File: | build/source/bolt/runtime/instr.cpp |
Warning: | line 1183, column 20 Value stored to 'Parent' during its initialization is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- bolt/runtime/instr.cpp ---------------------------------------------===// |
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 | // BOLT runtime instrumentation library for x86 Linux. Currently, BOLT does |
10 | // not support linking modules with dependencies on one another into the final |
11 | // binary (TODO?), which means this library has to be self-contained in a single |
12 | // module. |
13 | // |
14 | // All extern declarations here need to be defined by BOLT itself. Those will be |
15 | // undefined symbols that BOLT needs to resolve by emitting these symbols with |
16 | // MCStreamer. Currently, Passes/Instrumentation.cpp is the pass responsible |
17 | // for defining the symbols here and these two files have a tight coupling: one |
18 | // working statically when you run BOLT and another during program runtime when |
19 | // you run an instrumented binary. The main goal here is to output an fdata file |
20 | // (BOLT profile) with the instrumentation counters inserted by the static pass. |
21 | // Counters for indirect calls are an exception, as we can't know them |
22 | // statically. These counters are created and managed here. To allow this, we |
23 | // need a minimal framework for allocating memory dynamically. We provide this |
24 | // with the BumpPtrAllocator class (not LLVM's, but our own version of it). |
25 | // |
26 | // Since this code is intended to be inserted into any executable, we decided to |
27 | // make it standalone and do not depend on any external libraries (i.e. language |
28 | // support libraries, such as glibc or stdc++). To allow this, we provide a few |
29 | // light implementations of common OS interacting functionalities using direct |
30 | // syscall wrappers. Our simple allocator doesn't manage deallocations that |
31 | // fragment the memory space, so it's stack based. This is the minimal framework |
32 | // provided here to allow processing instrumented counters and writing fdata. |
33 | // |
34 | // In the C++ idiom used here, we never use or rely on constructors or |
35 | // destructors for global objects. That's because those need support from the |
36 | // linker in initialization/finalization code, and we want to keep our linker |
37 | // very simple. Similarly, we don't create any global objects that are zero |
38 | // initialized, since those would need to go .bss, which our simple linker also |
39 | // don't support (TODO?). |
40 | // |
41 | //===----------------------------------------------------------------------===// |
42 | |
43 | #if defined (__x86_64__1) |
44 | #include "common.h" |
45 | |
46 | // Enables a very verbose logging to stderr useful when debugging |
47 | //#define ENABLE_DEBUG |
48 | |
49 | #ifdef ENABLE_DEBUG |
50 | #define DEBUG(X){} \ |
51 | { X; } |
52 | #else |
53 | #define DEBUG(X){} \ |
54 | {} |
55 | #endif |
56 | |
57 | #pragma GCC visibility push(hidden) |
58 | |
59 | extern "C" { |
60 | |
61 | #if defined(__APPLE__) |
62 | extern uint64_t* _bolt_instr_locations_getter(); |
63 | extern uint32_t _bolt_num_counters_getter(); |
64 | |
65 | extern uint8_t* _bolt_instr_tables_getter(); |
66 | extern uint32_t _bolt_instr_num_funcs_getter(); |
67 | |
68 | #else |
69 | |
70 | // Main counters inserted by instrumentation, incremented during runtime when |
71 | // points of interest (locations) in the program are reached. Those are direct |
72 | // calls and direct and indirect branches (local ones). There are also counters |
73 | // for basic block execution if they are a spanning tree leaf and need to be |
74 | // counted in order to infer the execution count of other edges of the CFG. |
75 | extern uint64_t __bolt_instr_locations[]; |
76 | extern uint32_t __bolt_num_counters; |
77 | // Descriptions are serialized metadata about binary functions written by BOLT, |
78 | // so we have a minimal understanding about the program structure. For a |
79 | // reference on the exact format of this metadata, see *Description structs, |
80 | // Location, IntrumentedNode and EntryNode. |
81 | // Number of indirect call site descriptions |
82 | extern uint32_t __bolt_instr_num_ind_calls; |
83 | // Number of indirect call target descriptions |
84 | extern uint32_t __bolt_instr_num_ind_targets; |
85 | // Number of function descriptions |
86 | extern uint32_t __bolt_instr_num_funcs; |
87 | // Time to sleep across dumps (when we write the fdata profile to disk) |
88 | extern uint32_t __bolt_instr_sleep_time; |
89 | // Do not clear counters across dumps, rewrite file with the updated values |
90 | extern bool __bolt_instr_no_counters_clear; |
91 | // Wait until all forks of instrumented process will finish |
92 | extern bool __bolt_instr_wait_forks; |
93 | // Filename to dump data to |
94 | extern char __bolt_instr_filename[]; |
95 | // Instumented binary file path |
96 | extern char __bolt_instr_binpath[]; |
97 | // If true, append current PID to the fdata filename when creating it so |
98 | // different invocations of the same program can be differentiated. |
99 | extern bool __bolt_instr_use_pid; |
100 | // Functions that will be used to instrument indirect calls. BOLT static pass |
101 | // will identify indirect calls and modify them to load the address in these |
102 | // trampolines and call this address instead. BOLT can't use direct calls to |
103 | // our handlers because our addresses here are not known at analysis time. We |
104 | // only support resolving dependencies from this file to the output of BOLT, |
105 | // *not* the other way around. |
106 | // TODO: We need better linking support to make that happen. |
107 | extern void (*__bolt_ind_call_counter_func_pointer)(); |
108 | extern void (*__bolt_ind_tailcall_counter_func_pointer)(); |
109 | // Function pointers to init/fini trampoline routines in the binary, so we can |
110 | // resume regular execution of these functions that we hooked |
111 | extern void __bolt_start_trampoline(); |
112 | extern void __bolt_fini_trampoline(); |
113 | |
114 | #endif |
115 | } |
116 | |
117 | namespace { |
118 | |
119 | /// A simple allocator that mmaps a fixed size region and manages this space |
120 | /// in a stack fashion, meaning you always deallocate the last element that |
121 | /// was allocated. In practice, we don't need to deallocate individual elements. |
122 | /// We monotonically increase our usage and then deallocate everything once we |
123 | /// are done processing something. |
124 | class BumpPtrAllocator { |
125 | /// This is written before each allocation and act as a canary to detect when |
126 | /// a bug caused our program to cross allocation boundaries. |
127 | struct EntryMetadata { |
128 | uint64_t Magic; |
129 | uint64_t AllocSize; |
130 | }; |
131 | |
132 | public: |
133 | void *allocate(size_t Size) { |
134 | Lock L(M); |
135 | |
136 | if (StackBase == nullptr) { |
137 | #if defined(__APPLE__) |
138 | int MAP_PRIVATE_MAP_ANONYMOUS = 0x1002; |
139 | #else |
140 | int MAP_PRIVATE_MAP_ANONYMOUS = 0x22; |
141 | #endif |
142 | StackBase = reinterpret_cast<uint8_t *>( |
143 | __mmap(0, MaxSize, 0x3 /* PROT_READ | PROT_WRITE*/, |
144 | Shared ? 0x21 /*MAP_SHARED | MAP_ANONYMOUS*/ |
145 | : MAP_PRIVATE_MAP_ANONYMOUS /* MAP_PRIVATE | MAP_ANONYMOUS*/, |
146 | -1, 0)); |
147 | StackSize = 0; |
148 | } |
149 | |
150 | Size = alignTo(Size + sizeof(EntryMetadata), 16); |
151 | uint8_t *AllocAddress = StackBase + StackSize + sizeof(EntryMetadata); |
152 | auto *M = reinterpret_cast<EntryMetadata *>(StackBase + StackSize); |
153 | M->Magic = Magic; |
154 | M->AllocSize = Size; |
155 | StackSize += Size; |
156 | assert(StackSize < MaxSize, "allocator ran out of memory"); |
157 | return AllocAddress; |
158 | } |
159 | |
160 | #ifdef DEBUG |
161 | /// Element-wise deallocation is only used for debugging to catch memory |
162 | /// bugs by checking magic bytes. Ordinarily, we reset the allocator once |
163 | /// we are done with it. Reset is done with clear(). There's no need |
164 | /// to deallocate each element individually. |
165 | void deallocate(void *Ptr) { |
166 | Lock L(M); |
167 | uint8_t MetadataOffset = sizeof(EntryMetadata); |
168 | auto *M = reinterpret_cast<EntryMetadata *>( |
169 | reinterpret_cast<uint8_t *>(Ptr) - MetadataOffset); |
170 | const uint8_t *StackTop = StackBase + StackSize + MetadataOffset; |
171 | // Validate size |
172 | if (Ptr != StackTop - M->AllocSize) { |
173 | // Failed validation, check if it is a pointer returned by operator new [] |
174 | MetadataOffset += |
175 | sizeof(uint64_t); // Space for number of elements alloc'ed |
176 | M = reinterpret_cast<EntryMetadata *>(reinterpret_cast<uint8_t *>(Ptr) - |
177 | MetadataOffset); |
178 | // Ok, it failed both checks if this assertion fails. Stop the program, we |
179 | // have a memory bug. |
180 | assert(Ptr == StackTop - M->AllocSize, |
181 | "must deallocate the last element alloc'ed"); |
182 | } |
183 | assert(M->Magic == Magic, "allocator magic is corrupt"); |
184 | StackSize -= M->AllocSize; |
185 | } |
186 | #else |
187 | void deallocate(void *) {} |
188 | #endif |
189 | |
190 | void clear() { |
191 | Lock L(M); |
192 | StackSize = 0; |
193 | } |
194 | |
195 | /// Set mmap reservation size (only relevant before first allocation) |
196 | void setMaxSize(uint64_t Size) { MaxSize = Size; } |
197 | |
198 | /// Set mmap reservation privacy (only relevant before first allocation) |
199 | void setShared(bool S) { Shared = S; } |
200 | |
201 | void destroy() { |
202 | if (StackBase == nullptr) |
203 | return; |
204 | __munmap(StackBase, MaxSize); |
205 | } |
206 | |
207 | private: |
208 | static constexpr uint64_t Magic = 0x1122334455667788ull; |
209 | uint64_t MaxSize = 0xa00000; |
210 | uint8_t *StackBase{nullptr}; |
211 | uint64_t StackSize{0}; |
212 | bool Shared{false}; |
213 | Mutex M; |
214 | }; |
215 | |
216 | /// Used for allocating indirect call instrumentation counters. Initialized by |
217 | /// __bolt_instr_setup, our initialization routine. |
218 | BumpPtrAllocator GlobalAlloc; |
219 | } // anonymous namespace |
220 | |
221 | // User-defined placement new operators. We only use those (as opposed to |
222 | // overriding the regular operator new) so we can keep our allocator in the |
223 | // stack instead of in a data section (global). |
224 | void *operator new(size_t Sz, BumpPtrAllocator &A) { return A.allocate(Sz); } |
225 | void *operator new(size_t Sz, BumpPtrAllocator &A, char C) { |
226 | auto *Ptr = reinterpret_cast<char *>(A.allocate(Sz)); |
227 | memset(Ptr, C, Sz); |
228 | return Ptr; |
229 | } |
230 | void *operator new[](size_t Sz, BumpPtrAllocator &A) { |
231 | return A.allocate(Sz); |
232 | } |
233 | void *operator new[](size_t Sz, BumpPtrAllocator &A, char C) { |
234 | auto *Ptr = reinterpret_cast<char *>(A.allocate(Sz)); |
235 | memset(Ptr, C, Sz); |
236 | return Ptr; |
237 | } |
238 | // Only called during exception unwinding (useless). We must manually dealloc. |
239 | // C++ language weirdness |
240 | void operator delete(void *Ptr, BumpPtrAllocator &A) { A.deallocate(Ptr); } |
241 | |
242 | namespace { |
243 | |
244 | // Disable instrumentation optimizations that sacrifice profile accuracy |
245 | extern "C" bool __bolt_instr_conservative; |
246 | |
247 | /// Basic key-val atom stored in our hash |
248 | struct SimpleHashTableEntryBase { |
249 | uint64_t Key; |
250 | uint64_t Val; |
251 | }; |
252 | |
253 | /// This hash table implementation starts by allocating a table of size |
254 | /// InitialSize. When conflicts happen in this main table, it resolves |
255 | /// them by chaining a new table of size IncSize. It never reallocs as our |
256 | /// allocator doesn't support it. The key is intended to be function pointers. |
257 | /// There's no clever hash function (it's just x mod size, size being prime). |
258 | /// I never tuned the coefficientes in the modular equation (TODO) |
259 | /// This is used for indirect calls (each call site has one of this, so it |
260 | /// should have a small footprint) and for tallying call counts globally for |
261 | /// each target to check if we missed the origin of some calls (this one is a |
262 | /// large instantiation of this template, since it is global for all call sites) |
263 | template <typename T = SimpleHashTableEntryBase, uint32_t InitialSize = 7, |
264 | uint32_t IncSize = 7> |
265 | class SimpleHashTable { |
266 | public: |
267 | using MapEntry = T; |
268 | |
269 | /// Increment by 1 the value of \p Key. If it is not in this table, it will be |
270 | /// added to the table and its value set to 1. |
271 | void incrementVal(uint64_t Key, BumpPtrAllocator &Alloc) { |
272 | ++get(Key, Alloc).Val; |
273 | } |
274 | |
275 | /// Basic member accessing interface. Here we pass the allocator explicitly to |
276 | /// avoid storing a pointer to it as part of this table (remember there is one |
277 | /// hash for each indirect call site, so we wan't to minimize our footprint). |
278 | MapEntry &get(uint64_t Key, BumpPtrAllocator &Alloc) { |
279 | if (!__bolt_instr_conservative) { |
280 | TryLock L(M); |
281 | if (!L.isLocked()) |
282 | return NoEntry; |
283 | return getOrAllocEntry(Key, Alloc); |
284 | } |
285 | Lock L(M); |
286 | return getOrAllocEntry(Key, Alloc); |
287 | } |
288 | |
289 | /// Traverses all elements in the table |
290 | template <typename... Args> |
291 | void forEachElement(void (*Callback)(MapEntry &, Args...), Args... args) { |
292 | Lock L(M); |
293 | if (!TableRoot) |
294 | return; |
295 | return forEachElement(Callback, InitialSize, TableRoot, args...); |
296 | } |
297 | |
298 | void resetCounters(); |
299 | |
300 | private: |
301 | constexpr static uint64_t VacantMarker = 0; |
302 | constexpr static uint64_t FollowUpTableMarker = 0x8000000000000000ull; |
303 | |
304 | MapEntry *TableRoot{nullptr}; |
305 | MapEntry NoEntry; |
306 | Mutex M; |
307 | |
308 | template <typename... Args> |
309 | void forEachElement(void (*Callback)(MapEntry &, Args...), |
310 | uint32_t NumEntries, MapEntry *Entries, Args... args) { |
311 | for (uint32_t I = 0; I < NumEntries; ++I) { |
312 | MapEntry &Entry = Entries[I]; |
313 | if (Entry.Key == VacantMarker) |
314 | continue; |
315 | if (Entry.Key & FollowUpTableMarker) { |
316 | forEachElement(Callback, IncSize, |
317 | reinterpret_cast<MapEntry *>(Entry.Key & |
318 | ~FollowUpTableMarker), |
319 | args...); |
320 | continue; |
321 | } |
322 | Callback(Entry, args...); |
323 | } |
324 | } |
325 | |
326 | MapEntry &firstAllocation(uint64_t Key, BumpPtrAllocator &Alloc) { |
327 | TableRoot = new (Alloc, 0) MapEntry[InitialSize]; |
328 | MapEntry &Entry = TableRoot[Key % InitialSize]; |
329 | Entry.Key = Key; |
330 | return Entry; |
331 | } |
332 | |
333 | MapEntry &getEntry(MapEntry *Entries, uint64_t Key, uint64_t Selector, |
334 | BumpPtrAllocator &Alloc, int CurLevel) { |
335 | const uint32_t NumEntries = CurLevel == 0 ? InitialSize : IncSize; |
336 | uint64_t Remainder = Selector / NumEntries; |
337 | Selector = Selector % NumEntries; |
338 | MapEntry &Entry = Entries[Selector]; |
339 | |
340 | // A hit |
341 | if (Entry.Key == Key) { |
342 | return Entry; |
343 | } |
344 | |
345 | // Vacant - add new entry |
346 | if (Entry.Key == VacantMarker) { |
347 | Entry.Key = Key; |
348 | return Entry; |
349 | } |
350 | |
351 | // Defer to the next level |
352 | if (Entry.Key & FollowUpTableMarker) { |
353 | return getEntry( |
354 | reinterpret_cast<MapEntry *>(Entry.Key & ~FollowUpTableMarker), |
355 | Key, Remainder, Alloc, CurLevel + 1); |
356 | } |
357 | |
358 | // Conflict - create the next level |
359 | MapEntry *NextLevelTbl = new (Alloc, 0) MapEntry[IncSize]; |
360 | uint64_t CurEntrySelector = Entry.Key / InitialSize; |
361 | for (int I = 0; I < CurLevel; ++I) |
362 | CurEntrySelector /= IncSize; |
363 | CurEntrySelector = CurEntrySelector % IncSize; |
364 | NextLevelTbl[CurEntrySelector] = Entry; |
365 | Entry.Key = reinterpret_cast<uint64_t>(NextLevelTbl) | FollowUpTableMarker; |
366 | return getEntry(NextLevelTbl, Key, Remainder, Alloc, CurLevel + 1); |
367 | } |
368 | |
369 | MapEntry &getOrAllocEntry(uint64_t Key, BumpPtrAllocator &Alloc) { |
370 | if (TableRoot) |
371 | return getEntry(TableRoot, Key, Key, Alloc, 0); |
372 | return firstAllocation(Key, Alloc); |
373 | } |
374 | }; |
375 | |
376 | template <typename T> void resetIndCallCounter(T &Entry) { |
377 | Entry.Val = 0; |
378 | } |
379 | |
380 | template <typename T, uint32_t X, uint32_t Y> |
381 | void SimpleHashTable<T, X, Y>::resetCounters() { |
382 | forEachElement(resetIndCallCounter); |
383 | } |
384 | |
385 | /// Represents a hash table mapping a function target address to its counter. |
386 | using IndirectCallHashTable = SimpleHashTable<>; |
387 | |
388 | /// Initialize with number 1 instead of 0 so we don't go into .bss. This is the |
389 | /// global array of all hash tables storing indirect call destinations happening |
390 | /// during runtime, one table per call site. |
391 | IndirectCallHashTable *GlobalIndCallCounters{ |
392 | reinterpret_cast<IndirectCallHashTable *>(1)}; |
393 | |
394 | /// Don't allow reentrancy in the fdata writing phase - only one thread writes |
395 | /// it |
396 | Mutex *GlobalWriteProfileMutex{reinterpret_cast<Mutex *>(1)}; |
397 | |
398 | /// Store number of calls in additional to target address (Key) and frequency |
399 | /// as perceived by the basic block counter (Val). |
400 | struct CallFlowEntryBase : public SimpleHashTableEntryBase { |
401 | uint64_t Calls; |
402 | }; |
403 | |
404 | using CallFlowHashTableBase = SimpleHashTable<CallFlowEntryBase, 11939, 233>; |
405 | |
406 | /// This is a large table indexing all possible call targets (indirect and |
407 | /// direct ones). The goal is to find mismatches between number of calls (for |
408 | /// those calls we were able to track) and the entry basic block counter of the |
409 | /// callee. In most cases, these two should be equal. If not, there are two |
410 | /// possible scenarios here: |
411 | /// |
412 | /// * Entry BB has higher frequency than all known calls to this function. |
413 | /// In this case, we have dynamic library code or any uninstrumented code |
414 | /// calling this function. We will write the profile for these untracked |
415 | /// calls as having source "0 [unknown] 0" in the fdata file. |
416 | /// |
417 | /// * Number of known calls is higher than the frequency of entry BB |
418 | /// This only happens when there is no counter for the entry BB / callee |
419 | /// function is not simple (in BOLT terms). We don't do anything special |
420 | /// here and just ignore those (we still report all calls to the non-simple |
421 | /// function, though). |
422 | /// |
423 | class CallFlowHashTable : public CallFlowHashTableBase { |
424 | public: |
425 | CallFlowHashTable(BumpPtrAllocator &Alloc) : Alloc(Alloc) {} |
426 | |
427 | MapEntry &get(uint64_t Key) { return CallFlowHashTableBase::get(Key, Alloc); } |
428 | |
429 | private: |
430 | // Different than the hash table for indirect call targets, we do store the |
431 | // allocator here since there is only one call flow hash and space overhead |
432 | // is negligible. |
433 | BumpPtrAllocator &Alloc; |
434 | }; |
435 | |
436 | /// |
437 | /// Description metadata emitted by BOLT to describe the program - refer to |
438 | /// Passes/Instrumentation.cpp - Instrumentation::emitTablesAsELFNote() |
439 | /// |
440 | struct Location { |
441 | uint32_t FunctionName; |
442 | uint32_t Offset; |
443 | }; |
444 | |
445 | struct CallDescription { |
446 | Location From; |
447 | uint32_t FromNode; |
448 | Location To; |
449 | uint32_t Counter; |
450 | uint64_t TargetAddress; |
451 | }; |
452 | |
453 | using IndCallDescription = Location; |
454 | |
455 | struct IndCallTargetDescription { |
456 | Location Loc; |
457 | uint64_t Address; |
458 | }; |
459 | |
460 | struct EdgeDescription { |
461 | Location From; |
462 | uint32_t FromNode; |
463 | Location To; |
464 | uint32_t ToNode; |
465 | uint32_t Counter; |
466 | }; |
467 | |
468 | struct InstrumentedNode { |
469 | uint32_t Node; |
470 | uint32_t Counter; |
471 | }; |
472 | |
473 | struct EntryNode { |
474 | uint64_t Node; |
475 | uint64_t Address; |
476 | }; |
477 | |
478 | struct FunctionDescription { |
479 | uint32_t NumLeafNodes; |
480 | const InstrumentedNode *LeafNodes; |
481 | uint32_t NumEdges; |
482 | const EdgeDescription *Edges; |
483 | uint32_t NumCalls; |
484 | const CallDescription *Calls; |
485 | uint32_t NumEntryNodes; |
486 | const EntryNode *EntryNodes; |
487 | |
488 | /// Constructor will parse the serialized function metadata written by BOLT |
489 | FunctionDescription(const uint8_t *FuncDesc); |
490 | |
491 | uint64_t getSize() const { |
492 | return 16 + NumLeafNodes * sizeof(InstrumentedNode) + |
493 | NumEdges * sizeof(EdgeDescription) + |
494 | NumCalls * sizeof(CallDescription) + |
495 | NumEntryNodes * sizeof(EntryNode); |
496 | } |
497 | }; |
498 | |
499 | /// The context is created when the fdata profile needs to be written to disk |
500 | /// and we need to interpret our runtime counters. It contains pointers to the |
501 | /// mmaped binary (only the BOLT written metadata section). Deserialization |
502 | /// should be straightforward as most data is POD or an array of POD elements. |
503 | /// This metadata is used to reconstruct function CFGs. |
504 | struct ProfileWriterContext { |
505 | IndCallDescription *IndCallDescriptions; |
506 | IndCallTargetDescription *IndCallTargets; |
507 | uint8_t *FuncDescriptions; |
508 | char *Strings; // String table with function names used in this binary |
509 | int FileDesc; // File descriptor for the file on disk backing this |
510 | // information in memory via mmap |
511 | void *MMapPtr; // The mmap ptr |
512 | int MMapSize; // The mmap size |
513 | |
514 | /// Hash table storing all possible call destinations to detect untracked |
515 | /// calls and correctly report them as [unknown] in output fdata. |
516 | CallFlowHashTable *CallFlowTable; |
517 | |
518 | /// Lookup the sorted indirect call target vector to fetch function name and |
519 | /// offset for an arbitrary function pointer. |
520 | const IndCallTargetDescription *lookupIndCallTarget(uint64_t Target) const; |
521 | }; |
522 | |
523 | /// Perform a string comparison and returns zero if Str1 matches Str2. Compares |
524 | /// at most Size characters. |
525 | int compareStr(const char *Str1, const char *Str2, int Size) { |
526 | while (*Str1 == *Str2) { |
527 | if (*Str1 == '\0' || --Size == 0) |
528 | return 0; |
529 | ++Str1; |
530 | ++Str2; |
531 | } |
532 | return 1; |
533 | } |
534 | |
535 | /// Output Location to the fdata file |
536 | char *serializeLoc(const ProfileWriterContext &Ctx, char *OutBuf, |
537 | const Location Loc, uint32_t BufSize) { |
538 | // fdata location format: Type Name Offset |
539 | // Type 1 - regular symbol |
540 | OutBuf = strCopy(OutBuf, "1 "); |
541 | const char *Str = Ctx.Strings + Loc.FunctionName; |
542 | uint32_t Size = 25; |
543 | while (*Str) { |
544 | *OutBuf++ = *Str++; |
545 | if (++Size >= BufSize) |
546 | break; |
547 | } |
548 | assert(!*Str, "buffer overflow, function name too large"); |
549 | *OutBuf++ = ' '; |
550 | OutBuf = intToStr(OutBuf, Loc.Offset, 16); |
551 | *OutBuf++ = ' '; |
552 | return OutBuf; |
553 | } |
554 | |
555 | /// Read and deserialize a function description written by BOLT. \p FuncDesc |
556 | /// points at the beginning of the function metadata structure in the file. |
557 | /// See Instrumentation::emitTablesAsELFNote() |
558 | FunctionDescription::FunctionDescription(const uint8_t *FuncDesc) { |
559 | NumLeafNodes = *reinterpret_cast<const uint32_t *>(FuncDesc); |
560 | DEBUG(reportNumber("NumLeafNodes = ", NumLeafNodes, 10)){}; |
561 | LeafNodes = reinterpret_cast<const InstrumentedNode *>(FuncDesc + 4); |
562 | |
563 | NumEdges = *reinterpret_cast<const uint32_t *>( |
564 | FuncDesc + 4 + NumLeafNodes * sizeof(InstrumentedNode)); |
565 | DEBUG(reportNumber("NumEdges = ", NumEdges, 10)){}; |
566 | Edges = reinterpret_cast<const EdgeDescription *>( |
567 | FuncDesc + 8 + NumLeafNodes * sizeof(InstrumentedNode)); |
568 | |
569 | NumCalls = *reinterpret_cast<const uint32_t *>( |
570 | FuncDesc + 8 + NumLeafNodes * sizeof(InstrumentedNode) + |
571 | NumEdges * sizeof(EdgeDescription)); |
572 | DEBUG(reportNumber("NumCalls = ", NumCalls, 10)){}; |
573 | Calls = reinterpret_cast<const CallDescription *>( |
574 | FuncDesc + 12 + NumLeafNodes * sizeof(InstrumentedNode) + |
575 | NumEdges * sizeof(EdgeDescription)); |
576 | NumEntryNodes = *reinterpret_cast<const uint32_t *>( |
577 | FuncDesc + 12 + NumLeafNodes * sizeof(InstrumentedNode) + |
578 | NumEdges * sizeof(EdgeDescription) + NumCalls * sizeof(CallDescription)); |
579 | DEBUG(reportNumber("NumEntryNodes = ", NumEntryNodes, 10)){}; |
580 | EntryNodes = reinterpret_cast<const EntryNode *>( |
581 | FuncDesc + 16 + NumLeafNodes * sizeof(InstrumentedNode) + |
582 | NumEdges * sizeof(EdgeDescription) + NumCalls * sizeof(CallDescription)); |
583 | } |
584 | |
585 | /// Read and mmap descriptions written by BOLT from the executable's notes |
586 | /// section |
587 | #if defined(HAVE_ELF_H) and !defined(__APPLE__) |
588 | |
589 | void *__attribute__((noinline)) __get_pc() { |
590 | return __builtin_extract_return_addr(__builtin_return_address(0)); |
591 | } |
592 | |
593 | /// Get string with address and parse it to hex pair <StartAddress, EndAddress> |
594 | bool parseAddressRange(const char *Str, uint64_t &StartAddress, |
595 | uint64_t &EndAddress) { |
596 | if (!Str) |
597 | return false; |
598 | // Parsed string format: <hex1>-<hex2> |
599 | StartAddress = hexToLong(Str, '-'); |
600 | while (*Str && *Str != '-') |
601 | ++Str; |
602 | if (!*Str) |
603 | return false; |
604 | ++Str; // swallow '-' |
605 | EndAddress = hexToLong(Str); |
606 | return true; |
607 | } |
608 | |
609 | /// Get full path to the real binary by getting current virtual address |
610 | /// and searching for the appropriate link in address range in |
611 | /// /proc/self/map_files |
612 | static char *getBinaryPath() { |
613 | const uint32_t BufSize = 1024; |
614 | const uint32_t NameMax = 4096; |
615 | const char DirPath[] = "/proc/self/map_files/"; |
616 | static char TargetPath[NameMax] = {}; |
617 | char Buf[BufSize]; |
618 | |
619 | if (__bolt_instr_binpath[0] != '\0') |
620 | return __bolt_instr_binpath; |
621 | |
622 | if (TargetPath[0] != '\0') |
623 | return TargetPath; |
624 | |
625 | unsigned long CurAddr = (unsigned long)__get_pc(); |
626 | uint64_t FDdir = __open(DirPath, |
627 | /*flags=*/0 /*O_RDONLY*/, |
628 | /*mode=*/0666); |
629 | assert(static_cast<int64_t>(FDdir) >= 0, |
630 | "failed to open /proc/self/map_files"); |
631 | |
632 | while (long Nread = __getdents(FDdir, (struct dirent *)Buf, BufSize)) { |
633 | assert(static_cast<int64_t>(Nread) != -1, "failed to get folder entries"); |
634 | |
635 | struct dirent *d; |
636 | for (long Bpos = 0; Bpos < Nread; Bpos += d->d_reclen) { |
637 | d = (struct dirent *)(Buf + Bpos); |
638 | |
639 | uint64_t StartAddress, EndAddress; |
640 | if (!parseAddressRange(d->d_name, StartAddress, EndAddress)) |
641 | continue; |
642 | if (CurAddr < StartAddress || CurAddr > EndAddress) |
643 | continue; |
644 | char FindBuf[NameMax]; |
645 | char *C = strCopy(FindBuf, DirPath, NameMax); |
646 | C = strCopy(C, d->d_name, NameMax - (C - FindBuf)); |
647 | *C = '\0'; |
648 | uint32_t Ret = __readlink(FindBuf, TargetPath, sizeof(TargetPath)); |
649 | assert(Ret != -1 && Ret != BufSize, "readlink error"); |
650 | TargetPath[Ret] = '\0'; |
651 | return TargetPath; |
652 | } |
653 | } |
654 | return nullptr; |
655 | } |
656 | |
657 | ProfileWriterContext readDescriptions() { |
658 | ProfileWriterContext Result; |
659 | char *BinPath = getBinaryPath(); |
660 | assert(BinPath && BinPath[0] != '\0', "failed to find binary path"); |
661 | |
662 | uint64_t FD = __open(BinPath, |
663 | /*flags=*/0 /*O_RDONLY*/, |
664 | /*mode=*/0666); |
665 | assert(static_cast<int64_t>(FD) >= 0, "failed to open binary path"); |
666 | |
667 | Result.FileDesc = FD; |
668 | |
669 | // mmap our binary to memory |
670 | uint64_t Size = __lseek(FD, 0, 2 /*SEEK_END*/); |
671 | uint8_t *BinContents = reinterpret_cast<uint8_t *>( |
672 | __mmap(0, Size, 0x1 /* PROT_READ*/, 0x2 /* MAP_PRIVATE*/, FD, 0)); |
673 | Result.MMapPtr = BinContents; |
674 | Result.MMapSize = Size; |
675 | Elf64_Ehdr *Hdr = reinterpret_cast<Elf64_Ehdr *>(BinContents); |
676 | Elf64_Shdr *Shdr = reinterpret_cast<Elf64_Shdr *>(BinContents + Hdr->e_shoff); |
677 | Elf64_Shdr *StringTblHeader = reinterpret_cast<Elf64_Shdr *>( |
678 | BinContents + Hdr->e_shoff + Hdr->e_shstrndx * Hdr->e_shentsize); |
679 | |
680 | // Find .bolt.instr.tables with the data we need and set pointers to it |
681 | for (int I = 0; I < Hdr->e_shnum; ++I) { |
682 | char *SecName = reinterpret_cast<char *>( |
683 | BinContents + StringTblHeader->sh_offset + Shdr->sh_name); |
684 | if (compareStr(SecName, ".bolt.instr.tables", 64) != 0) { |
685 | Shdr = reinterpret_cast<Elf64_Shdr *>(BinContents + Hdr->e_shoff + |
686 | (I + 1) * Hdr->e_shentsize); |
687 | continue; |
688 | } |
689 | // Actual contents of the ELF note start after offset 20 decimal: |
690 | // Offset 0: Producer name size (4 bytes) |
691 | // Offset 4: Contents size (4 bytes) |
692 | // Offset 8: Note type (4 bytes) |
693 | // Offset 12: Producer name (BOLT\0) (5 bytes + align to 4-byte boundary) |
694 | // Offset 20: Contents |
695 | uint32_t IndCallDescSize = |
696 | *reinterpret_cast<uint32_t *>(BinContents + Shdr->sh_offset + 20); |
697 | uint32_t IndCallTargetDescSize = *reinterpret_cast<uint32_t *>( |
698 | BinContents + Shdr->sh_offset + 24 + IndCallDescSize); |
699 | uint32_t FuncDescSize = |
700 | *reinterpret_cast<uint32_t *>(BinContents + Shdr->sh_offset + 28 + |
701 | IndCallDescSize + IndCallTargetDescSize); |
702 | Result.IndCallDescriptions = reinterpret_cast<IndCallDescription *>( |
703 | BinContents + Shdr->sh_offset + 24); |
704 | Result.IndCallTargets = reinterpret_cast<IndCallTargetDescription *>( |
705 | BinContents + Shdr->sh_offset + 28 + IndCallDescSize); |
706 | Result.FuncDescriptions = BinContents + Shdr->sh_offset + 32 + |
707 | IndCallDescSize + IndCallTargetDescSize; |
708 | Result.Strings = reinterpret_cast<char *>( |
709 | BinContents + Shdr->sh_offset + 32 + IndCallDescSize + |
710 | IndCallTargetDescSize + FuncDescSize); |
711 | return Result; |
712 | } |
713 | const char ErrMsg[] = |
714 | "BOLT instrumentation runtime error: could not find section " |
715 | ".bolt.instr.tables\n"; |
716 | reportError(ErrMsg, sizeof(ErrMsg)); |
717 | return Result; |
718 | } |
719 | |
720 | #else |
721 | |
722 | ProfileWriterContext readDescriptions() { |
723 | ProfileWriterContext Result; |
724 | uint8_t *Tables = _bolt_instr_tables_getter(); |
725 | uint32_t IndCallDescSize = *reinterpret_cast<uint32_t *>(Tables); |
726 | uint32_t IndCallTargetDescSize = |
727 | *reinterpret_cast<uint32_t *>(Tables + 4 + IndCallDescSize); |
728 | uint32_t FuncDescSize = *reinterpret_cast<uint32_t *>( |
729 | Tables + 8 + IndCallDescSize + IndCallTargetDescSize); |
730 | Result.IndCallDescriptions = |
731 | reinterpret_cast<IndCallDescription *>(Tables + 4); |
732 | Result.IndCallTargets = reinterpret_cast<IndCallTargetDescription *>( |
733 | Tables + 8 + IndCallDescSize); |
734 | Result.FuncDescriptions = |
735 | Tables + 12 + IndCallDescSize + IndCallTargetDescSize; |
736 | Result.Strings = reinterpret_cast<char *>( |
737 | Tables + 12 + IndCallDescSize + IndCallTargetDescSize + FuncDescSize); |
738 | return Result; |
739 | } |
740 | |
741 | #endif |
742 | |
743 | #if !defined(__APPLE__) |
744 | /// Debug by printing overall metadata global numbers to check it is sane |
745 | void printStats(const ProfileWriterContext &Ctx) { |
746 | char StatMsg[BufSize]; |
747 | char *StatPtr = StatMsg; |
748 | StatPtr = |
749 | strCopy(StatPtr, |
750 | "\nBOLT INSTRUMENTATION RUNTIME STATISTICS\n\nIndCallDescSize: "); |
751 | StatPtr = intToStr(StatPtr, |
752 | Ctx.FuncDescriptions - |
753 | reinterpret_cast<uint8_t *>(Ctx.IndCallDescriptions), |
754 | 10); |
755 | StatPtr = strCopy(StatPtr, "\nFuncDescSize: "); |
756 | StatPtr = intToStr( |
757 | StatPtr, |
758 | reinterpret_cast<uint8_t *>(Ctx.Strings) - Ctx.FuncDescriptions, 10); |
759 | StatPtr = strCopy(StatPtr, "\n__bolt_instr_num_ind_calls: "); |
760 | StatPtr = intToStr(StatPtr, __bolt_instr_num_ind_calls, 10); |
761 | StatPtr = strCopy(StatPtr, "\n__bolt_instr_num_funcs: "); |
762 | StatPtr = intToStr(StatPtr, __bolt_instr_num_funcs, 10); |
763 | StatPtr = strCopy(StatPtr, "\n"); |
764 | __write(2, StatMsg, StatPtr - StatMsg); |
765 | } |
766 | #endif |
767 | |
768 | |
769 | /// This is part of a simple CFG representation in memory, where we store |
770 | /// a dynamically sized array of input and output edges per node, and store |
771 | /// a dynamically sized array of nodes per graph. We also store the spanning |
772 | /// tree edges for that CFG in a separate array of nodes in |
773 | /// \p SpanningTreeNodes, while the regular nodes live in \p CFGNodes. |
774 | struct Edge { |
775 | uint32_t Node; // Index in nodes array regarding the destination of this edge |
776 | uint32_t ID; // Edge index in an array comprising all edges of the graph |
777 | }; |
778 | |
779 | /// A regular graph node or a spanning tree node |
780 | struct Node { |
781 | uint32_t NumInEdges{0}; // Input edge count used to size InEdge |
782 | uint32_t NumOutEdges{0}; // Output edge count used to size OutEdges |
783 | Edge *InEdges{nullptr}; // Created and managed by \p Graph |
784 | Edge *OutEdges{nullptr}; // ditto |
785 | }; |
786 | |
787 | /// Main class for CFG representation in memory. Manages object creation and |
788 | /// destruction, populates an array of CFG nodes as well as corresponding |
789 | /// spanning tree nodes. |
790 | struct Graph { |
791 | uint32_t NumNodes; |
792 | Node *CFGNodes; |
793 | Node *SpanningTreeNodes; |
794 | uint64_t *EdgeFreqs; |
795 | uint64_t *CallFreqs; |
796 | BumpPtrAllocator &Alloc; |
797 | const FunctionDescription &D; |
798 | |
799 | /// Reads a list of edges from function description \p D and builds |
800 | /// the graph from it. Allocates several internal dynamic structures that are |
801 | /// later destroyed by ~Graph() and uses \p Alloc. D.LeafNodes contain all |
802 | /// spanning tree leaf nodes descriptions (their counters). They are the seed |
803 | /// used to compute the rest of the missing edge counts in a bottom-up |
804 | /// traversal of the spanning tree. |
805 | Graph(BumpPtrAllocator &Alloc, const FunctionDescription &D, |
806 | const uint64_t *Counters, ProfileWriterContext &Ctx); |
807 | ~Graph(); |
808 | void dump() const; |
809 | |
810 | private: |
811 | void computeEdgeFrequencies(const uint64_t *Counters, |
812 | ProfileWriterContext &Ctx); |
813 | void dumpEdgeFreqs() const; |
814 | }; |
815 | |
816 | Graph::Graph(BumpPtrAllocator &Alloc, const FunctionDescription &D, |
817 | const uint64_t *Counters, ProfileWriterContext &Ctx) |
818 | : Alloc(Alloc), D(D) { |
819 | DEBUG(reportNumber("G = 0x", (uint64_t)this, 16)){}; |
820 | // First pass to determine number of nodes |
821 | int32_t MaxNodes = -1; |
822 | CallFreqs = nullptr; |
823 | EdgeFreqs = nullptr; |
824 | for (int I = 0; I < D.NumEdges; ++I) { |
825 | if (static_cast<int32_t>(D.Edges[I].FromNode) > MaxNodes) |
826 | MaxNodes = D.Edges[I].FromNode; |
827 | if (static_cast<int32_t>(D.Edges[I].ToNode) > MaxNodes) |
828 | MaxNodes = D.Edges[I].ToNode; |
829 | } |
830 | |
831 | for (int I = 0; I < D.NumLeafNodes; ++I) |
832 | if (static_cast<int32_t>(D.LeafNodes[I].Node) > MaxNodes) |
833 | MaxNodes = D.LeafNodes[I].Node; |
834 | |
835 | for (int I = 0; I < D.NumCalls; ++I) |
836 | if (static_cast<int32_t>(D.Calls[I].FromNode) > MaxNodes) |
837 | MaxNodes = D.Calls[I].FromNode; |
838 | |
839 | // No nodes? Nothing to do |
840 | if (MaxNodes < 0) { |
841 | DEBUG(report("No nodes!\n")){}; |
842 | CFGNodes = nullptr; |
843 | SpanningTreeNodes = nullptr; |
844 | NumNodes = 0; |
845 | return; |
846 | } |
847 | ++MaxNodes; |
848 | DEBUG(reportNumber("NumNodes = ", MaxNodes, 10)){}; |
849 | NumNodes = static_cast<uint32_t>(MaxNodes); |
850 | |
851 | // Initial allocations |
852 | CFGNodes = new (Alloc) Node[MaxNodes]; |
853 | |
854 | DEBUG(reportNumber("G->CFGNodes = 0x", (uint64_t)CFGNodes, 16)){}; |
855 | SpanningTreeNodes = new (Alloc) Node[MaxNodes]; |
856 | DEBUG(reportNumber("G->SpanningTreeNodes = 0x",{} |
857 | (uint64_t)SpanningTreeNodes, 16)){}; |
858 | |
859 | // Figure out how much to allocate to each vector (in/out edge sets) |
860 | for (int I = 0; I < D.NumEdges; ++I) { |
861 | CFGNodes[D.Edges[I].FromNode].NumOutEdges++; |
862 | CFGNodes[D.Edges[I].ToNode].NumInEdges++; |
863 | if (D.Edges[I].Counter != 0xffffffff) |
864 | continue; |
865 | |
866 | SpanningTreeNodes[D.Edges[I].FromNode].NumOutEdges++; |
867 | SpanningTreeNodes[D.Edges[I].ToNode].NumInEdges++; |
868 | } |
869 | |
870 | // Allocate in/out edge sets |
871 | for (int I = 0; I < MaxNodes; ++I) { |
872 | if (CFGNodes[I].NumInEdges > 0) |
873 | CFGNodes[I].InEdges = new (Alloc) Edge[CFGNodes[I].NumInEdges]; |
874 | if (CFGNodes[I].NumOutEdges > 0) |
875 | CFGNodes[I].OutEdges = new (Alloc) Edge[CFGNodes[I].NumOutEdges]; |
876 | if (SpanningTreeNodes[I].NumInEdges > 0) |
877 | SpanningTreeNodes[I].InEdges = |
878 | new (Alloc) Edge[SpanningTreeNodes[I].NumInEdges]; |
879 | if (SpanningTreeNodes[I].NumOutEdges > 0) |
880 | SpanningTreeNodes[I].OutEdges = |
881 | new (Alloc) Edge[SpanningTreeNodes[I].NumOutEdges]; |
882 | CFGNodes[I].NumInEdges = 0; |
883 | CFGNodes[I].NumOutEdges = 0; |
884 | SpanningTreeNodes[I].NumInEdges = 0; |
885 | SpanningTreeNodes[I].NumOutEdges = 0; |
886 | } |
887 | |
888 | // Fill in/out edge sets |
889 | for (int I = 0; I < D.NumEdges; ++I) { |
890 | const uint32_t Src = D.Edges[I].FromNode; |
891 | const uint32_t Dst = D.Edges[I].ToNode; |
892 | Edge *E = &CFGNodes[Src].OutEdges[CFGNodes[Src].NumOutEdges++]; |
893 | E->Node = Dst; |
894 | E->ID = I; |
895 | |
896 | E = &CFGNodes[Dst].InEdges[CFGNodes[Dst].NumInEdges++]; |
897 | E->Node = Src; |
898 | E->ID = I; |
899 | |
900 | if (D.Edges[I].Counter != 0xffffffff) |
901 | continue; |
902 | |
903 | E = &SpanningTreeNodes[Src] |
904 | .OutEdges[SpanningTreeNodes[Src].NumOutEdges++]; |
905 | E->Node = Dst; |
906 | E->ID = I; |
907 | |
908 | E = &SpanningTreeNodes[Dst] |
909 | .InEdges[SpanningTreeNodes[Dst].NumInEdges++]; |
910 | E->Node = Src; |
911 | E->ID = I; |
912 | } |
913 | |
914 | computeEdgeFrequencies(Counters, Ctx); |
915 | } |
916 | |
917 | Graph::~Graph() { |
918 | if (CallFreqs) |
919 | Alloc.deallocate(CallFreqs); |
920 | if (EdgeFreqs) |
921 | Alloc.deallocate(EdgeFreqs); |
922 | for (int I = NumNodes - 1; I >= 0; --I) { |
923 | if (SpanningTreeNodes[I].OutEdges) |
924 | Alloc.deallocate(SpanningTreeNodes[I].OutEdges); |
925 | if (SpanningTreeNodes[I].InEdges) |
926 | Alloc.deallocate(SpanningTreeNodes[I].InEdges); |
927 | if (CFGNodes[I].OutEdges) |
928 | Alloc.deallocate(CFGNodes[I].OutEdges); |
929 | if (CFGNodes[I].InEdges) |
930 | Alloc.deallocate(CFGNodes[I].InEdges); |
931 | } |
932 | if (SpanningTreeNodes) |
933 | Alloc.deallocate(SpanningTreeNodes); |
934 | if (CFGNodes) |
935 | Alloc.deallocate(CFGNodes); |
936 | } |
937 | |
938 | void Graph::dump() const { |
939 | reportNumber("Dumping graph with number of nodes: ", NumNodes, 10); |
940 | report(" Full graph:\n"); |
941 | for (int I = 0; I < NumNodes; ++I) { |
942 | const Node *N = &CFGNodes[I]; |
943 | reportNumber(" Node #", I, 10); |
944 | reportNumber(" InEdges total ", N->NumInEdges, 10); |
945 | for (int J = 0; J < N->NumInEdges; ++J) |
946 | reportNumber(" ", N->InEdges[J].Node, 10); |
947 | reportNumber(" OutEdges total ", N->NumOutEdges, 10); |
948 | for (int J = 0; J < N->NumOutEdges; ++J) |
949 | reportNumber(" ", N->OutEdges[J].Node, 10); |
950 | report("\n"); |
951 | } |
952 | report(" Spanning tree:\n"); |
953 | for (int I = 0; I < NumNodes; ++I) { |
954 | const Node *N = &SpanningTreeNodes[I]; |
955 | reportNumber(" Node #", I, 10); |
956 | reportNumber(" InEdges total ", N->NumInEdges, 10); |
957 | for (int J = 0; J < N->NumInEdges; ++J) |
958 | reportNumber(" ", N->InEdges[J].Node, 10); |
959 | reportNumber(" OutEdges total ", N->NumOutEdges, 10); |
960 | for (int J = 0; J < N->NumOutEdges; ++J) |
961 | reportNumber(" ", N->OutEdges[J].Node, 10); |
962 | report("\n"); |
963 | } |
964 | } |
965 | |
966 | void Graph::dumpEdgeFreqs() const { |
967 | reportNumber( |
968 | "Dumping edge frequencies for graph with num edges: ", D.NumEdges, 10); |
969 | for (int I = 0; I < D.NumEdges; ++I) { |
970 | reportNumber("* Src: ", D.Edges[I].FromNode, 10); |
971 | reportNumber(" Dst: ", D.Edges[I].ToNode, 10); |
972 | reportNumber(" Cnt: ", EdgeFreqs[I], 10); |
973 | } |
974 | } |
975 | |
976 | /// Auxiliary map structure for fast lookups of which calls map to each node of |
977 | /// the function CFG |
978 | struct NodeToCallsMap { |
979 | struct MapEntry { |
980 | uint32_t NumCalls; |
981 | uint32_t *Calls; |
982 | }; |
983 | MapEntry *Entries; |
984 | BumpPtrAllocator &Alloc; |
985 | const uint32_t NumNodes; |
986 | |
987 | NodeToCallsMap(BumpPtrAllocator &Alloc, const FunctionDescription &D, |
988 | uint32_t NumNodes) |
989 | : Alloc(Alloc), NumNodes(NumNodes) { |
990 | Entries = new (Alloc, 0) MapEntry[NumNodes]; |
991 | for (int I = 0; I < D.NumCalls; ++I) { |
992 | DEBUG(reportNumber("Registering call in node ", D.Calls[I].FromNode, 10)){}; |
993 | ++Entries[D.Calls[I].FromNode].NumCalls; |
994 | } |
995 | for (int I = 0; I < NumNodes; ++I) { |
996 | Entries[I].Calls = Entries[I].NumCalls ? new (Alloc) |
997 | uint32_t[Entries[I].NumCalls] |
998 | : nullptr; |
999 | Entries[I].NumCalls = 0; |
1000 | } |
1001 | for (int I = 0; I < D.NumCalls; ++I) { |
1002 | MapEntry &Entry = Entries[D.Calls[I].FromNode]; |
1003 | Entry.Calls[Entry.NumCalls++] = I; |
1004 | } |
1005 | } |
1006 | |
1007 | /// Set the frequency of all calls in node \p NodeID to Freq. However, if |
1008 | /// the calls have their own counters and do not depend on the basic block |
1009 | /// counter, this means they have landing pads and throw exceptions. In this |
1010 | /// case, set their frequency with their counters and return the maximum |
1011 | /// value observed in such counters. This will be used as the new frequency |
1012 | /// at basic block entry. This is used to fix the CFG edge frequencies in the |
1013 | /// presence of exceptions. |
1014 | uint64_t visitAllCallsIn(uint32_t NodeID, uint64_t Freq, uint64_t *CallFreqs, |
1015 | const FunctionDescription &D, |
1016 | const uint64_t *Counters, |
1017 | ProfileWriterContext &Ctx) const { |
1018 | const MapEntry &Entry = Entries[NodeID]; |
1019 | uint64_t MaxValue = 0ull; |
1020 | for (int I = 0, E = Entry.NumCalls; I != E; ++I) { |
1021 | const uint32_t CallID = Entry.Calls[I]; |
1022 | DEBUG(reportNumber(" Setting freq for call ID: ", CallID, 10)){}; |
1023 | const CallDescription &CallDesc = D.Calls[CallID]; |
1024 | if (CallDesc.Counter == 0xffffffff) { |
1025 | CallFreqs[CallID] = Freq; |
1026 | DEBUG(reportNumber(" with : ", Freq, 10)){}; |
1027 | } else { |
1028 | const uint64_t CounterVal = Counters[CallDesc.Counter]; |
1029 | CallFreqs[CallID] = CounterVal; |
1030 | MaxValue = CounterVal > MaxValue ? CounterVal : MaxValue; |
1031 | DEBUG(reportNumber(" with (private counter) : ", CounterVal, 10)){}; |
1032 | } |
1033 | DEBUG(reportNumber(" Address: 0x", CallDesc.TargetAddress, 16)){}; |
1034 | if (CallFreqs[CallID] > 0) |
1035 | Ctx.CallFlowTable->get(CallDesc.TargetAddress).Calls += |
1036 | CallFreqs[CallID]; |
1037 | } |
1038 | return MaxValue; |
1039 | } |
1040 | |
1041 | ~NodeToCallsMap() { |
1042 | for (int I = NumNodes - 1; I >= 0; --I) |
1043 | if (Entries[I].Calls) |
1044 | Alloc.deallocate(Entries[I].Calls); |
1045 | Alloc.deallocate(Entries); |
1046 | } |
1047 | }; |
1048 | |
1049 | /// Fill an array with the frequency of each edge in the function represented |
1050 | /// by G, as well as another array for each call. |
1051 | void Graph::computeEdgeFrequencies(const uint64_t *Counters, |
1052 | ProfileWriterContext &Ctx) { |
1053 | if (NumNodes == 0) |
1054 | return; |
1055 | |
1056 | EdgeFreqs = D.NumEdges ? new (Alloc, 0) uint64_t [D.NumEdges] : nullptr; |
1057 | CallFreqs = D.NumCalls ? new (Alloc, 0) uint64_t [D.NumCalls] : nullptr; |
1058 | |
1059 | // Setup a lookup for calls present in each node (BB) |
1060 | NodeToCallsMap *CallMap = new (Alloc) NodeToCallsMap(Alloc, D, NumNodes); |
1061 | |
1062 | // Perform a bottom-up, BFS traversal of the spanning tree in G. Edges in the |
1063 | // spanning tree don't have explicit counters. We must infer their value using |
1064 | // a linear combination of other counters (sum of counters of the outgoing |
1065 | // edges minus sum of counters of the incoming edges). |
1066 | uint32_t *Stack = new (Alloc) uint32_t [NumNodes]; |
1067 | uint32_t StackTop = 0; |
1068 | enum Status : uint8_t { S_NEW = 0, S_VISITING, S_VISITED }; |
1069 | Status *Visited = new (Alloc, 0) Status[NumNodes]; |
1070 | uint64_t *LeafFrequency = new (Alloc, 0) uint64_t[NumNodes]; |
1071 | uint64_t *EntryAddress = new (Alloc, 0) uint64_t[NumNodes]; |
1072 | |
1073 | // Setup a fast lookup for frequency of leaf nodes, which have special |
1074 | // basic block frequency instrumentation (they are not edge profiled). |
1075 | for (int I = 0; I < D.NumLeafNodes; ++I) { |
1076 | LeafFrequency[D.LeafNodes[I].Node] = Counters[D.LeafNodes[I].Counter]; |
1077 | DEBUG({{} |
1078 | if (Counters[D.LeafNodes[I].Counter] > 0) {{} |
1079 | reportNumber("Leaf Node# ", D.LeafNodes[I].Node, 10);{} |
1080 | reportNumber(" Counter: ", Counters[D.LeafNodes[I].Counter], 10);{} |
1081 | }{} |
1082 | }){}; |
1083 | } |
1084 | for (int I = 0; I < D.NumEntryNodes; ++I) { |
1085 | EntryAddress[D.EntryNodes[I].Node] = D.EntryNodes[I].Address; |
1086 | DEBUG({{} |
1087 | reportNumber("Entry Node# ", D.EntryNodes[I].Node, 10);{} |
1088 | reportNumber(" Address: ", D.EntryNodes[I].Address, 16);{} |
1089 | }){}; |
1090 | } |
1091 | // Add all root nodes to the stack |
1092 | for (int I = 0; I < NumNodes; ++I) |
1093 | if (SpanningTreeNodes[I].NumInEdges == 0) |
1094 | Stack[StackTop++] = I; |
1095 | |
1096 | // Empty stack? |
1097 | if (StackTop == 0) { |
1098 | DEBUG(report("Empty stack!\n")){}; |
1099 | Alloc.deallocate(EntryAddress); |
1100 | Alloc.deallocate(LeafFrequency); |
1101 | Alloc.deallocate(Visited); |
1102 | Alloc.deallocate(Stack); |
1103 | CallMap->~NodeToCallsMap(); |
1104 | Alloc.deallocate(CallMap); |
1105 | if (CallFreqs) |
1106 | Alloc.deallocate(CallFreqs); |
1107 | if (EdgeFreqs) |
1108 | Alloc.deallocate(EdgeFreqs); |
1109 | EdgeFreqs = nullptr; |
1110 | CallFreqs = nullptr; |
1111 | return; |
1112 | } |
1113 | // Add all known edge counts, will infer the rest |
1114 | for (int I = 0; I < D.NumEdges; ++I) { |
1115 | const uint32_t C = D.Edges[I].Counter; |
1116 | if (C == 0xffffffff) // inferred counter - we will compute its value |
1117 | continue; |
1118 | EdgeFreqs[I] = Counters[C]; |
1119 | } |
1120 | |
1121 | while (StackTop > 0) { |
1122 | const uint32_t Cur = Stack[--StackTop]; |
1123 | DEBUG({{} |
1124 | if (Visited[Cur] == S_VISITING){} |
1125 | report("(visiting) ");{} |
1126 | else{} |
1127 | report("(new) ");{} |
1128 | reportNumber("Cur: ", Cur, 10);{} |
1129 | }){}; |
1130 | |
1131 | // This shouldn't happen in a tree |
1132 | assert(Visited[Cur] != S_VISITED, "should not have visited nodes in stack"); |
1133 | if (Visited[Cur] == S_NEW) { |
1134 | Visited[Cur] = S_VISITING; |
1135 | Stack[StackTop++] = Cur; |
1136 | assert(StackTop <= NumNodes, "stack grew too large"); |
1137 | for (int I = 0, E = SpanningTreeNodes[Cur].NumOutEdges; I < E; ++I) { |
1138 | const uint32_t Succ = SpanningTreeNodes[Cur].OutEdges[I].Node; |
1139 | Stack[StackTop++] = Succ; |
1140 | assert(StackTop <= NumNodes, "stack grew too large"); |
1141 | } |
1142 | continue; |
1143 | } |
1144 | Visited[Cur] = S_VISITED; |
1145 | |
1146 | // Establish our node frequency based on outgoing edges, which should all be |
1147 | // resolved by now. |
1148 | int64_t CurNodeFreq = LeafFrequency[Cur]; |
1149 | // Not a leaf? |
1150 | if (!CurNodeFreq) { |
1151 | for (int I = 0, E = CFGNodes[Cur].NumOutEdges; I != E; ++I) { |
1152 | const uint32_t SuccEdge = CFGNodes[Cur].OutEdges[I].ID; |
1153 | CurNodeFreq += EdgeFreqs[SuccEdge]; |
1154 | } |
1155 | } |
1156 | if (CurNodeFreq < 0) |
1157 | CurNodeFreq = 0; |
1158 | |
1159 | const uint64_t CallFreq = CallMap->visitAllCallsIn( |
1160 | Cur, CurNodeFreq > 0 ? CurNodeFreq : 0, CallFreqs, D, Counters, Ctx); |
1161 | |
1162 | // Exception handling affected our output flow? Fix with calls info |
1163 | DEBUG({{} |
1164 | if (CallFreq > CurNodeFreq){} |
1165 | report("Bumping node frequency with call info\n");{} |
1166 | }){}; |
1167 | CurNodeFreq = CallFreq > CurNodeFreq ? CallFreq : CurNodeFreq; |
1168 | |
1169 | if (CurNodeFreq > 0) { |
1170 | if (uint64_t Addr = EntryAddress[Cur]) { |
1171 | DEBUG({} |
1172 | reportNumber(" Setting flow at entry point address 0x", Addr, 16)){}; |
1173 | DEBUG(reportNumber(" with: ", CurNodeFreq, 10)){}; |
1174 | Ctx.CallFlowTable->get(Addr).Val = CurNodeFreq; |
1175 | } |
1176 | } |
1177 | |
1178 | // No parent? Reached a tree root, limit to call frequency updating. |
1179 | if (SpanningTreeNodes[Cur].NumInEdges == 0) |
1180 | continue; |
1181 | |
1182 | assert(SpanningTreeNodes[Cur].NumInEdges == 1, "must have 1 parent"); |
1183 | const uint32_t Parent = SpanningTreeNodes[Cur].InEdges[0].Node; |
Value stored to 'Parent' during its initialization is never read | |
1184 | const uint32_t ParentEdge = SpanningTreeNodes[Cur].InEdges[0].ID; |
1185 | |
1186 | // Calculate parent edge freq. |
1187 | int64_t ParentEdgeFreq = CurNodeFreq; |
1188 | for (int I = 0, E = CFGNodes[Cur].NumInEdges; I != E; ++I) { |
1189 | const uint32_t PredEdge = CFGNodes[Cur].InEdges[I].ID; |
1190 | ParentEdgeFreq -= EdgeFreqs[PredEdge]; |
1191 | } |
1192 | |
1193 | // Sometimes the conservative CFG that BOLT builds will lead to incorrect |
1194 | // flow computation. For example, in a BB that transitively calls the exit |
1195 | // syscall, BOLT will add a fall-through successor even though it should not |
1196 | // have any successors. So this block execution will likely be wrong. We |
1197 | // tolerate this imperfection since this case should be quite infrequent. |
1198 | if (ParentEdgeFreq < 0) { |
1199 | DEBUG(dumpEdgeFreqs()){}; |
1200 | DEBUG(report("WARNING: incorrect flow")){}; |
1201 | ParentEdgeFreq = 0; |
1202 | } |
1203 | DEBUG(reportNumber(" Setting freq for ParentEdge: ", ParentEdge, 10)){}; |
1204 | DEBUG(reportNumber(" with ParentEdgeFreq: ", ParentEdgeFreq, 10)){}; |
1205 | EdgeFreqs[ParentEdge] = ParentEdgeFreq; |
1206 | } |
1207 | |
1208 | Alloc.deallocate(EntryAddress); |
1209 | Alloc.deallocate(LeafFrequency); |
1210 | Alloc.deallocate(Visited); |
1211 | Alloc.deallocate(Stack); |
1212 | CallMap->~NodeToCallsMap(); |
1213 | Alloc.deallocate(CallMap); |
1214 | DEBUG(dumpEdgeFreqs()){}; |
1215 | } |
1216 | |
1217 | /// Write to \p FD all of the edge profiles for function \p FuncDesc. Uses |
1218 | /// \p Alloc to allocate helper dynamic structures used to compute profile for |
1219 | /// edges that we do not explictly instrument. |
1220 | const uint8_t *writeFunctionProfile(int FD, ProfileWriterContext &Ctx, |
1221 | const uint8_t *FuncDesc, |
1222 | BumpPtrAllocator &Alloc) { |
1223 | const FunctionDescription F(FuncDesc); |
1224 | const uint8_t *next = FuncDesc + F.getSize(); |
1225 | |
1226 | #if !defined(__APPLE__) |
1227 | uint64_t *bolt_instr_locations = __bolt_instr_locations; |
1228 | #else |
1229 | uint64_t *bolt_instr_locations = _bolt_instr_locations_getter(); |
1230 | #endif |
1231 | |
1232 | // Skip funcs we know are cold |
1233 | #ifndef ENABLE_DEBUG |
1234 | uint64_t CountersFreq = 0; |
1235 | for (int I = 0; I < F.NumLeafNodes; ++I) |
1236 | CountersFreq += bolt_instr_locations[F.LeafNodes[I].Counter]; |
1237 | |
1238 | if (CountersFreq == 0) { |
1239 | for (int I = 0; I < F.NumEdges; ++I) { |
1240 | const uint32_t C = F.Edges[I].Counter; |
1241 | if (C == 0xffffffff) |
1242 | continue; |
1243 | CountersFreq += bolt_instr_locations[C]; |
1244 | } |
1245 | if (CountersFreq == 0) { |
1246 | for (int I = 0; I < F.NumCalls; ++I) { |
1247 | const uint32_t C = F.Calls[I].Counter; |
1248 | if (C == 0xffffffff) |
1249 | continue; |
1250 | CountersFreq += bolt_instr_locations[C]; |
1251 | } |
1252 | if (CountersFreq == 0) |
1253 | return next; |
1254 | } |
1255 | } |
1256 | #endif |
1257 | |
1258 | Graph *G = new (Alloc) Graph(Alloc, F, bolt_instr_locations, Ctx); |
1259 | DEBUG(G->dump()){}; |
1260 | |
1261 | if (!G->EdgeFreqs && !G->CallFreqs) { |
1262 | G->~Graph(); |
1263 | Alloc.deallocate(G); |
1264 | return next; |
1265 | } |
1266 | |
1267 | for (int I = 0; I < F.NumEdges; ++I) { |
1268 | const uint64_t Freq = G->EdgeFreqs[I]; |
1269 | if (Freq == 0) |
1270 | continue; |
1271 | const EdgeDescription *Desc = &F.Edges[I]; |
1272 | char LineBuf[BufSize]; |
1273 | char *Ptr = LineBuf; |
1274 | Ptr = serializeLoc(Ctx, Ptr, Desc->From, BufSize); |
1275 | Ptr = serializeLoc(Ctx, Ptr, Desc->To, BufSize - (Ptr - LineBuf)); |
1276 | Ptr = strCopy(Ptr, "0 ", BufSize - (Ptr - LineBuf) - 22); |
1277 | Ptr = intToStr(Ptr, Freq, 10); |
1278 | *Ptr++ = '\n'; |
1279 | __write(FD, LineBuf, Ptr - LineBuf); |
1280 | } |
1281 | |
1282 | for (int I = 0; I < F.NumCalls; ++I) { |
1283 | const uint64_t Freq = G->CallFreqs[I]; |
1284 | if (Freq == 0) |
1285 | continue; |
1286 | char LineBuf[BufSize]; |
1287 | char *Ptr = LineBuf; |
1288 | const CallDescription *Desc = &F.Calls[I]; |
1289 | Ptr = serializeLoc(Ctx, Ptr, Desc->From, BufSize); |
1290 | Ptr = serializeLoc(Ctx, Ptr, Desc->To, BufSize - (Ptr - LineBuf)); |
1291 | Ptr = strCopy(Ptr, "0 ", BufSize - (Ptr - LineBuf) - 25); |
1292 | Ptr = intToStr(Ptr, Freq, 10); |
1293 | *Ptr++ = '\n'; |
1294 | __write(FD, LineBuf, Ptr - LineBuf); |
1295 | } |
1296 | |
1297 | G->~Graph(); |
1298 | Alloc.deallocate(G); |
1299 | return next; |
1300 | } |
1301 | |
1302 | #if !defined(__APPLE__) |
1303 | const IndCallTargetDescription * |
1304 | ProfileWriterContext::lookupIndCallTarget(uint64_t Target) const { |
1305 | uint32_t B = 0; |
1306 | uint32_t E = __bolt_instr_num_ind_targets; |
1307 | if (E == 0) |
1308 | return nullptr; |
1309 | do { |
1310 | uint32_t I = (E - B) / 2 + B; |
1311 | if (IndCallTargets[I].Address == Target) |
1312 | return &IndCallTargets[I]; |
1313 | if (IndCallTargets[I].Address < Target) |
1314 | B = I + 1; |
1315 | else |
1316 | E = I; |
1317 | } while (B < E); |
1318 | return nullptr; |
1319 | } |
1320 | |
1321 | /// Write a single indirect call <src, target> pair to the fdata file |
1322 | void visitIndCallCounter(IndirectCallHashTable::MapEntry &Entry, |
1323 | int FD, int CallsiteID, |
1324 | ProfileWriterContext *Ctx) { |
1325 | if (Entry.Val == 0) |
1326 | return; |
1327 | DEBUG(reportNumber("Target func 0x", Entry.Key, 16)){}; |
1328 | DEBUG(reportNumber("Target freq: ", Entry.Val, 10)){}; |
1329 | const IndCallDescription *CallsiteDesc = |
1330 | &Ctx->IndCallDescriptions[CallsiteID]; |
1331 | const IndCallTargetDescription *TargetDesc = |
1332 | Ctx->lookupIndCallTarget(Entry.Key); |
1333 | if (!TargetDesc) { |
1334 | DEBUG(report("Failed to lookup indirect call target\n")){}; |
1335 | char LineBuf[BufSize]; |
1336 | char *Ptr = LineBuf; |
1337 | Ptr = serializeLoc(*Ctx, Ptr, *CallsiteDesc, BufSize); |
1338 | Ptr = strCopy(Ptr, "0 [unknown] 0 0 ", BufSize - (Ptr - LineBuf) - 40); |
1339 | Ptr = intToStr(Ptr, Entry.Val, 10); |
1340 | *Ptr++ = '\n'; |
1341 | __write(FD, LineBuf, Ptr - LineBuf); |
1342 | return; |
1343 | } |
1344 | Ctx->CallFlowTable->get(TargetDesc->Address).Calls += Entry.Val; |
1345 | char LineBuf[BufSize]; |
1346 | char *Ptr = LineBuf; |
1347 | Ptr = serializeLoc(*Ctx, Ptr, *CallsiteDesc, BufSize); |
1348 | Ptr = serializeLoc(*Ctx, Ptr, TargetDesc->Loc, BufSize - (Ptr - LineBuf)); |
1349 | Ptr = strCopy(Ptr, "0 ", BufSize - (Ptr - LineBuf) - 25); |
1350 | Ptr = intToStr(Ptr, Entry.Val, 10); |
1351 | *Ptr++ = '\n'; |
1352 | __write(FD, LineBuf, Ptr - LineBuf); |
1353 | } |
1354 | |
1355 | /// Write to \p FD all of the indirect call profiles. |
1356 | void writeIndirectCallProfile(int FD, ProfileWriterContext &Ctx) { |
1357 | for (int I = 0; I < __bolt_instr_num_ind_calls; ++I) { |
1358 | DEBUG(reportNumber("IndCallsite #", I, 10)){}; |
1359 | GlobalIndCallCounters[I].forEachElement(visitIndCallCounter, FD, I, &Ctx); |
1360 | } |
1361 | } |
1362 | |
1363 | /// Check a single call flow for a callee versus all known callers. If there are |
1364 | /// less callers than what the callee expects, write the difference with source |
1365 | /// [unknown] in the profile. |
1366 | void visitCallFlowEntry(CallFlowHashTable::MapEntry &Entry, int FD, |
1367 | ProfileWriterContext *Ctx) { |
1368 | DEBUG(reportNumber("Call flow entry address: 0x", Entry.Key, 16)){}; |
1369 | DEBUG(reportNumber("Calls: ", Entry.Calls, 10)){}; |
1370 | DEBUG(reportNumber("Reported entry frequency: ", Entry.Val, 10)){}; |
1371 | DEBUG({{} |
1372 | if (Entry.Calls > Entry.Val){} |
1373 | report(" More calls than expected!\n");{} |
1374 | }){}; |
1375 | if (Entry.Val <= Entry.Calls) |
1376 | return; |
1377 | DEBUG(reportNumber({} |
1378 | " Balancing calls with traffic: ", Entry.Val - Entry.Calls, 10)){}; |
1379 | const IndCallTargetDescription *TargetDesc = |
1380 | Ctx->lookupIndCallTarget(Entry.Key); |
1381 | if (!TargetDesc) { |
1382 | // There is probably something wrong with this callee and this should be |
1383 | // investigated, but I don't want to assert and lose all data collected. |
1384 | DEBUG(report("WARNING: failed to look up call target!\n")){}; |
1385 | return; |
1386 | } |
1387 | char LineBuf[BufSize]; |
1388 | char *Ptr = LineBuf; |
1389 | Ptr = strCopy(Ptr, "0 [unknown] 0 ", BufSize); |
1390 | Ptr = serializeLoc(*Ctx, Ptr, TargetDesc->Loc, BufSize - (Ptr - LineBuf)); |
1391 | Ptr = strCopy(Ptr, "0 ", BufSize - (Ptr - LineBuf) - 25); |
1392 | Ptr = intToStr(Ptr, Entry.Val - Entry.Calls, 10); |
1393 | *Ptr++ = '\n'; |
1394 | __write(FD, LineBuf, Ptr - LineBuf); |
1395 | } |
1396 | |
1397 | /// Open fdata file for writing and return a valid file descriptor, aborting |
1398 | /// program upon failure. |
1399 | int openProfile() { |
1400 | // Build the profile name string by appending our PID |
1401 | char Buf[BufSize]; |
1402 | char *Ptr = Buf; |
1403 | uint64_t PID = __getpid(); |
1404 | Ptr = strCopy(Buf, __bolt_instr_filename, BufSize); |
1405 | if (__bolt_instr_use_pid) { |
1406 | Ptr = strCopy(Ptr, ".", BufSize - (Ptr - Buf + 1)); |
1407 | Ptr = intToStr(Ptr, PID, 10); |
1408 | Ptr = strCopy(Ptr, ".fdata", BufSize - (Ptr - Buf + 1)); |
1409 | } |
1410 | *Ptr++ = '\0'; |
1411 | uint64_t FD = __open(Buf, |
1412 | /*flags=*/0x241 /*O_WRONLY|O_TRUNC|O_CREAT*/, |
1413 | /*mode=*/0666); |
1414 | if (static_cast<int64_t>(FD) < 0) { |
1415 | report("Error while trying to open profile file for writing: "); |
1416 | report(Buf); |
1417 | reportNumber("\nFailed with error number: 0x", |
1418 | 0 - static_cast<int64_t>(FD), 16); |
1419 | __exit(1); |
1420 | } |
1421 | return FD; |
1422 | } |
1423 | |
1424 | #endif |
1425 | |
1426 | } // anonymous namespace |
1427 | |
1428 | #if !defined(__APPLE__) |
1429 | |
1430 | /// Reset all counters in case you want to start profiling a new phase of your |
1431 | /// program independently of prior phases. |
1432 | /// The address of this function is printed by BOLT and this can be called by |
1433 | /// any attached debugger during runtime. There is a useful oneliner for gdb: |
1434 | /// |
1435 | /// gdb -p $(pgrep -xo PROCESSNAME) -ex 'p ((void(*)())0xdeadbeef)()' \ |
1436 | /// -ex 'set confirm off' -ex quit |
1437 | /// |
1438 | /// Where 0xdeadbeef is this function address and PROCESSNAME your binary file |
1439 | /// name. |
1440 | extern "C" void __bolt_instr_clear_counters() { |
1441 | memset(reinterpret_cast<char *>(__bolt_instr_locations), 0, |
1442 | __bolt_num_counters * 8); |
1443 | for (int I = 0; I < __bolt_instr_num_ind_calls; ++I) |
1444 | GlobalIndCallCounters[I].resetCounters(); |
1445 | } |
1446 | |
1447 | /// This is the entry point for profile writing. |
1448 | /// There are three ways of getting here: |
1449 | /// |
1450 | /// * Program execution ended, finalization methods are running and BOLT |
1451 | /// hooked into FINI from your binary dynamic section; |
1452 | /// * You used the sleep timer option and during initialization we forked |
1453 | /// a separete process that will call this function periodically; |
1454 | /// * BOLT prints this function address so you can attach a debugger and |
1455 | /// call this function directly to get your profile written to disk |
1456 | /// on demand. |
1457 | /// |
1458 | extern "C" void __attribute((force_align_arg_pointer)) |
1459 | __bolt_instr_data_dump() { |
1460 | // Already dumping |
1461 | if (!GlobalWriteProfileMutex->acquire()) |
1462 | return; |
1463 | |
1464 | BumpPtrAllocator HashAlloc; |
1465 | HashAlloc.setMaxSize(0x6400000); |
1466 | ProfileWriterContext Ctx = readDescriptions(); |
1467 | Ctx.CallFlowTable = new (HashAlloc, 0) CallFlowHashTable(HashAlloc); |
1468 | |
1469 | DEBUG(printStats(Ctx)){}; |
1470 | |
1471 | int FD = openProfile(); |
1472 | |
1473 | BumpPtrAllocator Alloc; |
1474 | const uint8_t *FuncDesc = Ctx.FuncDescriptions; |
1475 | for (int I = 0, E = __bolt_instr_num_funcs; I < E; ++I) { |
1476 | FuncDesc = writeFunctionProfile(FD, Ctx, FuncDesc, Alloc); |
1477 | Alloc.clear(); |
1478 | DEBUG(reportNumber("FuncDesc now: ", (uint64_t)FuncDesc, 16)){}; |
1479 | } |
1480 | assert(FuncDesc == (void *)Ctx.Strings, |
1481 | "FuncDesc ptr must be equal to stringtable"); |
1482 | |
1483 | writeIndirectCallProfile(FD, Ctx); |
1484 | Ctx.CallFlowTable->forEachElement(visitCallFlowEntry, FD, &Ctx); |
1485 | |
1486 | __fsync(FD); |
1487 | __close(FD); |
1488 | __munmap(Ctx.MMapPtr, Ctx.MMapSize); |
1489 | __close(Ctx.FileDesc); |
1490 | HashAlloc.destroy(); |
1491 | GlobalWriteProfileMutex->release(); |
1492 | DEBUG(report("Finished writing profile.\n")){}; |
1493 | } |
1494 | |
1495 | /// Event loop for our child process spawned during setup to dump profile data |
1496 | /// at user-specified intervals |
1497 | void watchProcess() { |
1498 | timespec ts, rem; |
1499 | uint64_t Ellapsed = 0ull; |
1500 | uint64_t ppid; |
1501 | if (__bolt_instr_wait_forks) { |
1502 | // Store parent pgid |
1503 | ppid = -__getpgid(0); |
1504 | // And leave parent process group |
1505 | __setpgid(0, 0); |
1506 | } else { |
1507 | // Store parent pid |
1508 | ppid = __getppid(); |
1509 | if (ppid == 1) { |
1510 | // Parent already dead |
1511 | __bolt_instr_data_dump(); |
1512 | goto out; |
1513 | } |
1514 | } |
1515 | |
1516 | ts.tv_sec = 1; |
1517 | ts.tv_nsec = 0; |
1518 | while (1) { |
1519 | __nanosleep(&ts, &rem); |
1520 | // This means our parent process or all its forks are dead, |
1521 | // so no need for us to keep dumping. |
1522 | if (__kill(ppid, 0) < 0) { |
1523 | if (__bolt_instr_no_counters_clear) |
1524 | __bolt_instr_data_dump(); |
1525 | break; |
1526 | } |
1527 | |
1528 | if (++Ellapsed < __bolt_instr_sleep_time) |
1529 | continue; |
1530 | |
1531 | Ellapsed = 0; |
1532 | __bolt_instr_data_dump(); |
1533 | if (__bolt_instr_no_counters_clear == false) |
1534 | __bolt_instr_clear_counters(); |
1535 | } |
1536 | |
1537 | out:; |
1538 | DEBUG(report("My parent process is dead, bye!\n")){}; |
1539 | __exit(0); |
1540 | } |
1541 | |
1542 | extern "C" void __bolt_instr_indirect_call(); |
1543 | extern "C" void __bolt_instr_indirect_tailcall(); |
1544 | |
1545 | /// Initialization code |
1546 | extern "C" void __attribute((force_align_arg_pointer)) __bolt_instr_setup() { |
1547 | const uint64_t CountersStart = |
1548 | reinterpret_cast<uint64_t>(&__bolt_instr_locations[0]); |
1549 | const uint64_t CountersEnd = alignTo( |
1550 | reinterpret_cast<uint64_t>(&__bolt_instr_locations[__bolt_num_counters]), |
1551 | 0x1000); |
1552 | DEBUG(reportNumber("replace mmap start: ", CountersStart, 16)){}; |
1553 | DEBUG(reportNumber("replace mmap stop: ", CountersEnd, 16)){}; |
1554 | assert (CountersEnd > CountersStart, "no counters"); |
1555 | // Maps our counters to be shared instead of private, so we keep counting for |
1556 | // forked processes |
1557 | __mmap(CountersStart, CountersEnd - CountersStart, |
1558 | 0x3 /*PROT_READ|PROT_WRITE*/, |
1559 | 0x31 /*MAP_ANONYMOUS | MAP_SHARED | MAP_FIXED*/, -1, 0); |
1560 | |
1561 | __bolt_ind_call_counter_func_pointer = __bolt_instr_indirect_call; |
1562 | __bolt_ind_tailcall_counter_func_pointer = __bolt_instr_indirect_tailcall; |
1563 | // Conservatively reserve 100MiB shared pages |
1564 | GlobalAlloc.setMaxSize(0x6400000); |
1565 | GlobalAlloc.setShared(true); |
1566 | GlobalWriteProfileMutex = new (GlobalAlloc, 0) Mutex(); |
1567 | if (__bolt_instr_num_ind_calls > 0) |
1568 | GlobalIndCallCounters = |
1569 | new (GlobalAlloc, 0) IndirectCallHashTable[__bolt_instr_num_ind_calls]; |
1570 | |
1571 | if (__bolt_instr_sleep_time != 0) { |
1572 | // Separate instrumented process to the own process group |
1573 | if (__bolt_instr_wait_forks) |
1574 | __setpgid(0, 0); |
1575 | |
1576 | if (long PID = __fork()) |
1577 | return; |
1578 | watchProcess(); |
1579 | } |
1580 | } |
1581 | |
1582 | extern "C" __attribute((force_align_arg_pointer)) void |
1583 | instrumentIndirectCall(uint64_t Target, uint64_t IndCallID) { |
1584 | GlobalIndCallCounters[IndCallID].incrementVal(Target, GlobalAlloc); |
1585 | } |
1586 | |
1587 | /// We receive as in-stack arguments the identifier of the indirect call site |
1588 | /// as well as the target address for the call |
1589 | extern "C" __attribute((naked)) void __bolt_instr_indirect_call() |
1590 | { |
1591 | __asm__ __volatile__(SAVE_ALL"push %%rax\n" "push %%rbx\n" "push %%rcx\n" "push %%rdx\n" "push %%rdi\n" "push %%rsi\n" "push %%rbp\n" "push %%r8\n" "push %%r9\n" "push %%r10\n" "push %%r11\n" "push %%r12\n" "push %%r13\n" "push %%r14\n" "push %%r15\n" "sub $8, %%rsp\n" |
1592 | "mov 0xa0(%%rsp), %%rdi\n" |
1593 | "mov 0x98(%%rsp), %%rsi\n" |
1594 | "call instrumentIndirectCall\n" |
1595 | RESTORE_ALL"add $8, %%rsp\n" "pop %%r15\n" "pop %%r14\n" "pop %%r13\n" "pop %%r12\n" "pop %%r11\n" "pop %%r10\n" "pop %%r9\n" "pop %%r8\n" "pop %%rbp\n" "pop %%rsi\n" "pop %%rdi\n" "pop %%rdx\n" "pop %%rcx\n" "pop %%rbx\n" "pop %%rax\n" |
1596 | "ret\n" |
1597 | :::); |
1598 | } |
1599 | |
1600 | extern "C" __attribute((naked)) void __bolt_instr_indirect_tailcall() |
1601 | { |
1602 | __asm__ __volatile__(SAVE_ALL"push %%rax\n" "push %%rbx\n" "push %%rcx\n" "push %%rdx\n" "push %%rdi\n" "push %%rsi\n" "push %%rbp\n" "push %%r8\n" "push %%r9\n" "push %%r10\n" "push %%r11\n" "push %%r12\n" "push %%r13\n" "push %%r14\n" "push %%r15\n" "sub $8, %%rsp\n" |
1603 | "mov 0x98(%%rsp), %%rdi\n" |
1604 | "mov 0x90(%%rsp), %%rsi\n" |
1605 | "call instrumentIndirectCall\n" |
1606 | RESTORE_ALL"add $8, %%rsp\n" "pop %%r15\n" "pop %%r14\n" "pop %%r13\n" "pop %%r12\n" "pop %%r11\n" "pop %%r10\n" "pop %%r9\n" "pop %%r8\n" "pop %%rbp\n" "pop %%rsi\n" "pop %%rdi\n" "pop %%rdx\n" "pop %%rcx\n" "pop %%rbx\n" "pop %%rax\n" |
1607 | "ret\n" |
1608 | :::); |
1609 | } |
1610 | |
1611 | /// This is hooking ELF's entry, it needs to save all machine state. |
1612 | extern "C" __attribute((naked)) void __bolt_instr_start() |
1613 | { |
1614 | __asm__ __volatile__(SAVE_ALL"push %%rax\n" "push %%rbx\n" "push %%rcx\n" "push %%rdx\n" "push %%rdi\n" "push %%rsi\n" "push %%rbp\n" "push %%r8\n" "push %%r9\n" "push %%r10\n" "push %%r11\n" "push %%r12\n" "push %%r13\n" "push %%r14\n" "push %%r15\n" "sub $8, %%rsp\n" |
1615 | "call __bolt_instr_setup\n" |
1616 | RESTORE_ALL"add $8, %%rsp\n" "pop %%r15\n" "pop %%r14\n" "pop %%r13\n" "pop %%r12\n" "pop %%r11\n" "pop %%r10\n" "pop %%r9\n" "pop %%r8\n" "pop %%rbp\n" "pop %%rsi\n" "pop %%rdi\n" "pop %%rdx\n" "pop %%rcx\n" "pop %%rbx\n" "pop %%rax\n" |
1617 | "jmp __bolt_start_trampoline\n" |
1618 | :::); |
1619 | } |
1620 | |
1621 | /// This is hooking into ELF's DT_FINI |
1622 | extern "C" void __bolt_instr_fini() { |
1623 | __bolt_fini_trampoline(); |
1624 | if (__bolt_instr_sleep_time == 0) |
1625 | __bolt_instr_data_dump(); |
1626 | DEBUG(report("Finished.\n")){}; |
1627 | } |
1628 | |
1629 | #endif |
1630 | |
1631 | #if defined(__APPLE__) |
1632 | |
1633 | extern "C" void __bolt_instr_data_dump() { |
1634 | ProfileWriterContext Ctx = readDescriptions(); |
1635 | |
1636 | int FD = 2; |
1637 | BumpPtrAllocator Alloc; |
1638 | const uint8_t *FuncDesc = Ctx.FuncDescriptions; |
1639 | uint32_t bolt_instr_num_funcs = _bolt_instr_num_funcs_getter(); |
1640 | |
1641 | for (int I = 0, E = bolt_instr_num_funcs; I < E; ++I) { |
1642 | FuncDesc = writeFunctionProfile(FD, Ctx, FuncDesc, Alloc); |
1643 | Alloc.clear(); |
1644 | DEBUG(reportNumber("FuncDesc now: ", (uint64_t)FuncDesc, 16)){}; |
1645 | } |
1646 | assert(FuncDesc == (void *)Ctx.Strings, |
1647 | "FuncDesc ptr must be equal to stringtable"); |
1648 | } |
1649 | |
1650 | // On OSX/iOS the final symbol name of an extern "C" function/variable contains |
1651 | // one extra leading underscore: _bolt_instr_setup -> __bolt_instr_setup. |
1652 | extern "C" |
1653 | __attribute__((section("__TEXT,__setup"))) |
1654 | __attribute__((force_align_arg_pointer)) |
1655 | void _bolt_instr_setup() { |
1656 | __asm__ __volatile__(SAVE_ALL"push %%rax\n" "push %%rbx\n" "push %%rcx\n" "push %%rdx\n" "push %%rdi\n" "push %%rsi\n" "push %%rbp\n" "push %%r8\n" "push %%r9\n" "push %%r10\n" "push %%r11\n" "push %%r12\n" "push %%r13\n" "push %%r14\n" "push %%r15\n" "sub $8, %%rsp\n" :::); |
1657 | |
1658 | report("Hello!\n"); |
1659 | |
1660 | __asm__ __volatile__(RESTORE_ALL"add $8, %%rsp\n" "pop %%r15\n" "pop %%r14\n" "pop %%r13\n" "pop %%r12\n" "pop %%r11\n" "pop %%r10\n" "pop %%r9\n" "pop %%r8\n" "pop %%rbp\n" "pop %%rsi\n" "pop %%rdi\n" "pop %%rdx\n" "pop %%rcx\n" "pop %%rbx\n" "pop %%rax\n" :::); |
1661 | } |
1662 | |
1663 | extern "C" |
1664 | __attribute__((section("__TEXT,__fini"))) |
1665 | __attribute__((force_align_arg_pointer)) |
1666 | void _bolt_instr_fini() { |
1667 | report("Bye!\n"); |
1668 | __bolt_instr_data_dump(); |
1669 | } |
1670 | |
1671 | #endif |
1672 | #endif |