1 | //===-- nsan.cc -----------------------------------------------------------===// |
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 | // NumericalStabilitySanitizer runtime. |
10 | // |
11 | // This implements: |
12 | // - The public nsan interface (include/sanitizer/nsan_interface.h). |
13 | // - The private nsan interface (./nsan.h). |
14 | // - The internal instrumentation interface. These are function emitted by the |
15 | // instrumentation pass: |
16 | // * __nsan_get_shadow_ptr_for_{float,double,longdouble}_load |
17 | // These return the shadow memory pointer for loading the shadow value, |
18 | // after checking that the types are consistent. If the types are not |
19 | // consistent, returns nullptr. |
20 | // * __nsan_get_shadow_ptr_for_{float,double,longdouble}_store |
21 | // Sets the shadow types appropriately and returns the shadow memory |
22 | // pointer for storing the shadow value. |
23 | // * __nsan_internal_check_{float,double,long double}_{f,d,l} checks the |
24 | // accuracy of a value against its shadow and emits a warning depending |
25 | // on the runtime configuration. The middle part indicates the type of |
26 | // the application value, the suffix (f,d,l) indicates the type of the |
27 | // shadow, and depends on the instrumentation configuration. |
28 | // * __nsan_fcmp_fail_* emits a warning for a fcmp instruction whose |
29 | // corresponding shadow fcmp result differs. |
30 | // |
31 | //===----------------------------------------------------------------------===// |
32 | |
33 | #include "nsan.h" |
34 | #include "nsan_flags.h" |
35 | #include "nsan_stats.h" |
36 | #include "nsan_suppressions.h" |
37 | #include "nsan_thread.h" |
38 | |
39 | #include <assert.h> |
40 | #include <math.h> |
41 | #include <stdint.h> |
42 | #include <stdio.h> |
43 | #include <stdlib.h> |
44 | |
45 | #include "sanitizer_common/sanitizer_atomic.h" |
46 | #include "sanitizer_common/sanitizer_common.h" |
47 | #include "sanitizer_common/sanitizer_libc.h" |
48 | #include "sanitizer_common/sanitizer_report_decorator.h" |
49 | #include "sanitizer_common/sanitizer_stacktrace.h" |
50 | #include "sanitizer_common/sanitizer_symbolizer.h" |
51 | |
52 | using namespace __sanitizer; |
53 | using namespace __nsan; |
54 | |
55 | constexpr int kMaxVectorWidth = 8; |
56 | |
57 | // When copying application memory, we also copy its shadow and shadow type. |
58 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
59 | __nsan_copy_values(const void *daddr, const void *saddr, uptr size) { |
60 | internal_memmove(dest: GetShadowTypeAddrFor(ptr: daddr), src: GetShadowTypeAddrFor(ptr: saddr), |
61 | n: size); |
62 | internal_memmove(dest: GetShadowAddrFor(ptr: daddr), src: GetShadowAddrFor(ptr: saddr), |
63 | n: size * kShadowScale); |
64 | } |
65 | |
66 | #define NSAN_COPY_VALUES_N(N) \ |
67 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_copy_##N( \ |
68 | const u8 *daddr, const u8 *saddr) { \ |
69 | __builtin_memmove(GetShadowTypeAddrFor(daddr), \ |
70 | GetShadowTypeAddrFor(saddr), N); \ |
71 | __builtin_memmove(GetShadowAddrFor(daddr), GetShadowAddrFor(saddr), \ |
72 | N *kShadowScale); \ |
73 | } |
74 | |
75 | NSAN_COPY_VALUES_N(4) |
76 | NSAN_COPY_VALUES_N(8) |
77 | NSAN_COPY_VALUES_N(16) |
78 | |
79 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
80 | __nsan_set_value_unknown(const void *addr, uptr size) { |
81 | internal_memset(s: GetShadowTypeAddrFor(ptr: addr), c: 0, n: size); |
82 | } |
83 | |
84 | #define NSAN_SET_VALUE_UNKNOWN_N(N) \ |
85 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_set_value_unknown_##N( \ |
86 | const u8 *daddr) { \ |
87 | __builtin_memset(GetShadowTypeAddrFor(daddr), 0, N); \ |
88 | } |
89 | |
90 | NSAN_SET_VALUE_UNKNOWN_N(4) |
91 | NSAN_SET_VALUE_UNKNOWN_N(8) |
92 | NSAN_SET_VALUE_UNKNOWN_N(16) |
93 | |
94 | const char *FTInfo<float>::kCppTypeName = "float" ; |
95 | const char *FTInfo<double>::kCppTypeName = "double" ; |
96 | const char *FTInfo<long double>::kCppTypeName = "long double" ; |
97 | const char *FTInfo<__float128>::kCppTypeName = "__float128" ; |
98 | |
99 | const char FTInfo<float>::kTypePattern[sizeof(float)]; |
100 | const char FTInfo<double>::kTypePattern[sizeof(double)]; |
101 | const char FTInfo<long double>::kTypePattern[sizeof(long double)]; |
102 | |
103 | // Helper for __nsan_dump_shadow_mem: Reads the value at address `ptr`, |
104 | // identified by its type id. |
105 | template <typename ShadowFT> |
106 | static __float128 ReadShadowInternal(const u8 *ptr) { |
107 | ShadowFT Shadow; |
108 | __builtin_memcpy(&Shadow, ptr, sizeof(Shadow)); |
109 | return Shadow; |
110 | } |
111 | |
112 | static __float128 ReadShadow(const u8 *ptr, const char ShadowTypeId) { |
113 | switch (ShadowTypeId) { |
114 | case 'd': |
115 | return ReadShadowInternal<double>(ptr); |
116 | case 'l': |
117 | return ReadShadowInternal<long double>(ptr); |
118 | case 'q': |
119 | return ReadShadowInternal<__float128>(ptr); |
120 | default: |
121 | return 0.0; |
122 | } |
123 | } |
124 | |
125 | namespace { |
126 | class Decorator : public __sanitizer::SanitizerCommonDecorator { |
127 | public: |
128 | Decorator() : SanitizerCommonDecorator() {} |
129 | const char *Warning() { return Red(); } |
130 | const char *Name() { return Green(); } |
131 | const char *End() { return Default(); } |
132 | }; |
133 | |
134 | // Workaround for the fact that Printf() does not support floats. |
135 | struct PrintBuffer { |
136 | char Buffer[64]; |
137 | }; |
138 | template <typename FT> struct FTPrinter {}; |
139 | |
140 | template <> struct FTPrinter<double> { |
141 | static PrintBuffer dec(double value) { |
142 | PrintBuffer result; |
143 | snprintf(s: result.Buffer, maxlen: sizeof(result.Buffer) - 1, format: "%.20f" , value); |
144 | return result; |
145 | } |
146 | static PrintBuffer hex(double value) { |
147 | PrintBuffer result; |
148 | snprintf(s: result.Buffer, maxlen: sizeof(result.Buffer) - 1, format: "%.20a" , value); |
149 | return result; |
150 | } |
151 | }; |
152 | |
153 | template <> struct FTPrinter<float> : FTPrinter<double> {}; |
154 | |
155 | template <> struct FTPrinter<long double> { |
156 | static PrintBuffer dec(long double value) { |
157 | PrintBuffer result; |
158 | snprintf(s: result.Buffer, maxlen: sizeof(result.Buffer) - 1, format: "%.20Lf" , value); |
159 | return result; |
160 | } |
161 | static PrintBuffer hex(long double value) { |
162 | PrintBuffer result; |
163 | snprintf(s: result.Buffer, maxlen: sizeof(result.Buffer) - 1, format: "%.20La" , value); |
164 | return result; |
165 | } |
166 | }; |
167 | |
168 | // FIXME: print with full precision. |
169 | template <> struct FTPrinter<__float128> : FTPrinter<long double> {}; |
170 | |
171 | // This is a template so that there are no implicit conversions. |
172 | template <typename FT> inline FT ftAbs(FT v); |
173 | |
174 | template <> inline long double ftAbs(long double v) { return fabsl(x: v); } |
175 | template <> inline double ftAbs(double v) { return fabs(x: v); } |
176 | |
177 | // We don't care about nans. |
178 | // std::abs(__float128) code is suboptimal and generates a function call to |
179 | // __getf2(). |
180 | template <typename FT> inline FT ftAbs(FT v) { return v >= FT{0} ? v : -v; } |
181 | |
182 | template <typename FT1, typename FT2, bool Enable> struct LargestFTImpl { |
183 | using type = FT2; |
184 | }; |
185 | |
186 | template <typename FT1, typename FT2> struct LargestFTImpl<FT1, FT2, true> { |
187 | using type = FT1; |
188 | }; |
189 | |
190 | template <typename FT1, typename FT2> |
191 | using LargestFT = |
192 | typename LargestFTImpl<FT1, FT2, (sizeof(FT1) > sizeof(FT2))>::type; |
193 | |
194 | template <typename T> T max(T a, T b) { return a < b ? b : a; } |
195 | |
196 | } // end anonymous namespace |
197 | |
198 | void __sanitizer::BufferedStackTrace::UnwindImpl(uptr pc, uptr bp, |
199 | void *context, |
200 | bool request_fast, |
201 | u32 max_depth) { |
202 | using namespace __nsan; |
203 | NsanThread *t = GetCurrentThread(); |
204 | if (!t || !StackTrace::WillUseFastUnwind(request_fast_unwind: request_fast)) |
205 | return Unwind(max_depth, pc, bp, context, stack_top: t ? t->stack_top() : 0, |
206 | stack_bottom: t ? t->stack_bottom() : 0, request_fast_unwind: false); |
207 | if (StackTrace::WillUseFastUnwind(request_fast_unwind: request_fast)) |
208 | Unwind(max_depth, pc, bp, context: nullptr, stack_top: t->stack_top(), stack_bottom: t->stack_bottom(), request_fast_unwind: true); |
209 | else |
210 | Unwind(max_depth, pc, bp: 0, context, stack_top: 0, stack_bottom: 0, request_fast_unwind: false); |
211 | } |
212 | |
213 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_print_accumulated_stats() { |
214 | if (nsan_stats) |
215 | nsan_stats->Print(); |
216 | } |
217 | |
218 | static void NsanAtexit() { |
219 | Printf(format: "Numerical Sanitizer exit stats:\n" ); |
220 | __nsan_print_accumulated_stats(); |
221 | nsan_stats = nullptr; |
222 | } |
223 | |
224 | // The next three functions return a pointer for storing a shadow value for `n` |
225 | // values, after setting the shadow types. We return the pointer instead of |
226 | // storing ourselves because it avoids having to rely on the calling convention |
227 | // around long double being the same for nsan and the target application. |
228 | // We have to have 3 versions because we need to know which type we are storing |
229 | // since we are setting the type shadow memory. |
230 | template <typename FT> static u8 *getShadowPtrForStore(u8 *store_addr, uptr n) { |
231 | unsigned char *shadow_type = GetShadowTypeAddrFor(ptr: store_addr); |
232 | for (uptr i = 0; i < n; ++i) { |
233 | __builtin_memcpy(shadow_type + i * sizeof(FT), FTInfo<FT>::kTypePattern, |
234 | sizeof(FTInfo<FT>::kTypePattern)); |
235 | } |
236 | return GetShadowAddrFor(ptr: store_addr); |
237 | } |
238 | |
239 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE u8 * |
240 | __nsan_get_shadow_ptr_for_float_store(u8 *store_addr, uptr n) { |
241 | return getShadowPtrForStore<float>(store_addr, n); |
242 | } |
243 | |
244 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE u8 * |
245 | __nsan_get_shadow_ptr_for_double_store(u8 *store_addr, uptr n) { |
246 | return getShadowPtrForStore<double>(store_addr, n); |
247 | } |
248 | |
249 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE u8 * |
250 | __nsan_get_shadow_ptr_for_longdouble_store(u8 *store_addr, uptr n) { |
251 | return getShadowPtrForStore<long double>(store_addr, n); |
252 | } |
253 | |
254 | template <typename FT> static bool IsValidShadowType(const u8 *shadow_type) { |
255 | return __builtin_memcmp(shadow_type, FTInfo<FT>::kTypePattern, sizeof(FT)) == |
256 | 0; |
257 | } |
258 | |
259 | template <int kSize, typename T> static bool IsZero(const T *ptr) { |
260 | constexpr const char kZeros[kSize] = {}; // Zero initialized. |
261 | return __builtin_memcmp(ptr, kZeros, kSize) == 0; |
262 | } |
263 | |
264 | template <typename FT> static bool IsUnknownShadowType(const u8 *shadow_type) { |
265 | return IsZero<sizeof(FTInfo<FT>::kTypePattern)>(shadow_type); |
266 | } |
267 | |
268 | // The three folowing functions check that the address stores a complete |
269 | // shadow value of the given type and return a pointer for loading. |
270 | // They return nullptr if the type of the value is unknown or incomplete. |
271 | template <typename FT> |
272 | static const u8 *getShadowPtrForLoad(const u8 *load_addr, uptr n) { |
273 | const u8 *const shadow_type = GetShadowTypeAddrFor(ptr: load_addr); |
274 | for (uptr i = 0; i < n; ++i) { |
275 | if (!IsValidShadowType<FT>(shadow_type + i * sizeof(FT))) { |
276 | // If loadtracking stats are enabled, log loads with invalid types |
277 | // (tampered with through type punning). |
278 | if (flags().enable_loadtracking_stats) { |
279 | if (IsUnknownShadowType<FT>(shadow_type + i * sizeof(FT))) { |
280 | // Warn only if the value is non-zero. Zero is special because |
281 | // applications typically initialize large buffers to zero in an |
282 | // untyped way. |
283 | if (!IsZero<sizeof(FT)>(load_addr)) { |
284 | GET_CALLER_PC_BP; |
285 | nsan_stats->AddUnknownLoadTrackingEvent(pc, bp); |
286 | } |
287 | } else { |
288 | GET_CALLER_PC_BP; |
289 | nsan_stats->AddInvalidLoadTrackingEvent(pc, bp); |
290 | } |
291 | } |
292 | return nullptr; |
293 | } |
294 | } |
295 | return GetShadowAddrFor(ptr: load_addr); |
296 | } |
297 | |
298 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE const u8 * |
299 | __nsan_get_shadow_ptr_for_float_load(const u8 *load_addr, uptr n) { |
300 | return getShadowPtrForLoad<float>(load_addr, n); |
301 | } |
302 | |
303 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE const u8 * |
304 | __nsan_get_shadow_ptr_for_double_load(const u8 *load_addr, uptr n) { |
305 | return getShadowPtrForLoad<double>(load_addr, n); |
306 | } |
307 | |
308 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE const u8 * |
309 | __nsan_get_shadow_ptr_for_longdouble_load(const u8 *load_addr, uptr n) { |
310 | return getShadowPtrForLoad<long double>(load_addr, n); |
311 | } |
312 | |
313 | // Returns the raw shadow pointer. The returned pointer should be considered |
314 | // opaque. |
315 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE u8 * |
316 | __nsan_internal_get_raw_shadow_ptr(const u8 *addr) { |
317 | return GetShadowAddrFor(ptr: addr); |
318 | } |
319 | |
320 | // Returns the raw shadow type pointer. The returned pointer should be |
321 | // considered opaque. |
322 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE u8 * |
323 | __nsan_internal_get_raw_shadow_type_ptr(const u8 *addr) { |
324 | return reinterpret_cast<u8 *>(GetShadowTypeAddrFor(ptr: addr)); |
325 | } |
326 | |
327 | static ValueType getValueType(u8 c) { return static_cast<ValueType>(c & 0x3); } |
328 | |
329 | static int getValuePos(u8 c) { return c >> kValueSizeSizeBits; } |
330 | |
331 | // Checks the consistency of the value types at the given type pointer. |
332 | // If the value is inconsistent, returns ValueType::kUnknown. Else, return the |
333 | // consistent type. |
334 | template <typename FT> |
335 | static bool checkValueConsistency(const u8 *shadow_type) { |
336 | const int pos = getValuePos(c: *shadow_type); |
337 | // Check that all bytes from the start of the value are ordered. |
338 | for (uptr i = 0; i < sizeof(FT); ++i) { |
339 | const u8 T = *(shadow_type - pos + i); |
340 | if (!(getValueType(c: T) == FTInfo<FT>::kValueType && getValuePos(c: T) == i)) |
341 | return false; |
342 | } |
343 | return true; |
344 | } |
345 | |
346 | // The instrumentation automatically appends `shadow_value_type_ids`, see |
347 | // maybeAddSuffixForNsanInterface. |
348 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
349 | __nsan_dump_shadow_mem(const u8 *addr, size_t size_bytes, size_t bytes_per_line, |
350 | size_t shadow_value_type_ids) { |
351 | const u8 *const shadow_type = GetShadowTypeAddrFor(ptr: addr); |
352 | const u8 *const shadow = GetShadowAddrFor(ptr: addr); |
353 | |
354 | constexpr int kMaxNumDecodedValues = 16; |
355 | __float128 decoded_values[kMaxNumDecodedValues]; |
356 | int num_decoded_values = 0; |
357 | if (bytes_per_line > 4 * kMaxNumDecodedValues) |
358 | bytes_per_line = 4 * kMaxNumDecodedValues; |
359 | |
360 | // We keep track of the current type and position as we go. |
361 | ValueType LastValueTy = kUnknownValueType; |
362 | int LastPos = -1; |
363 | size_t Offset = 0; |
364 | for (size_t R = 0; R < (size_bytes + bytes_per_line - 1) / bytes_per_line; |
365 | ++R) { |
366 | printf(format: "%p: " , (void *)(addr + R * bytes_per_line)); |
367 | for (size_t C = 0; C < bytes_per_line && Offset < size_bytes; ++C) { |
368 | const ValueType ValueTy = getValueType(c: shadow_type[Offset]); |
369 | const int pos = getValuePos(c: shadow_type[Offset]); |
370 | if (ValueTy == LastValueTy && pos == LastPos + 1) { |
371 | ++LastPos; |
372 | } else { |
373 | LastValueTy = ValueTy; |
374 | LastPos = pos == 0 ? 0 : -1; |
375 | } |
376 | |
377 | switch (ValueTy) { |
378 | case kUnknownValueType: |
379 | printf(format: "__ " ); |
380 | break; |
381 | case kFloatValueType: |
382 | printf(format: "f%x " , pos); |
383 | if (LastPos == sizeof(float) - 1) { |
384 | decoded_values[num_decoded_values] = |
385 | ReadShadow(ptr: shadow + kShadowScale * (Offset + 1 - sizeof(float)), |
386 | ShadowTypeId: static_cast<char>(shadow_value_type_ids & 0xff)); |
387 | ++num_decoded_values; |
388 | } |
389 | break; |
390 | case kDoubleValueType: |
391 | printf(format: "d%x " , pos); |
392 | if (LastPos == sizeof(double) - 1) { |
393 | decoded_values[num_decoded_values] = ReadShadow( |
394 | ptr: shadow + kShadowScale * (Offset + 1 - sizeof(double)), |
395 | ShadowTypeId: static_cast<char>((shadow_value_type_ids >> 8) & 0xff)); |
396 | ++num_decoded_values; |
397 | } |
398 | break; |
399 | case kFp80ValueType: |
400 | printf(format: "l%x " , pos); |
401 | if (LastPos == sizeof(long double) - 1) { |
402 | decoded_values[num_decoded_values] = ReadShadow( |
403 | ptr: shadow + kShadowScale * (Offset + 1 - sizeof(long double)), |
404 | ShadowTypeId: static_cast<char>((shadow_value_type_ids >> 16) & 0xff)); |
405 | ++num_decoded_values; |
406 | } |
407 | break; |
408 | } |
409 | ++Offset; |
410 | } |
411 | for (int i = 0; i < num_decoded_values; ++i) { |
412 | printf(format: " (%s)" , FTPrinter<__float128>::dec(value: decoded_values[i]).Buffer); |
413 | } |
414 | num_decoded_values = 0; |
415 | printf(format: "\n" ); |
416 | } |
417 | } |
418 | |
419 | alignas(64) SANITIZER_INTERFACE_ATTRIBUTE |
420 | thread_local uptr __nsan_shadow_ret_tag = 0; |
421 | |
422 | alignas(64) SANITIZER_INTERFACE_ATTRIBUTE |
423 | thread_local char __nsan_shadow_ret_ptr[kMaxVectorWidth * |
424 | sizeof(__float128)]; |
425 | |
426 | alignas(64) SANITIZER_INTERFACE_ATTRIBUTE |
427 | thread_local uptr __nsan_shadow_args_tag = 0; |
428 | |
429 | // Maximum number of args. This should be enough for anyone (tm). An alternate |
430 | // scheme is to have the generated code create an alloca and make |
431 | // __nsan_shadow_args_ptr point ot the alloca. |
432 | constexpr const int kMaxNumArgs = 128; |
433 | alignas(64) SANITIZER_INTERFACE_ATTRIBUTE |
434 | thread_local char __nsan_shadow_args_ptr[kMaxVectorWidth * kMaxNumArgs * |
435 | sizeof(__float128)]; |
436 | |
437 | enum ContinuationType { // Keep in sync with instrumentation pass. |
438 | kContinueWithShadow = 0, |
439 | kResumeFromValue = 1, |
440 | }; |
441 | |
442 | // Checks the consistency between application and shadow value. Returns true |
443 | // when the instrumented code should resume computations from the original value |
444 | // rather than the shadow value. This prevents one error to propagate to all |
445 | // subsequent operations. This behaviour is tunable with flags. |
446 | template <typename FT, typename ShadowFT> |
447 | int32_t checkFT(const FT value, ShadowFT Shadow, CheckTypeT CheckType, |
448 | uptr CheckArg) { |
449 | // We do all comparisons in the InternalFT domain, which is the largest FT |
450 | // type. |
451 | using InternalFT = LargestFT<FT, ShadowFT>; |
452 | const InternalFT check_value = value; |
453 | const InternalFT check_shadow = Shadow; |
454 | |
455 | // We only check for NaNs in the value, not the shadow. |
456 | if (flags().check_nan && isnan(value)) { |
457 | GET_CALLER_PC_BP; |
458 | BufferedStackTrace stack; |
459 | stack.Unwind(pc, bp, context: nullptr, request_fast: false); |
460 | if (GetSuppressionForStack(Stack: &stack, K: CheckKind::Consistency)) { |
461 | // FIXME: optionally print. |
462 | return flags().resume_after_suppression ? kResumeFromValue |
463 | : kContinueWithShadow; |
464 | } |
465 | Decorator D; |
466 | Printf(format: "%s" , D.Warning()); |
467 | Printf(format: "WARNING: NumericalStabilitySanitizer: NaN detected\n" ); |
468 | Printf(format: "%s" , D.Default()); |
469 | stack.Print(); |
470 | if (flags().halt_on_error) { |
471 | if (common_flags()->abort_on_error) |
472 | Printf(format: "ABORTING\n" ); |
473 | else |
474 | Printf(format: "Exiting\n" ); |
475 | Die(); |
476 | } |
477 | // Performing other tests for NaN values is meaningless when dealing with numbers. |
478 | return kResumeFromValue; |
479 | } |
480 | |
481 | // See this article for an interesting discussion of how to compare floats: |
482 | // https://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/ |
483 | static constexpr const FT Eps = FTInfo<FT>::kEpsilon; |
484 | |
485 | const InternalFT abs_err = ftAbs(check_value - check_shadow); |
486 | |
487 | if (flags().enable_check_stats) { |
488 | GET_CALLER_PC_BP; |
489 | // We are re-computing `largest` here because this is a cold branch, and we |
490 | // want to avoid having to move the computation of `largest` before the |
491 | // absolute value check when this branch is not taken. |
492 | const InternalFT largest = max(ftAbs(check_value), ftAbs(check_shadow)); |
493 | nsan_stats->AddCheck(check_ty: CheckType, pc, bp, rel_err: abs_err / largest); |
494 | } |
495 | |
496 | // Note: writing the comparison that way ensures that when `abs_err` is Nan |
497 | // (value and shadow are inf or -inf), we pass the test. |
498 | if (!(abs_err >= flags().cached_absolute_error_threshold)) |
499 | return kContinueWithShadow; |
500 | |
501 | const InternalFT largest = max(ftAbs(check_value), ftAbs(check_shadow)); |
502 | if (abs_err * (1ull << flags().log2_max_relative_error) <= largest) |
503 | return kContinueWithShadow; // No problem here. |
504 | |
505 | if (!flags().disable_warnings) { |
506 | GET_CALLER_PC_BP; |
507 | UNINITIALIZED BufferedStackTrace stack; |
508 | stack.Unwind(pc, bp, context: nullptr, request_fast: false); |
509 | if (GetSuppressionForStack(Stack: &stack, K: CheckKind::Consistency)) { |
510 | // FIXME: optionally print. |
511 | return flags().resume_after_suppression ? kResumeFromValue |
512 | : kContinueWithShadow; |
513 | } |
514 | |
515 | Decorator D; |
516 | Printf(format: "%s" , D.Warning()); |
517 | // Printf does not support float formatting. |
518 | char RelErrBuf[64] = "inf" ; |
519 | if (largest > Eps) { |
520 | snprintf(s: RelErrBuf, maxlen: sizeof(RelErrBuf) - 1, format: "%.20Lf%% (2^%.0Lf epsilons)" , |
521 | static_cast<long double>(100.0 * abs_err / largest), |
522 | log2l(x: static_cast<long double>(abs_err / largest / Eps))); |
523 | } |
524 | char ulp_err_buf[128] = "" ; |
525 | const double shadow_ulp_diff = GetULPDiff(check_value, check_shadow); |
526 | if (shadow_ulp_diff != kMaxULPDiff) { |
527 | // This is the ULP diff in the internal domain. The user actually cares |
528 | // about that in the original domain. |
529 | const double ulp_diff = |
530 | shadow_ulp_diff / (u64{1} << (FTInfo<InternalFT>::kMantissaBits - |
531 | FTInfo<FT>::kMantissaBits)); |
532 | snprintf(s: ulp_err_buf, maxlen: sizeof(ulp_err_buf) - 1, |
533 | format: "(%.0f ULPs == %.1f digits == %.1f bits)" , ulp_diff, |
534 | log10(x: ulp_diff), log2(x: ulp_diff)); |
535 | } |
536 | Printf(format: "WARNING: NumericalStabilitySanitizer: inconsistent shadow results" ); |
537 | switch (CheckType) { |
538 | case CheckTypeT::kUnknown: |
539 | case CheckTypeT::kFcmp: |
540 | case CheckTypeT::kMaxCheckType: |
541 | break; |
542 | case CheckTypeT::kRet: |
543 | Printf(format: " while checking return value" ); |
544 | break; |
545 | case CheckTypeT::kArg: |
546 | Printf(format: " while checking call argument #%d" , static_cast<int>(CheckArg)); |
547 | break; |
548 | case CheckTypeT::kLoad: |
549 | Printf( |
550 | format: " while checking load from address 0x%lx. This is due to incorrect " |
551 | "shadow memory tracking, typically due to uninstrumented code " |
552 | "writing to memory." , |
553 | CheckArg); |
554 | break; |
555 | case CheckTypeT::kStore: |
556 | Printf(format: " while checking store to address 0x%lx" , CheckArg); |
557 | break; |
558 | case CheckTypeT::kInsert: |
559 | Printf(format: " while checking vector insert" ); |
560 | break; |
561 | case CheckTypeT::kUser: |
562 | Printf(format: " in user-initiated check" ); |
563 | break; |
564 | } |
565 | using ValuePrinter = FTPrinter<FT>; |
566 | using ShadowPrinter = FTPrinter<ShadowFT>; |
567 | Printf(format: "%s" , D.Default()); |
568 | |
569 | Printf("\n" |
570 | "%-12s precision (native): dec: %s hex: %s\n" |
571 | "%-12s precision (shadow): dec: %s hex: %s\n" |
572 | "shadow truncated to %-12s: dec: %s hex: %s\n" |
573 | "Relative error: %s\n" |
574 | "Absolute error: %s\n" |
575 | "%s\n" , |
576 | FTInfo<FT>::kCppTypeName, ValuePrinter::dec(value).Buffer, |
577 | ValuePrinter::hex(value).Buffer, FTInfo<ShadowFT>::kCppTypeName, |
578 | ShadowPrinter::dec(Shadow).Buffer, ShadowPrinter::hex(Shadow).Buffer, |
579 | FTInfo<FT>::kCppTypeName, ValuePrinter::dec(Shadow).Buffer, |
580 | ValuePrinter::hex(Shadow).Buffer, RelErrBuf, |
581 | ValuePrinter::hex(abs_err).Buffer, ulp_err_buf); |
582 | stack.Print(); |
583 | } |
584 | |
585 | if (flags().enable_warning_stats) { |
586 | GET_CALLER_PC_BP; |
587 | nsan_stats->AddWarning(check_ty: CheckType, pc, bp, rel_err: abs_err / largest); |
588 | } |
589 | |
590 | if (flags().halt_on_error) { |
591 | if (common_flags()->abort_on_error) |
592 | Printf(format: "ABORTING\n" ); |
593 | else |
594 | Printf(format: "Exiting\n" ); |
595 | Die(); |
596 | } |
597 | return flags().resume_after_warning ? kResumeFromValue : kContinueWithShadow; |
598 | } |
599 | |
600 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE int32_t __nsan_internal_check_float_d( |
601 | float value, double shadow, int32_t check_type, uptr check_arg) { |
602 | return checkFT(value, Shadow: shadow, CheckType: static_cast<CheckTypeT>(check_type), CheckArg: check_arg); |
603 | } |
604 | |
605 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE int32_t __nsan_internal_check_double_l( |
606 | double value, long double shadow, int32_t check_type, uptr check_arg) { |
607 | return checkFT(value, Shadow: shadow, CheckType: static_cast<CheckTypeT>(check_type), CheckArg: check_arg); |
608 | } |
609 | |
610 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE int32_t __nsan_internal_check_double_q( |
611 | double value, __float128 shadow, int32_t check_type, uptr check_arg) { |
612 | return checkFT(value, Shadow: shadow, CheckType: static_cast<CheckTypeT>(check_type), CheckArg: check_arg); |
613 | } |
614 | |
615 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE int32_t |
616 | __nsan_internal_check_longdouble_q(long double value, __float128 shadow, |
617 | int32_t check_type, uptr check_arg) { |
618 | return checkFT(value, Shadow: shadow, CheckType: static_cast<CheckTypeT>(check_type), CheckArg: check_arg); |
619 | } |
620 | |
621 | static const char *GetTruthValueName(bool v) { return v ? "true" : "false" ; } |
622 | |
623 | // This uses the same values as CmpInst::Predicate. |
624 | static const char *GetPredicateName(int v) { |
625 | switch (v) { |
626 | case 0: |
627 | return "(false)" ; |
628 | case 1: |
629 | return "==" ; |
630 | case 2: |
631 | return ">" ; |
632 | case 3: |
633 | return ">=" ; |
634 | case 4: |
635 | return "<" ; |
636 | case 5: |
637 | return "<=" ; |
638 | case 6: |
639 | return "!=" ; |
640 | case 7: |
641 | return "(ordered)" ; |
642 | case 8: |
643 | return "(unordered)" ; |
644 | case 9: |
645 | return "==" ; |
646 | case 10: |
647 | return ">" ; |
648 | case 11: |
649 | return ">=" ; |
650 | case 12: |
651 | return "<" ; |
652 | case 13: |
653 | return "<=" ; |
654 | case 14: |
655 | return "!=" ; |
656 | case 15: |
657 | return "(true)" ; |
658 | } |
659 | return "??" ; |
660 | } |
661 | |
662 | template <typename FT, typename ShadowFT> |
663 | void fCmpFailFT(const FT Lhs, const FT Rhs, ShadowFT LhsShadow, |
664 | ShadowFT RhsShadow, int Predicate, bool result, |
665 | bool ShadowResult) { |
666 | if (result == ShadowResult) { |
667 | // When a vector comparison fails, we fail each element of the comparison |
668 | // to simplify instrumented code. Skip elements where the shadow comparison |
669 | // gave the same result as the original one. |
670 | return; |
671 | } |
672 | |
673 | GET_CALLER_PC_BP; |
674 | UNINITIALIZED BufferedStackTrace stack; |
675 | stack.Unwind(pc, bp, context: nullptr, request_fast: false); |
676 | |
677 | if (GetSuppressionForStack(Stack: &stack, K: CheckKind::Fcmp)) { |
678 | // FIXME: optionally print. |
679 | return; |
680 | } |
681 | |
682 | if (flags().enable_warning_stats) |
683 | nsan_stats->AddWarning(check_ty: CheckTypeT::kFcmp, pc, bp, rel_err: 0.0); |
684 | |
685 | if (flags().disable_warnings || !flags().check_cmp) |
686 | return; |
687 | |
688 | // FIXME: ideally we would print the shadow value as FP128. Right now because |
689 | // we truncate to long double we can sometimes see stuff like: |
690 | // shadow <value> == <value> (false) |
691 | using ValuePrinter = FTPrinter<FT>; |
692 | using ShadowPrinter = FTPrinter<ShadowFT>; |
693 | Decorator D; |
694 | const char *const PredicateName = GetPredicateName(v: Predicate); |
695 | Printf(format: "%s" , D.Warning()); |
696 | Printf(format: "WARNING: NumericalStabilitySanitizer: floating-point comparison " |
697 | "results depend on precision\n" ); |
698 | Printf(format: "%s" , D.Default()); |
699 | Printf("%-12s precision dec (native): %s %s %s (%s)\n" |
700 | "%-12s precision dec (shadow): %s %s %s (%s)\n" |
701 | "%-12s precision hex (native): %s %s %s (%s)\n" |
702 | "%-12s precision hex (shadow): %s %s %s (%s)\n" |
703 | "%s" , |
704 | // Native, decimal. |
705 | FTInfo<FT>::kCppTypeName, ValuePrinter::dec(Lhs).Buffer, PredicateName, |
706 | ValuePrinter::dec(Rhs).Buffer, GetTruthValueName(v: result), |
707 | // Shadow, decimal |
708 | FTInfo<ShadowFT>::kCppTypeName, ShadowPrinter::dec(LhsShadow).Buffer, |
709 | PredicateName, ShadowPrinter::dec(RhsShadow).Buffer, |
710 | GetTruthValueName(v: ShadowResult), |
711 | // Native, hex. |
712 | FTInfo<FT>::kCppTypeName, ValuePrinter::hex(Lhs).Buffer, PredicateName, |
713 | ValuePrinter::hex(Rhs).Buffer, GetTruthValueName(v: result), |
714 | // Shadow, hex |
715 | FTInfo<ShadowFT>::kCppTypeName, ShadowPrinter::hex(LhsShadow).Buffer, |
716 | PredicateName, ShadowPrinter::hex(RhsShadow).Buffer, |
717 | GetTruthValueName(v: ShadowResult), D.End()); |
718 | stack.Print(); |
719 | if (flags().halt_on_error) { |
720 | Printf(format: "Exiting\n" ); |
721 | Die(); |
722 | } |
723 | } |
724 | |
725 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
726 | __nsan_fcmp_fail_float_d(float lhs, float rhs, double lhs_shadow, |
727 | double rhs_shadow, int predicate, bool result, |
728 | bool shadow_result) { |
729 | fCmpFailFT(Lhs: lhs, Rhs: rhs, LhsShadow: lhs_shadow, RhsShadow: rhs_shadow, Predicate: predicate, result, |
730 | ShadowResult: shadow_result); |
731 | } |
732 | |
733 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
734 | __nsan_fcmp_fail_double_q(double lhs, double rhs, __float128 lhs_shadow, |
735 | __float128 rhs_shadow, int predicate, bool result, |
736 | bool shadow_result) { |
737 | fCmpFailFT(Lhs: lhs, Rhs: rhs, LhsShadow: lhs_shadow, RhsShadow: rhs_shadow, Predicate: predicate, result, |
738 | ShadowResult: shadow_result); |
739 | } |
740 | |
741 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
742 | __nsan_fcmp_fail_double_l(double lhs, double rhs, long double lhs_shadow, |
743 | long double rhs_shadow, int predicate, bool result, |
744 | bool shadow_result) { |
745 | fCmpFailFT(Lhs: lhs, Rhs: rhs, LhsShadow: lhs_shadow, RhsShadow: rhs_shadow, Predicate: predicate, result, |
746 | ShadowResult: shadow_result); |
747 | } |
748 | |
749 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
750 | __nsan_fcmp_fail_longdouble_q(long double lhs, long double rhs, |
751 | __float128 lhs_shadow, __float128 rhs_shadow, |
752 | int predicate, bool result, bool shadow_result) { |
753 | fCmpFailFT(Lhs: lhs, Rhs: rhs, LhsShadow: lhs_shadow, RhsShadow: rhs_shadow, Predicate: predicate, result, |
754 | ShadowResult: shadow_result); |
755 | } |
756 | |
757 | template <typename FT> void checkFTFromShadowStack(const FT value) { |
758 | // Get the shadow 2FT value from the shadow stack. Note that |
759 | // __nsan_check_{float,double,long double} is a function like any other, so |
760 | // the instrumentation will have placed the shadow value on the shadow stack. |
761 | using ShadowFT = typename FTInfo<FT>::shadow_type; |
762 | ShadowFT Shadow; |
763 | __builtin_memcpy(&Shadow, __nsan_shadow_args_ptr, sizeof(ShadowFT)); |
764 | checkFT(value, Shadow, CheckTypeT::kUser, 0); |
765 | } |
766 | |
767 | // FIXME: Add suffixes and let the instrumentation pass automatically add |
768 | // suffixes. |
769 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_check_float(float value) { |
770 | assert(__nsan_shadow_args_tag == (uptr)&__nsan_check_float && |
771 | "__nsan_check_float called from non-instrumented function" ); |
772 | checkFTFromShadowStack(value); |
773 | } |
774 | |
775 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
776 | __nsan_check_double(double value) { |
777 | assert(__nsan_shadow_args_tag == (uptr)&__nsan_check_double && |
778 | "__nsan_check_double called from non-instrumented function" ); |
779 | checkFTFromShadowStack(value); |
780 | } |
781 | |
782 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
783 | __nsan_check_longdouble(long double value) { |
784 | assert(__nsan_shadow_args_tag == (uptr)&__nsan_check_longdouble && |
785 | "__nsan_check_longdouble called from non-instrumented function" ); |
786 | checkFTFromShadowStack(value); |
787 | } |
788 | |
789 | template <typename FT> static void dumpFTFromShadowStack(const FT value) { |
790 | // Get the shadow 2FT value from the shadow stack. Note that |
791 | // __nsan_dump_{float,double,long double} is a function like any other, so |
792 | // the instrumentation will have placed the shadow value on the shadow stack. |
793 | using ShadowFT = typename FTInfo<FT>::shadow_type; |
794 | ShadowFT shadow; |
795 | __builtin_memcpy(&shadow, __nsan_shadow_args_ptr, sizeof(ShadowFT)); |
796 | using ValuePrinter = FTPrinter<FT>; |
797 | using ShadowPrinter = FTPrinter<typename FTInfo<FT>::shadow_type>; |
798 | printf("value dec:%s hex:%s\n" |
799 | "shadow dec:%s hex:%s\n" , |
800 | ValuePrinter::dec(value).Buffer, ValuePrinter::hex(value).Buffer, |
801 | ShadowPrinter::dec(shadow).Buffer, ShadowPrinter::hex(shadow).Buffer); |
802 | } |
803 | |
804 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_dump_float(float value) { |
805 | assert(__nsan_shadow_args_tag == (uptr)&__nsan_dump_float && |
806 | "__nsan_dump_float called from non-instrumented function" ); |
807 | dumpFTFromShadowStack(value); |
808 | } |
809 | |
810 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_dump_double(double value) { |
811 | assert(__nsan_shadow_args_tag == (uptr)&__nsan_dump_double && |
812 | "__nsan_dump_double called from non-instrumented function" ); |
813 | dumpFTFromShadowStack(value); |
814 | } |
815 | |
816 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void |
817 | __nsan_dump_longdouble(long double value) { |
818 | assert(__nsan_shadow_args_tag == (uptr)&__nsan_dump_longdouble && |
819 | "__nsan_dump_longdouble called from non-instrumented function" ); |
820 | dumpFTFromShadowStack(value); |
821 | } |
822 | |
823 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_dump_shadow_ret() { |
824 | printf(format: "ret tag: %lx\n" , __nsan_shadow_ret_tag); |
825 | double v; |
826 | __builtin_memcpy(&v, __nsan_shadow_ret_ptr, sizeof(double)); |
827 | printf(format: "double value: %f\n" , v); |
828 | // FIXME: float128 value. |
829 | } |
830 | |
831 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_dump_shadow_args() { |
832 | printf(format: "args tag: %lx\n" , __nsan_shadow_args_tag); |
833 | } |
834 | |
835 | bool __nsan::nsan_initialized; |
836 | bool __nsan::nsan_init_is_running; |
837 | |
838 | extern "C" SANITIZER_INTERFACE_ATTRIBUTE void __nsan_init() { |
839 | CHECK(!nsan_init_is_running); |
840 | if (nsan_initialized) |
841 | return; |
842 | nsan_init_is_running = true; |
843 | SanitizerToolName = "NumericalStabilitySanitizer" ; |
844 | |
845 | InitializeFlags(); |
846 | InitializeSuppressions(); |
847 | InitializePlatformEarly(); |
848 | |
849 | DisableCoreDumperIfNecessary(); |
850 | |
851 | if (!MmapFixedNoReserve(fixed_addr: TypesAddr(), size: AllocatorAddr() - TypesAddr())) |
852 | Die(); |
853 | |
854 | InitializeInterceptors(); |
855 | NsanTSDInit(destructor: NsanTSDDtor); |
856 | NsanAllocatorInit(); |
857 | |
858 | NsanThread *main_thread = NsanThread::Create(start_routine: nullptr, arg: nullptr); |
859 | SetCurrentThread(main_thread); |
860 | main_thread->Init(); |
861 | |
862 | InitializeStats(); |
863 | if (flags().print_stats_on_exit) |
864 | Atexit(function: NsanAtexit); |
865 | |
866 | nsan_init_is_running = false; |
867 | nsan_initialized = true; |
868 | } |
869 | |