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 sanity checker.
10 // Details of the algorithm:
11 // https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm
12 //
13 // FIXME: This sanitizer does not yet handle scalable vectors
14 //
15 //===----------------------------------------------------------------------===//
16
17 #include "llvm/Transforms/Instrumentation/AddressSanitizer.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/ADT/Triple.h"
27 #include "llvm/ADT/Twine.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/BinaryFormat/MachO.h"
32 #include "llvm/IR/Argument.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/Comdat.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DIBuilder.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DebugInfoMetadata.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/GlobalAlias.h"
46 #include "llvm/IR/GlobalValue.h"
47 #include "llvm/IR/GlobalVariable.h"
48 #include "llvm/IR/IRBuilder.h"
49 #include "llvm/IR/InlineAsm.h"
50 #include "llvm/IR/InstVisitor.h"
51 #include "llvm/IR/InstrTypes.h"
52 #include "llvm/IR/Instruction.h"
53 #include "llvm/IR/Instructions.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Intrinsics.h"
56 #include "llvm/IR/LLVMContext.h"
57 #include "llvm/IR/MDBuilder.h"
58 #include "llvm/IR/Metadata.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/Use.h"
62 #include "llvm/IR/Value.h"
63 #include "llvm/InitializePasses.h"
64 #include "llvm/MC/MCSectionMachO.h"
65 #include "llvm/Pass.h"
66 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/ErrorHandling.h"
70 #include "llvm/Support/MathExtras.h"
71 #include "llvm/Support/ScopedPrinter.h"
72 #include "llvm/Support/raw_ostream.h"
73 #include "llvm/Transforms/Instrumentation.h"
74 #include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h"
75 #include "llvm/Transforms/Utils/ASanStackFrameLayout.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
78 #include "llvm/Transforms/Utils/ModuleUtils.h"
79 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
80 #include <algorithm>
81 #include <cassert>
82 #include <cstddef>
83 #include <cstdint>
84 #include <iomanip>
85 #include <limits>
86 #include <memory>
87 #include <sstream>
88 #include <string>
89 #include <tuple>
90
91 using namespace llvm;
92
93 #define DEBUG_TYPE "asan"
94
95 static const uint64_t kDefaultShadowScale = 3;
96 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
97 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
98 static const uint64_t kDynamicShadowSentinel =
99 std::numeric_limits<uint64_t>::max();
100 static const uint64_t kSmallX86_64ShadowOffsetBase = 0x7FFFFFFF; // < 2G.
101 static const uint64_t kSmallX86_64ShadowOffsetAlignMask = ~0xFFFULL;
102 static const uint64_t kLinuxKasan_ShadowOffset64 = 0xdffffc0000000000;
103 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 44;
104 static const uint64_t kSystemZ_ShadowOffset64 = 1ULL << 52;
105 static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa0000;
106 static const uint64_t kMIPS64_ShadowOffset64 = 1ULL << 37;
107 static const uint64_t kAArch64_ShadowOffset64 = 1ULL << 36;
108 static const uint64_t kRISCV64_ShadowOffset64 = 0x20000000;
109 static const uint64_t kFreeBSD_ShadowOffset32 = 1ULL << 30;
110 static const uint64_t kFreeBSD_ShadowOffset64 = 1ULL << 46;
111 static const uint64_t kNetBSD_ShadowOffset32 = 1ULL << 30;
112 static const uint64_t kNetBSD_ShadowOffset64 = 1ULL << 46;
113 static const uint64_t kNetBSDKasan_ShadowOffset64 = 0xdfff900000000000;
114 static const uint64_t kPS4CPU_ShadowOffset64 = 1ULL << 40;
115 static const uint64_t kWindowsShadowOffset32 = 3ULL << 28;
116 static const uint64_t kEmscriptenShadowOffset = 0;
117
118 static const uint64_t kMyriadShadowScale = 5;
119 static const uint64_t kMyriadMemoryOffset32 = 0x80000000ULL;
120 static const uint64_t kMyriadMemorySize32 = 0x20000000ULL;
121 static const uint64_t kMyriadTagShift = 29;
122 static const uint64_t kMyriadDDRTag = 4;
123 static const uint64_t kMyriadCacheBitMask32 = 0x40000000ULL;
124
125 // The shadow memory space is dynamically allocated.
126 static const uint64_t kWindowsShadowOffset64 = kDynamicShadowSentinel;
127
128 static const size_t kMinStackMallocSize = 1 << 6; // 64B
129 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
130 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
131 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
132
133 static const char *const kAsanModuleCtorName = "asan.module_ctor";
134 static const char *const kAsanModuleDtorName = "asan.module_dtor";
135 static const uint64_t kAsanCtorAndDtorPriority = 1;
136 // On Emscripten, the system needs more than one priorities for constructors.
137 static const uint64_t kAsanEmscriptenCtorAndDtorPriority = 50;
138 static const char *const kAsanReportErrorTemplate = "__asan_report_";
139 static const char *const kAsanRegisterGlobalsName = "__asan_register_globals";
140 static const char *const kAsanUnregisterGlobalsName =
141 "__asan_unregister_globals";
142 static const char *const kAsanRegisterImageGlobalsName =
143 "__asan_register_image_globals";
144 static const char *const kAsanUnregisterImageGlobalsName =
145 "__asan_unregister_image_globals";
146 static const char *const kAsanRegisterElfGlobalsName =
147 "__asan_register_elf_globals";
148 static const char *const kAsanUnregisterElfGlobalsName =
149 "__asan_unregister_elf_globals";
150 static const char *const kAsanPoisonGlobalsName = "__asan_before_dynamic_init";
151 static const char *const kAsanUnpoisonGlobalsName = "__asan_after_dynamic_init";
152 static const char *const kAsanInitName = "__asan_init";
153 static const char *const kAsanVersionCheckNamePrefix =
154 "__asan_version_mismatch_check_v";
155 static const char *const kAsanPtrCmp = "__sanitizer_ptr_cmp";
156 static const char *const kAsanPtrSub = "__sanitizer_ptr_sub";
157 static const char *const kAsanHandleNoReturnName = "__asan_handle_no_return";
158 static const int kMaxAsanStackMallocSizeClass = 10;
159 static const char *const kAsanStackMallocNameTemplate = "__asan_stack_malloc_";
160 static const char *const kAsanStackFreeNameTemplate = "__asan_stack_free_";
161 static const char *const kAsanGenPrefix = "___asan_gen_";
162 static const char *const kODRGenPrefix = "__odr_asan_gen_";
163 static const char *const kSanCovGenPrefix = "__sancov_gen_";
164 static const char *const kAsanSetShadowPrefix = "__asan_set_shadow_";
165 static const char *const kAsanPoisonStackMemoryName =
166 "__asan_poison_stack_memory";
167 static const char *const kAsanUnpoisonStackMemoryName =
168 "__asan_unpoison_stack_memory";
169
170 // ASan version script has __asan_* wildcard. Triple underscore prevents a
171 // linker (gold) warning about attempting to export a local symbol.
172 static const char *const kAsanGlobalsRegisteredFlagName =
173 "___asan_globals_registered";
174
175 static const char *const kAsanOptionDetectUseAfterReturn =
176 "__asan_option_detect_stack_use_after_return";
177
178 static const char *const kAsanShadowMemoryDynamicAddress =
179 "__asan_shadow_memory_dynamic_address";
180
181 static const char *const kAsanAllocaPoison = "__asan_alloca_poison";
182 static const char *const kAsanAllocasUnpoison = "__asan_allocas_unpoison";
183
184 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
185 static const size_t kNumberOfAccessSizes = 5;
186
187 static const unsigned kAllocaRzSize = 32;
188
189 // Command-line flags.
190
191 static cl::opt<bool> ClEnableKasan(
192 "asan-kernel", cl::desc("Enable KernelAddressSanitizer instrumentation"),
193 cl::Hidden, cl::init(false));
194
195 static cl::opt<bool> ClRecover(
196 "asan-recover",
197 cl::desc("Enable recovery mode (continue-after-error)."),
198 cl::Hidden, cl::init(false));
199
200 static cl::opt<bool> ClInsertVersionCheck(
201 "asan-guard-against-version-mismatch",
202 cl::desc("Guard against compiler/runtime version mismatch."),
203 cl::Hidden, cl::init(true));
204
205 // This flag may need to be replaced with -f[no-]asan-reads.
206 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
207 cl::desc("instrument read instructions"),
208 cl::Hidden, cl::init(true));
209
210 static cl::opt<bool> ClInstrumentWrites(
211 "asan-instrument-writes", cl::desc("instrument write instructions"),
212 cl::Hidden, cl::init(true));
213
214 static cl::opt<bool> ClInstrumentAtomics(
215 "asan-instrument-atomics",
216 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden,
217 cl::init(true));
218
219 static cl::opt<bool>
220 ClInstrumentByval("asan-instrument-byval",
221 cl::desc("instrument byval call arguments"), cl::Hidden,
222 cl::init(true));
223
224 static cl::opt<bool> ClAlwaysSlowPath(
225 "asan-always-slow-path",
226 cl::desc("use instrumentation with slow path for all accesses"), cl::Hidden,
227 cl::init(false));
228
229 static cl::opt<bool> ClForceDynamicShadow(
230 "asan-force-dynamic-shadow",
231 cl::desc("Load shadow address into a local variable for each function"),
232 cl::Hidden, cl::init(false));
233
234 static cl::opt<bool>
235 ClWithIfunc("asan-with-ifunc",
236 cl::desc("Access dynamic shadow through an ifunc global on "
237 "platforms that support this"),
238 cl::Hidden, cl::init(true));
239
240 static cl::opt<bool> ClWithIfuncSuppressRemat(
241 "asan-with-ifunc-suppress-remat",
242 cl::desc("Suppress rematerialization of dynamic shadow address by passing "
243 "it through inline asm in prologue."),
244 cl::Hidden, cl::init(true));
245
246 // This flag limits the number of instructions to be instrumented
247 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
248 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
249 // set it to 10000.
250 static cl::opt<int> ClMaxInsnsToInstrumentPerBB(
251 "asan-max-ins-per-bb", cl::init(10000),
252 cl::desc("maximal number of instructions to instrument in any given BB"),
253 cl::Hidden);
254
255 // This flag may need to be replaced with -f[no]asan-stack.
256 static cl::opt<bool> ClStack("asan-stack", cl::desc("Handle stack memory"),
257 cl::Hidden, cl::init(true));
258 static cl::opt<uint32_t> ClMaxInlinePoisoningSize(
259 "asan-max-inline-poisoning-size",
260 cl::desc(
261 "Inline shadow poisoning for blocks up to the given size in bytes."),
262 cl::Hidden, cl::init(64));
263
264 static cl::opt<bool> ClUseAfterReturn("asan-use-after-return",
265 cl::desc("Check stack-use-after-return"),
266 cl::Hidden, cl::init(true));
267
268 static cl::opt<bool> ClRedzoneByvalArgs("asan-redzone-byval-args",
269 cl::desc("Create redzones for byval "
270 "arguments (extra copy "
271 "required)"), cl::Hidden,
272 cl::init(true));
273
274 static cl::opt<bool> ClUseAfterScope("asan-use-after-scope",
275 cl::desc("Check stack-use-after-scope"),
276 cl::Hidden, cl::init(false));
277
278 // This flag may need to be replaced with -f[no]asan-globals.
279 static cl::opt<bool> ClGlobals("asan-globals",
280 cl::desc("Handle global objects"), cl::Hidden,
281 cl::init(true));
282
283 static cl::opt<bool> ClInitializers("asan-initialization-order",
284 cl::desc("Handle C++ initializer order"),
285 cl::Hidden, cl::init(true));
286
287 static cl::opt<bool> ClInvalidPointerPairs(
288 "asan-detect-invalid-pointer-pair",
289 cl::desc("Instrument <, <=, >, >=, - with pointer operands"), cl::Hidden,
290 cl::init(false));
291
292 static cl::opt<bool> ClInvalidPointerCmp(
293 "asan-detect-invalid-pointer-cmp",
294 cl::desc("Instrument <, <=, >, >= with pointer operands"), cl::Hidden,
295 cl::init(false));
296
297 static cl::opt<bool> ClInvalidPointerSub(
298 "asan-detect-invalid-pointer-sub",
299 cl::desc("Instrument - operations with pointer operands"), cl::Hidden,
300 cl::init(false));
301
302 static cl::opt<unsigned> ClRealignStack(
303 "asan-realign-stack",
304 cl::desc("Realign stack to the value of this flag (power of two)"),
305 cl::Hidden, cl::init(32));
306
307 static cl::opt<int> ClInstrumentationWithCallsThreshold(
308 "asan-instrumentation-with-call-threshold",
309 cl::desc(
310 "If the function being instrumented contains more than "
311 "this number of memory accesses, use callbacks instead of "
312 "inline checks (-1 means never use callbacks)."),
313 cl::Hidden, cl::init(7000));
314
315 static cl::opt<std::string> ClMemoryAccessCallbackPrefix(
316 "asan-memory-access-callback-prefix",
317 cl::desc("Prefix for memory access callbacks"), cl::Hidden,
318 cl::init("__asan_"));
319
320 static cl::opt<bool>
321 ClInstrumentDynamicAllocas("asan-instrument-dynamic-allocas",
322 cl::desc("instrument dynamic allocas"),
323 cl::Hidden, cl::init(true));
324
325 static cl::opt<bool> ClSkipPromotableAllocas(
326 "asan-skip-promotable-allocas",
327 cl::desc("Do not instrument promotable allocas"), cl::Hidden,
328 cl::init(true));
329
330 // These flags allow to change the shadow mapping.
331 // The shadow mapping looks like
332 // Shadow = (Mem >> scale) + offset
333
334 static cl::opt<int> ClMappingScale("asan-mapping-scale",
335 cl::desc("scale of asan shadow mapping"),
336 cl::Hidden, cl::init(0));
337
338 static cl::opt<uint64_t>
339 ClMappingOffset("asan-mapping-offset",
340 cl::desc("offset of asan shadow mapping [EXPERIMENTAL]"),
341 cl::Hidden, cl::init(0));
342
343 // Optimization flags. Not user visible, used mostly for testing
344 // and benchmarking the tool.
345
346 static cl::opt<bool> ClOpt("asan-opt", cl::desc("Optimize instrumentation"),
347 cl::Hidden, cl::init(true));
348
349 static cl::opt<bool> ClOptSameTemp(
350 "asan-opt-same-temp", cl::desc("Instrument the same temp just once"),
351 cl::Hidden, cl::init(true));
352
353 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
354 cl::desc("Don't instrument scalar globals"),
355 cl::Hidden, cl::init(true));
356
357 static cl::opt<bool> ClOptStack(
358 "asan-opt-stack", cl::desc("Don't instrument scalar stack variables"),
359 cl::Hidden, cl::init(false));
360
361 static cl::opt<bool> ClDynamicAllocaStack(
362 "asan-stack-dynamic-alloca",
363 cl::desc("Use dynamic alloca to represent stack variables"), cl::Hidden,
364 cl::init(true));
365
366 static cl::opt<uint32_t> ClForceExperiment(
367 "asan-force-experiment",
368 cl::desc("Force optimization experiment (for testing)"), cl::Hidden,
369 cl::init(0));
370
371 static cl::opt<bool>
372 ClUsePrivateAlias("asan-use-private-alias",
373 cl::desc("Use private aliases for global variables"),
374 cl::Hidden, cl::init(false));
375
376 static cl::opt<bool>
377 ClUseOdrIndicator("asan-use-odr-indicator",
378 cl::desc("Use odr indicators to improve ODR reporting"),
379 cl::Hidden, cl::init(false));
380
381 static cl::opt<bool>
382 ClUseGlobalsGC("asan-globals-live-support",
383 cl::desc("Use linker features to support dead "
384 "code stripping of globals"),
385 cl::Hidden, cl::init(true));
386
387 // This is on by default even though there is a bug in gold:
388 // https://sourceware.org/bugzilla/show_bug.cgi?id=19002
389 static cl::opt<bool>
390 ClWithComdat("asan-with-comdat",
391 cl::desc("Place ASan constructors in comdat sections"),
392 cl::Hidden, cl::init(true));
393
394 // Debug flags.
395
396 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
397 cl::init(0));
398
399 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
400 cl::Hidden, cl::init(0));
401
402 static cl::opt<std::string> ClDebugFunc("asan-debug-func", cl::Hidden,
403 cl::desc("Debug func"));
404
405 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
406 cl::Hidden, cl::init(-1));
407
408 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug max inst"),
409 cl::Hidden, cl::init(-1));
410
411 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
412 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
413 STATISTIC(NumOptimizedAccessesToGlobalVar,
414 "Number of optimized accesses to global vars");
415 STATISTIC(NumOptimizedAccessesToStackVar,
416 "Number of optimized accesses to stack vars");
417
418 namespace {
419
420 /// This struct defines the shadow mapping using the rule:
421 /// shadow = (mem >> Scale) ADD-or-OR Offset.
422 /// If InGlobal is true, then
423 /// extern char __asan_shadow[];
424 /// shadow = (mem >> Scale) + &__asan_shadow
425 struct ShadowMapping {
426 int Scale;
427 uint64_t Offset;
428 bool OrShadowOffset;
429 bool InGlobal;
430 };
431
432 } // end anonymous namespace
433
getShadowMapping(Triple & TargetTriple,int LongSize,bool IsKasan)434 static ShadowMapping getShadowMapping(Triple &TargetTriple, int LongSize,
435 bool IsKasan) {
436 bool IsAndroid = TargetTriple.isAndroid();
437 bool IsIOS = TargetTriple.isiOS() || TargetTriple.isWatchOS();
438 bool IsMacOS = TargetTriple.isMacOSX();
439 bool IsFreeBSD = TargetTriple.isOSFreeBSD();
440 bool IsNetBSD = TargetTriple.isOSNetBSD();
441 bool IsPS4CPU = TargetTriple.isPS4CPU();
442 bool IsLinux = TargetTriple.isOSLinux();
443 bool IsPPC64 = TargetTriple.getArch() == Triple::ppc64 ||
444 TargetTriple.getArch() == Triple::ppc64le;
445 bool IsSystemZ = TargetTriple.getArch() == Triple::systemz;
446 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64;
447 bool IsMIPS32 = TargetTriple.isMIPS32();
448 bool IsMIPS64 = TargetTriple.isMIPS64();
449 bool IsArmOrThumb = TargetTriple.isARM() || TargetTriple.isThumb();
450 bool IsAArch64 = TargetTriple.getArch() == Triple::aarch64;
451 bool IsRISCV64 = TargetTriple.getArch() == Triple::riscv64;
452 bool IsWindows = TargetTriple.isOSWindows();
453 bool IsFuchsia = TargetTriple.isOSFuchsia();
454 bool IsMyriad = TargetTriple.getVendor() == llvm::Triple::Myriad;
455 bool IsEmscripten = TargetTriple.isOSEmscripten();
456
457 ShadowMapping Mapping;
458
459 Mapping.Scale = IsMyriad ? kMyriadShadowScale : kDefaultShadowScale;
460 if (ClMappingScale.getNumOccurrences() > 0) {
461 Mapping.Scale = ClMappingScale;
462 }
463
464 if (LongSize == 32) {
465 if (IsAndroid)
466 Mapping.Offset = kDynamicShadowSentinel;
467 else if (IsMIPS32)
468 Mapping.Offset = kMIPS32_ShadowOffset32;
469 else if (IsFreeBSD)
470 Mapping.Offset = kFreeBSD_ShadowOffset32;
471 else if (IsNetBSD)
472 Mapping.Offset = kNetBSD_ShadowOffset32;
473 else if (IsIOS)
474 Mapping.Offset = kDynamicShadowSentinel;
475 else if (IsWindows)
476 Mapping.Offset = kWindowsShadowOffset32;
477 else if (IsEmscripten)
478 Mapping.Offset = kEmscriptenShadowOffset;
479 else if (IsMyriad) {
480 uint64_t ShadowOffset = (kMyriadMemoryOffset32 + kMyriadMemorySize32 -
481 (kMyriadMemorySize32 >> Mapping.Scale));
482 Mapping.Offset = ShadowOffset - (kMyriadMemoryOffset32 >> Mapping.Scale);
483 }
484 else
485 Mapping.Offset = kDefaultShadowOffset32;
486 } else { // LongSize == 64
487 // Fuchsia is always PIE, which means that the beginning of the address
488 // space is always available.
489 if (IsFuchsia)
490 Mapping.Offset = 0;
491 else if (IsPPC64)
492 Mapping.Offset = kPPC64_ShadowOffset64;
493 else if (IsSystemZ)
494 Mapping.Offset = kSystemZ_ShadowOffset64;
495 else if (IsFreeBSD && !IsMIPS64)
496 Mapping.Offset = kFreeBSD_ShadowOffset64;
497 else if (IsNetBSD) {
498 if (IsKasan)
499 Mapping.Offset = kNetBSDKasan_ShadowOffset64;
500 else
501 Mapping.Offset = kNetBSD_ShadowOffset64;
502 } else if (IsPS4CPU)
503 Mapping.Offset = kPS4CPU_ShadowOffset64;
504 else if (IsLinux && IsX86_64) {
505 if (IsKasan)
506 Mapping.Offset = kLinuxKasan_ShadowOffset64;
507 else
508 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
509 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
510 } else if (IsWindows && IsX86_64) {
511 Mapping.Offset = kWindowsShadowOffset64;
512 } else if (IsMIPS64)
513 Mapping.Offset = kMIPS64_ShadowOffset64;
514 else if (IsIOS)
515 Mapping.Offset = kDynamicShadowSentinel;
516 else if (IsMacOS && IsAArch64)
517 Mapping.Offset = kDynamicShadowSentinel;
518 else if (IsAArch64)
519 Mapping.Offset = kAArch64_ShadowOffset64;
520 else if (IsRISCV64)
521 Mapping.Offset = kRISCV64_ShadowOffset64;
522 else
523 Mapping.Offset = kDefaultShadowOffset64;
524 }
525
526 if (ClForceDynamicShadow) {
527 Mapping.Offset = kDynamicShadowSentinel;
528 }
529
530 if (ClMappingOffset.getNumOccurrences() > 0) {
531 Mapping.Offset = ClMappingOffset;
532 }
533
534 // OR-ing shadow offset if more efficient (at least on x86) if the offset
535 // is a power of two, but on ppc64 we have to use add since the shadow
536 // offset is not necessary 1/8-th of the address space. On SystemZ,
537 // we could OR the constant in a single instruction, but it's more
538 // efficient to load it once and use indexed addressing.
539 Mapping.OrShadowOffset = !IsAArch64 && !IsPPC64 && !IsSystemZ && !IsPS4CPU &&
540 !IsRISCV64 &&
541 !(Mapping.Offset & (Mapping.Offset - 1)) &&
542 Mapping.Offset != kDynamicShadowSentinel;
543 bool IsAndroidWithIfuncSupport =
544 IsAndroid && !TargetTriple.isAndroidVersionLT(21);
545 Mapping.InGlobal = ClWithIfunc && IsAndroidWithIfuncSupport && IsArmOrThumb;
546
547 return Mapping;
548 }
549
getRedzoneSizeForScale(int MappingScale)550 static uint64_t getRedzoneSizeForScale(int MappingScale) {
551 // Redzone used for stack and globals is at least 32 bytes.
552 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
553 return std::max(32U, 1U << MappingScale);
554 }
555
GetCtorAndDtorPriority(Triple & TargetTriple)556 static uint64_t GetCtorAndDtorPriority(Triple &TargetTriple) {
557 if (TargetTriple.isOSEmscripten()) {
558 return kAsanEmscriptenCtorAndDtorPriority;
559 } else {
560 return kAsanCtorAndDtorPriority;
561 }
562 }
563
564 namespace {
565
566 /// Module analysis for getting various metadata about the module.
567 class ASanGlobalsMetadataWrapperPass : public ModulePass {
568 public:
569 static char ID;
570
ASanGlobalsMetadataWrapperPass()571 ASanGlobalsMetadataWrapperPass() : ModulePass(ID) {
572 initializeASanGlobalsMetadataWrapperPassPass(
573 *PassRegistry::getPassRegistry());
574 }
575
runOnModule(Module & M)576 bool runOnModule(Module &M) override {
577 GlobalsMD = GlobalsMetadata(M);
578 return false;
579 }
580
getPassName() const581 StringRef getPassName() const override {
582 return "ASanGlobalsMetadataWrapperPass";
583 }
584
getAnalysisUsage(AnalysisUsage & AU) const585 void getAnalysisUsage(AnalysisUsage &AU) const override {
586 AU.setPreservesAll();
587 }
588
getGlobalsMD()589 GlobalsMetadata &getGlobalsMD() { return GlobalsMD; }
590
591 private:
592 GlobalsMetadata GlobalsMD;
593 };
594
595 char ASanGlobalsMetadataWrapperPass::ID = 0;
596
597 /// AddressSanitizer: instrument the code in module to find memory bugs.
598 struct AddressSanitizer {
AddressSanitizer__anon9cb762c90211::AddressSanitizer599 AddressSanitizer(Module &M, const GlobalsMetadata *GlobalsMD,
600 bool CompileKernel = false, bool Recover = false,
601 bool UseAfterScope = false)
602 : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
603 : CompileKernel),
604 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
605 UseAfterScope(UseAfterScope || ClUseAfterScope), GlobalsMD(*GlobalsMD) {
606 C = &(M.getContext());
607 LongSize = M.getDataLayout().getPointerSizeInBits();
608 IntptrTy = Type::getIntNTy(*C, LongSize);
609 TargetTriple = Triple(M.getTargetTriple());
610
611 Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
612 }
613
getAllocaSizeInBytes__anon9cb762c90211::AddressSanitizer614 uint64_t getAllocaSizeInBytes(const AllocaInst &AI) const {
615 uint64_t ArraySize = 1;
616 if (AI.isArrayAllocation()) {
617 const ConstantInt *CI = dyn_cast<ConstantInt>(AI.getArraySize());
618 assert(CI && "non-constant array size");
619 ArraySize = CI->getZExtValue();
620 }
621 Type *Ty = AI.getAllocatedType();
622 uint64_t SizeInBytes =
623 AI.getModule()->getDataLayout().getTypeAllocSize(Ty);
624 return SizeInBytes * ArraySize;
625 }
626
627 /// Check if we want (and can) handle this alloca.
628 bool isInterestingAlloca(const AllocaInst &AI);
629
630 bool ignoreAccess(Value *Ptr);
631 void getInterestingMemoryOperands(
632 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting);
633
634 void instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
635 InterestingMemoryOperand &O, bool UseCalls,
636 const DataLayout &DL);
637 void instrumentPointerComparisonOrSubtraction(Instruction *I);
638 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
639 Value *Addr, uint32_t TypeSize, bool IsWrite,
640 Value *SizeArgument, bool UseCalls, uint32_t Exp);
641 void instrumentUnusualSizeOrAlignment(Instruction *I,
642 Instruction *InsertBefore, Value *Addr,
643 uint32_t TypeSize, bool IsWrite,
644 Value *SizeArgument, bool UseCalls,
645 uint32_t Exp);
646 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
647 Value *ShadowValue, uint32_t TypeSize);
648 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
649 bool IsWrite, size_t AccessSizeIndex,
650 Value *SizeArgument, uint32_t Exp);
651 void instrumentMemIntrinsic(MemIntrinsic *MI);
652 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
653 bool suppressInstrumentationSiteForDebug(int &Instrumented);
654 bool instrumentFunction(Function &F, const TargetLibraryInfo *TLI);
655 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
656 bool maybeInsertDynamicShadowAtFunctionEntry(Function &F);
657 void markEscapedLocalAllocas(Function &F);
658
659 private:
660 friend struct FunctionStackPoisoner;
661
662 void initializeCallbacks(Module &M);
663
664 bool LooksLikeCodeInBug11395(Instruction *I);
665 bool GlobalIsLinkerInitialized(GlobalVariable *G);
666 bool isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, Value *Addr,
667 uint64_t TypeSize) const;
668
669 /// Helper to cleanup per-function state.
670 struct FunctionStateRAII {
671 AddressSanitizer *Pass;
672
FunctionStateRAII__anon9cb762c90211::AddressSanitizer::FunctionStateRAII673 FunctionStateRAII(AddressSanitizer *Pass) : Pass(Pass) {
674 assert(Pass->ProcessedAllocas.empty() &&
675 "last pass forgot to clear cache");
676 assert(!Pass->LocalDynamicShadow);
677 }
678
~FunctionStateRAII__anon9cb762c90211::AddressSanitizer::FunctionStateRAII679 ~FunctionStateRAII() {
680 Pass->LocalDynamicShadow = nullptr;
681 Pass->ProcessedAllocas.clear();
682 }
683 };
684
685 LLVMContext *C;
686 Triple TargetTriple;
687 int LongSize;
688 bool CompileKernel;
689 bool Recover;
690 bool UseAfterScope;
691 Type *IntptrTy;
692 ShadowMapping Mapping;
693 FunctionCallee AsanHandleNoReturnFunc;
694 FunctionCallee AsanPtrCmpFunction, AsanPtrSubFunction;
695 Constant *AsanShadowGlobal;
696
697 // These arrays is indexed by AccessIsWrite, Experiment and log2(AccessSize).
698 FunctionCallee AsanErrorCallback[2][2][kNumberOfAccessSizes];
699 FunctionCallee AsanMemoryAccessCallback[2][2][kNumberOfAccessSizes];
700
701 // These arrays is indexed by AccessIsWrite and Experiment.
702 FunctionCallee AsanErrorCallbackSized[2][2];
703 FunctionCallee AsanMemoryAccessCallbackSized[2][2];
704
705 FunctionCallee AsanMemmove, AsanMemcpy, AsanMemset;
706 Value *LocalDynamicShadow = nullptr;
707 const GlobalsMetadata &GlobalsMD;
708 DenseMap<const AllocaInst *, bool> ProcessedAllocas;
709 };
710
711 class AddressSanitizerLegacyPass : public FunctionPass {
712 public:
713 static char ID;
714
AddressSanitizerLegacyPass(bool CompileKernel=false,bool Recover=false,bool UseAfterScope=false)715 explicit AddressSanitizerLegacyPass(bool CompileKernel = false,
716 bool Recover = false,
717 bool UseAfterScope = false)
718 : FunctionPass(ID), CompileKernel(CompileKernel), Recover(Recover),
719 UseAfterScope(UseAfterScope) {
720 initializeAddressSanitizerLegacyPassPass(*PassRegistry::getPassRegistry());
721 }
722
getPassName() const723 StringRef getPassName() const override {
724 return "AddressSanitizerFunctionPass";
725 }
726
getAnalysisUsage(AnalysisUsage & AU) const727 void getAnalysisUsage(AnalysisUsage &AU) const override {
728 AU.addRequired<ASanGlobalsMetadataWrapperPass>();
729 AU.addRequired<TargetLibraryInfoWrapperPass>();
730 }
731
runOnFunction(Function & F)732 bool runOnFunction(Function &F) override {
733 GlobalsMetadata &GlobalsMD =
734 getAnalysis<ASanGlobalsMetadataWrapperPass>().getGlobalsMD();
735 const TargetLibraryInfo *TLI =
736 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
737 AddressSanitizer ASan(*F.getParent(), &GlobalsMD, CompileKernel, Recover,
738 UseAfterScope);
739 return ASan.instrumentFunction(F, TLI);
740 }
741
742 private:
743 bool CompileKernel;
744 bool Recover;
745 bool UseAfterScope;
746 };
747
748 class ModuleAddressSanitizer {
749 public:
ModuleAddressSanitizer(Module & M,const GlobalsMetadata * GlobalsMD,bool CompileKernel=false,bool Recover=false,bool UseGlobalsGC=true,bool UseOdrIndicator=false)750 ModuleAddressSanitizer(Module &M, const GlobalsMetadata *GlobalsMD,
751 bool CompileKernel = false, bool Recover = false,
752 bool UseGlobalsGC = true, bool UseOdrIndicator = false)
753 : GlobalsMD(*GlobalsMD),
754 CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
755 : CompileKernel),
756 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
757 UseGlobalsGC(UseGlobalsGC && ClUseGlobalsGC && !this->CompileKernel),
758 // Enable aliases as they should have no downside with ODR indicators.
759 UsePrivateAlias(UseOdrIndicator || ClUsePrivateAlias),
760 UseOdrIndicator(UseOdrIndicator || ClUseOdrIndicator),
761 // Not a typo: ClWithComdat is almost completely pointless without
762 // ClUseGlobalsGC (because then it only works on modules without
763 // globals, which are rare); it is a prerequisite for ClUseGlobalsGC;
764 // and both suffer from gold PR19002 for which UseGlobalsGC constructor
765 // argument is designed as workaround. Therefore, disable both
766 // ClWithComdat and ClUseGlobalsGC unless the frontend says it's ok to
767 // do globals-gc.
768 UseCtorComdat(UseGlobalsGC && ClWithComdat && !this->CompileKernel) {
769 C = &(M.getContext());
770 int LongSize = M.getDataLayout().getPointerSizeInBits();
771 IntptrTy = Type::getIntNTy(*C, LongSize);
772 TargetTriple = Triple(M.getTargetTriple());
773 Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
774 }
775
776 bool instrumentModule(Module &);
777
778 private:
779 void initializeCallbacks(Module &M);
780
781 bool InstrumentGlobals(IRBuilder<> &IRB, Module &M, bool *CtorComdat);
782 void InstrumentGlobalsCOFF(IRBuilder<> &IRB, Module &M,
783 ArrayRef<GlobalVariable *> ExtendedGlobals,
784 ArrayRef<Constant *> MetadataInitializers);
785 void InstrumentGlobalsELF(IRBuilder<> &IRB, Module &M,
786 ArrayRef<GlobalVariable *> ExtendedGlobals,
787 ArrayRef<Constant *> MetadataInitializers,
788 const std::string &UniqueModuleId);
789 void InstrumentGlobalsMachO(IRBuilder<> &IRB, Module &M,
790 ArrayRef<GlobalVariable *> ExtendedGlobals,
791 ArrayRef<Constant *> MetadataInitializers);
792 void
793 InstrumentGlobalsWithMetadataArray(IRBuilder<> &IRB, Module &M,
794 ArrayRef<GlobalVariable *> ExtendedGlobals,
795 ArrayRef<Constant *> MetadataInitializers);
796
797 GlobalVariable *CreateMetadataGlobal(Module &M, Constant *Initializer,
798 StringRef OriginalName);
799 void SetComdatForGlobalMetadata(GlobalVariable *G, GlobalVariable *Metadata,
800 StringRef InternalSuffix);
801 Instruction *CreateAsanModuleDtor(Module &M);
802
803 bool canInstrumentAliasedGlobal(const GlobalAlias &GA) const;
804 bool shouldInstrumentGlobal(GlobalVariable *G) const;
805 bool ShouldUseMachOGlobalsSection() const;
806 StringRef getGlobalMetadataSection() const;
807 void poisonOneInitializer(Function &GlobalInit, GlobalValue *ModuleName);
808 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
getMinRedzoneSizeForGlobal() const809 uint64_t getMinRedzoneSizeForGlobal() const {
810 return getRedzoneSizeForScale(Mapping.Scale);
811 }
812 uint64_t getRedzoneSizeForGlobal(uint64_t SizeInBytes) const;
813 int GetAsanVersion(const Module &M) const;
814
815 const GlobalsMetadata &GlobalsMD;
816 bool CompileKernel;
817 bool Recover;
818 bool UseGlobalsGC;
819 bool UsePrivateAlias;
820 bool UseOdrIndicator;
821 bool UseCtorComdat;
822 Type *IntptrTy;
823 LLVMContext *C;
824 Triple TargetTriple;
825 ShadowMapping Mapping;
826 FunctionCallee AsanPoisonGlobals;
827 FunctionCallee AsanUnpoisonGlobals;
828 FunctionCallee AsanRegisterGlobals;
829 FunctionCallee AsanUnregisterGlobals;
830 FunctionCallee AsanRegisterImageGlobals;
831 FunctionCallee AsanUnregisterImageGlobals;
832 FunctionCallee AsanRegisterElfGlobals;
833 FunctionCallee AsanUnregisterElfGlobals;
834
835 Function *AsanCtorFunction = nullptr;
836 Function *AsanDtorFunction = nullptr;
837 };
838
839 class ModuleAddressSanitizerLegacyPass : public ModulePass {
840 public:
841 static char ID;
842
ModuleAddressSanitizerLegacyPass(bool CompileKernel=false,bool Recover=false,bool UseGlobalGC=true,bool UseOdrIndicator=false)843 explicit ModuleAddressSanitizerLegacyPass(bool CompileKernel = false,
844 bool Recover = false,
845 bool UseGlobalGC = true,
846 bool UseOdrIndicator = false)
847 : ModulePass(ID), CompileKernel(CompileKernel), Recover(Recover),
848 UseGlobalGC(UseGlobalGC), UseOdrIndicator(UseOdrIndicator) {
849 initializeModuleAddressSanitizerLegacyPassPass(
850 *PassRegistry::getPassRegistry());
851 }
852
getPassName() const853 StringRef getPassName() const override { return "ModuleAddressSanitizer"; }
854
getAnalysisUsage(AnalysisUsage & AU) const855 void getAnalysisUsage(AnalysisUsage &AU) const override {
856 AU.addRequired<ASanGlobalsMetadataWrapperPass>();
857 }
858
runOnModule(Module & M)859 bool runOnModule(Module &M) override {
860 GlobalsMetadata &GlobalsMD =
861 getAnalysis<ASanGlobalsMetadataWrapperPass>().getGlobalsMD();
862 ModuleAddressSanitizer ASanModule(M, &GlobalsMD, CompileKernel, Recover,
863 UseGlobalGC, UseOdrIndicator);
864 return ASanModule.instrumentModule(M);
865 }
866
867 private:
868 bool CompileKernel;
869 bool Recover;
870 bool UseGlobalGC;
871 bool UseOdrIndicator;
872 };
873
874 // Stack poisoning does not play well with exception handling.
875 // When an exception is thrown, we essentially bypass the code
876 // that unpoisones the stack. This is why the run-time library has
877 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
878 // stack in the interceptor. This however does not work inside the
879 // actual function which catches the exception. Most likely because the
880 // compiler hoists the load of the shadow value somewhere too high.
881 // This causes asan to report a non-existing bug on 453.povray.
882 // It sounds like an LLVM bug.
883 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
884 Function &F;
885 AddressSanitizer &ASan;
886 DIBuilder DIB;
887 LLVMContext *C;
888 Type *IntptrTy;
889 Type *IntptrPtrTy;
890 ShadowMapping Mapping;
891
892 SmallVector<AllocaInst *, 16> AllocaVec;
893 SmallVector<AllocaInst *, 16> StaticAllocasToMoveUp;
894 SmallVector<Instruction *, 8> RetVec;
895 unsigned StackAlignment;
896
897 FunctionCallee AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1],
898 AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1];
899 FunctionCallee AsanSetShadowFunc[0x100] = {};
900 FunctionCallee AsanPoisonStackMemoryFunc, AsanUnpoisonStackMemoryFunc;
901 FunctionCallee AsanAllocaPoisonFunc, AsanAllocasUnpoisonFunc;
902
903 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
904 struct AllocaPoisonCall {
905 IntrinsicInst *InsBefore;
906 AllocaInst *AI;
907 uint64_t Size;
908 bool DoPoison;
909 };
910 SmallVector<AllocaPoisonCall, 8> DynamicAllocaPoisonCallVec;
911 SmallVector<AllocaPoisonCall, 8> StaticAllocaPoisonCallVec;
912 bool HasUntracedLifetimeIntrinsic = false;
913
914 SmallVector<AllocaInst *, 1> DynamicAllocaVec;
915 SmallVector<IntrinsicInst *, 1> StackRestoreVec;
916 AllocaInst *DynamicAllocaLayout = nullptr;
917 IntrinsicInst *LocalEscapeCall = nullptr;
918
919 bool HasInlineAsm = false;
920 bool HasReturnsTwiceCall = false;
921
FunctionStackPoisoner__anon9cb762c90211::FunctionStackPoisoner922 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
923 : F(F), ASan(ASan), DIB(*F.getParent(), /*AllowUnresolved*/ false),
924 C(ASan.C), IntptrTy(ASan.IntptrTy),
925 IntptrPtrTy(PointerType::get(IntptrTy, 0)), Mapping(ASan.Mapping),
926 StackAlignment(1 << Mapping.Scale) {}
927
runOnFunction__anon9cb762c90211::FunctionStackPoisoner928 bool runOnFunction() {
929 if (!ClStack) return false;
930
931 if (ClRedzoneByvalArgs)
932 copyArgsPassedByValToAllocas();
933
934 // Collect alloca, ret, lifetime instructions etc.
935 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) visit(*BB);
936
937 if (AllocaVec.empty() && DynamicAllocaVec.empty()) return false;
938
939 initializeCallbacks(*F.getParent());
940
941 if (HasUntracedLifetimeIntrinsic) {
942 // If there are lifetime intrinsics which couldn't be traced back to an
943 // alloca, we may not know exactly when a variable enters scope, and
944 // therefore should "fail safe" by not poisoning them.
945 StaticAllocaPoisonCallVec.clear();
946 DynamicAllocaPoisonCallVec.clear();
947 }
948
949 processDynamicAllocas();
950 processStaticAllocas();
951
952 if (ClDebugStack) {
953 LLVM_DEBUG(dbgs() << F);
954 }
955 return true;
956 }
957
958 // Arguments marked with the "byval" attribute are implicitly copied without
959 // using an alloca instruction. To produce redzones for those arguments, we
960 // copy them a second time into memory allocated with an alloca instruction.
961 void copyArgsPassedByValToAllocas();
962
963 // Finds all Alloca instructions and puts
964 // poisoned red zones around all of them.
965 // Then unpoison everything back before the function returns.
966 void processStaticAllocas();
967 void processDynamicAllocas();
968
969 void createDynamicAllocasInitStorage();
970
971 // ----------------------- Visitors.
972 /// Collect all Ret instructions, or the musttail call instruction if it
973 /// precedes the return instruction.
visitReturnInst__anon9cb762c90211::FunctionStackPoisoner974 void visitReturnInst(ReturnInst &RI) {
975 if (CallInst *CI = RI.getParent()->getTerminatingMustTailCall())
976 RetVec.push_back(CI);
977 else
978 RetVec.push_back(&RI);
979 }
980
981 /// Collect all Resume instructions.
visitResumeInst__anon9cb762c90211::FunctionStackPoisoner982 void visitResumeInst(ResumeInst &RI) { RetVec.push_back(&RI); }
983
984 /// Collect all CatchReturnInst instructions.
visitCleanupReturnInst__anon9cb762c90211::FunctionStackPoisoner985 void visitCleanupReturnInst(CleanupReturnInst &CRI) { RetVec.push_back(&CRI); }
986
unpoisonDynamicAllocasBeforeInst__anon9cb762c90211::FunctionStackPoisoner987 void unpoisonDynamicAllocasBeforeInst(Instruction *InstBefore,
988 Value *SavedStack) {
989 IRBuilder<> IRB(InstBefore);
990 Value *DynamicAreaPtr = IRB.CreatePtrToInt(SavedStack, IntptrTy);
991 // When we insert _asan_allocas_unpoison before @llvm.stackrestore, we
992 // need to adjust extracted SP to compute the address of the most recent
993 // alloca. We have a special @llvm.get.dynamic.area.offset intrinsic for
994 // this purpose.
995 if (!isa<ReturnInst>(InstBefore)) {
996 Function *DynamicAreaOffsetFunc = Intrinsic::getDeclaration(
997 InstBefore->getModule(), Intrinsic::get_dynamic_area_offset,
998 {IntptrTy});
999
1000 Value *DynamicAreaOffset = IRB.CreateCall(DynamicAreaOffsetFunc, {});
1001
1002 DynamicAreaPtr = IRB.CreateAdd(IRB.CreatePtrToInt(SavedStack, IntptrTy),
1003 DynamicAreaOffset);
1004 }
1005
1006 IRB.CreateCall(
1007 AsanAllocasUnpoisonFunc,
1008 {IRB.CreateLoad(IntptrTy, DynamicAllocaLayout), DynamicAreaPtr});
1009 }
1010
1011 // Unpoison dynamic allocas redzones.
unpoisonDynamicAllocas__anon9cb762c90211::FunctionStackPoisoner1012 void unpoisonDynamicAllocas() {
1013 for (Instruction *Ret : RetVec)
1014 unpoisonDynamicAllocasBeforeInst(Ret, DynamicAllocaLayout);
1015
1016 for (Instruction *StackRestoreInst : StackRestoreVec)
1017 unpoisonDynamicAllocasBeforeInst(StackRestoreInst,
1018 StackRestoreInst->getOperand(0));
1019 }
1020
1021 // Deploy and poison redzones around dynamic alloca call. To do this, we
1022 // should replace this call with another one with changed parameters and
1023 // replace all its uses with new address, so
1024 // addr = alloca type, old_size, align
1025 // is replaced by
1026 // new_size = (old_size + additional_size) * sizeof(type)
1027 // tmp = alloca i8, new_size, max(align, 32)
1028 // addr = tmp + 32 (first 32 bytes are for the left redzone).
1029 // Additional_size is added to make new memory allocation contain not only
1030 // requested memory, but also left, partial and right redzones.
1031 void handleDynamicAllocaCall(AllocaInst *AI);
1032
1033 /// Collect Alloca instructions we want (and can) handle.
visitAllocaInst__anon9cb762c90211::FunctionStackPoisoner1034 void visitAllocaInst(AllocaInst &AI) {
1035 if (!ASan.isInterestingAlloca(AI)) {
1036 if (AI.isStaticAlloca()) {
1037 // Skip over allocas that are present *before* the first instrumented
1038 // alloca, we don't want to move those around.
1039 if (AllocaVec.empty())
1040 return;
1041
1042 StaticAllocasToMoveUp.push_back(&AI);
1043 }
1044 return;
1045 }
1046
1047 StackAlignment = std::max(StackAlignment, AI.getAlignment());
1048 if (!AI.isStaticAlloca())
1049 DynamicAllocaVec.push_back(&AI);
1050 else
1051 AllocaVec.push_back(&AI);
1052 }
1053
1054 /// Collect lifetime intrinsic calls to check for use-after-scope
1055 /// errors.
visitIntrinsicInst__anon9cb762c90211::FunctionStackPoisoner1056 void visitIntrinsicInst(IntrinsicInst &II) {
1057 Intrinsic::ID ID = II.getIntrinsicID();
1058 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(&II);
1059 if (ID == Intrinsic::localescape) LocalEscapeCall = &II;
1060 if (!ASan.UseAfterScope)
1061 return;
1062 if (!II.isLifetimeStartOrEnd())
1063 return;
1064 // Found lifetime intrinsic, add ASan instrumentation if necessary.
1065 auto *Size = cast<ConstantInt>(II.getArgOperand(0));
1066 // If size argument is undefined, don't do anything.
1067 if (Size->isMinusOne()) return;
1068 // Check that size doesn't saturate uint64_t and can
1069 // be stored in IntptrTy.
1070 const uint64_t SizeValue = Size->getValue().getLimitedValue();
1071 if (SizeValue == ~0ULL ||
1072 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
1073 return;
1074 // Find alloca instruction that corresponds to llvm.lifetime argument.
1075 // Currently we can only handle lifetime markers pointing to the
1076 // beginning of the alloca.
1077 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1), true);
1078 if (!AI) {
1079 HasUntracedLifetimeIntrinsic = true;
1080 return;
1081 }
1082 // We're interested only in allocas we can handle.
1083 if (!ASan.isInterestingAlloca(*AI))
1084 return;
1085 bool DoPoison = (ID == Intrinsic::lifetime_end);
1086 AllocaPoisonCall APC = {&II, AI, SizeValue, DoPoison};
1087 if (AI->isStaticAlloca())
1088 StaticAllocaPoisonCallVec.push_back(APC);
1089 else if (ClInstrumentDynamicAllocas)
1090 DynamicAllocaPoisonCallVec.push_back(APC);
1091 }
1092
visitCallBase__anon9cb762c90211::FunctionStackPoisoner1093 void visitCallBase(CallBase &CB) {
1094 if (CallInst *CI = dyn_cast<CallInst>(&CB)) {
1095 HasInlineAsm |= CI->isInlineAsm() && &CB != ASan.LocalDynamicShadow;
1096 HasReturnsTwiceCall |= CI->canReturnTwice();
1097 }
1098 }
1099
1100 // ---------------------- Helpers.
1101 void initializeCallbacks(Module &M);
1102
1103 // Copies bytes from ShadowBytes into shadow memory for indexes where
1104 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that
1105 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten.
1106 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1107 IRBuilder<> &IRB, Value *ShadowBase);
1108 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1109 size_t Begin, size_t End, IRBuilder<> &IRB,
1110 Value *ShadowBase);
1111 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
1112 ArrayRef<uint8_t> ShadowBytes, size_t Begin,
1113 size_t End, IRBuilder<> &IRB, Value *ShadowBase);
1114
1115 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
1116
1117 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L,
1118 bool Dynamic);
1119 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue,
1120 Instruction *ThenTerm, Value *ValueIfFalse);
1121 };
1122
1123 } // end anonymous namespace
1124
parse(MDNode * MDN)1125 void LocationMetadata::parse(MDNode *MDN) {
1126 assert(MDN->getNumOperands() == 3);
1127 MDString *DIFilename = cast<MDString>(MDN->getOperand(0));
1128 Filename = DIFilename->getString();
1129 LineNo = mdconst::extract<ConstantInt>(MDN->getOperand(1))->getLimitedValue();
1130 ColumnNo =
1131 mdconst::extract<ConstantInt>(MDN->getOperand(2))->getLimitedValue();
1132 }
1133
1134 // FIXME: It would be cleaner to instead attach relevant metadata to the globals
1135 // we want to sanitize instead and reading this metadata on each pass over a
1136 // function instead of reading module level metadata at first.
GlobalsMetadata(Module & M)1137 GlobalsMetadata::GlobalsMetadata(Module &M) {
1138 NamedMDNode *Globals = M.getNamedMetadata("llvm.asan.globals");
1139 if (!Globals)
1140 return;
1141 for (auto MDN : Globals->operands()) {
1142 // Metadata node contains the global and the fields of "Entry".
1143 assert(MDN->getNumOperands() == 5);
1144 auto *V = mdconst::extract_or_null<Constant>(MDN->getOperand(0));
1145 // The optimizer may optimize away a global entirely.
1146 if (!V)
1147 continue;
1148 auto *StrippedV = V->stripPointerCasts();
1149 auto *GV = dyn_cast<GlobalVariable>(StrippedV);
1150 if (!GV)
1151 continue;
1152 // We can already have an entry for GV if it was merged with another
1153 // global.
1154 Entry &E = Entries[GV];
1155 if (auto *Loc = cast_or_null<MDNode>(MDN->getOperand(1)))
1156 E.SourceLoc.parse(Loc);
1157 if (auto *Name = cast_or_null<MDString>(MDN->getOperand(2)))
1158 E.Name = Name->getString();
1159 ConstantInt *IsDynInit = mdconst::extract<ConstantInt>(MDN->getOperand(3));
1160 E.IsDynInit |= IsDynInit->isOne();
1161 ConstantInt *IsExcluded =
1162 mdconst::extract<ConstantInt>(MDN->getOperand(4));
1163 E.IsExcluded |= IsExcluded->isOne();
1164 }
1165 }
1166
1167 AnalysisKey ASanGlobalsMetadataAnalysis::Key;
1168
run(Module & M,ModuleAnalysisManager & AM)1169 GlobalsMetadata ASanGlobalsMetadataAnalysis::run(Module &M,
1170 ModuleAnalysisManager &AM) {
1171 return GlobalsMetadata(M);
1172 }
1173
AddressSanitizerPass(bool CompileKernel,bool Recover,bool UseAfterScope)1174 AddressSanitizerPass::AddressSanitizerPass(bool CompileKernel, bool Recover,
1175 bool UseAfterScope)
1176 : CompileKernel(CompileKernel), Recover(Recover),
1177 UseAfterScope(UseAfterScope) {}
1178
run(Function & F,AnalysisManager<Function> & AM)1179 PreservedAnalyses AddressSanitizerPass::run(Function &F,
1180 AnalysisManager<Function> &AM) {
1181 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
1182 Module &M = *F.getParent();
1183 if (auto *R = MAMProxy.getCachedResult<ASanGlobalsMetadataAnalysis>(M)) {
1184 const TargetLibraryInfo *TLI = &AM.getResult<TargetLibraryAnalysis>(F);
1185 AddressSanitizer Sanitizer(M, R, CompileKernel, Recover, UseAfterScope);
1186 if (Sanitizer.instrumentFunction(F, TLI))
1187 return PreservedAnalyses::none();
1188 return PreservedAnalyses::all();
1189 }
1190
1191 report_fatal_error(
1192 "The ASanGlobalsMetadataAnalysis is required to run before "
1193 "AddressSanitizer can run");
1194 return PreservedAnalyses::all();
1195 }
1196
ModuleAddressSanitizerPass(bool CompileKernel,bool Recover,bool UseGlobalGC,bool UseOdrIndicator)1197 ModuleAddressSanitizerPass::ModuleAddressSanitizerPass(bool CompileKernel,
1198 bool Recover,
1199 bool UseGlobalGC,
1200 bool UseOdrIndicator)
1201 : CompileKernel(CompileKernel), Recover(Recover), UseGlobalGC(UseGlobalGC),
1202 UseOdrIndicator(UseOdrIndicator) {}
1203
run(Module & M,AnalysisManager<Module> & AM)1204 PreservedAnalyses ModuleAddressSanitizerPass::run(Module &M,
1205 AnalysisManager<Module> &AM) {
1206 GlobalsMetadata &GlobalsMD = AM.getResult<ASanGlobalsMetadataAnalysis>(M);
1207 ModuleAddressSanitizer Sanitizer(M, &GlobalsMD, CompileKernel, Recover,
1208 UseGlobalGC, UseOdrIndicator);
1209 if (Sanitizer.instrumentModule(M))
1210 return PreservedAnalyses::none();
1211 return PreservedAnalyses::all();
1212 }
1213
1214 INITIALIZE_PASS(ASanGlobalsMetadataWrapperPass, "asan-globals-md",
1215 "Read metadata to mark which globals should be instrumented "
1216 "when running ASan.",
1217 false, true)
1218
1219 char AddressSanitizerLegacyPass::ID = 0;
1220
1221 INITIALIZE_PASS_BEGIN(
1222 AddressSanitizerLegacyPass, "asan",
1223 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false,
1224 false)
INITIALIZE_PASS_DEPENDENCY(ASanGlobalsMetadataWrapperPass)1225 INITIALIZE_PASS_DEPENDENCY(ASanGlobalsMetadataWrapperPass)
1226 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1227 INITIALIZE_PASS_END(
1228 AddressSanitizerLegacyPass, "asan",
1229 "AddressSanitizer: detects use-after-free and out-of-bounds bugs.", false,
1230 false)
1231
1232 FunctionPass *llvm::createAddressSanitizerFunctionPass(bool CompileKernel,
1233 bool Recover,
1234 bool UseAfterScope) {
1235 assert(!CompileKernel || Recover);
1236 return new AddressSanitizerLegacyPass(CompileKernel, Recover, UseAfterScope);
1237 }
1238
1239 char ModuleAddressSanitizerLegacyPass::ID = 0;
1240
1241 INITIALIZE_PASS(
1242 ModuleAddressSanitizerLegacyPass, "asan-module",
1243 "AddressSanitizer: detects use-after-free and out-of-bounds bugs."
1244 "ModulePass",
1245 false, false)
1246
createModuleAddressSanitizerLegacyPassPass(bool CompileKernel,bool Recover,bool UseGlobalsGC,bool UseOdrIndicator)1247 ModulePass *llvm::createModuleAddressSanitizerLegacyPassPass(
1248 bool CompileKernel, bool Recover, bool UseGlobalsGC, bool UseOdrIndicator) {
1249 assert(!CompileKernel || Recover);
1250 return new ModuleAddressSanitizerLegacyPass(CompileKernel, Recover,
1251 UseGlobalsGC, UseOdrIndicator);
1252 }
1253
TypeSizeToSizeIndex(uint32_t TypeSize)1254 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
1255 size_t Res = countTrailingZeros(TypeSize / 8);
1256 assert(Res < kNumberOfAccessSizes);
1257 return Res;
1258 }
1259
1260 /// Create a global describing a source location.
createPrivateGlobalForSourceLoc(Module & M,LocationMetadata MD)1261 static GlobalVariable *createPrivateGlobalForSourceLoc(Module &M,
1262 LocationMetadata MD) {
1263 Constant *LocData[] = {
1264 createPrivateGlobalForString(M, MD.Filename, true, kAsanGenPrefix),
1265 ConstantInt::get(Type::getInt32Ty(M.getContext()), MD.LineNo),
1266 ConstantInt::get(Type::getInt32Ty(M.getContext()), MD.ColumnNo),
1267 };
1268 auto LocStruct = ConstantStruct::getAnon(LocData);
1269 auto GV = new GlobalVariable(M, LocStruct->getType(), true,
1270 GlobalValue::PrivateLinkage, LocStruct,
1271 kAsanGenPrefix);
1272 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
1273 return GV;
1274 }
1275
1276 /// Check if \p G has been created by a trusted compiler pass.
GlobalWasGeneratedByCompiler(GlobalVariable * G)1277 static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) {
1278 // Do not instrument @llvm.global_ctors, @llvm.used, etc.
1279 if (G->getName().startswith("llvm."))
1280 return true;
1281
1282 // Do not instrument asan globals.
1283 if (G->getName().startswith(kAsanGenPrefix) ||
1284 G->getName().startswith(kSanCovGenPrefix) ||
1285 G->getName().startswith(kODRGenPrefix))
1286 return true;
1287
1288 // Do not instrument gcov counter arrays.
1289 if (G->getName() == "__llvm_gcov_ctr")
1290 return true;
1291
1292 return false;
1293 }
1294
memToShadow(Value * Shadow,IRBuilder<> & IRB)1295 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
1296 // Shadow >> scale
1297 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
1298 if (Mapping.Offset == 0) return Shadow;
1299 // (Shadow >> scale) | offset
1300 Value *ShadowBase;
1301 if (LocalDynamicShadow)
1302 ShadowBase = LocalDynamicShadow;
1303 else
1304 ShadowBase = ConstantInt::get(IntptrTy, Mapping.Offset);
1305 if (Mapping.OrShadowOffset)
1306 return IRB.CreateOr(Shadow, ShadowBase);
1307 else
1308 return IRB.CreateAdd(Shadow, ShadowBase);
1309 }
1310
1311 // Instrument memset/memmove/memcpy
instrumentMemIntrinsic(MemIntrinsic * MI)1312 void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
1313 IRBuilder<> IRB(MI);
1314 if (isa<MemTransferInst>(MI)) {
1315 IRB.CreateCall(
1316 isa<MemMoveInst>(MI) ? AsanMemmove : AsanMemcpy,
1317 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
1318 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()),
1319 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
1320 } else if (isa<MemSetInst>(MI)) {
1321 IRB.CreateCall(
1322 AsanMemset,
1323 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
1324 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false),
1325 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
1326 }
1327 MI->eraseFromParent();
1328 }
1329
1330 /// Check if we want (and can) handle this alloca.
isInterestingAlloca(const AllocaInst & AI)1331 bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1332 auto PreviouslySeenAllocaInfo = ProcessedAllocas.find(&AI);
1333
1334 if (PreviouslySeenAllocaInfo != ProcessedAllocas.end())
1335 return PreviouslySeenAllocaInfo->getSecond();
1336
1337 bool IsInteresting =
1338 (AI.getAllocatedType()->isSized() &&
1339 // alloca() may be called with 0 size, ignore it.
1340 ((!AI.isStaticAlloca()) || getAllocaSizeInBytes(AI) > 0) &&
1341 // We are only interested in allocas not promotable to registers.
1342 // Promotable allocas are common under -O0.
1343 (!ClSkipPromotableAllocas || !isAllocaPromotable(&AI)) &&
1344 // inalloca allocas are not treated as static, and we don't want
1345 // dynamic alloca instrumentation for them as well.
1346 !AI.isUsedWithInAlloca() &&
1347 // swifterror allocas are register promoted by ISel
1348 !AI.isSwiftError());
1349
1350 ProcessedAllocas[&AI] = IsInteresting;
1351 return IsInteresting;
1352 }
1353
ignoreAccess(Value * Ptr)1354 bool AddressSanitizer::ignoreAccess(Value *Ptr) {
1355 // Do not instrument acesses from different address spaces; we cannot deal
1356 // with them.
1357 Type *PtrTy = cast<PointerType>(Ptr->getType()->getScalarType());
1358 if (PtrTy->getPointerAddressSpace() != 0)
1359 return true;
1360
1361 // Ignore swifterror addresses.
1362 // swifterror memory addresses are mem2reg promoted by instruction
1363 // selection. As such they cannot have regular uses like an instrumentation
1364 // function and it makes no sense to track them as memory.
1365 if (Ptr->isSwiftError())
1366 return true;
1367
1368 // Treat memory accesses to promotable allocas as non-interesting since they
1369 // will not cause memory violations. This greatly speeds up the instrumented
1370 // executable at -O0.
1371 if (auto AI = dyn_cast_or_null<AllocaInst>(Ptr))
1372 if (ClSkipPromotableAllocas && !isInterestingAlloca(*AI))
1373 return true;
1374
1375 return false;
1376 }
1377
getInterestingMemoryOperands(Instruction * I,SmallVectorImpl<InterestingMemoryOperand> & Interesting)1378 void AddressSanitizer::getInterestingMemoryOperands(
1379 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting) {
1380 // Skip memory accesses inserted by another instrumentation.
1381 if (I->hasMetadata("nosanitize"))
1382 return;
1383
1384 // Do not instrument the load fetching the dynamic shadow address.
1385 if (LocalDynamicShadow == I)
1386 return;
1387
1388 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1389 if (!ClInstrumentReads || ignoreAccess(LI->getPointerOperand()))
1390 return;
1391 Interesting.emplace_back(I, LI->getPointerOperandIndex(), false,
1392 LI->getType(), LI->getAlign());
1393 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1394 if (!ClInstrumentWrites || ignoreAccess(SI->getPointerOperand()))
1395 return;
1396 Interesting.emplace_back(I, SI->getPointerOperandIndex(), true,
1397 SI->getValueOperand()->getType(), SI->getAlign());
1398 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
1399 if (!ClInstrumentAtomics || ignoreAccess(RMW->getPointerOperand()))
1400 return;
1401 Interesting.emplace_back(I, RMW->getPointerOperandIndex(), true,
1402 RMW->getValOperand()->getType(), None);
1403 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
1404 if (!ClInstrumentAtomics || ignoreAccess(XCHG->getPointerOperand()))
1405 return;
1406 Interesting.emplace_back(I, XCHG->getPointerOperandIndex(), true,
1407 XCHG->getCompareOperand()->getType(), None);
1408 } else if (auto CI = dyn_cast<CallInst>(I)) {
1409 auto *F = CI->getCalledFunction();
1410 if (F && (F->getName().startswith("llvm.masked.load.") ||
1411 F->getName().startswith("llvm.masked.store."))) {
1412 bool IsWrite = F->getName().startswith("llvm.masked.store.");
1413 // Masked store has an initial operand for the value.
1414 unsigned OpOffset = IsWrite ? 1 : 0;
1415 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1416 return;
1417
1418 auto BasePtr = CI->getOperand(OpOffset);
1419 if (ignoreAccess(BasePtr))
1420 return;
1421 auto Ty = cast<PointerType>(BasePtr->getType())->getElementType();
1422 MaybeAlign Alignment = Align(1);
1423 // Otherwise no alignment guarantees. We probably got Undef.
1424 if (auto *Op = dyn_cast<ConstantInt>(CI->getOperand(1 + OpOffset)))
1425 Alignment = Op->getMaybeAlignValue();
1426 Value *Mask = CI->getOperand(2 + OpOffset);
1427 Interesting.emplace_back(I, OpOffset, IsWrite, Ty, Alignment, Mask);
1428 } else {
1429 for (unsigned ArgNo = 0; ArgNo < CI->getNumArgOperands(); ArgNo++) {
1430 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1431 ignoreAccess(CI->getArgOperand(ArgNo)))
1432 continue;
1433 Type *Ty = CI->getParamByValType(ArgNo);
1434 Interesting.emplace_back(I, ArgNo, false, Ty, Align(1));
1435 }
1436 }
1437 }
1438 }
1439
isPointerOperand(Value * V)1440 static bool isPointerOperand(Value *V) {
1441 return V->getType()->isPointerTy() || isa<PtrToIntInst>(V);
1442 }
1443
1444 // This is a rough heuristic; it may cause both false positives and
1445 // false negatives. The proper implementation requires cooperation with
1446 // the frontend.
isInterestingPointerComparison(Instruction * I)1447 static bool isInterestingPointerComparison(Instruction *I) {
1448 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(I)) {
1449 if (!Cmp->isRelational())
1450 return false;
1451 } else {
1452 return false;
1453 }
1454 return isPointerOperand(I->getOperand(0)) &&
1455 isPointerOperand(I->getOperand(1));
1456 }
1457
1458 // This is a rough heuristic; it may cause both false positives and
1459 // false negatives. The proper implementation requires cooperation with
1460 // the frontend.
isInterestingPointerSubtraction(Instruction * I)1461 static bool isInterestingPointerSubtraction(Instruction *I) {
1462 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1463 if (BO->getOpcode() != Instruction::Sub)
1464 return false;
1465 } else {
1466 return false;
1467 }
1468 return isPointerOperand(I->getOperand(0)) &&
1469 isPointerOperand(I->getOperand(1));
1470 }
1471
GlobalIsLinkerInitialized(GlobalVariable * G)1472 bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1473 // If a global variable does not have dynamic initialization we don't
1474 // have to instrument it. However, if a global does not have initializer
1475 // at all, we assume it has dynamic initializer (in other TU).
1476 //
1477 // FIXME: Metadata should be attched directly to the global directly instead
1478 // of being added to llvm.asan.globals.
1479 return G->hasInitializer() && !GlobalsMD.get(G).IsDynInit;
1480 }
1481
instrumentPointerComparisonOrSubtraction(Instruction * I)1482 void AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1483 Instruction *I) {
1484 IRBuilder<> IRB(I);
1485 FunctionCallee F = isa<ICmpInst>(I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1486 Value *Param[2] = {I->getOperand(0), I->getOperand(1)};
1487 for (Value *&i : Param) {
1488 if (i->getType()->isPointerTy())
1489 i = IRB.CreatePointerCast(i, IntptrTy);
1490 }
1491 IRB.CreateCall(F, Param);
1492 }
1493
doInstrumentAddress(AddressSanitizer * Pass,Instruction * I,Instruction * InsertBefore,Value * Addr,MaybeAlign Alignment,unsigned Granularity,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1494 static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1495 Instruction *InsertBefore, Value *Addr,
1496 MaybeAlign Alignment, unsigned Granularity,
1497 uint32_t TypeSize, bool IsWrite,
1498 Value *SizeArgument, bool UseCalls,
1499 uint32_t Exp) {
1500 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1501 // if the data is properly aligned.
1502 if ((TypeSize == 8 || TypeSize == 16 || TypeSize == 32 || TypeSize == 64 ||
1503 TypeSize == 128) &&
1504 (!Alignment || *Alignment >= Granularity || *Alignment >= TypeSize / 8))
1505 return Pass->instrumentAddress(I, InsertBefore, Addr, TypeSize, IsWrite,
1506 nullptr, UseCalls, Exp);
1507 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeSize,
1508 IsWrite, nullptr, UseCalls, Exp);
1509 }
1510
instrumentMaskedLoadOrStore(AddressSanitizer * Pass,const DataLayout & DL,Type * IntptrTy,Value * Mask,Instruction * I,Value * Addr,MaybeAlign Alignment,unsigned Granularity,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1511 static void instrumentMaskedLoadOrStore(AddressSanitizer *Pass,
1512 const DataLayout &DL, Type *IntptrTy,
1513 Value *Mask, Instruction *I,
1514 Value *Addr, MaybeAlign Alignment,
1515 unsigned Granularity, uint32_t TypeSize,
1516 bool IsWrite, Value *SizeArgument,
1517 bool UseCalls, uint32_t Exp) {
1518 auto *VTy = cast<FixedVectorType>(
1519 cast<PointerType>(Addr->getType())->getElementType());
1520 uint64_t ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType());
1521 unsigned Num = VTy->getNumElements();
1522 auto Zero = ConstantInt::get(IntptrTy, 0);
1523 for (unsigned Idx = 0; Idx < Num; ++Idx) {
1524 Value *InstrumentedAddress = nullptr;
1525 Instruction *InsertBefore = I;
1526 if (auto *Vector = dyn_cast<ConstantVector>(Mask)) {
1527 // dyn_cast as we might get UndefValue
1528 if (auto *Masked = dyn_cast<ConstantInt>(Vector->getOperand(Idx))) {
1529 if (Masked->isZero())
1530 // Mask is constant false, so no instrumentation needed.
1531 continue;
1532 // If we have a true or undef value, fall through to doInstrumentAddress
1533 // with InsertBefore == I
1534 }
1535 } else {
1536 IRBuilder<> IRB(I);
1537 Value *MaskElem = IRB.CreateExtractElement(Mask, Idx);
1538 Instruction *ThenTerm = SplitBlockAndInsertIfThen(MaskElem, I, false);
1539 InsertBefore = ThenTerm;
1540 }
1541
1542 IRBuilder<> IRB(InsertBefore);
1543 InstrumentedAddress =
1544 IRB.CreateGEP(VTy, Addr, {Zero, ConstantInt::get(IntptrTy, Idx)});
1545 doInstrumentAddress(Pass, I, InsertBefore, InstrumentedAddress, Alignment,
1546 Granularity, ElemTypeSize, IsWrite, SizeArgument,
1547 UseCalls, Exp);
1548 }
1549 }
1550
instrumentMop(ObjectSizeOffsetVisitor & ObjSizeVis,InterestingMemoryOperand & O,bool UseCalls,const DataLayout & DL)1551 void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1552 InterestingMemoryOperand &O, bool UseCalls,
1553 const DataLayout &DL) {
1554 Value *Addr = O.getPtr();
1555
1556 // Optimization experiments.
1557 // The experiments can be used to evaluate potential optimizations that remove
1558 // instrumentation (assess false negatives). Instead of completely removing
1559 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1560 // experiments that want to remove instrumentation of this instruction).
1561 // If Exp is non-zero, this pass will emit special calls into runtime
1562 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1563 // make runtime terminate the program in a special way (with a different
1564 // exit status). Then you run the new compiler on a buggy corpus, collect
1565 // the special terminations (ideally, you don't see them at all -- no false
1566 // negatives) and make the decision on the optimization.
1567 uint32_t Exp = ClForceExperiment;
1568
1569 if (ClOpt && ClOptGlobals) {
1570 // If initialization order checking is disabled, a simple access to a
1571 // dynamically initialized global is always valid.
1572 GlobalVariable *G = dyn_cast<GlobalVariable>(getUnderlyingObject(Addr));
1573 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1574 isSafeAccess(ObjSizeVis, Addr, O.TypeSize)) {
1575 NumOptimizedAccessesToGlobalVar++;
1576 return;
1577 }
1578 }
1579
1580 if (ClOpt && ClOptStack) {
1581 // A direct inbounds access to a stack variable is always valid.
1582 if (isa<AllocaInst>(getUnderlyingObject(Addr)) &&
1583 isSafeAccess(ObjSizeVis, Addr, O.TypeSize)) {
1584 NumOptimizedAccessesToStackVar++;
1585 return;
1586 }
1587 }
1588
1589 if (O.IsWrite)
1590 NumInstrumentedWrites++;
1591 else
1592 NumInstrumentedReads++;
1593
1594 unsigned Granularity = 1 << Mapping.Scale;
1595 if (O.MaybeMask) {
1596 instrumentMaskedLoadOrStore(this, DL, IntptrTy, O.MaybeMask, O.getInsn(),
1597 Addr, O.Alignment, Granularity, O.TypeSize,
1598 O.IsWrite, nullptr, UseCalls, Exp);
1599 } else {
1600 doInstrumentAddress(this, O.getInsn(), O.getInsn(), Addr, O.Alignment,
1601 Granularity, O.TypeSize, O.IsWrite, nullptr, UseCalls,
1602 Exp);
1603 }
1604 }
1605
generateCrashCode(Instruction * InsertBefore,Value * Addr,bool IsWrite,size_t AccessSizeIndex,Value * SizeArgument,uint32_t Exp)1606 Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1607 Value *Addr, bool IsWrite,
1608 size_t AccessSizeIndex,
1609 Value *SizeArgument,
1610 uint32_t Exp) {
1611 IRBuilder<> IRB(InsertBefore);
1612 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(IRB.getInt32Ty(), Exp);
1613 CallInst *Call = nullptr;
1614 if (SizeArgument) {
1615 if (Exp == 0)
1616 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][0],
1617 {Addr, SizeArgument});
1618 else
1619 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][1],
1620 {Addr, SizeArgument, ExpVal});
1621 } else {
1622 if (Exp == 0)
1623 Call =
1624 IRB.CreateCall(AsanErrorCallback[IsWrite][0][AccessSizeIndex], Addr);
1625 else
1626 Call = IRB.CreateCall(AsanErrorCallback[IsWrite][1][AccessSizeIndex],
1627 {Addr, ExpVal});
1628 }
1629
1630 Call->setCannotMerge();
1631 return Call;
1632 }
1633
createSlowPathCmp(IRBuilder<> & IRB,Value * AddrLong,Value * ShadowValue,uint32_t TypeSize)1634 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1635 Value *ShadowValue,
1636 uint32_t TypeSize) {
1637 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1638 // Addr & (Granularity - 1)
1639 Value *LastAccessedByte =
1640 IRB.CreateAnd(AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
1641 // (Addr & (Granularity - 1)) + size - 1
1642 if (TypeSize / 8 > 1)
1643 LastAccessedByte = IRB.CreateAdd(
1644 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
1645 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1646 LastAccessedByte =
1647 IRB.CreateIntCast(LastAccessedByte, ShadowValue->getType(), false);
1648 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1649 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
1650 }
1651
instrumentAddress(Instruction * OrigIns,Instruction * InsertBefore,Value * Addr,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1652 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1653 Instruction *InsertBefore, Value *Addr,
1654 uint32_t TypeSize, bool IsWrite,
1655 Value *SizeArgument, bool UseCalls,
1656 uint32_t Exp) {
1657 bool IsMyriad = TargetTriple.getVendor() == llvm::Triple::Myriad;
1658
1659 IRBuilder<> IRB(InsertBefore);
1660 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
1661 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
1662
1663 if (UseCalls) {
1664 if (Exp == 0)
1665 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex],
1666 AddrLong);
1667 else
1668 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
1669 {AddrLong, ConstantInt::get(IRB.getInt32Ty(), Exp)});
1670 return;
1671 }
1672
1673 if (IsMyriad) {
1674 // Strip the cache bit and do range check.
1675 // AddrLong &= ~kMyriadCacheBitMask32
1676 AddrLong = IRB.CreateAnd(AddrLong, ~kMyriadCacheBitMask32);
1677 // Tag = AddrLong >> kMyriadTagShift
1678 Value *Tag = IRB.CreateLShr(AddrLong, kMyriadTagShift);
1679 // Tag == kMyriadDDRTag
1680 Value *TagCheck =
1681 IRB.CreateICmpEQ(Tag, ConstantInt::get(IntptrTy, kMyriadDDRTag));
1682
1683 Instruction *TagCheckTerm =
1684 SplitBlockAndInsertIfThen(TagCheck, InsertBefore, false,
1685 MDBuilder(*C).createBranchWeights(1, 100000));
1686 assert(cast<BranchInst>(TagCheckTerm)->isUnconditional());
1687 IRB.SetInsertPoint(TagCheckTerm);
1688 InsertBefore = TagCheckTerm;
1689 }
1690
1691 Type *ShadowTy =
1692 IntegerType::get(*C, std::max(8U, TypeSize >> Mapping.Scale));
1693 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
1694 Value *ShadowPtr = memToShadow(AddrLong, IRB);
1695 Value *CmpVal = Constant::getNullValue(ShadowTy);
1696 Value *ShadowValue =
1697 IRB.CreateLoad(ShadowTy, IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
1698
1699 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
1700 size_t Granularity = 1ULL << Mapping.Scale;
1701 Instruction *CrashTerm = nullptr;
1702
1703 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
1704 // We use branch weights for the slow path check, to indicate that the slow
1705 // path is rarely taken. This seems to be the case for SPEC benchmarks.
1706 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
1707 Cmp, InsertBefore, false, MDBuilder(*C).createBranchWeights(1, 100000));
1708 assert(cast<BranchInst>(CheckTerm)->isUnconditional());
1709 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
1710 IRB.SetInsertPoint(CheckTerm);
1711 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
1712 if (Recover) {
1713 CrashTerm = SplitBlockAndInsertIfThen(Cmp2, CheckTerm, false);
1714 } else {
1715 BasicBlock *CrashBlock =
1716 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
1717 CrashTerm = new UnreachableInst(*C, CrashBlock);
1718 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
1719 ReplaceInstWithInst(CheckTerm, NewTerm);
1720 }
1721 } else {
1722 CrashTerm = SplitBlockAndInsertIfThen(Cmp, InsertBefore, !Recover);
1723 }
1724
1725 Instruction *Crash = generateCrashCode(CrashTerm, AddrLong, IsWrite,
1726 AccessSizeIndex, SizeArgument, Exp);
1727 Crash->setDebugLoc(OrigIns->getDebugLoc());
1728 }
1729
1730 // Instrument unusual size or unusual alignment.
1731 // We can not do it with a single check, so we do 1-byte check for the first
1732 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
1733 // to report the actual access size.
instrumentUnusualSizeOrAlignment(Instruction * I,Instruction * InsertBefore,Value * Addr,uint32_t TypeSize,bool IsWrite,Value * SizeArgument,bool UseCalls,uint32_t Exp)1734 void AddressSanitizer::instrumentUnusualSizeOrAlignment(
1735 Instruction *I, Instruction *InsertBefore, Value *Addr, uint32_t TypeSize,
1736 bool IsWrite, Value *SizeArgument, bool UseCalls, uint32_t Exp) {
1737 IRBuilder<> IRB(InsertBefore);
1738 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
1739 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
1740 if (UseCalls) {
1741 if (Exp == 0)
1742 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][0],
1743 {AddrLong, Size});
1744 else
1745 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][1],
1746 {AddrLong, Size, ConstantInt::get(IRB.getInt32Ty(), Exp)});
1747 } else {
1748 Value *LastByte = IRB.CreateIntToPtr(
1749 IRB.CreateAdd(AddrLong, ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
1750 Addr->getType());
1751 instrumentAddress(I, InsertBefore, Addr, 8, IsWrite, Size, false, Exp);
1752 instrumentAddress(I, InsertBefore, LastByte, 8, IsWrite, Size, false, Exp);
1753 }
1754 }
1755
poisonOneInitializer(Function & GlobalInit,GlobalValue * ModuleName)1756 void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit,
1757 GlobalValue *ModuleName) {
1758 // Set up the arguments to our poison/unpoison functions.
1759 IRBuilder<> IRB(&GlobalInit.front(),
1760 GlobalInit.front().getFirstInsertionPt());
1761
1762 // Add a call to poison all external globals before the given function starts.
1763 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
1764 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
1765
1766 // Add calls to unpoison all globals before each return instruction.
1767 for (auto &BB : GlobalInit.getBasicBlockList())
1768 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB.getTerminator()))
1769 CallInst::Create(AsanUnpoisonGlobals, "", RI);
1770 }
1771
createInitializerPoisonCalls(Module & M,GlobalValue * ModuleName)1772 void ModuleAddressSanitizer::createInitializerPoisonCalls(
1773 Module &M, GlobalValue *ModuleName) {
1774 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1775 if (!GV)
1776 return;
1777
1778 ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
1779 if (!CA)
1780 return;
1781
1782 for (Use &OP : CA->operands()) {
1783 if (isa<ConstantAggregateZero>(OP)) continue;
1784 ConstantStruct *CS = cast<ConstantStruct>(OP);
1785
1786 // Must have a function or null ptr.
1787 if (Function *F = dyn_cast<Function>(CS->getOperand(1))) {
1788 if (F->getName() == kAsanModuleCtorName) continue;
1789 auto *Priority = cast<ConstantInt>(CS->getOperand(0));
1790 // Don't instrument CTORs that will run before asan.module_ctor.
1791 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
1792 continue;
1793 poisonOneInitializer(*F, ModuleName);
1794 }
1795 }
1796 }
1797
canInstrumentAliasedGlobal(const GlobalAlias & GA) const1798 bool ModuleAddressSanitizer::canInstrumentAliasedGlobal(
1799 const GlobalAlias &GA) const {
1800 // In case this function should be expanded to include rules that do not just
1801 // apply when CompileKernel is true, either guard all existing rules with an
1802 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
1803 // should also apply to user space.
1804 assert(CompileKernel && "Only expecting to be called when compiling kernel");
1805
1806 // When compiling the kernel, globals that are aliased by symbols prefixed
1807 // by "__" are special and cannot be padded with a redzone.
1808 if (GA.getName().startswith("__"))
1809 return false;
1810
1811 return true;
1812 }
1813
shouldInstrumentGlobal(GlobalVariable * G) const1814 bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
1815 Type *Ty = G->getValueType();
1816 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
1817
1818 // FIXME: Metadata should be attched directly to the global directly instead
1819 // of being added to llvm.asan.globals.
1820 if (GlobalsMD.get(G).IsExcluded) return false;
1821 if (!Ty->isSized()) return false;
1822 if (!G->hasInitializer()) return false;
1823 // Only instrument globals of default address spaces
1824 if (G->getAddressSpace()) return false;
1825 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
1826 // Two problems with thread-locals:
1827 // - The address of the main thread's copy can't be computed at link-time.
1828 // - Need to poison all copies, not just the main thread's one.
1829 if (G->isThreadLocal()) return false;
1830 // For now, just ignore this Global if the alignment is large.
1831 if (G->getAlignment() > getMinRedzoneSizeForGlobal()) return false;
1832
1833 // For non-COFF targets, only instrument globals known to be defined by this
1834 // TU.
1835 // FIXME: We can instrument comdat globals on ELF if we are using the
1836 // GC-friendly metadata scheme.
1837 if (!TargetTriple.isOSBinFormatCOFF()) {
1838 if (!G->hasExactDefinition() || G->hasComdat())
1839 return false;
1840 } else {
1841 // On COFF, don't instrument non-ODR linkages.
1842 if (G->isInterposable())
1843 return false;
1844 }
1845
1846 // If a comdat is present, it must have a selection kind that implies ODR
1847 // semantics: no duplicates, any, or exact match.
1848 if (Comdat *C = G->getComdat()) {
1849 switch (C->getSelectionKind()) {
1850 case Comdat::Any:
1851 case Comdat::ExactMatch:
1852 case Comdat::NoDuplicates:
1853 break;
1854 case Comdat::Largest:
1855 case Comdat::SameSize:
1856 return false;
1857 }
1858 }
1859
1860 if (G->hasSection()) {
1861 // The kernel uses explicit sections for mostly special global variables
1862 // that we should not instrument. E.g. the kernel may rely on their layout
1863 // without redzones, or remove them at link time ("discard.*"), etc.
1864 if (CompileKernel)
1865 return false;
1866
1867 StringRef Section = G->getSection();
1868
1869 // Globals from llvm.metadata aren't emitted, do not instrument them.
1870 if (Section == "llvm.metadata") return false;
1871 // Do not instrument globals from special LLVM sections.
1872 if (Section.find("__llvm") != StringRef::npos || Section.find("__LLVM") != StringRef::npos) return false;
1873
1874 // Do not instrument function pointers to initialization and termination
1875 // routines: dynamic linker will not properly handle redzones.
1876 if (Section.startswith(".preinit_array") ||
1877 Section.startswith(".init_array") ||
1878 Section.startswith(".fini_array")) {
1879 return false;
1880 }
1881
1882 // Do not instrument user-defined sections (with names resembling
1883 // valid C identifiers)
1884 if (TargetTriple.isOSBinFormatELF()) {
1885 if (std::all_of(Section.begin(), Section.end(),
1886 [](char c) { return llvm::isAlnum(c) || c == '_'; }))
1887 return false;
1888 }
1889
1890 // On COFF, if the section name contains '$', it is highly likely that the
1891 // user is using section sorting to create an array of globals similar to
1892 // the way initialization callbacks are registered in .init_array and
1893 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
1894 // to such globals is counterproductive, because the intent is that they
1895 // will form an array, and out-of-bounds accesses are expected.
1896 // See https://github.com/google/sanitizers/issues/305
1897 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
1898 if (TargetTriple.isOSBinFormatCOFF() && Section.contains('$')) {
1899 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
1900 << *G << "\n");
1901 return false;
1902 }
1903
1904 if (TargetTriple.isOSBinFormatMachO()) {
1905 StringRef ParsedSegment, ParsedSection;
1906 unsigned TAA = 0, StubSize = 0;
1907 bool TAAParsed;
1908 std::string ErrorCode = MCSectionMachO::ParseSectionSpecifier(
1909 Section, ParsedSegment, ParsedSection, TAA, TAAParsed, StubSize);
1910 assert(ErrorCode.empty() && "Invalid section specifier.");
1911
1912 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
1913 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
1914 // them.
1915 if (ParsedSegment == "__OBJC" ||
1916 (ParsedSegment == "__DATA" && ParsedSection.startswith("__objc_"))) {
1917 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
1918 return false;
1919 }
1920 // See https://github.com/google/sanitizers/issues/32
1921 // Constant CFString instances are compiled in the following way:
1922 // -- the string buffer is emitted into
1923 // __TEXT,__cstring,cstring_literals
1924 // -- the constant NSConstantString structure referencing that buffer
1925 // is placed into __DATA,__cfstring
1926 // Therefore there's no point in placing redzones into __DATA,__cfstring.
1927 // Moreover, it causes the linker to crash on OS X 10.7
1928 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
1929 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
1930 return false;
1931 }
1932 // The linker merges the contents of cstring_literals and removes the
1933 // trailing zeroes.
1934 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
1935 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
1936 return false;
1937 }
1938 }
1939 }
1940
1941 if (CompileKernel) {
1942 // Globals that prefixed by "__" are special and cannot be padded with a
1943 // redzone.
1944 if (G->getName().startswith("__"))
1945 return false;
1946 }
1947
1948 return true;
1949 }
1950
1951 // On Mach-O platforms, we emit global metadata in a separate section of the
1952 // binary in order to allow the linker to properly dead strip. This is only
1953 // supported on recent versions of ld64.
ShouldUseMachOGlobalsSection() const1954 bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
1955 if (!TargetTriple.isOSBinFormatMachO())
1956 return false;
1957
1958 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(10, 11))
1959 return true;
1960 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(9))
1961 return true;
1962 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(2))
1963 return true;
1964
1965 return false;
1966 }
1967
getGlobalMetadataSection() const1968 StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
1969 switch (TargetTriple.getObjectFormat()) {
1970 case Triple::COFF: return ".ASAN$GL";
1971 case Triple::ELF: return "asan_globals";
1972 case Triple::MachO: return "__DATA,__asan_globals,regular";
1973 case Triple::Wasm:
1974 case Triple::GOFF:
1975 case Triple::XCOFF:
1976 report_fatal_error(
1977 "ModuleAddressSanitizer not implemented for object file format");
1978 case Triple::UnknownObjectFormat:
1979 break;
1980 }
1981 llvm_unreachable("unsupported object format");
1982 }
1983
initializeCallbacks(Module & M)1984 void ModuleAddressSanitizer::initializeCallbacks(Module &M) {
1985 IRBuilder<> IRB(*C);
1986
1987 // Declare our poisoning and unpoisoning functions.
1988 AsanPoisonGlobals =
1989 M.getOrInsertFunction(kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy);
1990 AsanUnpoisonGlobals =
1991 M.getOrInsertFunction(kAsanUnpoisonGlobalsName, IRB.getVoidTy());
1992
1993 // Declare functions that register/unregister globals.
1994 AsanRegisterGlobals = M.getOrInsertFunction(
1995 kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy);
1996 AsanUnregisterGlobals = M.getOrInsertFunction(
1997 kAsanUnregisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy);
1998
1999 // Declare the functions that find globals in a shared object and then invoke
2000 // the (un)register function on them.
2001 AsanRegisterImageGlobals = M.getOrInsertFunction(
2002 kAsanRegisterImageGlobalsName, IRB.getVoidTy(), IntptrTy);
2003 AsanUnregisterImageGlobals = M.getOrInsertFunction(
2004 kAsanUnregisterImageGlobalsName, IRB.getVoidTy(), IntptrTy);
2005
2006 AsanRegisterElfGlobals =
2007 M.getOrInsertFunction(kAsanRegisterElfGlobalsName, IRB.getVoidTy(),
2008 IntptrTy, IntptrTy, IntptrTy);
2009 AsanUnregisterElfGlobals =
2010 M.getOrInsertFunction(kAsanUnregisterElfGlobalsName, IRB.getVoidTy(),
2011 IntptrTy, IntptrTy, IntptrTy);
2012 }
2013
2014 // Put the metadata and the instrumented global in the same group. This ensures
2015 // that the metadata is discarded if the instrumented global is discarded.
SetComdatForGlobalMetadata(GlobalVariable * G,GlobalVariable * Metadata,StringRef InternalSuffix)2016 void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
2017 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
2018 Module &M = *G->getParent();
2019 Comdat *C = G->getComdat();
2020 if (!C) {
2021 if (!G->hasName()) {
2022 // If G is unnamed, it must be internal. Give it an artificial name
2023 // so we can put it in a comdat.
2024 assert(G->hasLocalLinkage());
2025 G->setName(Twine(kAsanGenPrefix) + "_anon_global");
2026 }
2027
2028 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
2029 std::string Name = std::string(G->getName());
2030 Name += InternalSuffix;
2031 C = M.getOrInsertComdat(Name);
2032 } else {
2033 C = M.getOrInsertComdat(G->getName());
2034 }
2035
2036 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
2037 // linkage to internal linkage so that a symbol table entry is emitted. This
2038 // is necessary in order to create the comdat group.
2039 if (TargetTriple.isOSBinFormatCOFF()) {
2040 C->setSelectionKind(Comdat::NoDuplicates);
2041 if (G->hasPrivateLinkage())
2042 G->setLinkage(GlobalValue::InternalLinkage);
2043 }
2044 G->setComdat(C);
2045 }
2046
2047 assert(G->hasComdat());
2048 Metadata->setComdat(G->getComdat());
2049 }
2050
2051 // Create a separate metadata global and put it in the appropriate ASan
2052 // global registration section.
2053 GlobalVariable *
CreateMetadataGlobal(Module & M,Constant * Initializer,StringRef OriginalName)2054 ModuleAddressSanitizer::CreateMetadataGlobal(Module &M, Constant *Initializer,
2055 StringRef OriginalName) {
2056 auto Linkage = TargetTriple.isOSBinFormatMachO()
2057 ? GlobalVariable::InternalLinkage
2058 : GlobalVariable::PrivateLinkage;
2059 GlobalVariable *Metadata = new GlobalVariable(
2060 M, Initializer->getType(), false, Linkage, Initializer,
2061 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(OriginalName));
2062 Metadata->setSection(getGlobalMetadataSection());
2063 return Metadata;
2064 }
2065
CreateAsanModuleDtor(Module & M)2066 Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor(Module &M) {
2067 AsanDtorFunction =
2068 Function::Create(FunctionType::get(Type::getVoidTy(*C), false),
2069 GlobalValue::InternalLinkage, kAsanModuleDtorName, &M);
2070 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
2071
2072 return ReturnInst::Create(*C, AsanDtorBB);
2073 }
2074
InstrumentGlobalsCOFF(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers)2075 void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
2076 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2077 ArrayRef<Constant *> MetadataInitializers) {
2078 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2079 auto &DL = M.getDataLayout();
2080
2081 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2082 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2083 Constant *Initializer = MetadataInitializers[i];
2084 GlobalVariable *G = ExtendedGlobals[i];
2085 GlobalVariable *Metadata =
2086 CreateMetadataGlobal(M, Initializer, G->getName());
2087 MDNode *MD = MDNode::get(M.getContext(), ValueAsMetadata::get(G));
2088 Metadata->setMetadata(LLVMContext::MD_associated, MD);
2089 MetadataGlobals[i] = Metadata;
2090
2091 // The MSVC linker always inserts padding when linking incrementally. We
2092 // cope with that by aligning each struct to its size, which must be a power
2093 // of two.
2094 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Initializer->getType());
2095 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2096 "global metadata will not be padded appropriately");
2097 Metadata->setAlignment(assumeAligned(SizeOfGlobalStruct));
2098
2099 SetComdatForGlobalMetadata(G, Metadata, "");
2100 }
2101
2102 // Update llvm.compiler.used, adding the new metadata globals. This is
2103 // needed so that during LTO these variables stay alive.
2104 if (!MetadataGlobals.empty())
2105 appendToCompilerUsed(M, MetadataGlobals);
2106 }
2107
InstrumentGlobalsELF(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers,const std::string & UniqueModuleId)2108 void ModuleAddressSanitizer::InstrumentGlobalsELF(
2109 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2110 ArrayRef<Constant *> MetadataInitializers,
2111 const std::string &UniqueModuleId) {
2112 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2113
2114 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2115 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2116 GlobalVariable *G = ExtendedGlobals[i];
2117 GlobalVariable *Metadata =
2118 CreateMetadataGlobal(M, MetadataInitializers[i], G->getName());
2119 MDNode *MD = MDNode::get(M.getContext(), ValueAsMetadata::get(G));
2120 Metadata->setMetadata(LLVMContext::MD_associated, MD);
2121 MetadataGlobals[i] = Metadata;
2122
2123 SetComdatForGlobalMetadata(G, Metadata, UniqueModuleId);
2124 }
2125
2126 // Update llvm.compiler.used, adding the new metadata globals. This is
2127 // needed so that during LTO these variables stay alive.
2128 if (!MetadataGlobals.empty())
2129 appendToCompilerUsed(M, MetadataGlobals);
2130
2131 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2132 // to look up the loaded image that contains it. Second, we can store in it
2133 // whether registration has already occurred, to prevent duplicate
2134 // registration.
2135 //
2136 // Common linkage ensures that there is only one global per shared library.
2137 GlobalVariable *RegisteredFlag = new GlobalVariable(
2138 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2139 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName);
2140 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2141
2142 // Create start and stop symbols.
2143 GlobalVariable *StartELFMetadata = new GlobalVariable(
2144 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2145 "__start_" + getGlobalMetadataSection());
2146 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2147 GlobalVariable *StopELFMetadata = new GlobalVariable(
2148 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2149 "__stop_" + getGlobalMetadataSection());
2150 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2151
2152 // Create a call to register the globals with the runtime.
2153 IRB.CreateCall(AsanRegisterElfGlobals,
2154 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy),
2155 IRB.CreatePointerCast(StartELFMetadata, IntptrTy),
2156 IRB.CreatePointerCast(StopELFMetadata, IntptrTy)});
2157
2158 // We also need to unregister globals at the end, e.g., when a shared library
2159 // gets closed.
2160 IRBuilder<> IRB_Dtor(CreateAsanModuleDtor(M));
2161 IRB_Dtor.CreateCall(AsanUnregisterElfGlobals,
2162 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy),
2163 IRB.CreatePointerCast(StartELFMetadata, IntptrTy),
2164 IRB.CreatePointerCast(StopELFMetadata, IntptrTy)});
2165 }
2166
InstrumentGlobalsMachO(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers)2167 void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2168 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2169 ArrayRef<Constant *> MetadataInitializers) {
2170 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2171
2172 // On recent Mach-O platforms, use a structure which binds the liveness of
2173 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2174 // created to be added to llvm.compiler.used
2175 StructType *LivenessTy = StructType::get(IntptrTy, IntptrTy);
2176 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2177
2178 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2179 Constant *Initializer = MetadataInitializers[i];
2180 GlobalVariable *G = ExtendedGlobals[i];
2181 GlobalVariable *Metadata =
2182 CreateMetadataGlobal(M, Initializer, G->getName());
2183
2184 // On recent Mach-O platforms, we emit the global metadata in a way that
2185 // allows the linker to properly strip dead globals.
2186 auto LivenessBinder =
2187 ConstantStruct::get(LivenessTy, Initializer->getAggregateElement(0u),
2188 ConstantExpr::getPointerCast(Metadata, IntptrTy));
2189 GlobalVariable *Liveness = new GlobalVariable(
2190 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2191 Twine("__asan_binder_") + G->getName());
2192 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2193 LivenessGlobals[i] = Liveness;
2194 }
2195
2196 // Update llvm.compiler.used, adding the new liveness globals. This is
2197 // needed so that during LTO these variables stay alive. The alternative
2198 // would be to have the linker handling the LTO symbols, but libLTO
2199 // current API does not expose access to the section for each symbol.
2200 if (!LivenessGlobals.empty())
2201 appendToCompilerUsed(M, LivenessGlobals);
2202
2203 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2204 // to look up the loaded image that contains it. Second, we can store in it
2205 // whether registration has already occurred, to prevent duplicate
2206 // registration.
2207 //
2208 // common linkage ensures that there is only one global per shared library.
2209 GlobalVariable *RegisteredFlag = new GlobalVariable(
2210 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2211 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName);
2212 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2213
2214 IRB.CreateCall(AsanRegisterImageGlobals,
2215 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)});
2216
2217 // We also need to unregister globals at the end, e.g., when a shared library
2218 // gets closed.
2219 IRBuilder<> IRB_Dtor(CreateAsanModuleDtor(M));
2220 IRB_Dtor.CreateCall(AsanUnregisterImageGlobals,
2221 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)});
2222 }
2223
InstrumentGlobalsWithMetadataArray(IRBuilder<> & IRB,Module & M,ArrayRef<GlobalVariable * > ExtendedGlobals,ArrayRef<Constant * > MetadataInitializers)2224 void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2225 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2226 ArrayRef<Constant *> MetadataInitializers) {
2227 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2228 unsigned N = ExtendedGlobals.size();
2229 assert(N > 0);
2230
2231 // On platforms that don't have a custom metadata section, we emit an array
2232 // of global metadata structures.
2233 ArrayType *ArrayOfGlobalStructTy =
2234 ArrayType::get(MetadataInitializers[0]->getType(), N);
2235 auto AllGlobals = new GlobalVariable(
2236 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2237 ConstantArray::get(ArrayOfGlobalStructTy, MetadataInitializers), "");
2238 if (Mapping.Scale > 3)
2239 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2240
2241 IRB.CreateCall(AsanRegisterGlobals,
2242 {IRB.CreatePointerCast(AllGlobals, IntptrTy),
2243 ConstantInt::get(IntptrTy, N)});
2244
2245 // We also need to unregister globals at the end, e.g., when a shared library
2246 // gets closed.
2247 IRBuilder<> IRB_Dtor(CreateAsanModuleDtor(M));
2248 IRB_Dtor.CreateCall(AsanUnregisterGlobals,
2249 {IRB.CreatePointerCast(AllGlobals, IntptrTy),
2250 ConstantInt::get(IntptrTy, N)});
2251 }
2252
2253 // This function replaces all global variables with new variables that have
2254 // trailing redzones. It also creates a function that poisons
2255 // redzones and inserts this function into llvm.global_ctors.
2256 // Sets *CtorComdat to true if the global registration code emitted into the
2257 // asan constructor is comdat-compatible.
InstrumentGlobals(IRBuilder<> & IRB,Module & M,bool * CtorComdat)2258 bool ModuleAddressSanitizer::InstrumentGlobals(IRBuilder<> &IRB, Module &M,
2259 bool *CtorComdat) {
2260 *CtorComdat = false;
2261
2262 // Build set of globals that are aliased by some GA, where
2263 // canInstrumentAliasedGlobal(GA) returns false.
2264 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2265 if (CompileKernel) {
2266 for (auto &GA : M.aliases()) {
2267 if (const auto *GV = dyn_cast<GlobalVariable>(GA.getAliasee())) {
2268 if (!canInstrumentAliasedGlobal(GA))
2269 AliasedGlobalExclusions.insert(GV);
2270 }
2271 }
2272 }
2273
2274 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2275 for (auto &G : M.globals()) {
2276 if (!AliasedGlobalExclusions.count(&G) && shouldInstrumentGlobal(&G))
2277 GlobalsToChange.push_back(&G);
2278 }
2279
2280 size_t n = GlobalsToChange.size();
2281 if (n == 0) {
2282 *CtorComdat = true;
2283 return false;
2284 }
2285
2286 auto &DL = M.getDataLayout();
2287
2288 // A global is described by a structure
2289 // size_t beg;
2290 // size_t size;
2291 // size_t size_with_redzone;
2292 // const char *name;
2293 // const char *module_name;
2294 // size_t has_dynamic_init;
2295 // void *source_location;
2296 // size_t odr_indicator;
2297 // We initialize an array of such structures and pass it to a run-time call.
2298 StructType *GlobalStructTy =
2299 StructType::get(IntptrTy, IntptrTy, IntptrTy, IntptrTy, IntptrTy,
2300 IntptrTy, IntptrTy, IntptrTy);
2301 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2302 SmallVector<Constant *, 16> Initializers(n);
2303
2304 bool HasDynamicallyInitializedGlobals = false;
2305
2306 // We shouldn't merge same module names, as this string serves as unique
2307 // module ID in runtime.
2308 GlobalVariable *ModuleName = createPrivateGlobalForString(
2309 M, M.getModuleIdentifier(), /*AllowMerging*/ false, kAsanGenPrefix);
2310
2311 for (size_t i = 0; i < n; i++) {
2312 GlobalVariable *G = GlobalsToChange[i];
2313
2314 // FIXME: Metadata should be attched directly to the global directly instead
2315 // of being added to llvm.asan.globals.
2316 auto MD = GlobalsMD.get(G);
2317 StringRef NameForGlobal = G->getName();
2318 // Create string holding the global name (use global name from metadata
2319 // if it's available, otherwise just write the name of global variable).
2320 GlobalVariable *Name = createPrivateGlobalForString(
2321 M, MD.Name.empty() ? NameForGlobal : MD.Name,
2322 /*AllowMerging*/ true, kAsanGenPrefix);
2323
2324 Type *Ty = G->getValueType();
2325 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2326 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2327 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
2328
2329 StructType *NewTy = StructType::get(Ty, RightRedZoneTy);
2330 Constant *NewInitializer = ConstantStruct::get(
2331 NewTy, G->getInitializer(), Constant::getNullValue(RightRedZoneTy));
2332
2333 // Create a new global variable with enough space for a redzone.
2334 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2335 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2336 Linkage = GlobalValue::InternalLinkage;
2337 GlobalVariable *NewGlobal =
2338 new GlobalVariable(M, NewTy, G->isConstant(), Linkage, NewInitializer,
2339 "", G, G->getThreadLocalMode());
2340 NewGlobal->copyAttributesFrom(G);
2341 NewGlobal->setComdat(G->getComdat());
2342 NewGlobal->setAlignment(MaybeAlign(getMinRedzoneSizeForGlobal()));
2343 // Don't fold globals with redzones. ODR violation detector and redzone
2344 // poisoning implicitly creates a dependence on the global's address, so it
2345 // is no longer valid for it to be marked unnamed_addr.
2346 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2347
2348 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2349 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2350 G->isConstant()) {
2351 auto Seq = dyn_cast<ConstantDataSequential>(G->getInitializer());
2352 if (Seq && Seq->isCString())
2353 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2354 }
2355
2356 // Transfer the debug info and type metadata. The payload starts at offset
2357 // zero so we can copy the metadata over as is.
2358 NewGlobal->copyMetadata(G, 0);
2359
2360 Value *Indices2[2];
2361 Indices2[0] = IRB.getInt32(0);
2362 Indices2[1] = IRB.getInt32(0);
2363
2364 G->replaceAllUsesWith(
2365 ConstantExpr::getGetElementPtr(NewTy, NewGlobal, Indices2, true));
2366 NewGlobal->takeName(G);
2367 G->eraseFromParent();
2368 NewGlobals[i] = NewGlobal;
2369
2370 Constant *SourceLoc;
2371 if (!MD.SourceLoc.empty()) {
2372 auto SourceLocGlobal = createPrivateGlobalForSourceLoc(M, MD.SourceLoc);
2373 SourceLoc = ConstantExpr::getPointerCast(SourceLocGlobal, IntptrTy);
2374 } else {
2375 SourceLoc = ConstantInt::get(IntptrTy, 0);
2376 }
2377
2378 Constant *ODRIndicator = ConstantExpr::getNullValue(IRB.getInt8PtrTy());
2379 GlobalValue *InstrumentedGlobal = NewGlobal;
2380
2381 bool CanUsePrivateAliases =
2382 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2383 TargetTriple.isOSBinFormatWasm();
2384 if (CanUsePrivateAliases && UsePrivateAlias) {
2385 // Create local alias for NewGlobal to avoid crash on ODR between
2386 // instrumented and non-instrumented libraries.
2387 InstrumentedGlobal =
2388 GlobalAlias::create(GlobalValue::PrivateLinkage, "", NewGlobal);
2389 }
2390
2391 // ODR should not happen for local linkage.
2392 if (NewGlobal->hasLocalLinkage()) {
2393 ODRIndicator = ConstantExpr::getIntToPtr(ConstantInt::get(IntptrTy, -1),
2394 IRB.getInt8PtrTy());
2395 } else if (UseOdrIndicator) {
2396 // With local aliases, we need to provide another externally visible
2397 // symbol __odr_asan_XXX to detect ODR violation.
2398 auto *ODRIndicatorSym =
2399 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2400 Constant::getNullValue(IRB.getInt8Ty()),
2401 kODRGenPrefix + NameForGlobal, nullptr,
2402 NewGlobal->getThreadLocalMode());
2403
2404 // Set meaningful attributes for indicator symbol.
2405 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2406 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2407 ODRIndicatorSym->setAlignment(Align(1));
2408 ODRIndicator = ODRIndicatorSym;
2409 }
2410
2411 Constant *Initializer = ConstantStruct::get(
2412 GlobalStructTy,
2413 ConstantExpr::getPointerCast(InstrumentedGlobal, IntptrTy),
2414 ConstantInt::get(IntptrTy, SizeInBytes),
2415 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
2416 ConstantExpr::getPointerCast(Name, IntptrTy),
2417 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
2418 ConstantInt::get(IntptrTy, MD.IsDynInit), SourceLoc,
2419 ConstantExpr::getPointerCast(ODRIndicator, IntptrTy));
2420
2421 if (ClInitializers && MD.IsDynInit) HasDynamicallyInitializedGlobals = true;
2422
2423 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2424
2425 Initializers[i] = Initializer;
2426 }
2427
2428 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2429 // ConstantMerge'ing them.
2430 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2431 for (size_t i = 0; i < n; i++) {
2432 GlobalVariable *G = NewGlobals[i];
2433 if (G->getName().empty()) continue;
2434 GlobalsToAddToUsedList.push_back(G);
2435 }
2436 appendToCompilerUsed(M, ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2437
2438 std::string ELFUniqueModuleId =
2439 (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) ? getUniqueModuleId(&M)
2440 : "";
2441
2442 if (!ELFUniqueModuleId.empty()) {
2443 InstrumentGlobalsELF(IRB, M, NewGlobals, Initializers, ELFUniqueModuleId);
2444 *CtorComdat = true;
2445 } else if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2446 InstrumentGlobalsCOFF(IRB, M, NewGlobals, Initializers);
2447 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2448 InstrumentGlobalsMachO(IRB, M, NewGlobals, Initializers);
2449 } else {
2450 InstrumentGlobalsWithMetadataArray(IRB, M, NewGlobals, Initializers);
2451 }
2452
2453 // Create calls for poisoning before initializers run and unpoisoning after.
2454 if (HasDynamicallyInitializedGlobals)
2455 createInitializerPoisonCalls(M, ModuleName);
2456
2457 LLVM_DEBUG(dbgs() << M);
2458 return true;
2459 }
2460
2461 uint64_t
getRedzoneSizeForGlobal(uint64_t SizeInBytes) const2462 ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2463 constexpr uint64_t kMaxRZ = 1 << 18;
2464 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2465
2466 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2467 uint64_t RZ =
2468 std::max(MinRZ, std::min(kMaxRZ, (SizeInBytes / MinRZ / 4) * MinRZ));
2469
2470 // Round up to multiple of MinRZ.
2471 if (SizeInBytes % MinRZ)
2472 RZ += MinRZ - (SizeInBytes % MinRZ);
2473 assert((RZ + SizeInBytes) % MinRZ == 0);
2474
2475 return RZ;
2476 }
2477
GetAsanVersion(const Module & M) const2478 int ModuleAddressSanitizer::GetAsanVersion(const Module &M) const {
2479 int LongSize = M.getDataLayout().getPointerSizeInBits();
2480 bool isAndroid = Triple(M.getTargetTriple()).isAndroid();
2481 int Version = 8;
2482 // 32-bit Android is one version ahead because of the switch to dynamic
2483 // shadow.
2484 Version += (LongSize == 32 && isAndroid);
2485 return Version;
2486 }
2487
instrumentModule(Module & M)2488 bool ModuleAddressSanitizer::instrumentModule(Module &M) {
2489 initializeCallbacks(M);
2490
2491 // Create a module constructor. A destructor is created lazily because not all
2492 // platforms, and not all modules need it.
2493 if (CompileKernel) {
2494 // The kernel always builds with its own runtime, and therefore does not
2495 // need the init and version check calls.
2496 AsanCtorFunction = createSanitizerCtor(M, kAsanModuleCtorName);
2497 } else {
2498 std::string AsanVersion = std::to_string(GetAsanVersion(M));
2499 std::string VersionCheckName =
2500 ClInsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2501 std::tie(AsanCtorFunction, std::ignore) =
2502 createSanitizerCtorAndInitFunctions(M, kAsanModuleCtorName,
2503 kAsanInitName, /*InitArgTypes=*/{},
2504 /*InitArgs=*/{}, VersionCheckName);
2505 }
2506
2507 bool CtorComdat = true;
2508 if (ClGlobals) {
2509 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2510 InstrumentGlobals(IRB, M, &CtorComdat);
2511 }
2512
2513 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2514
2515 // Put the constructor and destructor in comdat if both
2516 // (1) global instrumentation is not TU-specific
2517 // (2) target is ELF.
2518 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2519 AsanCtorFunction->setComdat(M.getOrInsertComdat(kAsanModuleCtorName));
2520 appendToGlobalCtors(M, AsanCtorFunction, Priority, AsanCtorFunction);
2521 if (AsanDtorFunction) {
2522 AsanDtorFunction->setComdat(M.getOrInsertComdat(kAsanModuleDtorName));
2523 appendToGlobalDtors(M, AsanDtorFunction, Priority, AsanDtorFunction);
2524 }
2525 } else {
2526 appendToGlobalCtors(M, AsanCtorFunction, Priority);
2527 if (AsanDtorFunction)
2528 appendToGlobalDtors(M, AsanDtorFunction, Priority);
2529 }
2530
2531 return true;
2532 }
2533
initializeCallbacks(Module & M)2534 void AddressSanitizer::initializeCallbacks(Module &M) {
2535 IRBuilder<> IRB(*C);
2536 // Create __asan_report* callbacks.
2537 // IsWrite, TypeSize and Exp are encoded in the function name.
2538 for (int Exp = 0; Exp < 2; Exp++) {
2539 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2540 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2541 const std::string ExpStr = Exp ? "exp_" : "";
2542 const std::string EndingStr = Recover ? "_noabort" : "";
2543
2544 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2545 SmallVector<Type *, 2> Args1{1, IntptrTy};
2546 if (Exp) {
2547 Type *ExpType = Type::getInt32Ty(*C);
2548 Args2.push_back(ExpType);
2549 Args1.push_back(ExpType);
2550 }
2551 AsanErrorCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2552 kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2553 FunctionType::get(IRB.getVoidTy(), Args2, false));
2554
2555 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2556 ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2557 FunctionType::get(IRB.getVoidTy(), Args2, false));
2558
2559 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2560 AccessSizeIndex++) {
2561 const std::string Suffix = TypeStr + itostr(1ULL << AccessSizeIndex);
2562 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2563 M.getOrInsertFunction(
2564 kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2565 FunctionType::get(IRB.getVoidTy(), Args1, false));
2566
2567 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2568 M.getOrInsertFunction(
2569 ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2570 FunctionType::get(IRB.getVoidTy(), Args1, false));
2571 }
2572 }
2573 }
2574
2575 const std::string MemIntrinCallbackPrefix =
2576 CompileKernel ? std::string("") : ClMemoryAccessCallbackPrefix;
2577 AsanMemmove = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memmove",
2578 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2579 IRB.getInt8PtrTy(), IntptrTy);
2580 AsanMemcpy = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memcpy",
2581 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2582 IRB.getInt8PtrTy(), IntptrTy);
2583 AsanMemset = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memset",
2584 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2585 IRB.getInt32Ty(), IntptrTy);
2586
2587 AsanHandleNoReturnFunc =
2588 M.getOrInsertFunction(kAsanHandleNoReturnName, IRB.getVoidTy());
2589
2590 AsanPtrCmpFunction =
2591 M.getOrInsertFunction(kAsanPtrCmp, IRB.getVoidTy(), IntptrTy, IntptrTy);
2592 AsanPtrSubFunction =
2593 M.getOrInsertFunction(kAsanPtrSub, IRB.getVoidTy(), IntptrTy, IntptrTy);
2594 if (Mapping.InGlobal)
2595 AsanShadowGlobal = M.getOrInsertGlobal("__asan_shadow",
2596 ArrayType::get(IRB.getInt8Ty(), 0));
2597 }
2598
maybeInsertAsanInitAtFunctionEntry(Function & F)2599 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
2600 // For each NSObject descendant having a +load method, this method is invoked
2601 // by the ObjC runtime before any of the static constructors is called.
2602 // Therefore we need to instrument such methods with a call to __asan_init
2603 // at the beginning in order to initialize our runtime before any access to
2604 // the shadow memory.
2605 // We cannot just ignore these methods, because they may call other
2606 // instrumented functions.
2607 if (F.getName().find(" load]") != std::string::npos) {
2608 FunctionCallee AsanInitFunction =
2609 declareSanitizerInitFunction(*F.getParent(), kAsanInitName, {});
2610 IRBuilder<> IRB(&F.front(), F.front().begin());
2611 IRB.CreateCall(AsanInitFunction, {});
2612 return true;
2613 }
2614 return false;
2615 }
2616
maybeInsertDynamicShadowAtFunctionEntry(Function & F)2617 bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
2618 // Generate code only when dynamic addressing is needed.
2619 if (Mapping.Offset != kDynamicShadowSentinel)
2620 return false;
2621
2622 IRBuilder<> IRB(&F.front().front());
2623 if (Mapping.InGlobal) {
2624 if (ClWithIfuncSuppressRemat) {
2625 // An empty inline asm with input reg == output reg.
2626 // An opaque pointer-to-int cast, basically.
2627 InlineAsm *Asm = InlineAsm::get(
2628 FunctionType::get(IntptrTy, {AsanShadowGlobal->getType()}, false),
2629 StringRef(""), StringRef("=r,0"),
2630 /*hasSideEffects=*/false);
2631 LocalDynamicShadow =
2632 IRB.CreateCall(Asm, {AsanShadowGlobal}, ".asan.shadow");
2633 } else {
2634 LocalDynamicShadow =
2635 IRB.CreatePointerCast(AsanShadowGlobal, IntptrTy, ".asan.shadow");
2636 }
2637 } else {
2638 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
2639 kAsanShadowMemoryDynamicAddress, IntptrTy);
2640 LocalDynamicShadow = IRB.CreateLoad(IntptrTy, GlobalDynamicAddress);
2641 }
2642 return true;
2643 }
2644
markEscapedLocalAllocas(Function & F)2645 void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
2646 // Find the one possible call to llvm.localescape and pre-mark allocas passed
2647 // to it as uninteresting. This assumes we haven't started processing allocas
2648 // yet. This check is done up front because iterating the use list in
2649 // isInterestingAlloca would be algorithmically slower.
2650 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
2651
2652 // Try to get the declaration of llvm.localescape. If it's not in the module,
2653 // we can exit early.
2654 if (!F.getParent()->getFunction("llvm.localescape")) return;
2655
2656 // Look for a call to llvm.localescape call in the entry block. It can't be in
2657 // any other block.
2658 for (Instruction &I : F.getEntryBlock()) {
2659 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I);
2660 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
2661 // We found a call. Mark all the allocas passed in as uninteresting.
2662 for (Value *Arg : II->arg_operands()) {
2663 AllocaInst *AI = dyn_cast<AllocaInst>(Arg->stripPointerCasts());
2664 assert(AI && AI->isStaticAlloca() &&
2665 "non-static alloca arg to localescape");
2666 ProcessedAllocas[AI] = false;
2667 }
2668 break;
2669 }
2670 }
2671 }
2672
suppressInstrumentationSiteForDebug(int & Instrumented)2673 bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
2674 bool ShouldInstrument =
2675 ClDebugMin < 0 || ClDebugMax < 0 ||
2676 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
2677 Instrumented++;
2678 return !ShouldInstrument;
2679 }
2680
instrumentFunction(Function & F,const TargetLibraryInfo * TLI)2681 bool AddressSanitizer::instrumentFunction(Function &F,
2682 const TargetLibraryInfo *TLI) {
2683 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
2684 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName()) return false;
2685 if (F.getName().startswith("__asan_")) return false;
2686
2687 bool FunctionModified = false;
2688
2689 // If needed, insert __asan_init before checking for SanitizeAddress attr.
2690 // This function needs to be called even if the function body is not
2691 // instrumented.
2692 if (maybeInsertAsanInitAtFunctionEntry(F))
2693 FunctionModified = true;
2694
2695 // Leave if the function doesn't need instrumentation.
2696 if (!F.hasFnAttribute(Attribute::SanitizeAddress)) return FunctionModified;
2697
2698 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
2699
2700 initializeCallbacks(*F.getParent());
2701
2702 FunctionStateRAII CleanupObj(this);
2703
2704 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
2705
2706 // We can't instrument allocas used with llvm.localescape. Only static allocas
2707 // can be passed to that intrinsic.
2708 markEscapedLocalAllocas(F);
2709
2710 // We want to instrument every address only once per basic block (unless there
2711 // are calls between uses).
2712 SmallPtrSet<Value *, 16> TempsToInstrument;
2713 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
2714 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
2715 SmallVector<Instruction *, 8> NoReturnCalls;
2716 SmallVector<BasicBlock *, 16> AllBlocks;
2717 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
2718 int NumAllocas = 0;
2719
2720 // Fill the set of memory operations to instrument.
2721 for (auto &BB : F) {
2722 AllBlocks.push_back(&BB);
2723 TempsToInstrument.clear();
2724 int NumInsnsPerBB = 0;
2725 for (auto &Inst : BB) {
2726 if (LooksLikeCodeInBug11395(&Inst)) return false;
2727 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
2728 getInterestingMemoryOperands(&Inst, InterestingOperands);
2729
2730 if (!InterestingOperands.empty()) {
2731 for (auto &Operand : InterestingOperands) {
2732 if (ClOpt && ClOptSameTemp) {
2733 Value *Ptr = Operand.getPtr();
2734 // If we have a mask, skip instrumentation if we've already
2735 // instrumented the full object. But don't add to TempsToInstrument
2736 // because we might get another load/store with a different mask.
2737 if (Operand.MaybeMask) {
2738 if (TempsToInstrument.count(Ptr))
2739 continue; // We've seen this (whole) temp in the current BB.
2740 } else {
2741 if (!TempsToInstrument.insert(Ptr).second)
2742 continue; // We've seen this temp in the current BB.
2743 }
2744 }
2745 OperandsToInstrument.push_back(Operand);
2746 NumInsnsPerBB++;
2747 }
2748 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
2749 isInterestingPointerComparison(&Inst)) ||
2750 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
2751 isInterestingPointerSubtraction(&Inst))) {
2752 PointerComparisonsOrSubtracts.push_back(&Inst);
2753 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&Inst)) {
2754 // ok, take it.
2755 IntrinToInstrument.push_back(MI);
2756 NumInsnsPerBB++;
2757 } else {
2758 if (isa<AllocaInst>(Inst)) NumAllocas++;
2759 if (auto *CB = dyn_cast<CallBase>(&Inst)) {
2760 // A call inside BB.
2761 TempsToInstrument.clear();
2762 if (CB->doesNotReturn() && !CB->hasMetadata("nosanitize"))
2763 NoReturnCalls.push_back(CB);
2764 }
2765 if (CallInst *CI = dyn_cast<CallInst>(&Inst))
2766 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
2767 }
2768 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
2769 }
2770 }
2771
2772 bool UseCalls = (ClInstrumentationWithCallsThreshold >= 0 &&
2773 OperandsToInstrument.size() + IntrinToInstrument.size() >
2774 (unsigned)ClInstrumentationWithCallsThreshold);
2775 const DataLayout &DL = F.getParent()->getDataLayout();
2776 ObjectSizeOpts ObjSizeOpts;
2777 ObjSizeOpts.RoundToAlign = true;
2778 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext(), ObjSizeOpts);
2779
2780 // Instrument.
2781 int NumInstrumented = 0;
2782 for (auto &Operand : OperandsToInstrument) {
2783 if (!suppressInstrumentationSiteForDebug(NumInstrumented))
2784 instrumentMop(ObjSizeVis, Operand, UseCalls,
2785 F.getParent()->getDataLayout());
2786 FunctionModified = true;
2787 }
2788 for (auto Inst : IntrinToInstrument) {
2789 if (!suppressInstrumentationSiteForDebug(NumInstrumented))
2790 instrumentMemIntrinsic(Inst);
2791 FunctionModified = true;
2792 }
2793
2794 FunctionStackPoisoner FSP(F, *this);
2795 bool ChangedStack = FSP.runOnFunction();
2796
2797 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
2798 // See e.g. https://github.com/google/sanitizers/issues/37
2799 for (auto CI : NoReturnCalls) {
2800 IRBuilder<> IRB(CI);
2801 IRB.CreateCall(AsanHandleNoReturnFunc, {});
2802 }
2803
2804 for (auto Inst : PointerComparisonsOrSubtracts) {
2805 instrumentPointerComparisonOrSubtraction(Inst);
2806 FunctionModified = true;
2807 }
2808
2809 if (ChangedStack || !NoReturnCalls.empty())
2810 FunctionModified = true;
2811
2812 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
2813 << F << "\n");
2814
2815 return FunctionModified;
2816 }
2817
2818 // Workaround for bug 11395: we don't want to instrument stack in functions
2819 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
2820 // FIXME: remove once the bug 11395 is fixed.
LooksLikeCodeInBug11395(Instruction * I)2821 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
2822 if (LongSize != 32) return false;
2823 CallInst *CI = dyn_cast<CallInst>(I);
2824 if (!CI || !CI->isInlineAsm()) return false;
2825 if (CI->getNumArgOperands() <= 5) return false;
2826 // We have inline assembly with quite a few arguments.
2827 return true;
2828 }
2829
initializeCallbacks(Module & M)2830 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
2831 IRBuilder<> IRB(*C);
2832 for (int i = 0; i <= kMaxAsanStackMallocSizeClass; i++) {
2833 std::string Suffix = itostr(i);
2834 AsanStackMallocFunc[i] = M.getOrInsertFunction(
2835 kAsanStackMallocNameTemplate + Suffix, IntptrTy, IntptrTy);
2836 AsanStackFreeFunc[i] =
2837 M.getOrInsertFunction(kAsanStackFreeNameTemplate + Suffix,
2838 IRB.getVoidTy(), IntptrTy, IntptrTy);
2839 }
2840 if (ASan.UseAfterScope) {
2841 AsanPoisonStackMemoryFunc = M.getOrInsertFunction(
2842 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy);
2843 AsanUnpoisonStackMemoryFunc = M.getOrInsertFunction(
2844 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy);
2845 }
2846
2847 for (size_t Val : {0x00, 0xf1, 0xf2, 0xf3, 0xf5, 0xf8}) {
2848 std::ostringstream Name;
2849 Name << kAsanSetShadowPrefix;
2850 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
2851 AsanSetShadowFunc[Val] =
2852 M.getOrInsertFunction(Name.str(), IRB.getVoidTy(), IntptrTy, IntptrTy);
2853 }
2854
2855 AsanAllocaPoisonFunc = M.getOrInsertFunction(
2856 kAsanAllocaPoison, IRB.getVoidTy(), IntptrTy, IntptrTy);
2857 AsanAllocasUnpoisonFunc = M.getOrInsertFunction(
2858 kAsanAllocasUnpoison, IRB.getVoidTy(), IntptrTy, IntptrTy);
2859 }
2860
copyToShadowInline(ArrayRef<uint8_t> ShadowMask,ArrayRef<uint8_t> ShadowBytes,size_t Begin,size_t End,IRBuilder<> & IRB,Value * ShadowBase)2861 void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
2862 ArrayRef<uint8_t> ShadowBytes,
2863 size_t Begin, size_t End,
2864 IRBuilder<> &IRB,
2865 Value *ShadowBase) {
2866 if (Begin >= End)
2867 return;
2868
2869 const size_t LargestStoreSizeInBytes =
2870 std::min<size_t>(sizeof(uint64_t), ASan.LongSize / 8);
2871
2872 const bool IsLittleEndian = F.getParent()->getDataLayout().isLittleEndian();
2873
2874 // Poison given range in shadow using larges store size with out leading and
2875 // trailing zeros in ShadowMask. Zeros never change, so they need neither
2876 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
2877 // middle of a store.
2878 for (size_t i = Begin; i < End;) {
2879 if (!ShadowMask[i]) {
2880 assert(!ShadowBytes[i]);
2881 ++i;
2882 continue;
2883 }
2884
2885 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
2886 // Fit store size into the range.
2887 while (StoreSizeInBytes > End - i)
2888 StoreSizeInBytes /= 2;
2889
2890 // Minimize store size by trimming trailing zeros.
2891 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
2892 while (j <= StoreSizeInBytes / 2)
2893 StoreSizeInBytes /= 2;
2894 }
2895
2896 uint64_t Val = 0;
2897 for (size_t j = 0; j < StoreSizeInBytes; j++) {
2898 if (IsLittleEndian)
2899 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
2900 else
2901 Val = (Val << 8) | ShadowBytes[i + j];
2902 }
2903
2904 Value *Ptr = IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i));
2905 Value *Poison = IRB.getIntN(StoreSizeInBytes * 8, Val);
2906 IRB.CreateAlignedStore(
2907 Poison, IRB.CreateIntToPtr(Ptr, Poison->getType()->getPointerTo()),
2908 Align(1));
2909
2910 i += StoreSizeInBytes;
2911 }
2912 }
2913
copyToShadow(ArrayRef<uint8_t> ShadowMask,ArrayRef<uint8_t> ShadowBytes,IRBuilder<> & IRB,Value * ShadowBase)2914 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
2915 ArrayRef<uint8_t> ShadowBytes,
2916 IRBuilder<> &IRB, Value *ShadowBase) {
2917 copyToShadow(ShadowMask, ShadowBytes, 0, ShadowMask.size(), IRB, ShadowBase);
2918 }
2919
copyToShadow(ArrayRef<uint8_t> ShadowMask,ArrayRef<uint8_t> ShadowBytes,size_t Begin,size_t End,IRBuilder<> & IRB,Value * ShadowBase)2920 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
2921 ArrayRef<uint8_t> ShadowBytes,
2922 size_t Begin, size_t End,
2923 IRBuilder<> &IRB, Value *ShadowBase) {
2924 assert(ShadowMask.size() == ShadowBytes.size());
2925 size_t Done = Begin;
2926 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
2927 if (!ShadowMask[i]) {
2928 assert(!ShadowBytes[i]);
2929 continue;
2930 }
2931 uint8_t Val = ShadowBytes[i];
2932 if (!AsanSetShadowFunc[Val])
2933 continue;
2934
2935 // Skip same values.
2936 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
2937 }
2938
2939 if (j - i >= ClMaxInlinePoisoningSize) {
2940 copyToShadowInline(ShadowMask, ShadowBytes, Done, i, IRB, ShadowBase);
2941 IRB.CreateCall(AsanSetShadowFunc[Val],
2942 {IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i)),
2943 ConstantInt::get(IntptrTy, j - i)});
2944 Done = j;
2945 }
2946 }
2947
2948 copyToShadowInline(ShadowMask, ShadowBytes, Done, End, IRB, ShadowBase);
2949 }
2950
2951 // Fake stack allocator (asan_fake_stack.h) has 11 size classes
2952 // for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
StackMallocSizeClass(uint64_t LocalStackSize)2953 static int StackMallocSizeClass(uint64_t LocalStackSize) {
2954 assert(LocalStackSize <= kMaxStackMallocSize);
2955 uint64_t MaxSize = kMinStackMallocSize;
2956 for (int i = 0;; i++, MaxSize *= 2)
2957 if (LocalStackSize <= MaxSize) return i;
2958 llvm_unreachable("impossible LocalStackSize");
2959 }
2960
copyArgsPassedByValToAllocas()2961 void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
2962 Instruction *CopyInsertPoint = &F.front().front();
2963 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
2964 // Insert after the dynamic shadow location is determined
2965 CopyInsertPoint = CopyInsertPoint->getNextNode();
2966 assert(CopyInsertPoint);
2967 }
2968 IRBuilder<> IRB(CopyInsertPoint);
2969 const DataLayout &DL = F.getParent()->getDataLayout();
2970 for (Argument &Arg : F.args()) {
2971 if (Arg.hasByValAttr()) {
2972 Type *Ty = Arg.getParamByValType();
2973 const Align Alignment =
2974 DL.getValueOrABITypeAlignment(Arg.getParamAlign(), Ty);
2975
2976 AllocaInst *AI = IRB.CreateAlloca(
2977 Ty, nullptr,
2978 (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
2979 ".byval");
2980 AI->setAlignment(Alignment);
2981 Arg.replaceAllUsesWith(AI);
2982
2983 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2984 IRB.CreateMemCpy(AI, Alignment, &Arg, Alignment, AllocSize);
2985 }
2986 }
2987 }
2988
createPHI(IRBuilder<> & IRB,Value * Cond,Value * ValueIfTrue,Instruction * ThenTerm,Value * ValueIfFalse)2989 PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
2990 Value *ValueIfTrue,
2991 Instruction *ThenTerm,
2992 Value *ValueIfFalse) {
2993 PHINode *PHI = IRB.CreatePHI(IntptrTy, 2);
2994 BasicBlock *CondBlock = cast<Instruction>(Cond)->getParent();
2995 PHI->addIncoming(ValueIfFalse, CondBlock);
2996 BasicBlock *ThenBlock = ThenTerm->getParent();
2997 PHI->addIncoming(ValueIfTrue, ThenBlock);
2998 return PHI;
2999 }
3000
createAllocaForLayout(IRBuilder<> & IRB,const ASanStackFrameLayout & L,bool Dynamic)3001 Value *FunctionStackPoisoner::createAllocaForLayout(
3002 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
3003 AllocaInst *Alloca;
3004 if (Dynamic) {
3005 Alloca = IRB.CreateAlloca(IRB.getInt8Ty(),
3006 ConstantInt::get(IRB.getInt64Ty(), L.FrameSize),
3007 "MyAlloca");
3008 } else {
3009 Alloca = IRB.CreateAlloca(ArrayType::get(IRB.getInt8Ty(), L.FrameSize),
3010 nullptr, "MyAlloca");
3011 assert(Alloca->isStaticAlloca());
3012 }
3013 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
3014 size_t FrameAlignment = std::max(L.FrameAlignment, (size_t)ClRealignStack);
3015 Alloca->setAlignment(Align(FrameAlignment));
3016 return IRB.CreatePointerCast(Alloca, IntptrTy);
3017 }
3018
createDynamicAllocasInitStorage()3019 void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
3020 BasicBlock &FirstBB = *F.begin();
3021 IRBuilder<> IRB(dyn_cast<Instruction>(FirstBB.begin()));
3022 DynamicAllocaLayout = IRB.CreateAlloca(IntptrTy, nullptr);
3023 IRB.CreateStore(Constant::getNullValue(IntptrTy), DynamicAllocaLayout);
3024 DynamicAllocaLayout->setAlignment(Align(32));
3025 }
3026
processDynamicAllocas()3027 void FunctionStackPoisoner::processDynamicAllocas() {
3028 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
3029 assert(DynamicAllocaPoisonCallVec.empty());
3030 return;
3031 }
3032
3033 // Insert poison calls for lifetime intrinsics for dynamic allocas.
3034 for (const auto &APC : DynamicAllocaPoisonCallVec) {
3035 assert(APC.InsBefore);
3036 assert(APC.AI);
3037 assert(ASan.isInterestingAlloca(*APC.AI));
3038 assert(!APC.AI->isStaticAlloca());
3039
3040 IRBuilder<> IRB(APC.InsBefore);
3041 poisonAlloca(APC.AI, APC.Size, IRB, APC.DoPoison);
3042 // Dynamic allocas will be unpoisoned unconditionally below in
3043 // unpoisonDynamicAllocas.
3044 // Flag that we need unpoison static allocas.
3045 }
3046
3047 // Handle dynamic allocas.
3048 createDynamicAllocasInitStorage();
3049 for (auto &AI : DynamicAllocaVec)
3050 handleDynamicAllocaCall(AI);
3051 unpoisonDynamicAllocas();
3052 }
3053
3054 /// Collect instructions in the entry block after \p InsBefore which initialize
3055 /// permanent storage for a function argument. These instructions must remain in
3056 /// the entry block so that uninitialized values do not appear in backtraces. An
3057 /// added benefit is that this conserves spill slots. This does not move stores
3058 /// before instrumented / "interesting" allocas.
findStoresToUninstrumentedArgAllocas(AddressSanitizer & ASan,Instruction & InsBefore,SmallVectorImpl<Instruction * > & InitInsts)3059 static void findStoresToUninstrumentedArgAllocas(
3060 AddressSanitizer &ASan, Instruction &InsBefore,
3061 SmallVectorImpl<Instruction *> &InitInsts) {
3062 Instruction *Start = InsBefore.getNextNonDebugInstruction();
3063 for (Instruction *It = Start; It; It = It->getNextNonDebugInstruction()) {
3064 // Argument initialization looks like:
3065 // 1) store <Argument>, <Alloca> OR
3066 // 2) <CastArgument> = cast <Argument> to ...
3067 // store <CastArgument> to <Alloca>
3068 // Do not consider any other kind of instruction.
3069 //
3070 // Note: This covers all known cases, but may not be exhaustive. An
3071 // alternative to pattern-matching stores is to DFS over all Argument uses:
3072 // this might be more general, but is probably much more complicated.
3073 if (isa<AllocaInst>(It) || isa<CastInst>(It))
3074 continue;
3075 if (auto *Store = dyn_cast<StoreInst>(It)) {
3076 // The store destination must be an alloca that isn't interesting for
3077 // ASan to instrument. These are moved up before InsBefore, and they're
3078 // not interesting because allocas for arguments can be mem2reg'd.
3079 auto *Alloca = dyn_cast<AllocaInst>(Store->getPointerOperand());
3080 if (!Alloca || ASan.isInterestingAlloca(*Alloca))
3081 continue;
3082
3083 Value *Val = Store->getValueOperand();
3084 bool IsDirectArgInit = isa<Argument>(Val);
3085 bool IsArgInitViaCast =
3086 isa<CastInst>(Val) &&
3087 isa<Argument>(cast<CastInst>(Val)->getOperand(0)) &&
3088 // Check that the cast appears directly before the store. Otherwise
3089 // moving the cast before InsBefore may break the IR.
3090 Val == It->getPrevNonDebugInstruction();
3091 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3092 if (!IsArgInit)
3093 continue;
3094
3095 if (IsArgInitViaCast)
3096 InitInsts.push_back(cast<Instruction>(Val));
3097 InitInsts.push_back(Store);
3098 continue;
3099 }
3100
3101 // Do not reorder past unknown instructions: argument initialization should
3102 // only involve casts and stores.
3103 return;
3104 }
3105 }
3106
processStaticAllocas()3107 void FunctionStackPoisoner::processStaticAllocas() {
3108 if (AllocaVec.empty()) {
3109 assert(StaticAllocaPoisonCallVec.empty());
3110 return;
3111 }
3112
3113 int StackMallocIdx = -1;
3114 DebugLoc EntryDebugLocation;
3115 if (auto SP = F.getSubprogram())
3116 EntryDebugLocation = DebugLoc::get(SP->getScopeLine(), 0, SP);
3117
3118 Instruction *InsBefore = AllocaVec[0];
3119 IRBuilder<> IRB(InsBefore);
3120
3121 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3122 // debug info is broken, because only entry-block allocas are treated as
3123 // regular stack slots.
3124 auto InsBeforeB = InsBefore->getParent();
3125 assert(InsBeforeB == &F.getEntryBlock());
3126 for (auto *AI : StaticAllocasToMoveUp)
3127 if (AI->getParent() == InsBeforeB)
3128 AI->moveBefore(InsBefore);
3129
3130 // Move stores of arguments into entry-block allocas as well. This prevents
3131 // extra stack slots from being generated (to house the argument values until
3132 // they can be stored into the allocas). This also prevents uninitialized
3133 // values from being shown in backtraces.
3134 SmallVector<Instruction *, 8> ArgInitInsts;
3135 findStoresToUninstrumentedArgAllocas(ASan, *InsBefore, ArgInitInsts);
3136 for (Instruction *ArgInitInst : ArgInitInsts)
3137 ArgInitInst->moveBefore(InsBefore);
3138
3139 // If we have a call to llvm.localescape, keep it in the entry block.
3140 if (LocalEscapeCall) LocalEscapeCall->moveBefore(InsBefore);
3141
3142 SmallVector<ASanStackVariableDescription, 16> SVD;
3143 SVD.reserve(AllocaVec.size());
3144 for (AllocaInst *AI : AllocaVec) {
3145 ASanStackVariableDescription D = {AI->getName().data(),
3146 ASan.getAllocaSizeInBytes(*AI),
3147 0,
3148 AI->getAlignment(),
3149 AI,
3150 0,
3151 0};
3152 SVD.push_back(D);
3153 }
3154
3155 // Minimal header size (left redzone) is 4 pointers,
3156 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3157 size_t Granularity = 1ULL << Mapping.Scale;
3158 size_t MinHeaderSize = std::max((size_t)ASan.LongSize / 2, Granularity);
3159 const ASanStackFrameLayout &L =
3160 ComputeASanStackFrameLayout(SVD, Granularity, MinHeaderSize);
3161
3162 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3163 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3164 for (auto &Desc : SVD)
3165 AllocaToSVDMap[Desc.AI] = &Desc;
3166
3167 // Update SVD with information from lifetime intrinsics.
3168 for (const auto &APC : StaticAllocaPoisonCallVec) {
3169 assert(APC.InsBefore);
3170 assert(APC.AI);
3171 assert(ASan.isInterestingAlloca(*APC.AI));
3172 assert(APC.AI->isStaticAlloca());
3173
3174 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3175 Desc.LifetimeSize = Desc.Size;
3176 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3177 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3178 if (LifetimeLoc->getFile() == FnLoc->getFile())
3179 if (unsigned Line = LifetimeLoc->getLine())
3180 Desc.Line = std::min(Desc.Line ? Desc.Line : Line, Line);
3181 }
3182 }
3183 }
3184
3185 auto DescriptionString = ComputeASanStackFrameDescription(SVD);
3186 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3187 uint64_t LocalStackSize = L.FrameSize;
3188 bool DoStackMalloc = ClUseAfterReturn && !ASan.CompileKernel &&
3189 LocalStackSize <= kMaxStackMallocSize;
3190 bool DoDynamicAlloca = ClDynamicAllocaStack;
3191 // Don't do dynamic alloca or stack malloc if:
3192 // 1) There is inline asm: too often it makes assumptions on which registers
3193 // are available.
3194 // 2) There is a returns_twice call (typically setjmp), which is
3195 // optimization-hostile, and doesn't play well with introduced indirect
3196 // register-relative calculation of local variable addresses.
3197 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3198 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3199
3200 Value *StaticAlloca =
3201 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, false);
3202
3203 Value *FakeStack;
3204 Value *LocalStackBase;
3205 Value *LocalStackBaseAlloca;
3206 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3207
3208 if (DoStackMalloc) {
3209 LocalStackBaseAlloca =
3210 IRB.CreateAlloca(IntptrTy, nullptr, "asan_local_stack_base");
3211 // void *FakeStack = __asan_option_detect_stack_use_after_return
3212 // ? __asan_stack_malloc_N(LocalStackSize)
3213 // : nullptr;
3214 // void *LocalStackBase = (FakeStack) ? FakeStack : alloca(LocalStackSize);
3215 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3216 kAsanOptionDetectUseAfterReturn, IRB.getInt32Ty());
3217 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3218 IRB.CreateLoad(IRB.getInt32Ty(), OptionDetectUseAfterReturn),
3219 Constant::getNullValue(IRB.getInt32Ty()));
3220 Instruction *Term =
3221 SplitBlockAndInsertIfThen(UseAfterReturnIsEnabled, InsBefore, false);
3222 IRBuilder<> IRBIf(Term);
3223 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3224 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3225 Value *FakeStackValue =
3226 IRBIf.CreateCall(AsanStackMallocFunc[StackMallocIdx],
3227 ConstantInt::get(IntptrTy, LocalStackSize));
3228 IRB.SetInsertPoint(InsBefore);
3229 FakeStack = createPHI(IRB, UseAfterReturnIsEnabled, FakeStackValue, Term,
3230 ConstantInt::get(IntptrTy, 0));
3231
3232 Value *NoFakeStack =
3233 IRB.CreateICmpEQ(FakeStack, Constant::getNullValue(IntptrTy));
3234 Term = SplitBlockAndInsertIfThen(NoFakeStack, InsBefore, false);
3235 IRBIf.SetInsertPoint(Term);
3236 Value *AllocaValue =
3237 DoDynamicAlloca ? createAllocaForLayout(IRBIf, L, true) : StaticAlloca;
3238
3239 IRB.SetInsertPoint(InsBefore);
3240 LocalStackBase = createPHI(IRB, NoFakeStack, AllocaValue, Term, FakeStack);
3241 IRB.CreateStore(LocalStackBase, LocalStackBaseAlloca);
3242 DIExprFlags |= DIExpression::DerefBefore;
3243 } else {
3244 // void *FakeStack = nullptr;
3245 // void *LocalStackBase = alloca(LocalStackSize);
3246 FakeStack = ConstantInt::get(IntptrTy, 0);
3247 LocalStackBase =
3248 DoDynamicAlloca ? createAllocaForLayout(IRB, L, true) : StaticAlloca;
3249 LocalStackBaseAlloca = LocalStackBase;
3250 }
3251
3252 // It shouldn't matter whether we pass an `alloca` or a `ptrtoint` as the
3253 // dbg.declare address opereand, but passing a `ptrtoint` seems to confuse
3254 // later passes and can result in dropped variable coverage in debug info.
3255 Value *LocalStackBaseAllocaPtr =
3256 isa<PtrToIntInst>(LocalStackBaseAlloca)
3257 ? cast<PtrToIntInst>(LocalStackBaseAlloca)->getPointerOperand()
3258 : LocalStackBaseAlloca;
3259 assert(isa<AllocaInst>(LocalStackBaseAllocaPtr) &&
3260 "Variable descriptions relative to ASan stack base will be dropped");
3261
3262 // Replace Alloca instructions with base+offset.
3263 for (const auto &Desc : SVD) {
3264 AllocaInst *AI = Desc.AI;
3265 replaceDbgDeclare(AI, LocalStackBaseAllocaPtr, DIB, DIExprFlags,
3266 Desc.Offset);
3267 Value *NewAllocaPtr = IRB.CreateIntToPtr(
3268 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Desc.Offset)),
3269 AI->getType());
3270 AI->replaceAllUsesWith(NewAllocaPtr);
3271 }
3272
3273 // The left-most redzone has enough space for at least 4 pointers.
3274 // Write the Magic value to redzone[0].
3275 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
3276 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
3277 BasePlus0);
3278 // Write the frame description constant to redzone[1].
3279 Value *BasePlus1 = IRB.CreateIntToPtr(
3280 IRB.CreateAdd(LocalStackBase,
3281 ConstantInt::get(IntptrTy, ASan.LongSize / 8)),
3282 IntptrPtrTy);
3283 GlobalVariable *StackDescriptionGlobal =
3284 createPrivateGlobalForString(*F.getParent(), DescriptionString,
3285 /*AllowMerging*/ true, kAsanGenPrefix);
3286 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal, IntptrTy);
3287 IRB.CreateStore(Description, BasePlus1);
3288 // Write the PC to redzone[2].
3289 Value *BasePlus2 = IRB.CreateIntToPtr(
3290 IRB.CreateAdd(LocalStackBase,
3291 ConstantInt::get(IntptrTy, 2 * ASan.LongSize / 8)),
3292 IntptrPtrTy);
3293 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
3294
3295 const auto &ShadowAfterScope = GetShadowBytesAfterScope(SVD, L);
3296
3297 // Poison the stack red zones at the entry.
3298 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
3299 // As mask we must use most poisoned case: red zones and after scope.
3300 // As bytes we can use either the same or just red zones only.
3301 copyToShadow(ShadowAfterScope, ShadowAfterScope, IRB, ShadowBase);
3302
3303 if (!StaticAllocaPoisonCallVec.empty()) {
3304 const auto &ShadowInScope = GetShadowBytes(SVD, L);
3305
3306 // Poison static allocas near lifetime intrinsics.
3307 for (const auto &APC : StaticAllocaPoisonCallVec) {
3308 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3309 assert(Desc.Offset % L.Granularity == 0);
3310 size_t Begin = Desc.Offset / L.Granularity;
3311 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3312
3313 IRBuilder<> IRB(APC.InsBefore);
3314 copyToShadow(ShadowAfterScope,
3315 APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3316 IRB, ShadowBase);
3317 }
3318 }
3319
3320 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3321 SmallVector<uint8_t, 64> ShadowAfterReturn;
3322
3323 // (Un)poison the stack before all ret instructions.
3324 for (Instruction *Ret : RetVec) {
3325 IRBuilder<> IRBRet(Ret);
3326 // Mark the current frame as retired.
3327 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
3328 BasePlus0);
3329 if (DoStackMalloc) {
3330 assert(StackMallocIdx >= 0);
3331 // if FakeStack != 0 // LocalStackBase == FakeStack
3332 // // In use-after-return mode, poison the whole stack frame.
3333 // if StackMallocIdx <= 4
3334 // // For small sizes inline the whole thing:
3335 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3336 // **SavedFlagPtr(FakeStack) = 0
3337 // else
3338 // __asan_stack_free_N(FakeStack, LocalStackSize)
3339 // else
3340 // <This is not a fake stack; unpoison the redzones>
3341 Value *Cmp =
3342 IRBRet.CreateICmpNE(FakeStack, Constant::getNullValue(IntptrTy));
3343 Instruction *ThenTerm, *ElseTerm;
3344 SplitBlockAndInsertIfThenElse(Cmp, Ret, &ThenTerm, &ElseTerm);
3345
3346 IRBuilder<> IRBPoison(ThenTerm);
3347 if (StackMallocIdx <= 4) {
3348 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3349 ShadowAfterReturn.resize(ClassSize / L.Granularity,
3350 kAsanStackUseAfterReturnMagic);
3351 copyToShadow(ShadowAfterReturn, ShadowAfterReturn, IRBPoison,
3352 ShadowBase);
3353 Value *SavedFlagPtrPtr = IRBPoison.CreateAdd(
3354 FakeStack,
3355 ConstantInt::get(IntptrTy, ClassSize - ASan.LongSize / 8));
3356 Value *SavedFlagPtr = IRBPoison.CreateLoad(
3357 IntptrTy, IRBPoison.CreateIntToPtr(SavedFlagPtrPtr, IntptrPtrTy));
3358 IRBPoison.CreateStore(
3359 Constant::getNullValue(IRBPoison.getInt8Ty()),
3360 IRBPoison.CreateIntToPtr(SavedFlagPtr, IRBPoison.getInt8PtrTy()));
3361 } else {
3362 // For larger frames call __asan_stack_free_*.
3363 IRBPoison.CreateCall(
3364 AsanStackFreeFunc[StackMallocIdx],
3365 {FakeStack, ConstantInt::get(IntptrTy, LocalStackSize)});
3366 }
3367
3368 IRBuilder<> IRBElse(ElseTerm);
3369 copyToShadow(ShadowAfterScope, ShadowClean, IRBElse, ShadowBase);
3370 } else {
3371 copyToShadow(ShadowAfterScope, ShadowClean, IRBRet, ShadowBase);
3372 }
3373 }
3374
3375 // We are done. Remove the old unused alloca instructions.
3376 for (auto AI : AllocaVec) AI->eraseFromParent();
3377 }
3378
poisonAlloca(Value * V,uint64_t Size,IRBuilder<> & IRB,bool DoPoison)3379 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3380 IRBuilder<> &IRB, bool DoPoison) {
3381 // For now just insert the call to ASan runtime.
3382 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
3383 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
3384 IRB.CreateCall(
3385 DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3386 {AddrArg, SizeArg});
3387 }
3388
3389 // Handling llvm.lifetime intrinsics for a given %alloca:
3390 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3391 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3392 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3393 // could be poisoned by previous llvm.lifetime.end instruction, as the
3394 // variable may go in and out of scope several times, e.g. in loops).
3395 // (3) if we poisoned at least one %alloca in a function,
3396 // unpoison the whole stack frame at function exit.
handleDynamicAllocaCall(AllocaInst * AI)3397 void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3398 IRBuilder<> IRB(AI);
3399
3400 const unsigned Alignment = std::max(kAllocaRzSize, AI->getAlignment());
3401 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3402
3403 Value *Zero = Constant::getNullValue(IntptrTy);
3404 Value *AllocaRzSize = ConstantInt::get(IntptrTy, kAllocaRzSize);
3405 Value *AllocaRzMask = ConstantInt::get(IntptrTy, AllocaRedzoneMask);
3406
3407 // Since we need to extend alloca with additional memory to locate
3408 // redzones, and OldSize is number of allocated blocks with
3409 // ElementSize size, get allocated memory size in bytes by
3410 // OldSize * ElementSize.
3411 const unsigned ElementSize =
3412 F.getParent()->getDataLayout().getTypeAllocSize(AI->getAllocatedType());
3413 Value *OldSize =
3414 IRB.CreateMul(IRB.CreateIntCast(AI->getArraySize(), IntptrTy, false),
3415 ConstantInt::get(IntptrTy, ElementSize));
3416
3417 // PartialSize = OldSize % 32
3418 Value *PartialSize = IRB.CreateAnd(OldSize, AllocaRzMask);
3419
3420 // Misalign = kAllocaRzSize - PartialSize;
3421 Value *Misalign = IRB.CreateSub(AllocaRzSize, PartialSize);
3422
3423 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3424 Value *Cond = IRB.CreateICmpNE(Misalign, AllocaRzSize);
3425 Value *PartialPadding = IRB.CreateSelect(Cond, Misalign, Zero);
3426
3427 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3428 // Alignment is added to locate left redzone, PartialPadding for possible
3429 // partial redzone and kAllocaRzSize for right redzone respectively.
3430 Value *AdditionalChunkSize = IRB.CreateAdd(
3431 ConstantInt::get(IntptrTy, Alignment + kAllocaRzSize), PartialPadding);
3432
3433 Value *NewSize = IRB.CreateAdd(OldSize, AdditionalChunkSize);
3434
3435 // Insert new alloca with new NewSize and Alignment params.
3436 AllocaInst *NewAlloca = IRB.CreateAlloca(IRB.getInt8Ty(), NewSize);
3437 NewAlloca->setAlignment(Align(Alignment));
3438
3439 // NewAddress = Address + Alignment
3440 Value *NewAddress = IRB.CreateAdd(IRB.CreatePtrToInt(NewAlloca, IntptrTy),
3441 ConstantInt::get(IntptrTy, Alignment));
3442
3443 // Insert __asan_alloca_poison call for new created alloca.
3444 IRB.CreateCall(AsanAllocaPoisonFunc, {NewAddress, OldSize});
3445
3446 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3447 // for unpoisoning stuff.
3448 IRB.CreateStore(IRB.CreatePtrToInt(NewAlloca, IntptrTy), DynamicAllocaLayout);
3449
3450 Value *NewAddressPtr = IRB.CreateIntToPtr(NewAddress, AI->getType());
3451
3452 // Replace all uses of AddessReturnedByAlloca with NewAddressPtr.
3453 AI->replaceAllUsesWith(NewAddressPtr);
3454
3455 // We are done. Erase old alloca from parent.
3456 AI->eraseFromParent();
3457 }
3458
3459 // isSafeAccess returns true if Addr is always inbounds with respect to its
3460 // base object. For example, it is a field access or an array access with
3461 // constant inbounds index.
isSafeAccess(ObjectSizeOffsetVisitor & ObjSizeVis,Value * Addr,uint64_t TypeSize) const3462 bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3463 Value *Addr, uint64_t TypeSize) const {
3464 SizeOffsetType SizeOffset = ObjSizeVis.compute(Addr);
3465 if (!ObjSizeVis.bothKnown(SizeOffset)) return false;
3466 uint64_t Size = SizeOffset.first.getZExtValue();
3467 int64_t Offset = SizeOffset.second.getSExtValue();
3468 // Three checks are required to ensure safety:
3469 // . Offset >= 0 (since the offset is given from the base ptr)
3470 // . Size >= Offset (unsigned)
3471 // . Size - Offset >= NeededSize (unsigned)
3472 return Offset >= 0 && Size >= uint64_t(Offset) &&
3473 Size - uint64_t(Offset) >= TypeSize / 8;
3474 }
3475