 ++i;
3304 continue;
3305 }
3306
3307 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
3308 // Fit store size into the range.
3309 while (StoreSizeInBytes > End - i)
3310 StoreSizeInBytes /= 2;
3311
3312 // Minimize store size by trimming trailing zeros.
3313 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
3314 while (j <= StoreSizeInBytes / 2)
3315 StoreSizeInBytes /= 2;
3316 }
3317
3318 uint64_t Val = 0;
3319 for (size_t j = 0; j < StoreSizeInBytes; j++) {
3320 if (IsLittleEndian)
3321 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
3322 else
3323 Val = (Val << 8) | ShadowBytes[i + j];
3324 }
3325
3326 Value *Ptr = IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i));
3327 Value *Poison = IRB.getIntN(N: StoreSizeInBytes * 8, C: Val);
3328 IRB.CreateAlignedStore(
3329 Val: Poison, Ptr: IRB.CreateIntToPtr(V: Ptr, DestTy: PointerType::getUnqual(C&: Poison->getContext())),
3330 Align: Align(1));
3331
3332 i += StoreSizeInBytes;
3333 }
3334}
3335
3336void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3337 ArrayRef<uint8_t> ShadowBytes,
3338 IRBuilder<> &IRB, Value *ShadowBase) {
3339 copyToShadow(ShadowMask, ShadowBytes, Begin: 0, End: ShadowMask.size(), IRB, ShadowBase);
3340}
3341
3342void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
3343 ArrayRef<uint8_t> ShadowBytes,
3344 size_t Begin, size_t End,
3345 IRBuilder<> &IRB, Value *ShadowBase) {
3346 assert(ShadowMask.size() == ShadowBytes.size());
3347 size_t Done = Begin;
3348 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
3349 if (!ShadowMask[i]) {
3350 assert(!ShadowBytes[i]);
3351 continue;
3352 }
3353 uint8_t Val = ShadowBytes[i];
3354 if (!AsanSetShadowFunc[Val])
3355 continue;
3356
3357 // Skip same values.
3358 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
3359 }
3360
3361 if (j - i >= ASan.MaxInlinePoisoningSize) {
3362 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End: i, IRB, ShadowBase);
3363 RTCI.createRuntimeCall(
3364 IRB, Callee: AsanSetShadowFunc[Val],
3365 Args: {IRB.CreateAdd(LHS: ShadowBase, RHS: ConstantInt::get(Ty: IntptrTy, V: i)),
3366 ConstantInt::get(Ty: IntptrTy, V: j - i)});
3367 Done = j;
3368 }
3369 }
3370
3371 copyToShadowInline(ShadowMask, ShadowBytes, Begin: Done, End, IRB, ShadowBase);
3372}
3373
3374// Fake stack allocator (asan_fake_stack.h) has 11 size classes
3375// for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
3376static int StackMallocSizeClass(uint64_t LocalStackSize) {
3377 assert(LocalStackSize <= kMaxStackMallocSize);
3378 uint64_t MaxSize = kMinStackMallocSize;
3379 for (int i = 0;; i++, MaxSize *= 2)
3380 if (LocalStackSize <= MaxSize) return i;
3381 llvm_unreachable("impossible LocalStackSize");
3382}
3383
3384void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
3385 Instruction *CopyInsertPoint = &F.front().front();
3386 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
3387 // Insert after the dynamic shadow location is determined
3388 CopyInsertPoint = CopyInsertPoint->getNextNode();
3389 assert(CopyInsertPoint);
3390 }
3391 IRBuilder<> IRB(CopyInsertPoint);
3392 const DataLayout &DL = F.getDataLayout();
3393 for (Argument &Arg : F.args()) {
3394 if (Arg.hasByValAttr()) {
3395 Type *Ty = Arg.getParamByValType();
3396 const Align Alignment =
3397 DL.getValueOrABITypeAlignment(Alignment: Arg.getParamAlign(), Ty);
3398
3399 AllocaInst *AI = IRB.CreateAlloca(
3400 Ty, ArraySize: nullptr,
3401 Name: (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
3402 ".byval");
3403 AI->setAlignment(Alignment);
3404 Arg.replaceAllUsesWith(V: AI);
3405
3406 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3407 IRB.CreateMemCpy(Dst: AI, DstAlign: Alignment, Src: &Arg, SrcAlign: Alignment, Size: AllocSize);
3408 }
3409 }
3410}
3411
3412PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
3413 Value *ValueIfTrue,
3414 Instruction *ThenTerm,
3415 Value *ValueIfFalse) {
3416 PHINode *PHI = IRB.CreatePHI(Ty: ValueIfTrue->getType(), NumReservedValues: 2);
3417 BasicBlock *CondBlock = cast<Instruction>(Val: Cond)->getParent();
3418 PHI->addIncoming(V: ValueIfFalse, BB: CondBlock);
3419 BasicBlock *ThenBlock = ThenTerm->getParent();
3420 PHI->addIncoming(V: ValueIfTrue, BB: ThenBlock);
3421 return PHI;
3422}
3423
3424Value *FunctionStackPoisoner::createAllocaForLayout(
3425 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
3426 AllocaInst *Alloca;
3427 if (Dynamic) {
3428 Alloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(),
3429 ArraySize: ConstantInt::get(Ty: IRB.getInt64Ty(), V: L.FrameSize),
3430 Name: "MyAlloca");
3431 } else {
3432 Alloca = IRB.CreateAlloca(Ty: ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: L.FrameSize),
3433 ArraySize: nullptr, Name: "MyAlloca");
3434 assert(Alloca->isStaticAlloca());
3435 }
3436 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
3437 uint64_t FrameAlignment = std::max(a: L.FrameAlignment, b: uint64_t(ClRealignStack));
3438 Alloca->setAlignment(Align(FrameAlignment));
3439 return Alloca;
3440}
3441
3442void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
3443 BasicBlock &FirstBB = *F.begin();
3444 IRBuilder<> IRB(dyn_cast<Instruction>(Val: FirstBB.begin()));
3445 DynamicAllocaLayout = IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr);
3446 IRB.CreateStore(Val: Constant::getNullValue(Ty: IntptrTy), Ptr: DynamicAllocaLayout);
3447 DynamicAllocaLayout->setAlignment(Align(32));
3448}
3449
3450void FunctionStackPoisoner::processDynamicAllocas() {
3451 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
3452 assert(DynamicAllocaPoisonCallVec.empty());
3453 return;
3454 }
3455
3456 // Insert poison calls for lifetime intrinsics for dynamic allocas.
3457 for (const auto &APC : DynamicAllocaPoisonCallVec) {
3458 assert(APC.InsBefore);
3459 assert(APC.AI);
3460 assert(ASan.isInterestingAlloca(*APC.AI));
3461 assert(!APC.AI->isStaticAlloca());
3462
3463 IRBuilder<> IRB(APC.InsBefore);
3464 poisonAlloca(V: APC.AI, Size: APC.Size, IRB, DoPoison: APC.DoPoison);
3465 // Dynamic allocas will be unpoisoned unconditionally below in
3466 // unpoisonDynamicAllocas.
3467 // Flag that we need unpoison static allocas.
3468 }
3469
3470 // Handle dynamic allocas.
3471 createDynamicAllocasInitStorage();
3472 for (auto &AI : DynamicAllocaVec)
3473 handleDynamicAllocaCall(AI);
3474 unpoisonDynamicAllocas();
3475}
3476
3477/// Collect instructions in the entry block after \p InsBefore which initialize
3478/// permanent storage for a function argument. These instructions must remain in
3479/// the entry block so that uninitialized values do not appear in backtraces. An
3480/// added benefit is that this conserves spill slots. This does not move stores
3481/// before instrumented / "interesting" allocas.
3482static void findStoresToUninstrumentedArgAllocas(
3483 AddressSanitizer &ASan, Instruction &InsBefore,
3484 SmallVectorImpl<Instruction *> &InitInsts) {
3485 Instruction *Start = InsBefore.getNextNode();
3486 for (Instruction *It = Start; It; It = It->getNextNode()) {
3487 // Argument initialization looks like:
3488 // 1) store <Argument>, <Alloca> OR
3489 // 2) <CastArgument> = cast <Argument> to ...
3490 // store <CastArgument> to <Alloca>
3491 // Do not consider any other kind of instruction.
3492 //
3493 // Note: This covers all known cases, but may not be exhaustive. An
3494 // alternative to pattern-matching stores is to DFS over all Argument uses:
3495 // this might be more general, but is probably much more complicated.
3496 if (isa<AllocaInst>(Val: It) || isa<CastInst>(Val: It))
3497 continue;
3498 if (auto *Store = dyn_cast<StoreInst>(Val: It)) {
3499 // The store destination must be an alloca that isn't interesting for
3500 // ASan to instrument. These are moved up before InsBefore, and they're
3501 // not interesting because allocas for arguments can be mem2reg'd.
3502 auto *Alloca = dyn_cast<AllocaInst>(Val: Store->getPointerOperand());
3503 if (!Alloca || ASan.isInterestingAlloca(AI: *Alloca))
3504 continue;
3505
3506 Value *Val = Store->getValueOperand();
3507 bool IsDirectArgInit = isa<Argument>(Val);
3508 bool IsArgInitViaCast =
3509 isa<CastInst>(Val) &&
3510 isa<Argument>(Val: cast<CastInst>(Val)->getOperand(i_nocapture: 0)) &&
3511 // Check that the cast appears directly before the store. Otherwise
3512 // moving the cast before InsBefore may break the IR.
3513 Val == It->getPrevNode();
3514 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3515 if (!IsArgInit)
3516 continue;
3517
3518 if (IsArgInitViaCast)
3519 InitInsts.push_back(Elt: cast<Instruction>(Val));
3520 InitInsts.push_back(Elt: Store);
3521 continue;
3522 }
3523
3524 // Do not reorder past unknown instructions: argument initialization should
3525 // only involve casts and stores.
3526 return;
3527 }
3528}
3529
3530static StringRef getAllocaName(AllocaInst *AI) {
3531 // Alloca could have been renamed for uniqueness. Its true name will have been
3532 // recorded as an annotation.
3533 if (AI->hasMetadata(KindID: LLVMContext::MD_annotation)) {
3534 MDTuple *AllocaAnnotations =
3535 cast<MDTuple>(Val: AI->getMetadata(KindID: LLVMContext::MD_annotation));
3536 for (auto &Annotation : AllocaAnnotations->operands()) {
3537 if (!isa<MDTuple>(Val: Annotation))
3538 continue;
3539 auto AnnotationTuple = cast<MDTuple>(Val: Annotation);
3540 for (unsigned Index = 0; Index < AnnotationTuple->getNumOperands();
3541 Index++) {
3542 // All annotations are strings
3543 auto MetadataString =
3544 cast<MDString>(Val: AnnotationTuple->getOperand(I: Index));
3545 if (MetadataString->getString() == "alloca_name_altered")
3546 return cast<MDString>(Val: AnnotationTuple->getOperand(I: Index + 1))
3547 ->getString();
3548 }
3549 }
3550 }
3551 return AI->getName();
3552}
3553
3554void FunctionStackPoisoner::processStaticAllocas() {
3555 if (AllocaVec.empty()) {
3556 assert(StaticAllocaPoisonCallVec.empty());
3557 return;
3558 }
3559
3560 int StackMallocIdx = -1;
3561 DebugLoc EntryDebugLocation;
3562 if (auto SP = F.getSubprogram())
3563 EntryDebugLocation =
3564 DILocation::get(Context&: SP->getContext(), Line: SP->getScopeLine(), Column: 0, Scope: SP);
3565
3566 Instruction *InsBefore = AllocaVec[0];
3567 IRBuilder<> IRB(InsBefore);
3568
3569 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3570 // debug info is broken, because only entry-block allocas are treated as
3571 // regular stack slots.
3572 auto InsBeforeB = InsBefore->getParent();
3573 assert(InsBeforeB == &F.getEntryBlock());
3574 for (auto *AI : StaticAllocasToMoveUp)
3575 if (AI->getParent() == InsBeforeB)
3576 AI->moveBefore(InsertPos: InsBefore->getIterator());
3577
3578 // Move stores of arguments into entry-block allocas as well. This prevents
3579 // extra stack slots from being generated (to house the argument values until
3580 // they can be stored into the allocas). This also prevents uninitialized
3581 // values from being shown in backtraces.
3582 SmallVector<Instruction *, 8> ArgInitInsts;
3583 findStoresToUninstrumentedArgAllocas(ASan, InsBefore&: *InsBefore, InitInsts&: ArgInitInsts);
3584 for (Instruction *ArgInitInst : ArgInitInsts)
3585 ArgInitInst->moveBefore(InsertPos: InsBefore->getIterator());
3586
3587 // If we have a call to llvm.localescape, keep it in the entry block.
3588 if (LocalEscapeCall)
3589 LocalEscapeCall->moveBefore(InsertPos: InsBefore->getIterator());
3590
3591 SmallVector<ASanStackVariableDescription, 16> SVD;
3592 SVD.reserve(N: AllocaVec.size());
3593 for (AllocaInst *AI : AllocaVec) {
3594 StringRef Name = getAllocaName(AI);
3595 ASanStackVariableDescription D = {.Name: Name.data(),
3596 .Size: ASan.getAllocaSizeInBytes(AI: *AI),
3597 .LifetimeSize: 0,
3598 .Alignment: AI->getAlign().value(),
3599 .AI: AI,
3600 .Offset: 0,
3601 .Line: 0};
3602 SVD.push_back(Elt: D);
3603 }
3604
3605 // Minimal header size (left redzone) is 4 pointers,
3606 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3607 uint64_t Granularity = 1ULL << Mapping.Scale;
3608 uint64_t MinHeaderSize = std::max(a: (uint64_t)ASan.LongSize / 2, b: Granularity);
3609 const ASanStackFrameLayout &L =
3610 ComputeASanStackFrameLayout(Vars&: SVD, Granularity, MinHeaderSize);
3611
3612 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3613 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3614 for (auto &Desc : SVD)
3615 AllocaToSVDMap[Desc.AI] = &Desc;
3616
3617 // Update SVD with information from lifetime intrinsics.
3618 for (const auto &APC : StaticAllocaPoisonCallVec) {
3619 assert(APC.InsBefore);
3620 assert(APC.AI);
3621 assert(ASan.isInterestingAlloca(*APC.AI));
3622 assert(APC.AI->isStaticAlloca());
3623
3624 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3625 Desc.LifetimeSize = Desc.Size;
3626 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3627 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3628 if (LifetimeLoc->getFile() == FnLoc->getFile())
3629 if (unsigned Line = LifetimeLoc->getLine())
3630 Desc.Line = std::min(a: Desc.Line ? Desc.Line : Line, b: Line);
3631 }
3632 }
3633 }
3634
3635 auto DescriptionString = ComputeASanStackFrameDescription(Vars: SVD);
3636 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3637 uint64_t LocalStackSize = L.FrameSize;
3638 bool DoStackMalloc =
3639 ASan.UseAfterReturn != AsanDetectStackUseAfterReturnMode::Never &&
3640 !ASan.CompileKernel && LocalStackSize <= kMaxStackMallocSize;
3641 bool DoDynamicAlloca = ClDynamicAllocaStack;
3642 // Don't do dynamic alloca or stack malloc if:
3643 // 1) There is inline asm: too often it makes assumptions on which registers
3644 // are available.
3645 // 2) There is a returns_twice call (typically setjmp), which is
3646 // optimization-hostile, and doesn't play well with introduced indirect
3647 // register-relative calculation of local variable addresses.
3648 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3649 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3650
3651 Type *PtrTy = F.getDataLayout().getAllocaPtrType(Ctx&: F.getContext());
3652 Value *StaticAlloca =
3653 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, Dynamic: false);
3654
3655 Value *FakeStackPtr;
3656 Value *FakeStackInt;
3657 Value *LocalStackBase;
3658 Value *LocalStackBaseAlloca;
3659 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3660
3661 if (DoStackMalloc) {
3662 LocalStackBaseAlloca =
3663 IRB.CreateAlloca(Ty: IntptrTy, ArraySize: nullptr, Name: "asan_local_stack_base");
3664 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3665 // void *FakeStack = __asan_option_detect_stack_use_after_return
3666 // ? __asan_stack_malloc_N(LocalStackSize)
3667 // : nullptr;
3668 // void *LocalStackBase = (FakeStack) ? FakeStack :
3669 // alloca(LocalStackSize);
3670 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3671 Name: kAsanOptionDetectUseAfterReturn, Ty: IRB.getInt32Ty());
3672 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3673 LHS: IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: OptionDetectUseAfterReturn),
3674 RHS: Constant::getNullValue(Ty: IRB.getInt32Ty()));
3675 Instruction *Term =
3676 SplitBlockAndInsertIfThen(Cond: UseAfterReturnIsEnabled, SplitBefore: InsBefore, Unreachable: false);
3677 IRBuilder<> IRBIf(Term);
3678 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3679 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3680 Value *FakeStackValue =
3681 RTCI.createRuntimeCall(IRB&: IRBIf, Callee: AsanStackMallocFunc[StackMallocIdx],
3682 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3683 IRB.SetInsertPoint(InsBefore);
3684 FakeStackInt = createPHI(IRB, Cond: UseAfterReturnIsEnabled, ValueIfTrue: FakeStackValue,
3685 ThenTerm: Term, ValueIfFalse: ConstantInt::get(Ty: IntptrTy, V: 0));
3686 } else {
3687 // assert(ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode:Always)
3688 // void *FakeStack = __asan_stack_malloc_N(LocalStackSize);
3689 // void *LocalStackBase = (FakeStack) ? FakeStack :
3690 // alloca(LocalStackSize);
3691 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3692 FakeStackInt =
3693 RTCI.createRuntimeCall(IRB, Callee: AsanStackMallocFunc[StackMallocIdx],
3694 Args: ConstantInt::get(Ty: IntptrTy, V: LocalStackSize));
3695 }
3696 FakeStackPtr = IRB.CreateIntToPtr(V: FakeStackInt, DestTy: PtrTy);
3697 Value *NoFakeStack =
3698 IRB.CreateICmpEQ(LHS: FakeStackInt, RHS: Constant::getNullValue(Ty: IntptrTy));
3699 Instruction *Term =
3700 SplitBlockAndInsertIfThen(Cond: NoFakeStack, SplitBefore: InsBefore, Unreachable: false);
3701 IRBuilder<> IRBIf(Term);
3702 Value *AllocaValue =
3703 DoDynamicAlloca ? createAllocaForLayout(IRB&: IRBIf, L, Dynamic: true) : StaticAlloca;
3704
3705 IRB.SetInsertPoint(InsBefore);
3706 LocalStackBase =
3707 createPHI(IRB, Cond: NoFakeStack, ValueIfTrue: AllocaValue, ThenTerm: Term, ValueIfFalse: FakeStackPtr);
3708 IRB.CreateStore(Val: LocalStackBase, Ptr: LocalStackBaseAlloca);
3709 DIExprFlags |= DIExpression::DerefBefore;
3710 } else {
3711 // void *FakeStack = nullptr;
3712 // void *LocalStackBase = alloca(LocalStackSize);
3713 FakeStackInt = Constant::getNullValue(Ty: IntptrTy);
3714 FakeStackPtr = Constant::getNullValue(Ty: PtrTy);
3715 LocalStackBase =
3716 DoDynamicAlloca ? createAllocaForLayout(IRB, L, Dynamic: true) : StaticAlloca;
3717 LocalStackBaseAlloca = LocalStackBase;
3718 }
3719
3720 // Replace Alloca instructions with base+offset.
3721 SmallVector<Value *> NewAllocaPtrs;
3722 for (const auto &Desc : SVD) {
3723 AllocaInst *AI = Desc.AI;
3724 replaceDbgDeclare(Address: AI, NewAddress: LocalStackBaseAlloca, Builder&: DIB, DIExprFlags, Offset: Desc.Offset);
3725 Value *NewAllocaPtr = IRB.CreatePtrAdd(
3726 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: Desc.Offset));
3727 AI->replaceAllUsesWith(V: NewAllocaPtr);
3728 NewAllocaPtrs.push_back(Elt: NewAllocaPtr);
3729 }
3730
3731 // The left-most redzone has enough space for at least 4 pointers.
3732 // Write the Magic value to redzone[0].
3733 IRB.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kCurrentStackFrameMagic),
3734 Ptr: LocalStackBase);
3735 // Write the frame description constant to redzone[1].
3736 Value *BasePlus1 = IRB.CreatePtrAdd(
3737 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: ASan.LongSize / 8));
3738 GlobalVariable *StackDescriptionGlobal =
3739 createPrivateGlobalForString(M&: *F.getParent(), Str: DescriptionString,
3740 /*AllowMerging*/ true, NamePrefix: genName(suffix: "stack"));
3741 Value *Description = IRB.CreatePointerCast(V: StackDescriptionGlobal, DestTy: IntptrTy);
3742 IRB.CreateStore(Val: Description, Ptr: BasePlus1);
3743 // Write the PC to redzone[2].
3744 Value *BasePlus2 = IRB.CreatePtrAdd(
3745 Ptr: LocalStackBase, Offset: ConstantInt::get(Ty: IntptrTy, V: 2 * ASan.LongSize / 8));
3746 IRB.CreateStore(Val: IRB.CreatePointerCast(V: &F, DestTy: IntptrTy), Ptr: BasePlus2);
3747
3748 const auto &ShadowAfterScope = GetShadowBytesAfterScope(Vars: SVD, Layout: L);
3749
3750 // Poison the stack red zones at the entry.
3751 Value *ShadowBase =
3752 ASan.memToShadow(Shadow: IRB.CreatePtrToInt(V: LocalStackBase, DestTy: IntptrTy), IRB);
3753 // As mask we must use most poisoned case: red zones and after scope.
3754 // As bytes we can use either the same or just red zones only.
3755 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowAfterScope, IRB, ShadowBase);
3756
3757 if (!StaticAllocaPoisonCallVec.empty()) {
3758 const auto &ShadowInScope = GetShadowBytes(Vars: SVD, Layout: L);
3759
3760 // Poison static allocas near lifetime intrinsics.
3761 for (const auto &APC : StaticAllocaPoisonCallVec) {
3762 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3763 assert(Desc.Offset % L.Granularity == 0);
3764 size_t Begin = Desc.Offset / L.Granularity;
3765 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3766
3767 IRBuilder<> IRB(APC.InsBefore);
3768 copyToShadow(ShadowMask: ShadowAfterScope,
3769 ShadowBytes: APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3770 IRB, ShadowBase);
3771 }
3772 }
3773
3774 // Remove lifetime markers now that these are no longer allocas.
3775 for (Value *NewAllocaPtr : NewAllocaPtrs) {
3776 for (User *U : make_early_inc_range(Range: NewAllocaPtr->users())) {
3777 auto *I = cast<Instruction>(Val: U);
3778 if (I->isLifetimeStartOrEnd())
3779 I->eraseFromParent();
3780 }
3781 }
3782
3783 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3784 SmallVector<uint8_t, 64> ShadowAfterReturn;
3785
3786 // (Un)poison the stack before all ret instructions.
3787 for (Instruction *Ret : RetVec) {
3788 IRBuilder<> IRBRet(Ret);
3789 // Mark the current frame as retired.
3790 IRBRet.CreateStore(Val: ConstantInt::get(Ty: IntptrTy, V: kRetiredStackFrameMagic),
3791 Ptr: LocalStackBase);
3792 if (DoStackMalloc) {
3793 assert(StackMallocIdx >= 0);
3794 // if FakeStack != 0 // LocalStackBase == FakeStack
3795 // // In use-after-return mode, poison the whole stack frame.
3796 // if StackMallocIdx <= 4
3797 // // For small sizes inline the whole thing:
3798 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3799 // **SavedFlagPtr(FakeStack) = 0
3800 // else
3801 // __asan_stack_free_N(FakeStack, LocalStackSize)
3802 // else
3803 // <This is not a fake stack; unpoison the redzones>
3804 Value *Cmp =
3805 IRBRet.CreateICmpNE(LHS: FakeStackInt, RHS: Constant::getNullValue(Ty: IntptrTy));
3806 Instruction *ThenTerm, *ElseTerm;
3807 SplitBlockAndInsertIfThenElse(Cond: Cmp, SplitBefore: Ret, ThenTerm: &ThenTerm, ElseTerm: &ElseTerm);
3808
3809 IRBuilder<> IRBPoison(ThenTerm);
3810 if (ASan.MaxInlinePoisoningSize != 0 && StackMallocIdx <= 4) {
3811 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3812 ShadowAfterReturn.resize(N: ClassSize / L.Granularity,
3813 NV: kAsanStackUseAfterReturnMagic);
3814 copyToShadow(ShadowMask: ShadowAfterReturn, ShadowBytes: ShadowAfterReturn, IRB&: IRBPoison,
3815 ShadowBase);
3816 Value *SavedFlagPtrPtr = IRBPoison.CreatePtrAdd(
3817 Ptr: FakeStackPtr,
3818 Offset: ConstantInt::get(Ty: IntptrTy, V: ClassSize - ASan.LongSize / 8));
3819 Value *SavedFlagPtr = IRBPoison.CreateLoad(Ty: IntptrTy, Ptr: SavedFlagPtrPtr);
3820 IRBPoison.CreateStore(
3821 Val: Constant::getNullValue(Ty: IRBPoison.getInt8Ty()),
3822 Ptr: IRBPoison.CreateIntToPtr(V: SavedFlagPtr, DestTy: IRBPoison.getPtrTy()));
3823 } else {
3824 // For larger frames call __asan_stack_free_*.
3825 RTCI.createRuntimeCall(
3826 IRB&: IRBPoison, Callee: AsanStackFreeFunc[StackMallocIdx],
3827 Args: {FakeStackInt, ConstantInt::get(Ty: IntptrTy, V: LocalStackSize)});
3828 }
3829
3830 IRBuilder<> IRBElse(ElseTerm);
3831 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBElse, ShadowBase);
3832 } else {
3833 copyToShadow(ShadowMask: ShadowAfterScope, ShadowBytes: ShadowClean, IRB&: IRBRet, ShadowBase);
3834 }
3835 }
3836
3837 // We are done. Remove the old unused alloca instructions.
3838 for (auto *AI : AllocaVec)
3839 AI->eraseFromParent();
3840}
3841
3842void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3843 IRBuilder<> &IRB, bool DoPoison) {
3844 // For now just insert the call to ASan runtime.
3845 Value *AddrArg = IRB.CreatePointerCast(V, DestTy: IntptrTy);
3846 Value *SizeArg = ConstantInt::get(Ty: IntptrTy, V: Size);
3847 RTCI.createRuntimeCall(
3848 IRB, Callee: DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3849 Args: {AddrArg, SizeArg});
3850}
3851
3852// Handling llvm.lifetime intrinsics for a given %alloca:
3853// (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3854// (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3855// invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3856// could be poisoned by previous llvm.lifetime.end instruction, as the
3857// variable may go in and out of scope several times, e.g. in loops).
3858// (3) if we poisoned at least one %alloca in a function,
3859// unpoison the whole stack frame at function exit.
3860void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3861 IRBuilder<> IRB(AI);
3862
3863 const Align Alignment = std::max(a: Align(kAllocaRzSize), b: AI->getAlign());
3864 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3865
3866 Value *Zero = Constant::getNullValue(Ty: IntptrTy);
3867 Value *AllocaRzSize = ConstantInt::get(Ty: IntptrTy, V: kAllocaRzSize);
3868 Value *AllocaRzMask = ConstantInt::get(Ty: IntptrTy, V: AllocaRedzoneMask);
3869
3870 // Since we need to extend alloca with additional memory to locate
3871 // redzones, and OldSize is number of allocated blocks with
3872 // ElementSize size, get allocated memory size in bytes by
3873 // OldSize * ElementSize.
3874 Value *OldSize = IRB.CreateAllocationSize(DestTy: IntptrTy, AI);
3875
3876 // PartialSize = OldSize % 32
3877 Value *PartialSize = IRB.CreateAnd(LHS: OldSize, RHS: AllocaRzMask);
3878
3879 // Misalign = kAllocaRzSize - PartialSize;
3880 Value *Misalign = IRB.CreateSub(LHS: AllocaRzSize, RHS: PartialSize);
3881
3882 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3883 Value *Cond = IRB.CreateICmpNE(LHS: Misalign, RHS: AllocaRzSize);
3884 Value *PartialPadding = IRB.CreateSelect(C: Cond, True: Misalign, False: Zero);
3885
3886 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3887 // Alignment is added to locate left redzone, PartialPadding for possible
3888 // partial redzone and kAllocaRzSize for right redzone respectively.
3889 Value *AdditionalChunkSize = IRB.CreateAdd(
3890 LHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value() + kAllocaRzSize),
3891 RHS: PartialPadding);
3892
3893 Value *NewSize = IRB.CreateAdd(LHS: OldSize, RHS: AdditionalChunkSize);
3894
3895 // Insert new alloca with new NewSize and Alignment params.
3896 AllocaInst *NewAlloca = IRB.CreateAlloca(Ty: IRB.getInt8Ty(), ArraySize: NewSize);
3897 NewAlloca->setAlignment(Alignment);
3898
3899 // NewAddress = Address + Alignment
3900 Value *NewAddress =
3901 IRB.CreateAdd(LHS: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy),
3902 RHS: ConstantInt::get(Ty: IntptrTy, V: Alignment.value()));
3903
3904 // Insert __asan_alloca_poison call for new created alloca.
3905 RTCI.createRuntimeCall(IRB, Callee: AsanAllocaPoisonFunc, Args: {NewAddress, OldSize});
3906
3907 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3908 // for unpoisoning stuff.
3909 IRB.CreateStore(Val: IRB.CreatePtrToInt(V: NewAlloca, DestTy: IntptrTy), Ptr: DynamicAllocaLayout);
3910
3911 Value *NewAddressPtr = IRB.CreateIntToPtr(V: NewAddress, DestTy: AI->getType());
3912
3913 // Remove lifetime markers now that this is no longer an alloca.
3914 for (User *U : make_early_inc_range(Range: AI->users())) {
3915 auto *I = cast<Instruction>(Val: U);
3916 if (I->isLifetimeStartOrEnd())
3917 I->eraseFromParent();
3918 }
3919
3920 // Replace all uses of AddressReturnedByAlloca with NewAddressPtr.
3921 AI->replaceAllUsesWith(V: NewAddressPtr);
3922
3923 // We are done. Erase old alloca from parent.
3924 AI->eraseFromParent();
3925}
3926
3927// isSafeAccess returns true if Addr is always inbounds with respect to its
3928// base object. For example, it is a field access or an array access with
3929// constant inbounds index.
3930bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3931 Value *Addr, TypeSize TypeStoreSize) const {
3932 if (TypeStoreSize.isScalable())
3933 // TODO: We can use vscale_range to convert a scalable value to an
3934 // upper bound on the access size.
3935 return false;
3936
3937 SizeOffsetAPInt SizeOffset = ObjSizeVis.compute(V: Addr);
3938 if (!SizeOffset.bothKnown())
3939 return false;
3940
3941 uint64_t Size = SizeOffset.Size.getZExtValue();
3942 int64_t Offset = SizeOffset.Offset.getSExtValue();
3943
3944 // Three checks are required to ensure safety:
3945 // . Offset >= 0 (since the offset is given from the base ptr)
3946 // . Size >= Offset (unsigned)
3947 // . Size - Offset >= NeededSize (unsigned)
3948 return Offset >= 0 && Size >= uint64_t(Offset) &&
3949 Size - uint64_t(Offset) >= TypeStoreSize / 8;
3950}
3951