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 | #include "common.h" |
44 | |
45 | // Enables a very verbose logging to stderr useful when debugging |
46 | //#define ENABLE_DEBUG |
47 | |
48 | #ifdef ENABLE_DEBUG |
49 | #define DEBUG(X) \ |
50 | { X; } |
51 | #else |
52 | #define DEBUG(X) \ |
53 | {} |
54 | #endif |
55 | |
56 | #pragma GCC visibility push(hidden) |
57 | |
58 | extern "C" { |
59 | |
60 | #if defined(__APPLE__) |
61 | extern uint64_t* _bolt_instr_locations_getter(); |
62 | extern uint32_t _bolt_num_counters_getter(); |
63 | |
64 | extern uint8_t* _bolt_instr_tables_getter(); |
65 | extern uint32_t _bolt_instr_num_funcs_getter(); |
66 | |
67 | #else |
68 | |
69 | // Main counters inserted by instrumentation, incremented during runtime when |
70 | // points of interest (locations) in the program are reached. Those are direct |
71 | // calls and direct and indirect branches (local ones). There are also counters |
72 | // for basic block execution if they are a spanning tree leaf and need to be |
73 | // counted in order to infer the execution count of other edges of the CFG. |
74 | extern uint64_t __bolt_instr_locations[]; |
75 | extern uint32_t __bolt_num_counters; |
76 | // Descriptions are serialized metadata about binary functions written by BOLT, |
77 | // so we have a minimal understanding about the program structure. For a |
78 | // reference on the exact format of this metadata, see *Description structs, |
79 | // Location, IntrumentedNode and EntryNode. |
80 | // Number of indirect call site descriptions |
81 | extern uint32_t __bolt_instr_num_ind_calls; |
82 | // Number of indirect call target descriptions |
83 | extern uint32_t __bolt_instr_num_ind_targets; |
84 | // Number of function descriptions |
85 | extern uint32_t __bolt_instr_num_funcs; |
86 | // Time to sleep across dumps (when we write the fdata profile to disk) |
87 | extern uint32_t __bolt_instr_sleep_time; |
88 | // Do not clear counters across dumps, rewrite file with the updated values |
89 | extern bool __bolt_instr_no_counters_clear; |
90 | // Wait until all forks of instrumented process will finish |
91 | extern bool __bolt_instr_wait_forks; |
92 | // Filename to dump data to |
93 | extern char __bolt_instr_filename[]; |
94 | // Instumented binary file path |
95 | extern char __bolt_instr_binpath[]; |
96 | // If true, append current PID to the fdata filename when creating it so |
97 | // different invocations of the same program can be differentiated. |
98 | extern bool __bolt_instr_use_pid; |
99 | // Functions that will be used to instrument indirect calls. BOLT static pass |
100 | // will identify indirect calls and modify them to load the address in these |
101 | // trampolines and call this address instead. BOLT can't use direct calls to |
102 | // our handlers because our addresses here are not known at analysis time. We |
103 | // only support resolving dependencies from this file to the output of BOLT, |
104 | // *not* the other way around. |
105 | // TODO: We need better linking support to make that happen. |
106 | extern void (*__bolt_ind_call_counter_func_pointer)(); |
107 | extern void (*__bolt_ind_tailcall_counter_func_pointer)(); |
108 | // Function pointers to init/fini trampoline routines in the binary, so we can |
109 | // resume regular execution of these functions that we hooked |
110 | extern void __bolt_start_trampoline(); |
111 | extern void __bolt_fini_trampoline(); |
112 | |
113 | #endif |
114 | } |
115 | |
116 | namespace { |
117 | |
118 | /// A simple allocator that mmaps a fixed size region and manages this space |
119 | /// in a stack fashion, meaning you always deallocate the last element that |
120 | /// was allocated. In practice, we don't need to deallocate individual elements. |
121 | /// We monotonically increase our usage and then deallocate everything once we |
122 | /// are done processing something. |
123 | class BumpPtrAllocator { |
124 | /// This is written before each allocation and act as a canary to detect when |
125 | /// a bug caused our program to cross allocation boundaries. |
126 | struct EntryMetadata { |
127 | uint64_t Magic; |
128 | uint64_t AllocSize; |
129 | }; |
130 | |
131 | public: |
132 | void *allocate(size_t Size) { |
133 | Lock L(M); |
134 | |
135 | if (StackBase == nullptr) { |
136 | StackBase = reinterpret_cast<uint8_t *>( |
137 | __mmap(addr: 0, size: MaxSize, PROT_READ | PROT_WRITE, |
138 | flags: (Shared ? MAP_SHARED : MAP_PRIVATE) | MAP_ANONYMOUS, fd: -1, offset: 0)); |
139 | assert(Assertion: StackBase != MAP_FAILED, |
140 | Msg: "BumpPtrAllocator: failed to mmap stack!" ); |
141 | StackSize = 0; |
142 | } |
143 | |
144 | Size = alignTo(Value: Size + sizeof(EntryMetadata), Align: 16); |
145 | uint8_t *AllocAddress = StackBase + StackSize + sizeof(EntryMetadata); |
146 | auto *M = reinterpret_cast<EntryMetadata *>(StackBase + StackSize); |
147 | M->Magic = Magic; |
148 | M->AllocSize = Size; |
149 | StackSize += Size; |
150 | assert(Assertion: StackSize < MaxSize, Msg: "allocator ran out of memory" ); |
151 | return AllocAddress; |
152 | } |
153 | |
154 | #ifdef DEBUG |
155 | /// Element-wise deallocation is only used for debugging to catch memory |
156 | /// bugs by checking magic bytes. Ordinarily, we reset the allocator once |
157 | /// we are done with it. Reset is done with clear(). There's no need |
158 | /// to deallocate each element individually. |
159 | void deallocate(void *Ptr) { |
160 | Lock L(M); |
161 | uint8_t MetadataOffset = sizeof(EntryMetadata); |
162 | auto *M = reinterpret_cast<EntryMetadata *>( |
163 | reinterpret_cast<uint8_t *>(Ptr) - MetadataOffset); |
164 | const uint8_t *StackTop = StackBase + StackSize + MetadataOffset; |
165 | // Validate size |
166 | if (Ptr != StackTop - M->AllocSize) { |
167 | // Failed validation, check if it is a pointer returned by operator new [] |
168 | MetadataOffset += |
169 | sizeof(uint64_t); // Space for number of elements alloc'ed |
170 | M = reinterpret_cast<EntryMetadata *>(reinterpret_cast<uint8_t *>(Ptr) - |
171 | MetadataOffset); |
172 | // Ok, it failed both checks if this assertion fails. Stop the program, we |
173 | // have a memory bug. |
174 | assert(Assertion: Ptr == StackTop - M->AllocSize, |
175 | Msg: "must deallocate the last element alloc'ed" ); |
176 | } |
177 | assert(Assertion: M->Magic == Magic, Msg: "allocator magic is corrupt" ); |
178 | StackSize -= M->AllocSize; |
179 | } |
180 | #else |
181 | void deallocate(void *) {} |
182 | #endif |
183 | |
184 | void clear() { |
185 | Lock L(M); |
186 | StackSize = 0; |
187 | } |
188 | |
189 | /// Set mmap reservation size (only relevant before first allocation) |
190 | void setMaxSize(uint64_t Size) { MaxSize = Size; } |
191 | |
192 | /// Set mmap reservation privacy (only relevant before first allocation) |
193 | void setShared(bool S) { Shared = S; } |
194 | |
195 | void destroy() { |
196 | if (StackBase == nullptr) |
197 | return; |
198 | __munmap(addr: StackBase, size: MaxSize); |
199 | } |
200 | |
201 | // Placement operator to construct allocator in possibly shared mmaped memory |
202 | static void *operator new(size_t, void *Ptr) { return Ptr; }; |
203 | |
204 | private: |
205 | static constexpr uint64_t Magic = 0x1122334455667788ull; |
206 | uint64_t MaxSize = 0xa00000; |
207 | uint8_t *StackBase{nullptr}; |
208 | uint64_t StackSize{0}; |
209 | bool Shared{false}; |
210 | Mutex M; |
211 | }; |
212 | |
213 | /// Used for allocating indirect call instrumentation counters. Initialized by |
214 | /// __bolt_instr_setup, our initialization routine. |
215 | BumpPtrAllocator *GlobalAlloc; |
216 | |
217 | // Base address which we substract from recorded PC values when searching for |
218 | // indirect call description entries. Needed because indCall descriptions are |
219 | // mapped read-only and contain static addresses. Initialized in |
220 | // __bolt_instr_setup. |
221 | uint64_t TextBaseAddress = 0; |
222 | |
223 | // Storage for GlobalAlloc which can be shared if not using |
224 | // instrumentation-file-append-pid. |
225 | void *GlobalMetadataStorage; |
226 | |
227 | } // anonymous namespace |
228 | |
229 | // User-defined placement new operators. We only use those (as opposed to |
230 | // overriding the regular operator new) so we can keep our allocator in the |
231 | // stack instead of in a data section (global). |
232 | void *operator new(size_t Sz, BumpPtrAllocator &A) { return A.allocate(Size: Sz); } |
233 | void *operator new(size_t Sz, BumpPtrAllocator &A, char C) { |
234 | auto *Ptr = reinterpret_cast<char *>(A.allocate(Size: Sz)); |
235 | memset(Buf: Ptr, C, Size: Sz); |
236 | return Ptr; |
237 | } |
238 | void *operator new[](size_t Sz, BumpPtrAllocator &A) { |
239 | return A.allocate(Size: Sz); |
240 | } |
241 | void *operator new[](size_t Sz, BumpPtrAllocator &A, char C) { |
242 | auto *Ptr = reinterpret_cast<char *>(A.allocate(Size: Sz)); |
243 | memset(Buf: Ptr, C, Size: Sz); |
244 | return Ptr; |
245 | } |
246 | // Only called during exception unwinding (useless). We must manually dealloc. |
247 | // C++ language weirdness |
248 | void operator delete(void *Ptr, BumpPtrAllocator &A) { A.deallocate(Ptr); } |
249 | |
250 | namespace { |
251 | |
252 | // Disable instrumentation optimizations that sacrifice profile accuracy |
253 | extern "C" bool __bolt_instr_conservative; |
254 | |
255 | /// Basic key-val atom stored in our hash |
256 | struct SimpleHashTableEntryBase { |
257 | uint64_t Key; |
258 | uint64_t Val; |
259 | void dump(const char *Msg = nullptr) { |
260 | // TODO: make some sort of formatting function |
261 | // Currently we have to do it the ugly way because |
262 | // we want every message to be printed atomically via a single call to |
263 | // __write. If we use reportNumber() and others nultiple times, we'll get |
264 | // garbage in mulithreaded environment |
265 | char Buf[BufSize]; |
266 | char *Ptr = Buf; |
267 | Ptr = intToStr(OutBuf: Ptr, Num: __getpid(), Base: 10); |
268 | *Ptr++ = ':'; |
269 | *Ptr++ = ' '; |
270 | if (Msg) |
271 | Ptr = strCopy(OutBuf: Ptr, Str: Msg, Size: strLen(Str: Msg)); |
272 | *Ptr++ = '0'; |
273 | *Ptr++ = 'x'; |
274 | Ptr = intToStr(OutBuf: Ptr, Num: (uint64_t)this, Base: 16); |
275 | *Ptr++ = ':'; |
276 | *Ptr++ = ' '; |
277 | Ptr = strCopy(OutBuf: Ptr, Str: "MapEntry(0x" , Size: sizeof("MapEntry(0x" ) - 1); |
278 | Ptr = intToStr(OutBuf: Ptr, Num: Key, Base: 16); |
279 | *Ptr++ = ','; |
280 | *Ptr++ = ' '; |
281 | *Ptr++ = '0'; |
282 | *Ptr++ = 'x'; |
283 | Ptr = intToStr(OutBuf: Ptr, Num: Val, Base: 16); |
284 | *Ptr++ = ')'; |
285 | *Ptr++ = '\n'; |
286 | assert(Assertion: Ptr - Buf < BufSize, Msg: "Buffer overflow!" ); |
287 | // print everything all at once for atomicity |
288 | __write(fd: 2, buf: Buf, count: Ptr - Buf); |
289 | } |
290 | }; |
291 | |
292 | /// This hash table implementation starts by allocating a table of size |
293 | /// InitialSize. When conflicts happen in this main table, it resolves |
294 | /// them by chaining a new table of size IncSize. It never reallocs as our |
295 | /// allocator doesn't support it. The key is intended to be function pointers. |
296 | /// There's no clever hash function (it's just x mod size, size being prime). |
297 | /// I never tuned the coefficientes in the modular equation (TODO) |
298 | /// This is used for indirect calls (each call site has one of this, so it |
299 | /// should have a small footprint) and for tallying call counts globally for |
300 | /// each target to check if we missed the origin of some calls (this one is a |
301 | /// large instantiation of this template, since it is global for all call sites) |
302 | template <typename T = SimpleHashTableEntryBase, uint32_t InitialSize = 7, |
303 | uint32_t IncSize = 7> |
304 | class SimpleHashTable { |
305 | public: |
306 | using MapEntry = T; |
307 | |
308 | /// Increment by 1 the value of \p Key. If it is not in this table, it will be |
309 | /// added to the table and its value set to 1. |
310 | void incrementVal(uint64_t Key, BumpPtrAllocator &Alloc) { |
311 | if (!__bolt_instr_conservative) { |
312 | TryLock L(M); |
313 | if (!L.isLocked()) |
314 | return; |
315 | auto &E = getOrAllocEntry(Key, Alloc); |
316 | ++E.Val; |
317 | return; |
318 | } |
319 | Lock L(M); |
320 | auto &E = getOrAllocEntry(Key, Alloc); |
321 | ++E.Val; |
322 | } |
323 | |
324 | /// Basic member accessing interface. Here we pass the allocator explicitly to |
325 | /// avoid storing a pointer to it as part of this table (remember there is one |
326 | /// hash for each indirect call site, so we want to minimize our footprint). |
327 | MapEntry &get(uint64_t Key, BumpPtrAllocator &Alloc) { |
328 | if (!__bolt_instr_conservative) { |
329 | TryLock L(M); |
330 | if (!L.isLocked()) |
331 | return NoEntry; |
332 | return getOrAllocEntry(Key, Alloc); |
333 | } |
334 | Lock L(M); |
335 | return getOrAllocEntry(Key, Alloc); |
336 | } |
337 | |
338 | /// Traverses all elements in the table |
339 | template <typename... Args> |
340 | void forEachElement(void (*Callback)(MapEntry &, Args...), Args... args) { |
341 | Lock L(M); |
342 | if (!TableRoot) |
343 | return; |
344 | return forEachElement(Callback, InitialSize, TableRoot, args...); |
345 | } |
346 | |
347 | void resetCounters(); |
348 | |
349 | private: |
350 | constexpr static uint64_t VacantMarker = 0; |
351 | constexpr static uint64_t FollowUpTableMarker = 0x8000000000000000ull; |
352 | |
353 | MapEntry *TableRoot{nullptr}; |
354 | MapEntry NoEntry; |
355 | Mutex M; |
356 | |
357 | template <typename... Args> |
358 | void forEachElement(void (*Callback)(MapEntry &, Args...), |
359 | uint32_t NumEntries, MapEntry *Entries, Args... args) { |
360 | for (uint32_t I = 0; I < NumEntries; ++I) { |
361 | MapEntry &Entry = Entries[I]; |
362 | if (Entry.Key == VacantMarker) |
363 | continue; |
364 | if (Entry.Key & FollowUpTableMarker) { |
365 | MapEntry *Next = |
366 | reinterpret_cast<MapEntry *>(Entry.Key & ~FollowUpTableMarker); |
367 | assert(Next != Entries, "Circular reference!" ); |
368 | forEachElement(Callback, IncSize, Next, args...); |
369 | continue; |
370 | } |
371 | Callback(Entry, args...); |
372 | } |
373 | } |
374 | |
375 | MapEntry &firstAllocation(uint64_t Key, BumpPtrAllocator &Alloc) { |
376 | TableRoot = new (Alloc, 0) MapEntry[InitialSize]; |
377 | MapEntry &Entry = TableRoot[Key % InitialSize]; |
378 | Entry.Key = Key; |
379 | // DEBUG(Entry.dump("Created root entry: ")); |
380 | return Entry; |
381 | } |
382 | |
383 | MapEntry &getEntry(MapEntry *Entries, uint64_t Key, uint64_t Selector, |
384 | BumpPtrAllocator &Alloc, int CurLevel) { |
385 | // DEBUG(reportNumber("getEntry called, level ", CurLevel, 10)); |
386 | const uint32_t NumEntries = CurLevel == 0 ? InitialSize : IncSize; |
387 | uint64_t Remainder = Selector / NumEntries; |
388 | Selector = Selector % NumEntries; |
389 | MapEntry &Entry = Entries[Selector]; |
390 | |
391 | // A hit |
392 | if (Entry.Key == Key) { |
393 | // DEBUG(Entry.dump("Hit: ")); |
394 | return Entry; |
395 | } |
396 | |
397 | // Vacant - add new entry |
398 | if (Entry.Key == VacantMarker) { |
399 | Entry.Key = Key; |
400 | // DEBUG(Entry.dump("Adding new entry: ")); |
401 | return Entry; |
402 | } |
403 | |
404 | // Defer to the next level |
405 | if (Entry.Key & FollowUpTableMarker) { |
406 | return getEntry( |
407 | Entries: reinterpret_cast<MapEntry *>(Entry.Key & ~FollowUpTableMarker), |
408 | Key, Selector: Remainder, Alloc, CurLevel: CurLevel + 1); |
409 | } |
410 | |
411 | // Conflict - create the next level |
412 | // DEBUG(Entry.dump("Creating new level: ")); |
413 | |
414 | MapEntry *NextLevelTbl = new (Alloc, 0) MapEntry[IncSize]; |
415 | // DEBUG( |
416 | // reportNumber("Newly allocated level: 0x", uint64_t(NextLevelTbl), |
417 | // 16)); |
418 | uint64_t CurEntrySelector = Entry.Key / InitialSize; |
419 | for (int I = 0; I < CurLevel; ++I) |
420 | CurEntrySelector /= IncSize; |
421 | CurEntrySelector = CurEntrySelector % IncSize; |
422 | NextLevelTbl[CurEntrySelector] = Entry; |
423 | Entry.Key = reinterpret_cast<uint64_t>(NextLevelTbl) | FollowUpTableMarker; |
424 | assert((NextLevelTbl[CurEntrySelector].Key & ~FollowUpTableMarker) != |
425 | uint64_t(Entries), |
426 | "circular reference created!\n" ); |
427 | // DEBUG(NextLevelTbl[CurEntrySelector].dump("New level entry: ")); |
428 | // DEBUG(Entry.dump("Updated old entry: ")); |
429 | return getEntry(Entries: NextLevelTbl, Key, Selector: Remainder, Alloc, CurLevel: CurLevel + 1); |
430 | } |
431 | |
432 | MapEntry &getOrAllocEntry(uint64_t Key, BumpPtrAllocator &Alloc) { |
433 | if (TableRoot) { |
434 | MapEntry &E = getEntry(Entries: TableRoot, Key, Selector: Key, Alloc, CurLevel: 0); |
435 | assert(!(E.Key & FollowUpTableMarker), "Invalid entry!" ); |
436 | return E; |
437 | } |
438 | return firstAllocation(Key, Alloc); |
439 | } |
440 | }; |
441 | |
442 | template <typename T> void resetIndCallCounter(T &Entry) { |
443 | Entry.Val = 0; |
444 | } |
445 | |
446 | template <typename T, uint32_t X, uint32_t Y> |
447 | void SimpleHashTable<T, X, Y>::resetCounters() { |
448 | forEachElement(resetIndCallCounter); |
449 | } |
450 | |
451 | /// Represents a hash table mapping a function target address to its counter. |
452 | using IndirectCallHashTable = SimpleHashTable<>; |
453 | |
454 | /// Initialize with number 1 instead of 0 so we don't go into .bss. This is the |
455 | /// global array of all hash tables storing indirect call destinations happening |
456 | /// during runtime, one table per call site. |
457 | IndirectCallHashTable *GlobalIndCallCounters{ |
458 | reinterpret_cast<IndirectCallHashTable *>(1)}; |
459 | |
460 | /// Don't allow reentrancy in the fdata writing phase - only one thread writes |
461 | /// it |
462 | Mutex *GlobalWriteProfileMutex{reinterpret_cast<Mutex *>(1)}; |
463 | |
464 | /// Store number of calls in additional to target address (Key) and frequency |
465 | /// as perceived by the basic block counter (Val). |
466 | struct CallFlowEntryBase : public SimpleHashTableEntryBase { |
467 | uint64_t Calls; |
468 | }; |
469 | |
470 | using CallFlowHashTableBase = SimpleHashTable<CallFlowEntryBase, 11939, 233>; |
471 | |
472 | /// This is a large table indexing all possible call targets (indirect and |
473 | /// direct ones). The goal is to find mismatches between number of calls (for |
474 | /// those calls we were able to track) and the entry basic block counter of the |
475 | /// callee. In most cases, these two should be equal. If not, there are two |
476 | /// possible scenarios here: |
477 | /// |
478 | /// * Entry BB has higher frequency than all known calls to this function. |
479 | /// In this case, we have dynamic library code or any uninstrumented code |
480 | /// calling this function. We will write the profile for these untracked |
481 | /// calls as having source "0 [unknown] 0" in the fdata file. |
482 | /// |
483 | /// * Number of known calls is higher than the frequency of entry BB |
484 | /// This only happens when there is no counter for the entry BB / callee |
485 | /// function is not simple (in BOLT terms). We don't do anything special |
486 | /// here and just ignore those (we still report all calls to the non-simple |
487 | /// function, though). |
488 | /// |
489 | class CallFlowHashTable : public CallFlowHashTableBase { |
490 | public: |
491 | CallFlowHashTable(BumpPtrAllocator &Alloc) : Alloc(Alloc) {} |
492 | |
493 | MapEntry &get(uint64_t Key) { return CallFlowHashTableBase::get(Key, Alloc); } |
494 | |
495 | private: |
496 | // Different than the hash table for indirect call targets, we do store the |
497 | // allocator here since there is only one call flow hash and space overhead |
498 | // is negligible. |
499 | BumpPtrAllocator &Alloc; |
500 | }; |
501 | |
502 | /// |
503 | /// Description metadata emitted by BOLT to describe the program - refer to |
504 | /// Passes/Instrumentation.cpp - Instrumentation::emitTablesAsELFNote() |
505 | /// |
506 | struct Location { |
507 | uint32_t FunctionName; |
508 | uint32_t Offset; |
509 | }; |
510 | |
511 | struct CallDescription { |
512 | Location From; |
513 | uint32_t FromNode; |
514 | Location To; |
515 | uint32_t Counter; |
516 | uint64_t TargetAddress; |
517 | }; |
518 | |
519 | using IndCallDescription = Location; |
520 | |
521 | struct IndCallTargetDescription { |
522 | Location Loc; |
523 | uint64_t Address; |
524 | }; |
525 | |
526 | struct EdgeDescription { |
527 | Location From; |
528 | uint32_t FromNode; |
529 | Location To; |
530 | uint32_t ToNode; |
531 | uint32_t Counter; |
532 | }; |
533 | |
534 | struct InstrumentedNode { |
535 | uint32_t Node; |
536 | uint32_t Counter; |
537 | }; |
538 | |
539 | struct EntryNode { |
540 | uint64_t Node; |
541 | uint64_t Address; |
542 | }; |
543 | |
544 | struct FunctionDescription { |
545 | uint32_t NumLeafNodes; |
546 | const InstrumentedNode *LeafNodes; |
547 | uint32_t NumEdges; |
548 | const EdgeDescription *Edges; |
549 | uint32_t NumCalls; |
550 | const CallDescription *Calls; |
551 | uint32_t NumEntryNodes; |
552 | const EntryNode *EntryNodes; |
553 | |
554 | /// Constructor will parse the serialized function metadata written by BOLT |
555 | FunctionDescription(const uint8_t *FuncDesc); |
556 | |
557 | uint64_t getSize() const { |
558 | return 16 + NumLeafNodes * sizeof(InstrumentedNode) + |
559 | NumEdges * sizeof(EdgeDescription) + |
560 | NumCalls * sizeof(CallDescription) + |
561 | NumEntryNodes * sizeof(EntryNode); |
562 | } |
563 | }; |
564 | |
565 | /// The context is created when the fdata profile needs to be written to disk |
566 | /// and we need to interpret our runtime counters. It contains pointers to the |
567 | /// mmaped binary (only the BOLT written metadata section). Deserialization |
568 | /// should be straightforward as most data is POD or an array of POD elements. |
569 | /// This metadata is used to reconstruct function CFGs. |
570 | struct ProfileWriterContext { |
571 | IndCallDescription *IndCallDescriptions; |
572 | IndCallTargetDescription *IndCallTargets; |
573 | uint8_t *FuncDescriptions; |
574 | char *Strings; // String table with function names used in this binary |
575 | int FileDesc; // File descriptor for the file on disk backing this |
576 | // information in memory via mmap |
577 | void *MMapPtr; // The mmap ptr |
578 | int MMapSize; // The mmap size |
579 | |
580 | /// Hash table storing all possible call destinations to detect untracked |
581 | /// calls and correctly report them as [unknown] in output fdata. |
582 | CallFlowHashTable *CallFlowTable; |
583 | |
584 | /// Lookup the sorted indirect call target vector to fetch function name and |
585 | /// offset for an arbitrary function pointer. |
586 | const IndCallTargetDescription *lookupIndCallTarget(uint64_t Target) const; |
587 | }; |
588 | |
589 | /// Perform a string comparison and returns zero if Str1 matches Str2. Compares |
590 | /// at most Size characters. |
591 | int compareStr(const char *Str1, const char *Str2, int Size) { |
592 | while (*Str1 == *Str2) { |
593 | if (*Str1 == '\0' || --Size == 0) |
594 | return 0; |
595 | ++Str1; |
596 | ++Str2; |
597 | } |
598 | return 1; |
599 | } |
600 | |
601 | /// Output Location to the fdata file |
602 | char *serializeLoc(const ProfileWriterContext &Ctx, char *OutBuf, |
603 | const Location Loc, uint32_t BufSize) { |
604 | // fdata location format: Type Name Offset |
605 | // Type 1 - regular symbol |
606 | OutBuf = strCopy(OutBuf, Str: "1 " ); |
607 | const char *Str = Ctx.Strings + Loc.FunctionName; |
608 | uint32_t Size = 25; |
609 | while (*Str) { |
610 | *OutBuf++ = *Str++; |
611 | if (++Size >= BufSize) |
612 | break; |
613 | } |
614 | assert(Assertion: !*Str, Msg: "buffer overflow, function name too large" ); |
615 | *OutBuf++ = ' '; |
616 | OutBuf = intToStr(OutBuf, Num: Loc.Offset, Base: 16); |
617 | *OutBuf++ = ' '; |
618 | return OutBuf; |
619 | } |
620 | |
621 | /// Read and deserialize a function description written by BOLT. \p FuncDesc |
622 | /// points at the beginning of the function metadata structure in the file. |
623 | /// See Instrumentation::emitTablesAsELFNote() |
624 | FunctionDescription::FunctionDescription(const uint8_t *FuncDesc) { |
625 | NumLeafNodes = *reinterpret_cast<const uint32_t *>(FuncDesc); |
626 | DEBUG(reportNumber("NumLeafNodes = " , NumLeafNodes, 10)); |
627 | LeafNodes = reinterpret_cast<const InstrumentedNode *>(FuncDesc + 4); |
628 | |
629 | NumEdges = *reinterpret_cast<const uint32_t *>( |
630 | FuncDesc + 4 + NumLeafNodes * sizeof(InstrumentedNode)); |
631 | DEBUG(reportNumber("NumEdges = " , NumEdges, 10)); |
632 | Edges = reinterpret_cast<const EdgeDescription *>( |
633 | FuncDesc + 8 + NumLeafNodes * sizeof(InstrumentedNode)); |
634 | |
635 | NumCalls = *reinterpret_cast<const uint32_t *>( |
636 | FuncDesc + 8 + NumLeafNodes * sizeof(InstrumentedNode) + |
637 | NumEdges * sizeof(EdgeDescription)); |
638 | DEBUG(reportNumber("NumCalls = " , NumCalls, 10)); |
639 | Calls = reinterpret_cast<const CallDescription *>( |
640 | FuncDesc + 12 + NumLeafNodes * sizeof(InstrumentedNode) + |
641 | NumEdges * sizeof(EdgeDescription)); |
642 | NumEntryNodes = *reinterpret_cast<const uint32_t *>( |
643 | FuncDesc + 12 + NumLeafNodes * sizeof(InstrumentedNode) + |
644 | NumEdges * sizeof(EdgeDescription) + NumCalls * sizeof(CallDescription)); |
645 | DEBUG(reportNumber("NumEntryNodes = " , NumEntryNodes, 10)); |
646 | EntryNodes = reinterpret_cast<const EntryNode *>( |
647 | FuncDesc + 16 + NumLeafNodes * sizeof(InstrumentedNode) + |
648 | NumEdges * sizeof(EdgeDescription) + NumCalls * sizeof(CallDescription)); |
649 | } |
650 | |
651 | /// Read and mmap descriptions written by BOLT from the executable's notes |
652 | /// section |
653 | #if defined(HAVE_ELF_H) and !defined(__APPLE__) |
654 | |
655 | void *__attribute__((noinline)) __get_pc() { |
656 | return __builtin_extract_return_addr(__builtin_return_address(0)); |
657 | } |
658 | |
659 | /// Get string with address and parse it to hex pair <StartAddress, EndAddress> |
660 | bool parseAddressRange(const char *Str, uint64_t &StartAddress, |
661 | uint64_t &EndAddress) { |
662 | if (!Str) |
663 | return false; |
664 | // Parsed string format: <hex1>-<hex2> |
665 | StartAddress = hexToLong(Str, '-'); |
666 | while (*Str && *Str != '-') |
667 | ++Str; |
668 | if (!*Str) |
669 | return false; |
670 | ++Str; // swallow '-' |
671 | EndAddress = hexToLong(Str); |
672 | return true; |
673 | } |
674 | |
675 | /// Get full path to the real binary by getting current virtual address |
676 | /// and searching for the appropriate link in address range in |
677 | /// /proc/self/map_files |
678 | static char *getBinaryPath() { |
679 | const uint32_t BufSize = 1024; |
680 | const uint32_t NameMax = 4096; |
681 | const char DirPath[] = "/proc/self/map_files/" ; |
682 | static char TargetPath[NameMax] = {}; |
683 | char Buf[BufSize]; |
684 | |
685 | if (__bolt_instr_binpath[0] != '\0') |
686 | return __bolt_instr_binpath; |
687 | |
688 | if (TargetPath[0] != '\0') |
689 | return TargetPath; |
690 | |
691 | unsigned long CurAddr = (unsigned long)__get_pc(); |
692 | uint64_t FDdir = __open(DirPath, O_RDONLY, |
693 | /*mode=*/0666); |
694 | assert(static_cast<int64_t>(FDdir) >= 0, |
695 | "failed to open /proc/self/map_files" ); |
696 | |
697 | while (long Nread = __getdents64(FDdir, (struct dirent64 *)Buf, BufSize)) { |
698 | assert(static_cast<int64_t>(Nread) != -1, "failed to get folder entries" ); |
699 | |
700 | struct dirent64 *d; |
701 | for (long Bpos = 0; Bpos < Nread; Bpos += d->d_reclen) { |
702 | d = (struct dirent64 *)(Buf + Bpos); |
703 | |
704 | uint64_t StartAddress, EndAddress; |
705 | if (!parseAddressRange(d->d_name, StartAddress, EndAddress)) |
706 | continue; |
707 | if (CurAddr < StartAddress || CurAddr > EndAddress) |
708 | continue; |
709 | char FindBuf[NameMax]; |
710 | char *C = strCopy(FindBuf, DirPath, NameMax); |
711 | C = strCopy(C, d->d_name, NameMax - (C - FindBuf)); |
712 | *C = '\0'; |
713 | uint32_t Ret = __readlink(FindBuf, TargetPath, sizeof(TargetPath)); |
714 | assert(Ret != -1 && Ret != BufSize, "readlink error" ); |
715 | TargetPath[Ret] = '\0'; |
716 | return TargetPath; |
717 | } |
718 | } |
719 | return nullptr; |
720 | } |
721 | |
722 | ProfileWriterContext readDescriptions() { |
723 | ProfileWriterContext Result; |
724 | char *BinPath = getBinaryPath(); |
725 | assert(BinPath && BinPath[0] != '\0', "failed to find binary path" ); |
726 | |
727 | uint64_t FD = __open(BinPath, O_RDONLY, |
728 | /*mode=*/0666); |
729 | assert(static_cast<int64_t>(FD) >= 0, "failed to open binary path" ); |
730 | |
731 | Result.FileDesc = FD; |
732 | |
733 | // mmap our binary to memory |
734 | uint64_t Size = __lseek(FD, 0, SEEK_END); |
735 | uint8_t *BinContents = reinterpret_cast<uint8_t *>( |
736 | __mmap(0, Size, PROT_READ, MAP_PRIVATE, FD, 0)); |
737 | assert(BinContents != MAP_FAILED, "readDescriptions: Failed to mmap self!" ); |
738 | Result.MMapPtr = BinContents; |
739 | Result.MMapSize = Size; |
740 | Elf64_Ehdr *Hdr = reinterpret_cast<Elf64_Ehdr *>(BinContents); |
741 | Elf64_Shdr *Shdr = reinterpret_cast<Elf64_Shdr *>(BinContents + Hdr->e_shoff); |
742 | Elf64_Shdr *StringTblHeader = reinterpret_cast<Elf64_Shdr *>( |
743 | BinContents + Hdr->e_shoff + Hdr->e_shstrndx * Hdr->e_shentsize); |
744 | |
745 | // Find .bolt.instr.tables with the data we need and set pointers to it |
746 | for (int I = 0; I < Hdr->e_shnum; ++I) { |
747 | char *SecName = reinterpret_cast<char *>( |
748 | BinContents + StringTblHeader->sh_offset + Shdr->sh_name); |
749 | if (compareStr(SecName, ".bolt.instr.tables" , 64) != 0) { |
750 | Shdr = reinterpret_cast<Elf64_Shdr *>(BinContents + Hdr->e_shoff + |
751 | (I + 1) * Hdr->e_shentsize); |
752 | continue; |
753 | } |
754 | // Actual contents of the ELF note start after offset 20 decimal: |
755 | // Offset 0: Producer name size (4 bytes) |
756 | // Offset 4: Contents size (4 bytes) |
757 | // Offset 8: Note type (4 bytes) |
758 | // Offset 12: Producer name (BOLT\0) (5 bytes + align to 4-byte boundary) |
759 | // Offset 20: Contents |
760 | uint32_t IndCallDescSize = |
761 | *reinterpret_cast<uint32_t *>(BinContents + Shdr->sh_offset + 20); |
762 | uint32_t IndCallTargetDescSize = *reinterpret_cast<uint32_t *>( |
763 | BinContents + Shdr->sh_offset + 24 + IndCallDescSize); |
764 | uint32_t FuncDescSize = |
765 | *reinterpret_cast<uint32_t *>(BinContents + Shdr->sh_offset + 28 + |
766 | IndCallDescSize + IndCallTargetDescSize); |
767 | Result.IndCallDescriptions = reinterpret_cast<IndCallDescription *>( |
768 | BinContents + Shdr->sh_offset + 24); |
769 | Result.IndCallTargets = reinterpret_cast<IndCallTargetDescription *>( |
770 | BinContents + Shdr->sh_offset + 28 + IndCallDescSize); |
771 | Result.FuncDescriptions = BinContents + Shdr->sh_offset + 32 + |
772 | IndCallDescSize + IndCallTargetDescSize; |
773 | Result.Strings = reinterpret_cast<char *>( |
774 | BinContents + Shdr->sh_offset + 32 + IndCallDescSize + |
775 | IndCallTargetDescSize + FuncDescSize); |
776 | return Result; |
777 | } |
778 | const char ErrMsg[] = |
779 | "BOLT instrumentation runtime error: could not find section " |
780 | ".bolt.instr.tables\n" ; |
781 | reportError(ErrMsg, sizeof(ErrMsg)); |
782 | return Result; |
783 | } |
784 | |
785 | #else |
786 | |
787 | ProfileWriterContext readDescriptions() { |
788 | ProfileWriterContext Result; |
789 | uint8_t *Tables = _bolt_instr_tables_getter(); |
790 | uint32_t IndCallDescSize = *reinterpret_cast<uint32_t *>(Tables); |
791 | uint32_t IndCallTargetDescSize = |
792 | *reinterpret_cast<uint32_t *>(Tables + 4 + IndCallDescSize); |
793 | uint32_t FuncDescSize = *reinterpret_cast<uint32_t *>( |
794 | Tables + 8 + IndCallDescSize + IndCallTargetDescSize); |
795 | Result.IndCallDescriptions = |
796 | reinterpret_cast<IndCallDescription *>(Tables + 4); |
797 | Result.IndCallTargets = reinterpret_cast<IndCallTargetDescription *>( |
798 | Tables + 8 + IndCallDescSize); |
799 | Result.FuncDescriptions = |
800 | Tables + 12 + IndCallDescSize + IndCallTargetDescSize; |
801 | Result.Strings = reinterpret_cast<char *>( |
802 | Tables + 12 + IndCallDescSize + IndCallTargetDescSize + FuncDescSize); |
803 | return Result; |
804 | } |
805 | |
806 | #endif |
807 | |
808 | #if !defined(__APPLE__) |
809 | /// Debug by printing overall metadata global numbers to check it is sane |
810 | void printStats(const ProfileWriterContext &Ctx) { |
811 | char StatMsg[BufSize]; |
812 | char *StatPtr = StatMsg; |
813 | StatPtr = |
814 | strCopy(OutBuf: StatPtr, |
815 | Str: "\nBOLT INSTRUMENTATION RUNTIME STATISTICS\n\nIndCallDescSize: " ); |
816 | StatPtr = intToStr(OutBuf: StatPtr, |
817 | Num: Ctx.FuncDescriptions - |
818 | reinterpret_cast<uint8_t *>(Ctx.IndCallDescriptions), |
819 | Base: 10); |
820 | StatPtr = strCopy(OutBuf: StatPtr, Str: "\nFuncDescSize: " ); |
821 | StatPtr = intToStr( |
822 | OutBuf: StatPtr, |
823 | Num: reinterpret_cast<uint8_t *>(Ctx.Strings) - Ctx.FuncDescriptions, Base: 10); |
824 | StatPtr = strCopy(OutBuf: StatPtr, Str: "\n__bolt_instr_num_ind_calls: " ); |
825 | StatPtr = intToStr(OutBuf: StatPtr, Num: __bolt_instr_num_ind_calls, Base: 10); |
826 | StatPtr = strCopy(OutBuf: StatPtr, Str: "\n__bolt_instr_num_funcs: " ); |
827 | StatPtr = intToStr(OutBuf: StatPtr, Num: __bolt_instr_num_funcs, Base: 10); |
828 | StatPtr = strCopy(OutBuf: StatPtr, Str: "\n" ); |
829 | __write(fd: 2, buf: StatMsg, count: StatPtr - StatMsg); |
830 | } |
831 | #endif |
832 | |
833 | |
834 | /// This is part of a simple CFG representation in memory, where we store |
835 | /// a dynamically sized array of input and output edges per node, and store |
836 | /// a dynamically sized array of nodes per graph. We also store the spanning |
837 | /// tree edges for that CFG in a separate array of nodes in |
838 | /// \p SpanningTreeNodes, while the regular nodes live in \p CFGNodes. |
839 | struct Edge { |
840 | uint32_t Node; // Index in nodes array regarding the destination of this edge |
841 | uint32_t ID; // Edge index in an array comprising all edges of the graph |
842 | }; |
843 | |
844 | /// A regular graph node or a spanning tree node |
845 | struct Node { |
846 | uint32_t NumInEdges{0}; // Input edge count used to size InEdge |
847 | uint32_t NumOutEdges{0}; // Output edge count used to size OutEdges |
848 | Edge *InEdges{nullptr}; // Created and managed by \p Graph |
849 | Edge *OutEdges{nullptr}; // ditto |
850 | }; |
851 | |
852 | /// Main class for CFG representation in memory. Manages object creation and |
853 | /// destruction, populates an array of CFG nodes as well as corresponding |
854 | /// spanning tree nodes. |
855 | struct Graph { |
856 | uint32_t NumNodes; |
857 | Node *CFGNodes; |
858 | Node *SpanningTreeNodes; |
859 | uint64_t *EdgeFreqs; |
860 | uint64_t *CallFreqs; |
861 | BumpPtrAllocator &Alloc; |
862 | const FunctionDescription &D; |
863 | |
864 | /// Reads a list of edges from function description \p D and builds |
865 | /// the graph from it. Allocates several internal dynamic structures that are |
866 | /// later destroyed by ~Graph() and uses \p Alloc. D.LeafNodes contain all |
867 | /// spanning tree leaf nodes descriptions (their counters). They are the seed |
868 | /// used to compute the rest of the missing edge counts in a bottom-up |
869 | /// traversal of the spanning tree. |
870 | Graph(BumpPtrAllocator &Alloc, const FunctionDescription &D, |
871 | const uint64_t *Counters, ProfileWriterContext &Ctx); |
872 | ~Graph(); |
873 | void dump() const; |
874 | |
875 | private: |
876 | void computeEdgeFrequencies(const uint64_t *Counters, |
877 | ProfileWriterContext &Ctx); |
878 | void dumpEdgeFreqs() const; |
879 | }; |
880 | |
881 | Graph::Graph(BumpPtrAllocator &Alloc, const FunctionDescription &D, |
882 | const uint64_t *Counters, ProfileWriterContext &Ctx) |
883 | : Alloc(Alloc), D(D) { |
884 | DEBUG(reportNumber("G = 0x" , (uint64_t)this, 16)); |
885 | // First pass to determine number of nodes |
886 | int32_t MaxNodes = -1; |
887 | CallFreqs = nullptr; |
888 | EdgeFreqs = nullptr; |
889 | for (int I = 0; I < D.NumEdges; ++I) { |
890 | if (static_cast<int32_t>(D.Edges[I].FromNode) > MaxNodes) |
891 | MaxNodes = D.Edges[I].FromNode; |
892 | if (static_cast<int32_t>(D.Edges[I].ToNode) > MaxNodes) |
893 | MaxNodes = D.Edges[I].ToNode; |
894 | } |
895 | |
896 | for (int I = 0; I < D.NumLeafNodes; ++I) |
897 | if (static_cast<int32_t>(D.LeafNodes[I].Node) > MaxNodes) |
898 | MaxNodes = D.LeafNodes[I].Node; |
899 | |
900 | for (int I = 0; I < D.NumCalls; ++I) |
901 | if (static_cast<int32_t>(D.Calls[I].FromNode) > MaxNodes) |
902 | MaxNodes = D.Calls[I].FromNode; |
903 | |
904 | // No nodes? Nothing to do |
905 | if (MaxNodes < 0) { |
906 | DEBUG(report("No nodes!\n" )); |
907 | CFGNodes = nullptr; |
908 | SpanningTreeNodes = nullptr; |
909 | NumNodes = 0; |
910 | return; |
911 | } |
912 | ++MaxNodes; |
913 | DEBUG(reportNumber("NumNodes = " , MaxNodes, 10)); |
914 | NumNodes = static_cast<uint32_t>(MaxNodes); |
915 | |
916 | // Initial allocations |
917 | CFGNodes = new (Alloc) Node[MaxNodes]; |
918 | |
919 | DEBUG(reportNumber("G->CFGNodes = 0x" , (uint64_t)CFGNodes, 16)); |
920 | SpanningTreeNodes = new (Alloc) Node[MaxNodes]; |
921 | DEBUG(reportNumber("G->SpanningTreeNodes = 0x" , |
922 | (uint64_t)SpanningTreeNodes, 16)); |
923 | |
924 | // Figure out how much to allocate to each vector (in/out edge sets) |
925 | for (int I = 0; I < D.NumEdges; ++I) { |
926 | CFGNodes[D.Edges[I].FromNode].NumOutEdges++; |
927 | CFGNodes[D.Edges[I].ToNode].NumInEdges++; |
928 | if (D.Edges[I].Counter != 0xffffffff) |
929 | continue; |
930 | |
931 | SpanningTreeNodes[D.Edges[I].FromNode].NumOutEdges++; |
932 | SpanningTreeNodes[D.Edges[I].ToNode].NumInEdges++; |
933 | } |
934 | |
935 | // Allocate in/out edge sets |
936 | for (int I = 0; I < MaxNodes; ++I) { |
937 | if (CFGNodes[I].NumInEdges > 0) |
938 | CFGNodes[I].InEdges = new (Alloc) Edge[CFGNodes[I].NumInEdges]; |
939 | if (CFGNodes[I].NumOutEdges > 0) |
940 | CFGNodes[I].OutEdges = new (Alloc) Edge[CFGNodes[I].NumOutEdges]; |
941 | if (SpanningTreeNodes[I].NumInEdges > 0) |
942 | SpanningTreeNodes[I].InEdges = |
943 | new (Alloc) Edge[SpanningTreeNodes[I].NumInEdges]; |
944 | if (SpanningTreeNodes[I].NumOutEdges > 0) |
945 | SpanningTreeNodes[I].OutEdges = |
946 | new (Alloc) Edge[SpanningTreeNodes[I].NumOutEdges]; |
947 | CFGNodes[I].NumInEdges = 0; |
948 | CFGNodes[I].NumOutEdges = 0; |
949 | SpanningTreeNodes[I].NumInEdges = 0; |
950 | SpanningTreeNodes[I].NumOutEdges = 0; |
951 | } |
952 | |
953 | // Fill in/out edge sets |
954 | for (int I = 0; I < D.NumEdges; ++I) { |
955 | const uint32_t Src = D.Edges[I].FromNode; |
956 | const uint32_t Dst = D.Edges[I].ToNode; |
957 | Edge *E = &CFGNodes[Src].OutEdges[CFGNodes[Src].NumOutEdges++]; |
958 | E->Node = Dst; |
959 | E->ID = I; |
960 | |
961 | E = &CFGNodes[Dst].InEdges[CFGNodes[Dst].NumInEdges++]; |
962 | E->Node = Src; |
963 | E->ID = I; |
964 | |
965 | if (D.Edges[I].Counter != 0xffffffff) |
966 | continue; |
967 | |
968 | E = &SpanningTreeNodes[Src] |
969 | .OutEdges[SpanningTreeNodes[Src].NumOutEdges++]; |
970 | E->Node = Dst; |
971 | E->ID = I; |
972 | |
973 | E = &SpanningTreeNodes[Dst] |
974 | .InEdges[SpanningTreeNodes[Dst].NumInEdges++]; |
975 | E->Node = Src; |
976 | E->ID = I; |
977 | } |
978 | |
979 | computeEdgeFrequencies(Counters, Ctx); |
980 | } |
981 | |
982 | Graph::~Graph() { |
983 | if (CallFreqs) |
984 | Alloc.deallocate(Ptr: CallFreqs); |
985 | if (EdgeFreqs) |
986 | Alloc.deallocate(Ptr: EdgeFreqs); |
987 | for (int I = NumNodes - 1; I >= 0; --I) { |
988 | if (SpanningTreeNodes[I].OutEdges) |
989 | Alloc.deallocate(Ptr: SpanningTreeNodes[I].OutEdges); |
990 | if (SpanningTreeNodes[I].InEdges) |
991 | Alloc.deallocate(Ptr: SpanningTreeNodes[I].InEdges); |
992 | if (CFGNodes[I].OutEdges) |
993 | Alloc.deallocate(Ptr: CFGNodes[I].OutEdges); |
994 | if (CFGNodes[I].InEdges) |
995 | Alloc.deallocate(Ptr: CFGNodes[I].InEdges); |
996 | } |
997 | if (SpanningTreeNodes) |
998 | Alloc.deallocate(Ptr: SpanningTreeNodes); |
999 | if (CFGNodes) |
1000 | Alloc.deallocate(Ptr: CFGNodes); |
1001 | } |
1002 | |
1003 | void Graph::dump() const { |
1004 | reportNumber(Msg: "Dumping graph with number of nodes: " , Num: NumNodes, Base: 10); |
1005 | report(Msg: " Full graph:\n" ); |
1006 | for (int I = 0; I < NumNodes; ++I) { |
1007 | const Node *N = &CFGNodes[I]; |
1008 | reportNumber(Msg: " Node #" , Num: I, Base: 10); |
1009 | reportNumber(Msg: " InEdges total " , Num: N->NumInEdges, Base: 10); |
1010 | for (int J = 0; J < N->NumInEdges; ++J) |
1011 | reportNumber(Msg: " " , Num: N->InEdges[J].Node, Base: 10); |
1012 | reportNumber(Msg: " OutEdges total " , Num: N->NumOutEdges, Base: 10); |
1013 | for (int J = 0; J < N->NumOutEdges; ++J) |
1014 | reportNumber(Msg: " " , Num: N->OutEdges[J].Node, Base: 10); |
1015 | report(Msg: "\n" ); |
1016 | } |
1017 | report(Msg: " Spanning tree:\n" ); |
1018 | for (int I = 0; I < NumNodes; ++I) { |
1019 | const Node *N = &SpanningTreeNodes[I]; |
1020 | reportNumber(Msg: " Node #" , Num: I, Base: 10); |
1021 | reportNumber(Msg: " InEdges total " , Num: N->NumInEdges, Base: 10); |
1022 | for (int J = 0; J < N->NumInEdges; ++J) |
1023 | reportNumber(Msg: " " , Num: N->InEdges[J].Node, Base: 10); |
1024 | reportNumber(Msg: " OutEdges total " , Num: N->NumOutEdges, Base: 10); |
1025 | for (int J = 0; J < N->NumOutEdges; ++J) |
1026 | reportNumber(Msg: " " , Num: N->OutEdges[J].Node, Base: 10); |
1027 | report(Msg: "\n" ); |
1028 | } |
1029 | } |
1030 | |
1031 | void Graph::dumpEdgeFreqs() const { |
1032 | reportNumber( |
1033 | Msg: "Dumping edge frequencies for graph with num edges: " , Num: D.NumEdges, Base: 10); |
1034 | for (int I = 0; I < D.NumEdges; ++I) { |
1035 | reportNumber(Msg: "* Src: " , Num: D.Edges[I].FromNode, Base: 10); |
1036 | reportNumber(Msg: " Dst: " , Num: D.Edges[I].ToNode, Base: 10); |
1037 | reportNumber(Msg: " Cnt: " , Num: EdgeFreqs[I], Base: 10); |
1038 | } |
1039 | } |
1040 | |
1041 | /// Auxiliary map structure for fast lookups of which calls map to each node of |
1042 | /// the function CFG |
1043 | struct NodeToCallsMap { |
1044 | struct MapEntry { |
1045 | uint32_t NumCalls; |
1046 | uint32_t *Calls; |
1047 | }; |
1048 | MapEntry *Entries; |
1049 | BumpPtrAllocator &Alloc; |
1050 | const uint32_t NumNodes; |
1051 | |
1052 | NodeToCallsMap(BumpPtrAllocator &Alloc, const FunctionDescription &D, |
1053 | uint32_t NumNodes) |
1054 | : Alloc(Alloc), NumNodes(NumNodes) { |
1055 | Entries = new (Alloc, 0) MapEntry[NumNodes]; |
1056 | for (int I = 0; I < D.NumCalls; ++I) { |
1057 | DEBUG(reportNumber("Registering call in node " , D.Calls[I].FromNode, 10)); |
1058 | ++Entries[D.Calls[I].FromNode].NumCalls; |
1059 | } |
1060 | for (int I = 0; I < NumNodes; ++I) { |
1061 | Entries[I].Calls = Entries[I].NumCalls ? new (Alloc) |
1062 | uint32_t[Entries[I].NumCalls] |
1063 | : nullptr; |
1064 | Entries[I].NumCalls = 0; |
1065 | } |
1066 | for (int I = 0; I < D.NumCalls; ++I) { |
1067 | MapEntry &Entry = Entries[D.Calls[I].FromNode]; |
1068 | Entry.Calls[Entry.NumCalls++] = I; |
1069 | } |
1070 | } |
1071 | |
1072 | /// Set the frequency of all calls in node \p NodeID to Freq. However, if |
1073 | /// the calls have their own counters and do not depend on the basic block |
1074 | /// counter, this means they have landing pads and throw exceptions. In this |
1075 | /// case, set their frequency with their counters and return the maximum |
1076 | /// value observed in such counters. This will be used as the new frequency |
1077 | /// at basic block entry. This is used to fix the CFG edge frequencies in the |
1078 | /// presence of exceptions. |
1079 | uint64_t visitAllCallsIn(uint32_t NodeID, uint64_t Freq, uint64_t *CallFreqs, |
1080 | const FunctionDescription &D, |
1081 | const uint64_t *Counters, |
1082 | ProfileWriterContext &Ctx) const { |
1083 | const MapEntry &Entry = Entries[NodeID]; |
1084 | uint64_t MaxValue = 0ull; |
1085 | for (int I = 0, E = Entry.NumCalls; I != E; ++I) { |
1086 | const uint32_t CallID = Entry.Calls[I]; |
1087 | DEBUG(reportNumber(" Setting freq for call ID: " , CallID, 10)); |
1088 | const CallDescription &CallDesc = D.Calls[CallID]; |
1089 | if (CallDesc.Counter == 0xffffffff) { |
1090 | CallFreqs[CallID] = Freq; |
1091 | DEBUG(reportNumber(" with : " , Freq, 10)); |
1092 | } else { |
1093 | const uint64_t CounterVal = Counters[CallDesc.Counter]; |
1094 | CallFreqs[CallID] = CounterVal; |
1095 | MaxValue = CounterVal > MaxValue ? CounterVal : MaxValue; |
1096 | DEBUG(reportNumber(" with (private counter) : " , CounterVal, 10)); |
1097 | } |
1098 | DEBUG(reportNumber(" Address: 0x" , CallDesc.TargetAddress, 16)); |
1099 | if (CallFreqs[CallID] > 0) |
1100 | Ctx.CallFlowTable->get(Key: CallDesc.TargetAddress).Calls += |
1101 | CallFreqs[CallID]; |
1102 | } |
1103 | return MaxValue; |
1104 | } |
1105 | |
1106 | ~NodeToCallsMap() { |
1107 | for (int I = NumNodes - 1; I >= 0; --I) |
1108 | if (Entries[I].Calls) |
1109 | Alloc.deallocate(Ptr: Entries[I].Calls); |
1110 | Alloc.deallocate(Ptr: Entries); |
1111 | } |
1112 | }; |
1113 | |
1114 | /// Fill an array with the frequency of each edge in the function represented |
1115 | /// by G, as well as another array for each call. |
1116 | void Graph::computeEdgeFrequencies(const uint64_t *Counters, |
1117 | ProfileWriterContext &Ctx) { |
1118 | if (NumNodes == 0) |
1119 | return; |
1120 | |
1121 | EdgeFreqs = D.NumEdges ? new (Alloc, 0) uint64_t [D.NumEdges] : nullptr; |
1122 | CallFreqs = D.NumCalls ? new (Alloc, 0) uint64_t [D.NumCalls] : nullptr; |
1123 | |
1124 | // Setup a lookup for calls present in each node (BB) |
1125 | NodeToCallsMap *CallMap = new (Alloc) NodeToCallsMap(Alloc, D, NumNodes); |
1126 | |
1127 | // Perform a bottom-up, BFS traversal of the spanning tree in G. Edges in the |
1128 | // spanning tree don't have explicit counters. We must infer their value using |
1129 | // a linear combination of other counters (sum of counters of the outgoing |
1130 | // edges minus sum of counters of the incoming edges). |
1131 | uint32_t *Stack = new (Alloc) uint32_t [NumNodes]; |
1132 | uint32_t StackTop = 0; |
1133 | enum Status : uint8_t { S_NEW = 0, S_VISITING, S_VISITED }; |
1134 | Status *Visited = new (Alloc, 0) Status[NumNodes]; |
1135 | uint64_t *LeafFrequency = new (Alloc, 0) uint64_t[NumNodes]; |
1136 | uint64_t *EntryAddress = new (Alloc, 0) uint64_t[NumNodes]; |
1137 | |
1138 | // Setup a fast lookup for frequency of leaf nodes, which have special |
1139 | // basic block frequency instrumentation (they are not edge profiled). |
1140 | for (int I = 0; I < D.NumLeafNodes; ++I) { |
1141 | LeafFrequency[D.LeafNodes[I].Node] = Counters[D.LeafNodes[I].Counter]; |
1142 | DEBUG({ |
1143 | if (Counters[D.LeafNodes[I].Counter] > 0) { |
1144 | reportNumber("Leaf Node# " , D.LeafNodes[I].Node, 10); |
1145 | reportNumber(" Counter: " , Counters[D.LeafNodes[I].Counter], 10); |
1146 | } |
1147 | }); |
1148 | } |
1149 | for (int I = 0; I < D.NumEntryNodes; ++I) { |
1150 | EntryAddress[D.EntryNodes[I].Node] = D.EntryNodes[I].Address; |
1151 | DEBUG({ |
1152 | reportNumber("Entry Node# " , D.EntryNodes[I].Node, 10); |
1153 | reportNumber(" Address: " , D.EntryNodes[I].Address, 16); |
1154 | }); |
1155 | } |
1156 | // Add all root nodes to the stack |
1157 | for (int I = 0; I < NumNodes; ++I) |
1158 | if (SpanningTreeNodes[I].NumInEdges == 0) |
1159 | Stack[StackTop++] = I; |
1160 | |
1161 | // Empty stack? |
1162 | if (StackTop == 0) { |
1163 | DEBUG(report("Empty stack!\n" )); |
1164 | Alloc.deallocate(Ptr: EntryAddress); |
1165 | Alloc.deallocate(Ptr: LeafFrequency); |
1166 | Alloc.deallocate(Ptr: Visited); |
1167 | Alloc.deallocate(Ptr: Stack); |
1168 | CallMap->~NodeToCallsMap(); |
1169 | Alloc.deallocate(Ptr: CallMap); |
1170 | if (CallFreqs) |
1171 | Alloc.deallocate(Ptr: CallFreqs); |
1172 | if (EdgeFreqs) |
1173 | Alloc.deallocate(Ptr: EdgeFreqs); |
1174 | EdgeFreqs = nullptr; |
1175 | CallFreqs = nullptr; |
1176 | return; |
1177 | } |
1178 | // Add all known edge counts, will infer the rest |
1179 | for (int I = 0; I < D.NumEdges; ++I) { |
1180 | const uint32_t C = D.Edges[I].Counter; |
1181 | if (C == 0xffffffff) // inferred counter - we will compute its value |
1182 | continue; |
1183 | EdgeFreqs[I] = Counters[C]; |
1184 | } |
1185 | |
1186 | while (StackTop > 0) { |
1187 | const uint32_t Cur = Stack[--StackTop]; |
1188 | DEBUG({ |
1189 | if (Visited[Cur] == S_VISITING) |
1190 | report("(visiting) " ); |
1191 | else |
1192 | report("(new) " ); |
1193 | reportNumber("Cur: " , Cur, 10); |
1194 | }); |
1195 | |
1196 | // This shouldn't happen in a tree |
1197 | assert(Assertion: Visited[Cur] != S_VISITED, Msg: "should not have visited nodes in stack" ); |
1198 | if (Visited[Cur] == S_NEW) { |
1199 | Visited[Cur] = S_VISITING; |
1200 | Stack[StackTop++] = Cur; |
1201 | assert(Assertion: StackTop <= NumNodes, Msg: "stack grew too large" ); |
1202 | for (int I = 0, E = SpanningTreeNodes[Cur].NumOutEdges; I < E; ++I) { |
1203 | const uint32_t Succ = SpanningTreeNodes[Cur].OutEdges[I].Node; |
1204 | Stack[StackTop++] = Succ; |
1205 | assert(Assertion: StackTop <= NumNodes, Msg: "stack grew too large" ); |
1206 | } |
1207 | continue; |
1208 | } |
1209 | Visited[Cur] = S_VISITED; |
1210 | |
1211 | // Establish our node frequency based on outgoing edges, which should all be |
1212 | // resolved by now. |
1213 | int64_t CurNodeFreq = LeafFrequency[Cur]; |
1214 | // Not a leaf? |
1215 | if (!CurNodeFreq) { |
1216 | for (int I = 0, E = CFGNodes[Cur].NumOutEdges; I != E; ++I) { |
1217 | const uint32_t SuccEdge = CFGNodes[Cur].OutEdges[I].ID; |
1218 | CurNodeFreq += EdgeFreqs[SuccEdge]; |
1219 | } |
1220 | } |
1221 | if (CurNodeFreq < 0) |
1222 | CurNodeFreq = 0; |
1223 | |
1224 | const uint64_t CallFreq = CallMap->visitAllCallsIn( |
1225 | NodeID: Cur, Freq: CurNodeFreq > 0 ? CurNodeFreq : 0, CallFreqs, D, Counters, Ctx); |
1226 | |
1227 | // Exception handling affected our output flow? Fix with calls info |
1228 | DEBUG({ |
1229 | if (CallFreq > CurNodeFreq) |
1230 | report("Bumping node frequency with call info\n" ); |
1231 | }); |
1232 | CurNodeFreq = CallFreq > CurNodeFreq ? CallFreq : CurNodeFreq; |
1233 | |
1234 | if (CurNodeFreq > 0) { |
1235 | if (uint64_t Addr = EntryAddress[Cur]) { |
1236 | DEBUG( |
1237 | reportNumber(" Setting flow at entry point address 0x" , Addr, 16)); |
1238 | DEBUG(reportNumber(" with: " , CurNodeFreq, 10)); |
1239 | Ctx.CallFlowTable->get(Key: Addr).Val = CurNodeFreq; |
1240 | } |
1241 | } |
1242 | |
1243 | // No parent? Reached a tree root, limit to call frequency updating. |
1244 | if (SpanningTreeNodes[Cur].NumInEdges == 0) |
1245 | continue; |
1246 | |
1247 | assert(Assertion: SpanningTreeNodes[Cur].NumInEdges == 1, Msg: "must have 1 parent" ); |
1248 | const uint32_t Parent = SpanningTreeNodes[Cur].InEdges[0].Node; |
1249 | const uint32_t ParentEdge = SpanningTreeNodes[Cur].InEdges[0].ID; |
1250 | |
1251 | // Calculate parent edge freq. |
1252 | int64_t ParentEdgeFreq = CurNodeFreq; |
1253 | for (int I = 0, E = CFGNodes[Cur].NumInEdges; I != E; ++I) { |
1254 | const uint32_t PredEdge = CFGNodes[Cur].InEdges[I].ID; |
1255 | ParentEdgeFreq -= EdgeFreqs[PredEdge]; |
1256 | } |
1257 | |
1258 | // Sometimes the conservative CFG that BOLT builds will lead to incorrect |
1259 | // flow computation. For example, in a BB that transitively calls the exit |
1260 | // syscall, BOLT will add a fall-through successor even though it should not |
1261 | // have any successors. So this block execution will likely be wrong. We |
1262 | // tolerate this imperfection since this case should be quite infrequent. |
1263 | if (ParentEdgeFreq < 0) { |
1264 | DEBUG(dumpEdgeFreqs()); |
1265 | DEBUG(report("WARNING: incorrect flow" )); |
1266 | ParentEdgeFreq = 0; |
1267 | } |
1268 | DEBUG(reportNumber(" Setting freq for ParentEdge: " , ParentEdge, 10)); |
1269 | DEBUG(reportNumber(" with ParentEdgeFreq: " , ParentEdgeFreq, 10)); |
1270 | EdgeFreqs[ParentEdge] = ParentEdgeFreq; |
1271 | } |
1272 | |
1273 | Alloc.deallocate(Ptr: EntryAddress); |
1274 | Alloc.deallocate(Ptr: LeafFrequency); |
1275 | Alloc.deallocate(Ptr: Visited); |
1276 | Alloc.deallocate(Ptr: Stack); |
1277 | CallMap->~NodeToCallsMap(); |
1278 | Alloc.deallocate(Ptr: CallMap); |
1279 | DEBUG(dumpEdgeFreqs()); |
1280 | } |
1281 | |
1282 | /// Write to \p FD all of the edge profiles for function \p FuncDesc. Uses |
1283 | /// \p Alloc to allocate helper dynamic structures used to compute profile for |
1284 | /// edges that we do not explicitly instrument. |
1285 | const uint8_t *writeFunctionProfile(int FD, ProfileWriterContext &Ctx, |
1286 | const uint8_t *FuncDesc, |
1287 | BumpPtrAllocator &Alloc) { |
1288 | const FunctionDescription F(FuncDesc); |
1289 | const uint8_t *next = FuncDesc + F.getSize(); |
1290 | |
1291 | #if !defined(__APPLE__) |
1292 | uint64_t *bolt_instr_locations = __bolt_instr_locations; |
1293 | #else |
1294 | uint64_t *bolt_instr_locations = _bolt_instr_locations_getter(); |
1295 | #endif |
1296 | |
1297 | // Skip funcs we know are cold |
1298 | #ifndef ENABLE_DEBUG |
1299 | uint64_t CountersFreq = 0; |
1300 | for (int I = 0; I < F.NumLeafNodes; ++I) |
1301 | CountersFreq += bolt_instr_locations[F.LeafNodes[I].Counter]; |
1302 | |
1303 | if (CountersFreq == 0) { |
1304 | for (int I = 0; I < F.NumEdges; ++I) { |
1305 | const uint32_t C = F.Edges[I].Counter; |
1306 | if (C == 0xffffffff) |
1307 | continue; |
1308 | CountersFreq += bolt_instr_locations[C]; |
1309 | } |
1310 | if (CountersFreq == 0) { |
1311 | for (int I = 0; I < F.NumCalls; ++I) { |
1312 | const uint32_t C = F.Calls[I].Counter; |
1313 | if (C == 0xffffffff) |
1314 | continue; |
1315 | CountersFreq += bolt_instr_locations[C]; |
1316 | } |
1317 | if (CountersFreq == 0) |
1318 | return next; |
1319 | } |
1320 | } |
1321 | #endif |
1322 | |
1323 | Graph *G = new (Alloc) Graph(Alloc, F, bolt_instr_locations, Ctx); |
1324 | DEBUG(G->dump()); |
1325 | |
1326 | if (!G->EdgeFreqs && !G->CallFreqs) { |
1327 | G->~Graph(); |
1328 | Alloc.deallocate(Ptr: G); |
1329 | return next; |
1330 | } |
1331 | |
1332 | for (int I = 0; I < F.NumEdges; ++I) { |
1333 | const uint64_t Freq = G->EdgeFreqs[I]; |
1334 | if (Freq == 0) |
1335 | continue; |
1336 | const EdgeDescription *Desc = &F.Edges[I]; |
1337 | char LineBuf[BufSize]; |
1338 | char *Ptr = LineBuf; |
1339 | Ptr = serializeLoc(Ctx, OutBuf: Ptr, Loc: Desc->From, BufSize); |
1340 | Ptr = serializeLoc(Ctx, OutBuf: Ptr, Loc: Desc->To, BufSize: BufSize - (Ptr - LineBuf)); |
1341 | Ptr = strCopy(OutBuf: Ptr, Str: "0 " , Size: BufSize - (Ptr - LineBuf) - 22); |
1342 | Ptr = intToStr(OutBuf: Ptr, Num: Freq, Base: 10); |
1343 | *Ptr++ = '\n'; |
1344 | __write(fd: FD, buf: LineBuf, count: Ptr - LineBuf); |
1345 | } |
1346 | |
1347 | for (int I = 0; I < F.NumCalls; ++I) { |
1348 | const uint64_t Freq = G->CallFreqs[I]; |
1349 | if (Freq == 0) |
1350 | continue; |
1351 | char LineBuf[BufSize]; |
1352 | char *Ptr = LineBuf; |
1353 | const CallDescription *Desc = &F.Calls[I]; |
1354 | Ptr = serializeLoc(Ctx, OutBuf: Ptr, Loc: Desc->From, BufSize); |
1355 | Ptr = serializeLoc(Ctx, OutBuf: Ptr, Loc: Desc->To, BufSize: BufSize - (Ptr - LineBuf)); |
1356 | Ptr = strCopy(OutBuf: Ptr, Str: "0 " , Size: BufSize - (Ptr - LineBuf) - 25); |
1357 | Ptr = intToStr(OutBuf: Ptr, Num: Freq, Base: 10); |
1358 | *Ptr++ = '\n'; |
1359 | __write(fd: FD, buf: LineBuf, count: Ptr - LineBuf); |
1360 | } |
1361 | |
1362 | G->~Graph(); |
1363 | Alloc.deallocate(Ptr: G); |
1364 | return next; |
1365 | } |
1366 | |
1367 | #if !defined(__APPLE__) |
1368 | const IndCallTargetDescription * |
1369 | ProfileWriterContext::lookupIndCallTarget(uint64_t Target) const { |
1370 | uint32_t B = 0; |
1371 | uint32_t E = __bolt_instr_num_ind_targets; |
1372 | if (E == 0) |
1373 | return nullptr; |
1374 | do { |
1375 | uint32_t I = (E - B) / 2 + B; |
1376 | if (IndCallTargets[I].Address == Target) |
1377 | return &IndCallTargets[I]; |
1378 | if (IndCallTargets[I].Address < Target) |
1379 | B = I + 1; |
1380 | else |
1381 | E = I; |
1382 | } while (B < E); |
1383 | return nullptr; |
1384 | } |
1385 | |
1386 | /// Write a single indirect call <src, target> pair to the fdata file |
1387 | void visitIndCallCounter(IndirectCallHashTable::MapEntry &Entry, |
1388 | int FD, int CallsiteID, |
1389 | ProfileWriterContext *Ctx) { |
1390 | if (Entry.Val == 0) |
1391 | return; |
1392 | DEBUG(reportNumber("Target func 0x" , Entry.Key, 16)); |
1393 | DEBUG(reportNumber("Target freq: " , Entry.Val, 10)); |
1394 | const IndCallDescription *CallsiteDesc = |
1395 | &Ctx->IndCallDescriptions[CallsiteID]; |
1396 | const IndCallTargetDescription *TargetDesc = |
1397 | Ctx->lookupIndCallTarget(Target: Entry.Key - TextBaseAddress); |
1398 | if (!TargetDesc) { |
1399 | DEBUG(report("Failed to lookup indirect call target\n" )); |
1400 | char LineBuf[BufSize]; |
1401 | char *Ptr = LineBuf; |
1402 | Ptr = serializeLoc(Ctx: *Ctx, OutBuf: Ptr, Loc: *CallsiteDesc, BufSize); |
1403 | Ptr = strCopy(OutBuf: Ptr, Str: "0 [unknown] 0 0 " , Size: BufSize - (Ptr - LineBuf) - 40); |
1404 | Ptr = intToStr(OutBuf: Ptr, Num: Entry.Val, Base: 10); |
1405 | *Ptr++ = '\n'; |
1406 | __write(fd: FD, buf: LineBuf, count: Ptr - LineBuf); |
1407 | return; |
1408 | } |
1409 | Ctx->CallFlowTable->get(Key: TargetDesc->Address).Calls += Entry.Val; |
1410 | char LineBuf[BufSize]; |
1411 | char *Ptr = LineBuf; |
1412 | Ptr = serializeLoc(Ctx: *Ctx, OutBuf: Ptr, Loc: *CallsiteDesc, BufSize); |
1413 | Ptr = serializeLoc(Ctx: *Ctx, OutBuf: Ptr, Loc: TargetDesc->Loc, BufSize: BufSize - (Ptr - LineBuf)); |
1414 | Ptr = strCopy(OutBuf: Ptr, Str: "0 " , Size: BufSize - (Ptr - LineBuf) - 25); |
1415 | Ptr = intToStr(OutBuf: Ptr, Num: Entry.Val, Base: 10); |
1416 | *Ptr++ = '\n'; |
1417 | __write(fd: FD, buf: LineBuf, count: Ptr - LineBuf); |
1418 | } |
1419 | |
1420 | /// Write to \p FD all of the indirect call profiles. |
1421 | void writeIndirectCallProfile(int FD, ProfileWriterContext &Ctx) { |
1422 | for (int I = 0; I < __bolt_instr_num_ind_calls; ++I) { |
1423 | DEBUG(reportNumber("IndCallsite #" , I, 10)); |
1424 | GlobalIndCallCounters[I].forEachElement(Callback: visitIndCallCounter, args: FD, args: I, args: &Ctx); |
1425 | } |
1426 | } |
1427 | |
1428 | /// Check a single call flow for a callee versus all known callers. If there are |
1429 | /// less callers than what the callee expects, write the difference with source |
1430 | /// [unknown] in the profile. |
1431 | void visitCallFlowEntry(CallFlowHashTable::MapEntry &Entry, int FD, |
1432 | ProfileWriterContext *Ctx) { |
1433 | DEBUG(reportNumber("Call flow entry address: 0x" , Entry.Key, 16)); |
1434 | DEBUG(reportNumber("Calls: " , Entry.Calls, 10)); |
1435 | DEBUG(reportNumber("Reported entry frequency: " , Entry.Val, 10)); |
1436 | DEBUG({ |
1437 | if (Entry.Calls > Entry.Val) |
1438 | report(" More calls than expected!\n" ); |
1439 | }); |
1440 | if (Entry.Val <= Entry.Calls) |
1441 | return; |
1442 | DEBUG(reportNumber( |
1443 | " Balancing calls with traffic: " , Entry.Val - Entry.Calls, 10)); |
1444 | const IndCallTargetDescription *TargetDesc = |
1445 | Ctx->lookupIndCallTarget(Target: Entry.Key); |
1446 | if (!TargetDesc) { |
1447 | // There is probably something wrong with this callee and this should be |
1448 | // investigated, but I don't want to assert and lose all data collected. |
1449 | DEBUG(report("WARNING: failed to look up call target!\n" )); |
1450 | return; |
1451 | } |
1452 | char LineBuf[BufSize]; |
1453 | char *Ptr = LineBuf; |
1454 | Ptr = strCopy(OutBuf: Ptr, Str: "0 [unknown] 0 " , Size: BufSize); |
1455 | Ptr = serializeLoc(Ctx: *Ctx, OutBuf: Ptr, Loc: TargetDesc->Loc, BufSize: BufSize - (Ptr - LineBuf)); |
1456 | Ptr = strCopy(OutBuf: Ptr, Str: "0 " , Size: BufSize - (Ptr - LineBuf) - 25); |
1457 | Ptr = intToStr(OutBuf: Ptr, Num: Entry.Val - Entry.Calls, Base: 10); |
1458 | *Ptr++ = '\n'; |
1459 | __write(fd: FD, buf: LineBuf, count: Ptr - LineBuf); |
1460 | } |
1461 | |
1462 | /// Open fdata file for writing and return a valid file descriptor, aborting |
1463 | /// program upon failure. |
1464 | int openProfile() { |
1465 | // Build the profile name string by appending our PID |
1466 | char Buf[BufSize]; |
1467 | char *Ptr = Buf; |
1468 | uint64_t PID = __getpid(); |
1469 | Ptr = strCopy(OutBuf: Buf, Str: __bolt_instr_filename, Size: BufSize); |
1470 | if (__bolt_instr_use_pid) { |
1471 | Ptr = strCopy(OutBuf: Ptr, Str: "." , Size: BufSize - (Ptr - Buf + 1)); |
1472 | Ptr = intToStr(OutBuf: Ptr, Num: PID, Base: 10); |
1473 | Ptr = strCopy(OutBuf: Ptr, Str: ".fdata" , Size: BufSize - (Ptr - Buf + 1)); |
1474 | } |
1475 | *Ptr++ = '\0'; |
1476 | uint64_t FD = __open(pathname: Buf, O_WRONLY | O_TRUNC | O_CREAT, |
1477 | /*mode=*/0666); |
1478 | if (static_cast<int64_t>(FD) < 0) { |
1479 | report(Msg: "Error while trying to open profile file for writing: " ); |
1480 | report(Msg: Buf); |
1481 | reportNumber(Msg: "\nFailed with error number: 0x" , |
1482 | Num: 0 - static_cast<int64_t>(FD), Base: 16); |
1483 | __exit(code: 1); |
1484 | } |
1485 | return FD; |
1486 | } |
1487 | |
1488 | #endif |
1489 | |
1490 | } // anonymous namespace |
1491 | |
1492 | #if !defined(__APPLE__) |
1493 | |
1494 | /// Reset all counters in case you want to start profiling a new phase of your |
1495 | /// program independently of prior phases. |
1496 | /// The address of this function is printed by BOLT and this can be called by |
1497 | /// any attached debugger during runtime. There is a useful oneliner for gdb: |
1498 | /// |
1499 | /// gdb -p $(pgrep -xo PROCESSNAME) -ex 'p ((void(*)())0xdeadbeef)()' \ |
1500 | /// -ex 'set confirm off' -ex quit |
1501 | /// |
1502 | /// Where 0xdeadbeef is this function address and PROCESSNAME your binary file |
1503 | /// name. |
1504 | extern "C" void __bolt_instr_clear_counters() { |
1505 | memset(Buf: reinterpret_cast<char *>(__bolt_instr_locations), C: 0, |
1506 | Size: __bolt_num_counters * 8); |
1507 | for (int I = 0; I < __bolt_instr_num_ind_calls; ++I) |
1508 | GlobalIndCallCounters[I].resetCounters(); |
1509 | } |
1510 | |
1511 | /// This is the entry point for profile writing. |
1512 | /// There are three ways of getting here: |
1513 | /// |
1514 | /// * Program execution ended, finalization methods are running and BOLT |
1515 | /// hooked into FINI from your binary dynamic section; |
1516 | /// * You used the sleep timer option and during initialization we forked |
1517 | /// a separate process that will call this function periodically; |
1518 | /// * BOLT prints this function address so you can attach a debugger and |
1519 | /// call this function directly to get your profile written to disk |
1520 | /// on demand. |
1521 | /// |
1522 | extern "C" void __attribute((force_align_arg_pointer)) |
1523 | __bolt_instr_data_dump(int FD) { |
1524 | // Already dumping |
1525 | if (!GlobalWriteProfileMutex->acquire()) |
1526 | return; |
1527 | |
1528 | int ret = __lseek(fd: FD, pos: 0, SEEK_SET); |
1529 | assert(Assertion: ret == 0, Msg: "Failed to lseek!" ); |
1530 | ret = __ftruncate(fd: FD, length: 0); |
1531 | assert(Assertion: ret == 0, Msg: "Failed to ftruncate!" ); |
1532 | BumpPtrAllocator HashAlloc; |
1533 | HashAlloc.setMaxSize(0x6400000); |
1534 | ProfileWriterContext Ctx = readDescriptions(); |
1535 | Ctx.CallFlowTable = new (HashAlloc, 0) CallFlowHashTable(HashAlloc); |
1536 | |
1537 | DEBUG(printStats(Ctx)); |
1538 | |
1539 | BumpPtrAllocator Alloc; |
1540 | Alloc.setMaxSize(0x6400000); |
1541 | const uint8_t *FuncDesc = Ctx.FuncDescriptions; |
1542 | for (int I = 0, E = __bolt_instr_num_funcs; I < E; ++I) { |
1543 | FuncDesc = writeFunctionProfile(FD, Ctx, FuncDesc, Alloc); |
1544 | Alloc.clear(); |
1545 | DEBUG(reportNumber("FuncDesc now: " , (uint64_t)FuncDesc, 16)); |
1546 | } |
1547 | assert(Assertion: FuncDesc == (void *)Ctx.Strings, |
1548 | Msg: "FuncDesc ptr must be equal to stringtable" ); |
1549 | |
1550 | writeIndirectCallProfile(FD, Ctx); |
1551 | Ctx.CallFlowTable->forEachElement(Callback: visitCallFlowEntry, args: FD, args: &Ctx); |
1552 | |
1553 | __fsync(fd: FD); |
1554 | __munmap(addr: Ctx.MMapPtr, size: Ctx.MMapSize); |
1555 | __close(fd: Ctx.FileDesc); |
1556 | HashAlloc.destroy(); |
1557 | GlobalWriteProfileMutex->release(); |
1558 | DEBUG(report("Finished writing profile.\n" )); |
1559 | } |
1560 | |
1561 | /// Event loop for our child process spawned during setup to dump profile data |
1562 | /// at user-specified intervals |
1563 | void watchProcess() { |
1564 | timespec ts, rem; |
1565 | uint64_t Ellapsed = 0ull; |
1566 | int FD = openProfile(); |
1567 | uint64_t ppid; |
1568 | if (__bolt_instr_wait_forks) { |
1569 | // Store parent pgid |
1570 | ppid = -__getpgid(pid: 0); |
1571 | // And leave parent process group |
1572 | __setpgid(pid: 0, pgid: 0); |
1573 | } else { |
1574 | // Store parent pid |
1575 | ppid = __getppid(); |
1576 | if (ppid == 1) { |
1577 | // Parent already dead |
1578 | __bolt_instr_data_dump(FD); |
1579 | goto out; |
1580 | } |
1581 | } |
1582 | |
1583 | ts.tv_sec = 1; |
1584 | ts.tv_nsec = 0; |
1585 | while (1) { |
1586 | __nanosleep(req: &ts, rem: &rem); |
1587 | // This means our parent process or all its forks are dead, |
1588 | // so no need for us to keep dumping. |
1589 | if (__kill(pid: ppid, sig: 0) < 0) { |
1590 | if (__bolt_instr_no_counters_clear) |
1591 | __bolt_instr_data_dump(FD); |
1592 | break; |
1593 | } |
1594 | |
1595 | if (++Ellapsed < __bolt_instr_sleep_time) |
1596 | continue; |
1597 | |
1598 | Ellapsed = 0; |
1599 | __bolt_instr_data_dump(FD); |
1600 | if (__bolt_instr_no_counters_clear == false) |
1601 | __bolt_instr_clear_counters(); |
1602 | } |
1603 | |
1604 | out:; |
1605 | DEBUG(report("My parent process is dead, bye!\n" )); |
1606 | __close(fd: FD); |
1607 | __exit(code: 0); |
1608 | } |
1609 | |
1610 | extern "C" void __bolt_instr_indirect_call(); |
1611 | extern "C" void __bolt_instr_indirect_tailcall(); |
1612 | |
1613 | /// Initialization code |
1614 | extern "C" void __attribute((force_align_arg_pointer)) __bolt_instr_setup() { |
1615 | __bolt_ind_call_counter_func_pointer = __bolt_instr_indirect_call; |
1616 | __bolt_ind_tailcall_counter_func_pointer = __bolt_instr_indirect_tailcall; |
1617 | TextBaseAddress = getTextBaseAddress(); |
1618 | |
1619 | const uint64_t = |
1620 | reinterpret_cast<uint64_t>(&__bolt_instr_locations[0]); |
1621 | const uint64_t CountersEnd = alignTo( |
1622 | Value: reinterpret_cast<uint64_t>(&__bolt_instr_locations[__bolt_num_counters]), |
1623 | Align: 0x1000); |
1624 | DEBUG(reportNumber("replace mmap start: " , CountersStart, 16)); |
1625 | DEBUG(reportNumber("replace mmap stop: " , CountersEnd, 16)); |
1626 | assert(Assertion: CountersEnd > CountersStart, Msg: "no counters" ); |
1627 | |
1628 | const bool Shared = !__bolt_instr_use_pid; |
1629 | const uint64_t MapPrivateOrShared = Shared ? MAP_SHARED : MAP_PRIVATE; |
1630 | |
1631 | void *Ret = |
1632 | __mmap(addr: CountersStart, size: CountersEnd - CountersStart, PROT_READ | PROT_WRITE, |
1633 | MAP_ANONYMOUS | MapPrivateOrShared | MAP_FIXED, fd: -1, offset: 0); |
1634 | assert(Assertion: Ret != MAP_FAILED, Msg: "__bolt_instr_setup: Failed to mmap counters!" ); |
1635 | |
1636 | GlobalMetadataStorage = __mmap(addr: 0, size: 4096, PROT_READ | PROT_WRITE, |
1637 | flags: MapPrivateOrShared | MAP_ANONYMOUS, fd: -1, offset: 0); |
1638 | assert(Assertion: GlobalMetadataStorage != MAP_FAILED, |
1639 | Msg: "__bolt_instr_setup: failed to mmap page for metadata!" ); |
1640 | |
1641 | GlobalAlloc = new (GlobalMetadataStorage) BumpPtrAllocator; |
1642 | // Conservatively reserve 100MiB |
1643 | GlobalAlloc->setMaxSize(0x6400000); |
1644 | GlobalAlloc->setShared(Shared); |
1645 | GlobalWriteProfileMutex = new (*GlobalAlloc, 0) Mutex(); |
1646 | if (__bolt_instr_num_ind_calls > 0) |
1647 | GlobalIndCallCounters = |
1648 | new (*GlobalAlloc, 0) IndirectCallHashTable[__bolt_instr_num_ind_calls]; |
1649 | |
1650 | if (__bolt_instr_sleep_time != 0) { |
1651 | // Separate instrumented process to the own process group |
1652 | if (__bolt_instr_wait_forks) |
1653 | __setpgid(pid: 0, pgid: 0); |
1654 | |
1655 | if (long PID = __fork()) |
1656 | return; |
1657 | watchProcess(); |
1658 | } |
1659 | } |
1660 | |
1661 | extern "C" __attribute((force_align_arg_pointer)) void |
1662 | instrumentIndirectCall(uint64_t Target, uint64_t IndCallID) { |
1663 | GlobalIndCallCounters[IndCallID].incrementVal(Key: Target, Alloc&: *GlobalAlloc); |
1664 | } |
1665 | |
1666 | /// We receive as in-stack arguments the identifier of the indirect call site |
1667 | /// as well as the target address for the call |
1668 | extern "C" __attribute((naked)) void __bolt_instr_indirect_call() |
1669 | { |
1670 | #if defined(__aarch64__) |
1671 | // clang-format off |
1672 | __asm__ __volatile__(SAVE_ALL |
1673 | "ldp x0, x1, [sp, #288]\n" |
1674 | "bl instrumentIndirectCall\n" |
1675 | RESTORE_ALL |
1676 | "ret\n" |
1677 | :::); |
1678 | // clang-format on |
1679 | #else |
1680 | // clang-format off |
1681 | __asm__ __volatile__(SAVE_ALL |
1682 | "mov 0xa0(%%rsp), %%rdi\n" |
1683 | "mov 0x98(%%rsp), %%rsi\n" |
1684 | "call instrumentIndirectCall\n" |
1685 | RESTORE_ALL |
1686 | "ret\n" |
1687 | :::); |
1688 | // clang-format on |
1689 | #endif |
1690 | } |
1691 | |
1692 | extern "C" __attribute((naked)) void __bolt_instr_indirect_tailcall() |
1693 | { |
1694 | #if defined(__aarch64__) |
1695 | // clang-format off |
1696 | __asm__ __volatile__(SAVE_ALL |
1697 | "ldp x0, x1, [sp, #288]\n" |
1698 | "bl instrumentIndirectCall\n" |
1699 | RESTORE_ALL |
1700 | "ret\n" |
1701 | :::); |
1702 | // clang-format on |
1703 | #else |
1704 | // clang-format off |
1705 | __asm__ __volatile__(SAVE_ALL |
1706 | "mov 0x98(%%rsp), %%rdi\n" |
1707 | "mov 0x90(%%rsp), %%rsi\n" |
1708 | "call instrumentIndirectCall\n" |
1709 | RESTORE_ALL |
1710 | "ret\n" |
1711 | :::); |
1712 | // clang-format on |
1713 | #endif |
1714 | } |
1715 | |
1716 | /// This is hooking ELF's entry, it needs to save all machine state. |
1717 | extern "C" __attribute((naked)) void __bolt_instr_start() |
1718 | { |
1719 | #if defined(__aarch64__) |
1720 | // clang-format off |
1721 | __asm__ __volatile__(SAVE_ALL |
1722 | "bl __bolt_instr_setup\n" |
1723 | RESTORE_ALL |
1724 | "adrp x16, __bolt_start_trampoline\n" |
1725 | "add x16, x16, #:lo12:__bolt_start_trampoline\n" |
1726 | "br x16\n" |
1727 | :::); |
1728 | // clang-format on |
1729 | #else |
1730 | // clang-format off |
1731 | __asm__ __volatile__(SAVE_ALL |
1732 | "call __bolt_instr_setup\n" |
1733 | RESTORE_ALL |
1734 | "jmp __bolt_start_trampoline\n" |
1735 | :::); |
1736 | // clang-format on |
1737 | #endif |
1738 | } |
1739 | |
1740 | /// This is hooking into ELF's DT_FINI |
1741 | extern "C" void __bolt_instr_fini() { |
1742 | #if defined(__aarch64__) |
1743 | // clang-format off |
1744 | __asm__ __volatile__(SAVE_ALL |
1745 | "adrp x16, __bolt_fini_trampoline\n" |
1746 | "add x16, x16, #:lo12:__bolt_fini_trampoline\n" |
1747 | "blr x16\n" |
1748 | RESTORE_ALL |
1749 | :::); |
1750 | // clang-format on |
1751 | #else |
1752 | __asm__ __volatile__("call __bolt_fini_trampoline\n" :::); |
1753 | #endif |
1754 | if (__bolt_instr_sleep_time == 0) { |
1755 | int FD = openProfile(); |
1756 | __bolt_instr_data_dump(FD); |
1757 | __close(fd: FD); |
1758 | } |
1759 | DEBUG(report("Finished.\n" )); |
1760 | } |
1761 | |
1762 | #endif |
1763 | |
1764 | #if defined(__APPLE__) |
1765 | |
1766 | extern "C" void __bolt_instr_data_dump() { |
1767 | ProfileWriterContext Ctx = readDescriptions(); |
1768 | |
1769 | int FD = 2; |
1770 | BumpPtrAllocator Alloc; |
1771 | const uint8_t *FuncDesc = Ctx.FuncDescriptions; |
1772 | uint32_t bolt_instr_num_funcs = _bolt_instr_num_funcs_getter(); |
1773 | |
1774 | for (int I = 0, E = bolt_instr_num_funcs; I < E; ++I) { |
1775 | FuncDesc = writeFunctionProfile(FD, Ctx, FuncDesc, Alloc); |
1776 | Alloc.clear(); |
1777 | DEBUG(reportNumber("FuncDesc now: " , (uint64_t)FuncDesc, 16)); |
1778 | } |
1779 | assert(FuncDesc == (void *)Ctx.Strings, |
1780 | "FuncDesc ptr must be equal to stringtable" ); |
1781 | } |
1782 | |
1783 | // On OSX/iOS the final symbol name of an extern "C" function/variable contains |
1784 | // one extra leading underscore: _bolt_instr_setup -> __bolt_instr_setup. |
1785 | extern "C" |
1786 | __attribute__((section("__TEXT,__setup" ))) |
1787 | __attribute__((force_align_arg_pointer)) |
1788 | void _bolt_instr_setup() { |
1789 | __asm__ __volatile__(SAVE_ALL :::); |
1790 | |
1791 | report("Hello!\n" ); |
1792 | |
1793 | __asm__ __volatile__(RESTORE_ALL :::); |
1794 | } |
1795 | |
1796 | extern "C" |
1797 | __attribute__((section("__TEXT,__fini" ))) |
1798 | __attribute__((force_align_arg_pointer)) |
1799 | void _bolt_instr_fini() { |
1800 | report("Bye!\n" ); |
1801 | __bolt_instr_data_dump(); |
1802 | } |
1803 | |
1804 | #endif |
1805 | |