1 | //===- llvm-profdata.cpp - LLVM profile data tool -------------------------===// |
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 | // llvm-profdata merges .profdata files. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/ADT/SmallSet.h" |
14 | #include "llvm/ADT/SmallVector.h" |
15 | #include "llvm/ADT/StringRef.h" |
16 | #include "llvm/IR/LLVMContext.h" |
17 | #include "llvm/Object/Binary.h" |
18 | #include "llvm/ProfileData/InstrProfCorrelator.h" |
19 | #include "llvm/ProfileData/InstrProfReader.h" |
20 | #include "llvm/ProfileData/InstrProfWriter.h" |
21 | #include "llvm/ProfileData/MemProf.h" |
22 | #include "llvm/ProfileData/MemProfReader.h" |
23 | #include "llvm/ProfileData/ProfileCommon.h" |
24 | #include "llvm/ProfileData/SampleProfReader.h" |
25 | #include "llvm/ProfileData/SampleProfWriter.h" |
26 | #include "llvm/Support/BalancedPartitioning.h" |
27 | #include "llvm/Support/CommandLine.h" |
28 | #include "llvm/Support/Discriminator.h" |
29 | #include "llvm/Support/Errc.h" |
30 | #include "llvm/Support/FileSystem.h" |
31 | #include "llvm/Support/Format.h" |
32 | #include "llvm/Support/FormattedStream.h" |
33 | #include "llvm/Support/LLVMDriver.h" |
34 | #include "llvm/Support/MD5.h" |
35 | #include "llvm/Support/MemoryBuffer.h" |
36 | #include "llvm/Support/Path.h" |
37 | #include "llvm/Support/Regex.h" |
38 | #include "llvm/Support/ThreadPool.h" |
39 | #include "llvm/Support/Threading.h" |
40 | #include "llvm/Support/VirtualFileSystem.h" |
41 | #include "llvm/Support/WithColor.h" |
42 | #include "llvm/Support/raw_ostream.h" |
43 | #include <algorithm> |
44 | #include <cmath> |
45 | #include <optional> |
46 | #include <queue> |
47 | |
48 | using namespace llvm; |
49 | using ProfCorrelatorKind = InstrProfCorrelator::ProfCorrelatorKind; |
50 | |
51 | // https://llvm.org/docs/CommandGuide/llvm-profdata.html has documentations |
52 | // on each subcommand. |
53 | cl::SubCommand ShowSubcommand( |
54 | "show" , |
55 | "Takes a profile data file and displays the profiles. See detailed " |
56 | "documentation in " |
57 | "https://llvm.org/docs/CommandGuide/llvm-profdata.html#profdata-show" ); |
58 | cl::SubCommand OrderSubcommand( |
59 | "order" , |
60 | "Reads temporal profiling traces from a profile and outputs a function " |
61 | "order that reduces the number of page faults for those traces. See " |
62 | "detailed documentation in " |
63 | "https://llvm.org/docs/CommandGuide/llvm-profdata.html#profdata-order" ); |
64 | cl::SubCommand OverlapSubcommand( |
65 | "overlap" , |
66 | "Computes and displays the overlap between two profiles. See detailed " |
67 | "documentation in " |
68 | "https://llvm.org/docs/CommandGuide/llvm-profdata.html#profdata-overlap" ); |
69 | cl::SubCommand MergeSubcommand( |
70 | "merge" , |
71 | "Takes several profiles and merge them together. See detailed " |
72 | "documentation in " |
73 | "https://llvm.org/docs/CommandGuide/llvm-profdata.html#profdata-merge" ); |
74 | |
75 | namespace { |
76 | enum ProfileKinds { instr, sample, memory }; |
77 | enum FailureMode { warnOnly, failIfAnyAreInvalid, failIfAllAreInvalid }; |
78 | |
79 | enum ProfileFormat { |
80 | PF_None = 0, |
81 | PF_Text, |
82 | PF_Compact_Binary, // Deprecated |
83 | PF_Ext_Binary, |
84 | PF_GCC, |
85 | PF_Binary |
86 | }; |
87 | |
88 | enum class ShowFormat { Text, Json, Yaml }; |
89 | } // namespace |
90 | |
91 | // Common options. |
92 | cl::opt<std::string> OutputFilename("output" , cl::value_desc("output" ), |
93 | cl::init(Val: "-" ), cl::desc("Output file" ), |
94 | cl::sub(ShowSubcommand), |
95 | cl::sub(OrderSubcommand), |
96 | cl::sub(OverlapSubcommand), |
97 | cl::sub(MergeSubcommand)); |
98 | // NOTE: cl::alias must not have cl::sub(), since aliased option's cl::sub() |
99 | // will be used. llvm::cl::alias::done() method asserts this condition. |
100 | cl::alias OutputFilenameA("o" , cl::desc("Alias for --output" ), |
101 | cl::aliasopt(OutputFilename)); |
102 | |
103 | // Options common to at least two commands. |
104 | cl::opt<ProfileKinds> ProfileKind( |
105 | cl::desc("Profile kind:" ), cl::sub(MergeSubcommand), |
106 | cl::sub(OverlapSubcommand), cl::init(Val: instr), |
107 | cl::values(clEnumVal(instr, "Instrumentation profile (default)" ), |
108 | clEnumVal(sample, "Sample profile" ))); |
109 | cl::opt<std::string> Filename(cl::Positional, cl::desc("<profdata-file>" ), |
110 | cl::sub(ShowSubcommand), |
111 | cl::sub(OrderSubcommand)); |
112 | cl::opt<unsigned> MaxDbgCorrelationWarnings( |
113 | "max-debug-info-correlation-warnings" , |
114 | cl::desc("The maximum number of warnings to emit when correlating " |
115 | "profile from debug info (0 = no limit)" ), |
116 | cl::sub(MergeSubcommand), cl::sub(ShowSubcommand), cl::init(Val: 5)); |
117 | cl::opt<std::string> ProfiledBinary( |
118 | "profiled-binary" , cl::init(Val: "" ), |
119 | cl::desc("Path to binary from which the profile was collected." ), |
120 | cl::sub(ShowSubcommand), cl::sub(MergeSubcommand)); |
121 | cl::opt<std::string> DebugInfoFilename( |
122 | "debug-info" , cl::init(Val: "" ), |
123 | cl::desc( |
124 | "For show, read and extract profile metadata from debug info and show " |
125 | "the functions it found. For merge, use the provided debug info to " |
126 | "correlate the raw profile." ), |
127 | cl::sub(ShowSubcommand), cl::sub(MergeSubcommand)); |
128 | cl::opt<std::string> |
129 | BinaryFilename("binary-file" , cl::init(Val: "" ), |
130 | cl::desc("For merge, use the provided unstripped bianry to " |
131 | "correlate the raw profile." ), |
132 | cl::sub(MergeSubcommand)); |
133 | cl::opt<std::string> FuncNameFilter( |
134 | "function" , |
135 | cl::desc("Only functions matching the filter are shown in the output. For " |
136 | "overlapping CSSPGO, this takes a function name with calling " |
137 | "context." ), |
138 | cl::sub(ShowSubcommand), cl::sub(OverlapSubcommand), |
139 | cl::sub(MergeSubcommand)); |
140 | |
141 | // TODO: Consider creating a template class (e.g., MergeOption, ShowOption) to |
142 | // factor out the common cl::sub in cl::opt constructor for subcommand-specific |
143 | // options. |
144 | |
145 | // Options specific to merge subcommand. |
146 | cl::list<std::string> InputFilenames(cl::Positional, cl::sub(MergeSubcommand), |
147 | cl::desc("<filename...>" )); |
148 | cl::list<std::string> WeightedInputFilenames("weighted-input" , |
149 | cl::sub(MergeSubcommand), |
150 | cl::desc("<weight>,<filename>" )); |
151 | cl::opt<ProfileFormat> OutputFormat( |
152 | cl::desc("Format of output profile" ), cl::sub(MergeSubcommand), |
153 | cl::init(Val: PF_Ext_Binary), |
154 | cl::values(clEnumValN(PF_Binary, "binary" , "Binary encoding" ), |
155 | clEnumValN(PF_Ext_Binary, "extbinary" , |
156 | "Extensible binary encoding " |
157 | "(default)" ), |
158 | clEnumValN(PF_Text, "text" , "Text encoding" ), |
159 | clEnumValN(PF_GCC, "gcc" , |
160 | "GCC encoding (only meaningful for -sample)" ))); |
161 | cl::opt<std::string> |
162 | InputFilenamesFile("input-files" , cl::init(Val: "" ), cl::sub(MergeSubcommand), |
163 | cl::desc("Path to file containing newline-separated " |
164 | "[<weight>,]<filename> entries" )); |
165 | cl::alias InputFilenamesFileA("f" , cl::desc("Alias for --input-files" ), |
166 | cl::aliasopt(InputFilenamesFile)); |
167 | cl::opt<bool> DumpInputFileList( |
168 | "dump-input-file-list" , cl::init(Val: false), cl::Hidden, |
169 | cl::sub(MergeSubcommand), |
170 | cl::desc("Dump the list of input files and their weights, then exit" )); |
171 | cl::opt<std::string> RemappingFile("remapping-file" , cl::value_desc("file" ), |
172 | cl::sub(MergeSubcommand), |
173 | cl::desc("Symbol remapping file" )); |
174 | cl::alias RemappingFileA("r" , cl::desc("Alias for --remapping-file" ), |
175 | cl::aliasopt(RemappingFile)); |
176 | cl::opt<bool> |
177 | UseMD5("use-md5" , cl::init(Val: false), cl::Hidden, |
178 | cl::desc("Choose to use MD5 to represent string in name table (only " |
179 | "meaningful for -extbinary)" ), |
180 | cl::sub(MergeSubcommand)); |
181 | cl::opt<bool> CompressAllSections( |
182 | "compress-all-sections" , cl::init(Val: false), cl::Hidden, |
183 | cl::sub(MergeSubcommand), |
184 | cl::desc("Compress all sections when writing the profile (only " |
185 | "meaningful for -extbinary)" )); |
186 | cl::opt<bool> SampleMergeColdContext( |
187 | "sample-merge-cold-context" , cl::init(Val: false), cl::Hidden, |
188 | cl::sub(MergeSubcommand), |
189 | cl::desc( |
190 | "Merge context sample profiles whose count is below cold threshold" )); |
191 | cl::opt<bool> SampleTrimColdContext( |
192 | "sample-trim-cold-context" , cl::init(Val: false), cl::Hidden, |
193 | cl::sub(MergeSubcommand), |
194 | cl::desc( |
195 | "Trim context sample profiles whose count is below cold threshold" )); |
196 | cl::opt<uint32_t> SampleColdContextFrameDepth( |
197 | "sample-frame-depth-for-cold-context" , cl::init(Val: 1), |
198 | cl::sub(MergeSubcommand), |
199 | cl::desc("Keep the last K frames while merging cold profile. 1 means the " |
200 | "context-less base profile" )); |
201 | cl::opt<size_t> OutputSizeLimit( |
202 | "output-size-limit" , cl::init(Val: 0), cl::Hidden, cl::sub(MergeSubcommand), |
203 | cl::desc("Trim cold functions until profile size is below specified " |
204 | "limit in bytes. This uses a heursitic and functions may be " |
205 | "excessively trimmed" )); |
206 | cl::opt<bool> GenPartialProfile( |
207 | "gen-partial-profile" , cl::init(Val: false), cl::Hidden, |
208 | cl::sub(MergeSubcommand), |
209 | cl::desc("Generate a partial profile (only meaningful for -extbinary)" )); |
210 | cl::opt<std::string> SupplInstrWithSample( |
211 | "supplement-instr-with-sample" , cl::init(Val: "" ), cl::Hidden, |
212 | cl::sub(MergeSubcommand), |
213 | cl::desc("Supplement an instr profile with sample profile, to correct " |
214 | "the profile unrepresentativeness issue. The sample " |
215 | "profile is the input of the flag. Output will be in instr " |
216 | "format (The flag only works with -instr)" )); |
217 | cl::opt<float> ZeroCounterThreshold( |
218 | "zero-counter-threshold" , cl::init(Val: 0.7), cl::Hidden, |
219 | cl::sub(MergeSubcommand), |
220 | cl::desc("For the function which is cold in instr profile but hot in " |
221 | "sample profile, if the ratio of the number of zero counters " |
222 | "divided by the total number of counters is above the " |
223 | "threshold, the profile of the function will be regarded as " |
224 | "being harmful for performance and will be dropped." )); |
225 | cl::opt<unsigned> SupplMinSizeThreshold( |
226 | "suppl-min-size-threshold" , cl::init(Val: 10), cl::Hidden, |
227 | cl::sub(MergeSubcommand), |
228 | cl::desc("If the size of a function is smaller than the threshold, " |
229 | "assume it can be inlined by PGO early inliner and it won't " |
230 | "be adjusted based on sample profile." )); |
231 | cl::opt<unsigned> InstrProfColdThreshold( |
232 | "instr-prof-cold-threshold" , cl::init(Val: 0), cl::Hidden, |
233 | cl::sub(MergeSubcommand), |
234 | cl::desc("User specified cold threshold for instr profile which will " |
235 | "override the cold threshold got from profile summary. " )); |
236 | // WARNING: This reservoir size value is propagated to any input indexed |
237 | // profiles for simplicity. Changing this value between invocations could |
238 | // result in sample bias. |
239 | cl::opt<uint64_t> TemporalProfTraceReservoirSize( |
240 | "temporal-profile-trace-reservoir-size" , cl::init(Val: 100), |
241 | cl::sub(MergeSubcommand), |
242 | cl::desc("The maximum number of stored temporal profile traces (default: " |
243 | "100)" )); |
244 | cl::opt<uint64_t> TemporalProfMaxTraceLength( |
245 | "temporal-profile-max-trace-length" , cl::init(Val: 10000), |
246 | cl::sub(MergeSubcommand), |
247 | cl::desc("The maximum length of a single temporal profile trace " |
248 | "(default: 10000)" )); |
249 | cl::opt<std::string> FuncNameNegativeFilter( |
250 | "no-function" , cl::init(Val: "" ), |
251 | cl::sub(MergeSubcommand), |
252 | cl::desc("Exclude functions matching the filter from the output." )); |
253 | |
254 | cl::opt<FailureMode> |
255 | FailMode("failure-mode" , cl::init(Val: failIfAnyAreInvalid), |
256 | cl::desc("Failure mode:" ), cl::sub(MergeSubcommand), |
257 | cl::values(clEnumValN(warnOnly, "warn" , |
258 | "Do not fail and just print warnings." ), |
259 | clEnumValN(failIfAnyAreInvalid, "any" , |
260 | "Fail if any profile is invalid." ), |
261 | clEnumValN(failIfAllAreInvalid, "all" , |
262 | "Fail only if all profiles are invalid." ))); |
263 | |
264 | cl::opt<bool> OutputSparse( |
265 | "sparse" , cl::init(Val: false), cl::sub(MergeSubcommand), |
266 | cl::desc("Generate a sparse profile (only meaningful for -instr)" )); |
267 | cl::opt<unsigned> NumThreads( |
268 | "num-threads" , cl::init(Val: 0), cl::sub(MergeSubcommand), |
269 | cl::desc("Number of merge threads to use (default: autodetect)" )); |
270 | cl::alias NumThreadsA("j" , cl::desc("Alias for --num-threads" ), |
271 | cl::aliasopt(NumThreads)); |
272 | |
273 | cl::opt<std::string> ProfileSymbolListFile( |
274 | "prof-sym-list" , cl::init(Val: "" ), cl::sub(MergeSubcommand), |
275 | cl::desc("Path to file containing the list of function symbols " |
276 | "used to populate profile symbol list" )); |
277 | |
278 | cl::opt<SampleProfileLayout> ProfileLayout( |
279 | "convert-sample-profile-layout" , |
280 | cl::desc("Convert the generated profile to a profile with a new layout" ), |
281 | cl::sub(MergeSubcommand), cl::init(Val: SPL_None), |
282 | cl::values( |
283 | clEnumValN(SPL_Nest, "nest" , |
284 | "Nested profile, the input should be CS flat profile" ), |
285 | clEnumValN(SPL_Flat, "flat" , |
286 | "Profile with nested inlinee flatten out" ))); |
287 | |
288 | cl::opt<bool> DropProfileSymbolList( |
289 | "drop-profile-symbol-list" , cl::init(Val: false), cl::Hidden, |
290 | cl::sub(MergeSubcommand), |
291 | cl::desc("Drop the profile symbol list when merging AutoFDO profiles " |
292 | "(only meaningful for -sample)" )); |
293 | |
294 | cl::opt<bool> KeepVTableSymbols( |
295 | "keep-vtable-symbols" , cl::init(Val: false), cl::Hidden, |
296 | cl::sub(MergeSubcommand), |
297 | cl::desc("If true, keep the vtable symbols in indexed profiles" )); |
298 | |
299 | // Temporary support for writing the previous version of the format, to enable |
300 | // some forward compatibility. |
301 | // TODO: Consider enabling this with future version changes as well, to ease |
302 | // deployment of newer versions of llvm-profdata. |
303 | cl::opt<bool> DoWritePrevVersion( |
304 | "write-prev-version" , cl::init(Val: false), cl::Hidden, |
305 | cl::desc("Write the previous version of indexed format, to enable " |
306 | "some forward compatibility." )); |
307 | |
308 | cl::opt<memprof::IndexedVersion> MemProfVersionRequested( |
309 | "memprof-version" , cl::Hidden, cl::sub(MergeSubcommand), |
310 | cl::desc("Specify the version of the memprof format to use" ), |
311 | cl::init(Val: memprof::Version0), |
312 | cl::values(clEnumValN(memprof::Version0, "0" , "version 0" ), |
313 | clEnumValN(memprof::Version1, "1" , "version 1" ), |
314 | clEnumValN(memprof::Version2, "2" , "version 2" ), |
315 | clEnumValN(memprof::Version3, "3" , "version 3" ))); |
316 | |
317 | cl::opt<bool> MemProfFullSchema( |
318 | "memprof-full-schema" , cl::Hidden, cl::sub(MergeSubcommand), |
319 | cl::desc("Use the full schema for serialization" ), cl::init(Val: false)); |
320 | |
321 | // Options specific to overlap subcommand. |
322 | cl::opt<std::string> BaseFilename(cl::Positional, cl::Required, |
323 | cl::desc("<base profile file>" ), |
324 | cl::sub(OverlapSubcommand)); |
325 | cl::opt<std::string> TestFilename(cl::Positional, cl::Required, |
326 | cl::desc("<test profile file>" ), |
327 | cl::sub(OverlapSubcommand)); |
328 | |
329 | cl::opt<unsigned long long> SimilarityCutoff( |
330 | "similarity-cutoff" , cl::init(Val: 0), |
331 | cl::desc("For sample profiles, list function names (with calling context " |
332 | "for csspgo) for overlapped functions " |
333 | "with similarities below the cutoff (percentage times 10000)." ), |
334 | cl::sub(OverlapSubcommand)); |
335 | |
336 | cl::opt<bool> IsCS( |
337 | "cs" , cl::init(Val: false), |
338 | cl::desc("For context sensitive PGO counts. Does not work with CSSPGO." ), |
339 | cl::sub(OverlapSubcommand)); |
340 | |
341 | cl::opt<unsigned long long> OverlapValueCutoff( |
342 | "value-cutoff" , cl::init(Val: -1), |
343 | cl::desc( |
344 | "Function level overlap information for every function (with calling " |
345 | "context for csspgo) in test " |
346 | "profile with max count value greater then the parameter value" ), |
347 | cl::sub(OverlapSubcommand)); |
348 | |
349 | // Options specific to show subcommand. |
350 | cl::opt<bool> ShowCounts("counts" , cl::init(Val: false), |
351 | cl::desc("Show counter values for shown functions" ), |
352 | cl::sub(ShowSubcommand)); |
353 | cl::opt<ShowFormat> |
354 | SFormat("show-format" , cl::init(Val: ShowFormat::Text), |
355 | cl::desc("Emit output in the selected format if supported" ), |
356 | cl::sub(ShowSubcommand), |
357 | cl::values(clEnumValN(ShowFormat::Text, "text" , |
358 | "emit normal text output (default)" ), |
359 | clEnumValN(ShowFormat::Json, "json" , "emit JSON" ), |
360 | clEnumValN(ShowFormat::Yaml, "yaml" , "emit YAML" ))); |
361 | // TODO: Consider replacing this with `--show-format=text-encoding`. |
362 | cl::opt<bool> |
363 | TextFormat("text" , cl::init(Val: false), |
364 | cl::desc("Show instr profile data in text dump format" ), |
365 | cl::sub(ShowSubcommand)); |
366 | cl::opt<bool> |
367 | JsonFormat("json" , |
368 | cl::desc("Show sample profile data in the JSON format " |
369 | "(deprecated, please use --show-format=json)" ), |
370 | cl::sub(ShowSubcommand)); |
371 | cl::opt<bool> ShowIndirectCallTargets( |
372 | "ic-targets" , cl::init(Val: false), |
373 | cl::desc("Show indirect call site target values for shown functions" ), |
374 | cl::sub(ShowSubcommand)); |
375 | cl::opt<bool> ShowVTables("show-vtables" , cl::init(Val: false), |
376 | cl::desc("Show vtable names for shown functions" ), |
377 | cl::sub(ShowSubcommand)); |
378 | cl::opt<bool> ShowMemOPSizes( |
379 | "memop-sizes" , cl::init(Val: false), |
380 | cl::desc("Show the profiled sizes of the memory intrinsic calls " |
381 | "for shown functions" ), |
382 | cl::sub(ShowSubcommand)); |
383 | cl::opt<bool> ShowDetailedSummary("detailed-summary" , cl::init(Val: false), |
384 | cl::desc("Show detailed profile summary" ), |
385 | cl::sub(ShowSubcommand)); |
386 | cl::list<uint32_t> DetailedSummaryCutoffs( |
387 | cl::CommaSeparated, "detailed-summary-cutoffs" , |
388 | cl::desc( |
389 | "Cutoff percentages (times 10000) for generating detailed summary" ), |
390 | cl::value_desc("800000,901000,999999" ), cl::sub(ShowSubcommand)); |
391 | cl::opt<bool> |
392 | ShowHotFuncList("hot-func-list" , cl::init(Val: false), |
393 | cl::desc("Show profile summary of a list of hot functions" ), |
394 | cl::sub(ShowSubcommand)); |
395 | cl::opt<bool> ShowAllFunctions("all-functions" , cl::init(Val: false), |
396 | cl::desc("Details for each and every function" ), |
397 | cl::sub(ShowSubcommand)); |
398 | cl::opt<bool> ShowCS("showcs" , cl::init(Val: false), |
399 | cl::desc("Show context sensitive counts" ), |
400 | cl::sub(ShowSubcommand)); |
401 | cl::opt<ProfileKinds> ShowProfileKind( |
402 | cl::desc("Profile kind supported by show:" ), cl::sub(ShowSubcommand), |
403 | cl::init(Val: instr), |
404 | cl::values(clEnumVal(instr, "Instrumentation profile (default)" ), |
405 | clEnumVal(sample, "Sample profile" ), |
406 | clEnumVal(memory, "MemProf memory access profile" ))); |
407 | cl::opt<uint32_t> TopNFunctions( |
408 | "topn" , cl::init(Val: 0), |
409 | cl::desc("Show the list of functions with the largest internal counts" ), |
410 | cl::sub(ShowSubcommand)); |
411 | cl::opt<uint32_t> ShowValueCutoff( |
412 | "value-cutoff" , cl::init(Val: 0), |
413 | cl::desc("Set the count value cutoff. Functions with the maximum count " |
414 | "less than this value will not be printed out. (Default is 0)" ), |
415 | cl::sub(ShowSubcommand)); |
416 | cl::opt<bool> OnlyListBelow( |
417 | "list-below-cutoff" , cl::init(Val: false), |
418 | cl::desc("Only output names of functions whose max count values are " |
419 | "below the cutoff value" ), |
420 | cl::sub(ShowSubcommand)); |
421 | cl::opt<bool> ShowProfileSymbolList( |
422 | "show-prof-sym-list" , cl::init(Val: false), |
423 | cl::desc("Show profile symbol list if it exists in the profile. " ), |
424 | cl::sub(ShowSubcommand)); |
425 | cl::opt<bool> ShowSectionInfoOnly( |
426 | "show-sec-info-only" , cl::init(Val: false), |
427 | cl::desc("Show the information of each section in the sample profile. " |
428 | "The flag is only usable when the sample profile is in " |
429 | "extbinary format" ), |
430 | cl::sub(ShowSubcommand)); |
431 | cl::opt<bool> ShowBinaryIds("binary-ids" , cl::init(Val: false), |
432 | cl::desc("Show binary ids in the profile. " ), |
433 | cl::sub(ShowSubcommand)); |
434 | cl::opt<bool> ShowTemporalProfTraces( |
435 | "temporal-profile-traces" , |
436 | cl::desc("Show temporal profile traces in the profile." ), |
437 | cl::sub(ShowSubcommand)); |
438 | |
439 | cl::opt<bool> |
440 | ShowCovered("covered" , cl::init(Val: false), |
441 | cl::desc("Show only the functions that have been executed." ), |
442 | cl::sub(ShowSubcommand)); |
443 | |
444 | cl::opt<bool> ShowProfileVersion("profile-version" , cl::init(Val: false), |
445 | cl::desc("Show profile version. " ), |
446 | cl::sub(ShowSubcommand)); |
447 | |
448 | // Options specific to order subcommand. |
449 | cl::opt<unsigned> |
450 | NumTestTraces("num-test-traces" , cl::init(Val: 0), |
451 | cl::desc("Keep aside the last <num-test-traces> traces in " |
452 | "the profile when computing the function order and " |
453 | "instead use them to evaluate that order" ), |
454 | cl::sub(OrderSubcommand)); |
455 | |
456 | // We use this string to indicate that there are |
457 | // multiple static functions map to the same name. |
458 | const std::string DuplicateNameStr = "----" ; |
459 | |
460 | static void warn(Twine Message, StringRef Whence = "" , StringRef Hint = "" ) { |
461 | WithColor::warning(); |
462 | if (!Whence.empty()) |
463 | errs() << Whence << ": " ; |
464 | errs() << Message << "\n" ; |
465 | if (!Hint.empty()) |
466 | WithColor::note() << Hint << "\n" ; |
467 | } |
468 | |
469 | static void warn(Error E, StringRef Whence = "" ) { |
470 | if (E.isA<InstrProfError>()) { |
471 | handleAllErrors(E: std::move(E), Handlers: [&](const InstrProfError &IPE) { |
472 | warn(Message: IPE.message(), Whence); |
473 | }); |
474 | } |
475 | } |
476 | |
477 | static void exitWithError(Twine Message, StringRef Whence = "" , |
478 | StringRef Hint = "" ) { |
479 | WithColor::error(); |
480 | if (!Whence.empty()) |
481 | errs() << Whence << ": " ; |
482 | errs() << Message << "\n" ; |
483 | if (!Hint.empty()) |
484 | WithColor::note() << Hint << "\n" ; |
485 | ::exit(status: 1); |
486 | } |
487 | |
488 | static void exitWithError(Error E, StringRef Whence = "" ) { |
489 | if (E.isA<InstrProfError>()) { |
490 | handleAllErrors(E: std::move(E), Handlers: [&](const InstrProfError &IPE) { |
491 | instrprof_error instrError = IPE.get(); |
492 | StringRef Hint = "" ; |
493 | if (instrError == instrprof_error::unrecognized_format) { |
494 | // Hint in case user missed specifying the profile type. |
495 | Hint = "Perhaps you forgot to use the --sample or --memory option?" ; |
496 | } |
497 | exitWithError(Message: IPE.message(), Whence, Hint); |
498 | }); |
499 | return; |
500 | } |
501 | |
502 | exitWithError(Message: toString(E: std::move(E)), Whence); |
503 | } |
504 | |
505 | static void exitWithErrorCode(std::error_code EC, StringRef Whence = "" ) { |
506 | exitWithError(Message: EC.message(), Whence); |
507 | } |
508 | |
509 | static void warnOrExitGivenError(FailureMode FailMode, std::error_code EC, |
510 | StringRef Whence = "" ) { |
511 | if (FailMode == failIfAnyAreInvalid) |
512 | exitWithErrorCode(EC, Whence); |
513 | else |
514 | warn(Message: EC.message(), Whence); |
515 | } |
516 | |
517 | static void handleMergeWriterError(Error E, StringRef WhenceFile = "" , |
518 | StringRef WhenceFunction = "" , |
519 | bool ShowHint = true) { |
520 | if (!WhenceFile.empty()) |
521 | errs() << WhenceFile << ": " ; |
522 | if (!WhenceFunction.empty()) |
523 | errs() << WhenceFunction << ": " ; |
524 | |
525 | auto IPE = instrprof_error::success; |
526 | E = handleErrors(E: std::move(E), |
527 | Hs: [&IPE](std::unique_ptr<InstrProfError> E) -> Error { |
528 | IPE = E->get(); |
529 | return Error(std::move(E)); |
530 | }); |
531 | errs() << toString(E: std::move(E)) << "\n" ; |
532 | |
533 | if (ShowHint) { |
534 | StringRef Hint = "" ; |
535 | if (IPE != instrprof_error::success) { |
536 | switch (IPE) { |
537 | case instrprof_error::hash_mismatch: |
538 | case instrprof_error::count_mismatch: |
539 | case instrprof_error::value_site_count_mismatch: |
540 | Hint = "Make sure that all profile data to be merged is generated " |
541 | "from the same binary." ; |
542 | break; |
543 | default: |
544 | break; |
545 | } |
546 | } |
547 | |
548 | if (!Hint.empty()) |
549 | errs() << Hint << "\n" ; |
550 | } |
551 | } |
552 | |
553 | namespace { |
554 | /// A remapper from original symbol names to new symbol names based on a file |
555 | /// containing a list of mappings from old name to new name. |
556 | class SymbolRemapper { |
557 | std::unique_ptr<MemoryBuffer> File; |
558 | DenseMap<StringRef, StringRef> RemappingTable; |
559 | |
560 | public: |
561 | /// Build a SymbolRemapper from a file containing a list of old/new symbols. |
562 | static std::unique_ptr<SymbolRemapper> create(StringRef InputFile) { |
563 | auto BufOrError = MemoryBuffer::getFileOrSTDIN(Filename: InputFile); |
564 | if (!BufOrError) |
565 | exitWithErrorCode(EC: BufOrError.getError(), Whence: InputFile); |
566 | |
567 | auto Remapper = std::make_unique<SymbolRemapper>(); |
568 | Remapper->File = std::move(BufOrError.get()); |
569 | |
570 | for (line_iterator LineIt(*Remapper->File, /*SkipBlanks=*/true, '#'); |
571 | !LineIt.is_at_eof(); ++LineIt) { |
572 | std::pair<StringRef, StringRef> Parts = LineIt->split(Separator: ' '); |
573 | if (Parts.first.empty() || Parts.second.empty() || |
574 | Parts.second.count(C: ' ')) { |
575 | exitWithError(Message: "unexpected line in remapping file" , |
576 | Whence: (InputFile + ":" + Twine(LineIt.line_number())).str(), |
577 | Hint: "expected 'old_symbol new_symbol'" ); |
578 | } |
579 | Remapper->RemappingTable.insert(KV: Parts); |
580 | } |
581 | return Remapper; |
582 | } |
583 | |
584 | /// Attempt to map the given old symbol into a new symbol. |
585 | /// |
586 | /// \return The new symbol, or \p Name if no such symbol was found. |
587 | StringRef operator()(StringRef Name) { |
588 | StringRef New = RemappingTable.lookup(Val: Name); |
589 | return New.empty() ? Name : New; |
590 | } |
591 | |
592 | FunctionId operator()(FunctionId Name) { |
593 | // MD5 name cannot be remapped. |
594 | if (!Name.isStringRef()) |
595 | return Name; |
596 | StringRef New = RemappingTable.lookup(Val: Name.stringRef()); |
597 | return New.empty() ? Name : FunctionId(New); |
598 | } |
599 | }; |
600 | } |
601 | |
602 | struct WeightedFile { |
603 | std::string Filename; |
604 | uint64_t Weight; |
605 | }; |
606 | typedef SmallVector<WeightedFile, 5> WeightedFileVector; |
607 | |
608 | /// Keep track of merged data and reported errors. |
609 | struct WriterContext { |
610 | std::mutex Lock; |
611 | InstrProfWriter Writer; |
612 | std::vector<std::pair<Error, std::string>> Errors; |
613 | std::mutex &ErrLock; |
614 | SmallSet<instrprof_error, 4> &WriterErrorCodes; |
615 | |
616 | WriterContext(bool IsSparse, std::mutex &ErrLock, |
617 | SmallSet<instrprof_error, 4> &WriterErrorCodes, |
618 | uint64_t ReservoirSize = 0, uint64_t MaxTraceLength = 0) |
619 | : Writer(IsSparse, ReservoirSize, MaxTraceLength, DoWritePrevVersion, |
620 | MemProfVersionRequested, MemProfFullSchema), |
621 | ErrLock(ErrLock), WriterErrorCodes(WriterErrorCodes) {} |
622 | }; |
623 | |
624 | /// Computer the overlap b/w profile BaseFilename and TestFileName, |
625 | /// and store the program level result to Overlap. |
626 | static void overlapInput(const std::string &BaseFilename, |
627 | const std::string &TestFilename, WriterContext *WC, |
628 | OverlapStats &Overlap, |
629 | const OverlapFuncFilters &FuncFilter, |
630 | raw_fd_ostream &OS, bool IsCS) { |
631 | auto FS = vfs::getRealFileSystem(); |
632 | auto ReaderOrErr = InstrProfReader::create(Path: TestFilename, FS&: *FS); |
633 | if (Error E = ReaderOrErr.takeError()) { |
634 | // Skip the empty profiles by returning sliently. |
635 | auto [ErrorCode, Msg] = InstrProfError::take(E: std::move(E)); |
636 | if (ErrorCode != instrprof_error::empty_raw_profile) |
637 | WC->Errors.emplace_back(args: make_error<InstrProfError>(Args&: ErrorCode, Args&: Msg), |
638 | args: TestFilename); |
639 | return; |
640 | } |
641 | |
642 | auto Reader = std::move(ReaderOrErr.get()); |
643 | for (auto &I : *Reader) { |
644 | OverlapStats FuncOverlap(OverlapStats::FunctionLevel); |
645 | FuncOverlap.setFuncInfo(Name: I.Name, Hash: I.Hash); |
646 | |
647 | WC->Writer.overlapRecord(Other: std::move(I), Overlap, FuncLevelOverlap&: FuncOverlap, FuncFilter); |
648 | FuncOverlap.dump(OS); |
649 | } |
650 | } |
651 | |
652 | /// Load an input into a writer context. |
653 | static void loadInput(const WeightedFile &Input, SymbolRemapper *Remapper, |
654 | const InstrProfCorrelator *Correlator, |
655 | const StringRef ProfiledBinary, WriterContext *WC) { |
656 | std::unique_lock<std::mutex> CtxGuard{WC->Lock}; |
657 | |
658 | // Copy the filename, because llvm::ThreadPool copied the input "const |
659 | // WeightedFile &" by value, making a reference to the filename within it |
660 | // invalid outside of this packaged task. |
661 | std::string Filename = Input.Filename; |
662 | |
663 | using ::llvm::memprof::RawMemProfReader; |
664 | if (RawMemProfReader::hasFormat(Path: Input.Filename)) { |
665 | auto ReaderOrErr = RawMemProfReader::create(Path: Input.Filename, ProfiledBinary); |
666 | if (!ReaderOrErr) { |
667 | exitWithError(E: ReaderOrErr.takeError(), Whence: Input.Filename); |
668 | } |
669 | std::unique_ptr<RawMemProfReader> Reader = std::move(ReaderOrErr.get()); |
670 | // Check if the profile types can be merged, e.g. clang frontend profiles |
671 | // should not be merged with memprof profiles. |
672 | if (Error E = WC->Writer.mergeProfileKind(Other: Reader->getProfileKind())) { |
673 | consumeError(Err: std::move(E)); |
674 | WC->Errors.emplace_back( |
675 | args: make_error<StringError>( |
676 | Args: "Cannot merge MemProf profile with Clang generated profile." , |
677 | Args: std::error_code()), |
678 | args&: Filename); |
679 | return; |
680 | } |
681 | |
682 | auto MemProfError = [&](Error E) { |
683 | auto [ErrorCode, Msg] = InstrProfError::take(E: std::move(E)); |
684 | WC->Errors.emplace_back(args: make_error<InstrProfError>(Args&: ErrorCode, Args&: Msg), |
685 | args&: Filename); |
686 | }; |
687 | |
688 | // Add the frame mappings into the writer context. |
689 | const auto &IdToFrame = Reader->getFrameMapping(); |
690 | for (const auto &I : IdToFrame) { |
691 | bool Succeeded = WC->Writer.addMemProfFrame( |
692 | /*Id=*/I.first, /*Frame=*/F: I.getSecond(), Warn: MemProfError); |
693 | // If we weren't able to add the frame mappings then it doesn't make sense |
694 | // to try to add the records from this profile. |
695 | if (!Succeeded) |
696 | return; |
697 | } |
698 | |
699 | // Add the call stacks into the writer context. |
700 | const auto &CSIdToCallStacks = Reader->getCallStacks(); |
701 | for (const auto &I : CSIdToCallStacks) { |
702 | bool Succeeded = WC->Writer.addMemProfCallStack( |
703 | /*Id=*/CSId: I.first, /*Frame=*/CallStack: I.getSecond(), Warn: MemProfError); |
704 | // If we weren't able to add the call stacks then it doesn't make sense |
705 | // to try to add the records from this profile. |
706 | if (!Succeeded) |
707 | return; |
708 | } |
709 | |
710 | const auto &FunctionProfileData = Reader->getProfileData(); |
711 | // Add the memprof records into the writer context. |
712 | for (const auto &[GUID, Record] : FunctionProfileData) { |
713 | WC->Writer.addMemProfRecord(Id: GUID, Record); |
714 | } |
715 | return; |
716 | } |
717 | |
718 | auto FS = vfs::getRealFileSystem(); |
719 | // TODO: This only saves the first non-fatal error from InstrProfReader, and |
720 | // then added to WriterContext::Errors. However, this is not extensible, if |
721 | // we have more non-fatal errors from InstrProfReader in the future. How |
722 | // should this interact with different -failure-mode? |
723 | std::optional<std::pair<Error, std::string>> ReaderWarning; |
724 | auto Warn = [&](Error E) { |
725 | if (ReaderWarning) { |
726 | consumeError(Err: std::move(E)); |
727 | return; |
728 | } |
729 | // Only show the first time an error occurs in this file. |
730 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
731 | ReaderWarning = {make_error<InstrProfError>(Args&: ErrCode, Args&: Msg), Filename}; |
732 | }; |
733 | auto ReaderOrErr = |
734 | InstrProfReader::create(Path: Input.Filename, FS&: *FS, Correlator, Warn); |
735 | if (Error E = ReaderOrErr.takeError()) { |
736 | // Skip the empty profiles by returning silently. |
737 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
738 | if (ErrCode != instrprof_error::empty_raw_profile) |
739 | WC->Errors.emplace_back(args: make_error<InstrProfError>(Args&: ErrCode, Args&: Msg), |
740 | args&: Filename); |
741 | return; |
742 | } |
743 | |
744 | auto Reader = std::move(ReaderOrErr.get()); |
745 | if (Error E = WC->Writer.mergeProfileKind(Other: Reader->getProfileKind())) { |
746 | consumeError(Err: std::move(E)); |
747 | WC->Errors.emplace_back( |
748 | args: make_error<StringError>( |
749 | Args: "Merge IR generated profile with Clang generated profile." , |
750 | Args: std::error_code()), |
751 | args&: Filename); |
752 | return; |
753 | } |
754 | |
755 | for (auto &I : *Reader) { |
756 | if (Remapper) |
757 | I.Name = (*Remapper)(I.Name); |
758 | const StringRef FuncName = I.Name; |
759 | bool Reported = false; |
760 | WC->Writer.addRecord(I: std::move(I), Weight: Input.Weight, Warn: [&](Error E) { |
761 | if (Reported) { |
762 | consumeError(Err: std::move(E)); |
763 | return; |
764 | } |
765 | Reported = true; |
766 | // Only show hint the first time an error occurs. |
767 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
768 | std::unique_lock<std::mutex> ErrGuard{WC->ErrLock}; |
769 | bool firstTime = WC->WriterErrorCodes.insert(V: ErrCode).second; |
770 | handleMergeWriterError(E: make_error<InstrProfError>(Args&: ErrCode, Args&: Msg), |
771 | WhenceFile: Input.Filename, WhenceFunction: FuncName, ShowHint: firstTime); |
772 | }); |
773 | } |
774 | |
775 | if (KeepVTableSymbols) { |
776 | const InstrProfSymtab &symtab = Reader->getSymtab(); |
777 | const auto &VTableNames = symtab.getVTableNames(); |
778 | |
779 | for (const auto &kv : VTableNames) |
780 | WC->Writer.addVTableName(VTableName: kv.getKey()); |
781 | } |
782 | |
783 | if (Reader->hasTemporalProfile()) { |
784 | auto &Traces = Reader->getTemporalProfTraces(Weight: Input.Weight); |
785 | if (!Traces.empty()) |
786 | WC->Writer.addTemporalProfileTraces( |
787 | SrcTraces&: Traces, SrcStreamSize: Reader->getTemporalProfTraceStreamSize()); |
788 | } |
789 | if (Reader->hasError()) { |
790 | if (Error E = Reader->getError()) { |
791 | WC->Errors.emplace_back(args: std::move(E), args&: Filename); |
792 | return; |
793 | } |
794 | } |
795 | |
796 | std::vector<llvm::object::BuildID> BinaryIds; |
797 | if (Error E = Reader->readBinaryIds(BinaryIds)) { |
798 | WC->Errors.emplace_back(args: std::move(E), args&: Filename); |
799 | return; |
800 | } |
801 | WC->Writer.addBinaryIds(BIs: BinaryIds); |
802 | |
803 | if (ReaderWarning) { |
804 | WC->Errors.emplace_back(args: std::move(ReaderWarning->first), |
805 | args&: ReaderWarning->second); |
806 | } |
807 | } |
808 | |
809 | /// Merge the \p Src writer context into \p Dst. |
810 | static void mergeWriterContexts(WriterContext *Dst, WriterContext *Src) { |
811 | for (auto &ErrorPair : Src->Errors) |
812 | Dst->Errors.push_back(x: std::move(ErrorPair)); |
813 | Src->Errors.clear(); |
814 | |
815 | if (Error E = Dst->Writer.mergeProfileKind(Other: Src->Writer.getProfileKind())) |
816 | exitWithError(E: std::move(E)); |
817 | |
818 | Dst->Writer.mergeRecordsFromWriter(IPW: std::move(Src->Writer), Warn: [&](Error E) { |
819 | auto [ErrorCode, Msg] = InstrProfError::take(E: std::move(E)); |
820 | std::unique_lock<std::mutex> ErrGuard{Dst->ErrLock}; |
821 | bool firstTime = Dst->WriterErrorCodes.insert(V: ErrorCode).second; |
822 | if (firstTime) |
823 | warn(Message: toString(E: make_error<InstrProfError>(Args&: ErrorCode, Args&: Msg))); |
824 | }); |
825 | } |
826 | |
827 | static StringRef |
828 | getFuncName(const StringMap<InstrProfWriter::ProfilingData>::value_type &Val) { |
829 | return Val.first(); |
830 | } |
831 | |
832 | static std::string |
833 | getFuncName(const SampleProfileMap::value_type &Val) { |
834 | return Val.second.getContext().toString(); |
835 | } |
836 | |
837 | template <typename T> |
838 | static void filterFunctions(T &ProfileMap) { |
839 | bool hasFilter = !FuncNameFilter.empty(); |
840 | bool hasNegativeFilter = !FuncNameNegativeFilter.empty(); |
841 | if (!hasFilter && !hasNegativeFilter) |
842 | return; |
843 | |
844 | // If filter starts with '?' it is MSVC mangled name, not a regex. |
845 | llvm::Regex ProbablyMSVCMangledName("[?@$_0-9A-Za-z]+" ); |
846 | if (hasFilter && FuncNameFilter[0] == '?' && |
847 | ProbablyMSVCMangledName.match(String: FuncNameFilter)) |
848 | FuncNameFilter = llvm::Regex::escape(String: FuncNameFilter); |
849 | if (hasNegativeFilter && FuncNameNegativeFilter[0] == '?' && |
850 | ProbablyMSVCMangledName.match(String: FuncNameNegativeFilter)) |
851 | FuncNameNegativeFilter = llvm::Regex::escape(String: FuncNameNegativeFilter); |
852 | |
853 | size_t Count = ProfileMap.size(); |
854 | llvm::Regex Pattern(FuncNameFilter); |
855 | llvm::Regex NegativePattern(FuncNameNegativeFilter); |
856 | std::string Error; |
857 | if (hasFilter && !Pattern.isValid(Error)) |
858 | exitWithError(Message: Error); |
859 | if (hasNegativeFilter && !NegativePattern.isValid(Error)) |
860 | exitWithError(Message: Error); |
861 | |
862 | // Handle MD5 profile, so it is still able to match using the original name. |
863 | std::string MD5Name = std::to_string(val: llvm::MD5Hash(Str: FuncNameFilter)); |
864 | std::string NegativeMD5Name = |
865 | std::to_string(val: llvm::MD5Hash(Str: FuncNameNegativeFilter)); |
866 | |
867 | for (auto I = ProfileMap.begin(); I != ProfileMap.end();) { |
868 | auto Tmp = I++; |
869 | const auto &FuncName = getFuncName(*Tmp); |
870 | // Negative filter has higher precedence than positive filter. |
871 | if ((hasNegativeFilter && |
872 | (NegativePattern.match(String: FuncName) || |
873 | (FunctionSamples::UseMD5 && NegativeMD5Name == FuncName))) || |
874 | (hasFilter && !(Pattern.match(String: FuncName) || |
875 | (FunctionSamples::UseMD5 && MD5Name == FuncName)))) |
876 | ProfileMap.erase(Tmp); |
877 | } |
878 | |
879 | llvm::dbgs() << Count - ProfileMap.size() << " of " << Count << " functions " |
880 | << "in the original profile are filtered.\n" ; |
881 | } |
882 | |
883 | static void writeInstrProfile(StringRef OutputFilename, |
884 | ProfileFormat OutputFormat, |
885 | InstrProfWriter &Writer) { |
886 | std::error_code EC; |
887 | raw_fd_ostream Output(OutputFilename.data(), EC, |
888 | OutputFormat == PF_Text ? sys::fs::OF_TextWithCRLF |
889 | : sys::fs::OF_None); |
890 | if (EC) |
891 | exitWithErrorCode(EC, Whence: OutputFilename); |
892 | |
893 | if (OutputFormat == PF_Text) { |
894 | if (Error E = Writer.writeText(OS&: Output)) |
895 | warn(E: std::move(E)); |
896 | } else { |
897 | if (Output.is_displayed()) |
898 | exitWithError(Message: "cannot write a non-text format profile to the terminal" ); |
899 | if (Error E = Writer.write(OS&: Output)) |
900 | warn(E: std::move(E)); |
901 | } |
902 | } |
903 | |
904 | static void mergeInstrProfile(const WeightedFileVector &Inputs, |
905 | SymbolRemapper *Remapper, |
906 | int MaxDbgCorrelationWarnings, |
907 | const StringRef ProfiledBinary) { |
908 | const uint64_t TraceReservoirSize = TemporalProfTraceReservoirSize.getValue(); |
909 | const uint64_t MaxTraceLength = TemporalProfMaxTraceLength.getValue(); |
910 | if (OutputFormat == PF_Compact_Binary) |
911 | exitWithError(Message: "Compact Binary is deprecated" ); |
912 | if (OutputFormat != PF_Binary && OutputFormat != PF_Ext_Binary && |
913 | OutputFormat != PF_Text) |
914 | exitWithError(Message: "unknown format is specified" ); |
915 | |
916 | // TODO: Maybe we should support correlation with mixture of different |
917 | // correlation modes(w/wo debug-info/object correlation). |
918 | if (!DebugInfoFilename.empty() && !BinaryFilename.empty()) |
919 | exitWithError(Message: "Expected only one of -debug-info, -binary-file" ); |
920 | std::string CorrelateFilename; |
921 | ProfCorrelatorKind CorrelateKind = ProfCorrelatorKind::NONE; |
922 | if (!DebugInfoFilename.empty()) { |
923 | CorrelateFilename = DebugInfoFilename; |
924 | CorrelateKind = ProfCorrelatorKind::DEBUG_INFO; |
925 | } else if (!BinaryFilename.empty()) { |
926 | CorrelateFilename = BinaryFilename; |
927 | CorrelateKind = ProfCorrelatorKind::BINARY; |
928 | } |
929 | |
930 | std::unique_ptr<InstrProfCorrelator> Correlator; |
931 | if (CorrelateKind != InstrProfCorrelator::NONE) { |
932 | if (auto Err = InstrProfCorrelator::get(Filename: CorrelateFilename, FileKind: CorrelateKind) |
933 | .moveInto(Value&: Correlator)) |
934 | exitWithError(E: std::move(Err), Whence: CorrelateFilename); |
935 | if (auto Err = Correlator->correlateProfileData(MaxWarnings: MaxDbgCorrelationWarnings)) |
936 | exitWithError(E: std::move(Err), Whence: CorrelateFilename); |
937 | } |
938 | |
939 | std::mutex ErrorLock; |
940 | SmallSet<instrprof_error, 4> WriterErrorCodes; |
941 | |
942 | // If NumThreads is not specified, auto-detect a good default. |
943 | if (NumThreads == 0) |
944 | NumThreads = std::min(a: hardware_concurrency().compute_thread_count(), |
945 | b: unsigned((Inputs.size() + 1) / 2)); |
946 | |
947 | // Initialize the writer contexts. |
948 | SmallVector<std::unique_ptr<WriterContext>, 4> Contexts; |
949 | for (unsigned I = 0; I < NumThreads; ++I) |
950 | Contexts.emplace_back(Args: std::make_unique<WriterContext>( |
951 | args&: OutputSparse, args&: ErrorLock, args&: WriterErrorCodes, args: TraceReservoirSize, |
952 | args: MaxTraceLength)); |
953 | |
954 | if (NumThreads == 1) { |
955 | for (const auto &Input : Inputs) |
956 | loadInput(Input, Remapper, Correlator: Correlator.get(), ProfiledBinary, |
957 | WC: Contexts[0].get()); |
958 | } else { |
959 | DefaultThreadPool Pool(hardware_concurrency(ThreadCount: NumThreads)); |
960 | |
961 | // Load the inputs in parallel (N/NumThreads serial steps). |
962 | unsigned Ctx = 0; |
963 | for (const auto &Input : Inputs) { |
964 | Pool.async(F&: loadInput, ArgList: Input, ArgList&: Remapper, ArgList: Correlator.get(), ArgList: ProfiledBinary, |
965 | ArgList: Contexts[Ctx].get()); |
966 | Ctx = (Ctx + 1) % NumThreads; |
967 | } |
968 | Pool.wait(); |
969 | |
970 | // Merge the writer contexts together (~ lg(NumThreads) serial steps). |
971 | unsigned Mid = Contexts.size() / 2; |
972 | unsigned End = Contexts.size(); |
973 | assert(Mid > 0 && "Expected more than one context" ); |
974 | do { |
975 | for (unsigned I = 0; I < Mid; ++I) |
976 | Pool.async(F&: mergeWriterContexts, ArgList: Contexts[I].get(), |
977 | ArgList: Contexts[I + Mid].get()); |
978 | Pool.wait(); |
979 | if (End & 1) { |
980 | Pool.async(F&: mergeWriterContexts, ArgList: Contexts[0].get(), |
981 | ArgList: Contexts[End - 1].get()); |
982 | Pool.wait(); |
983 | } |
984 | End = Mid; |
985 | Mid /= 2; |
986 | } while (Mid > 0); |
987 | } |
988 | |
989 | // Handle deferred errors encountered during merging. If the number of errors |
990 | // is equal to the number of inputs the merge failed. |
991 | unsigned NumErrors = 0; |
992 | for (std::unique_ptr<WriterContext> &WC : Contexts) { |
993 | for (auto &ErrorPair : WC->Errors) { |
994 | ++NumErrors; |
995 | warn(Message: toString(E: std::move(ErrorPair.first)), Whence: ErrorPair.second); |
996 | } |
997 | } |
998 | if ((NumErrors == Inputs.size() && FailMode == failIfAllAreInvalid) || |
999 | (NumErrors > 0 && FailMode == failIfAnyAreInvalid)) |
1000 | exitWithError(Message: "no profile can be merged" ); |
1001 | |
1002 | filterFunctions(ProfileMap&: Contexts[0]->Writer.getProfileData()); |
1003 | |
1004 | writeInstrProfile(OutputFilename, OutputFormat, Writer&: Contexts[0]->Writer); |
1005 | } |
1006 | |
1007 | /// The profile entry for a function in instrumentation profile. |
1008 | struct InstrProfileEntry { |
1009 | uint64_t MaxCount = 0; |
1010 | uint64_t NumEdgeCounters = 0; |
1011 | float ZeroCounterRatio = 0.0; |
1012 | InstrProfRecord *ProfRecord; |
1013 | InstrProfileEntry(InstrProfRecord *Record); |
1014 | InstrProfileEntry() = default; |
1015 | }; |
1016 | |
1017 | InstrProfileEntry::InstrProfileEntry(InstrProfRecord *Record) { |
1018 | ProfRecord = Record; |
1019 | uint64_t CntNum = Record->Counts.size(); |
1020 | uint64_t ZeroCntNum = 0; |
1021 | for (size_t I = 0; I < CntNum; ++I) { |
1022 | MaxCount = std::max(a: MaxCount, b: Record->Counts[I]); |
1023 | ZeroCntNum += !Record->Counts[I]; |
1024 | } |
1025 | ZeroCounterRatio = (float)ZeroCntNum / CntNum; |
1026 | NumEdgeCounters = CntNum; |
1027 | } |
1028 | |
1029 | /// Either set all the counters in the instr profile entry \p IFE to |
1030 | /// -1 / -2 /in order to drop the profile or scale up the |
1031 | /// counters in \p IFP to be above hot / cold threshold. We use |
1032 | /// the ratio of zero counters in the profile of a function to |
1033 | /// decide the profile is helpful or harmful for performance, |
1034 | /// and to choose whether to scale up or drop it. |
1035 | static void updateInstrProfileEntry(InstrProfileEntry &IFE, bool SetToHot, |
1036 | uint64_t HotInstrThreshold, |
1037 | uint64_t ColdInstrThreshold, |
1038 | float ZeroCounterThreshold) { |
1039 | InstrProfRecord *ProfRecord = IFE.ProfRecord; |
1040 | if (!IFE.MaxCount || IFE.ZeroCounterRatio > ZeroCounterThreshold) { |
1041 | // If all or most of the counters of the function are zero, the |
1042 | // profile is unaccountable and should be dropped. Reset all the |
1043 | // counters to be -1 / -2 and PGO profile-use will drop the profile. |
1044 | // All counters being -1 also implies that the function is hot so |
1045 | // PGO profile-use will also set the entry count metadata to be |
1046 | // above hot threshold. |
1047 | // All counters being -2 implies that the function is warm so |
1048 | // PGO profile-use will also set the entry count metadata to be |
1049 | // above cold threshold. |
1050 | auto Kind = |
1051 | (SetToHot ? InstrProfRecord::PseudoHot : InstrProfRecord::PseudoWarm); |
1052 | ProfRecord->setPseudoCount(Kind); |
1053 | return; |
1054 | } |
1055 | |
1056 | // Scale up the MaxCount to be multiple times above hot / cold threshold. |
1057 | const unsigned MultiplyFactor = 3; |
1058 | uint64_t Threshold = (SetToHot ? HotInstrThreshold : ColdInstrThreshold); |
1059 | uint64_t Numerator = Threshold * MultiplyFactor; |
1060 | |
1061 | // Make sure Threshold for warm counters is below the HotInstrThreshold. |
1062 | if (!SetToHot && Threshold >= HotInstrThreshold) { |
1063 | Threshold = (HotInstrThreshold + ColdInstrThreshold) / 2; |
1064 | } |
1065 | |
1066 | uint64_t Denominator = IFE.MaxCount; |
1067 | if (Numerator <= Denominator) |
1068 | return; |
1069 | ProfRecord->scale(N: Numerator, D: Denominator, Warn: [&](instrprof_error E) { |
1070 | warn(Message: toString(E: make_error<InstrProfError>(Args&: E))); |
1071 | }); |
1072 | } |
1073 | |
1074 | const uint64_t ColdPercentileIdx = 15; |
1075 | const uint64_t HotPercentileIdx = 11; |
1076 | |
1077 | using sampleprof::FSDiscriminatorPass; |
1078 | |
1079 | // Internal options to set FSDiscriminatorPass. Used in merge and show |
1080 | // commands. |
1081 | static cl::opt<FSDiscriminatorPass> FSDiscriminatorPassOption( |
1082 | "fs-discriminator-pass" , cl::init(Val: PassLast), cl::Hidden, |
1083 | cl::desc("Zero out the discriminator bits for the FS discrimiantor " |
1084 | "pass beyond this value. The enum values are defined in " |
1085 | "Support/Discriminator.h" ), |
1086 | cl::values(clEnumVal(Base, "Use base discriminators only" ), |
1087 | clEnumVal(Pass1, "Use base and pass 1 discriminators" ), |
1088 | clEnumVal(Pass2, "Use base and pass 1-2 discriminators" ), |
1089 | clEnumVal(Pass3, "Use base and pass 1-3 discriminators" ), |
1090 | clEnumVal(PassLast, "Use all discriminator bits (default)" ))); |
1091 | |
1092 | static unsigned getDiscriminatorMask() { |
1093 | return getN1Bits(N: getFSPassBitEnd(P: FSDiscriminatorPassOption.getValue())); |
1094 | } |
1095 | |
1096 | /// Adjust the instr profile in \p WC based on the sample profile in |
1097 | /// \p Reader. |
1098 | static void |
1099 | adjustInstrProfile(std::unique_ptr<WriterContext> &WC, |
1100 | std::unique_ptr<sampleprof::SampleProfileReader> &Reader, |
1101 | unsigned SupplMinSizeThreshold, float ZeroCounterThreshold, |
1102 | unsigned InstrProfColdThreshold) { |
1103 | // Function to its entry in instr profile. |
1104 | StringMap<InstrProfileEntry> InstrProfileMap; |
1105 | StringMap<StringRef> StaticFuncMap; |
1106 | InstrProfSummaryBuilder IPBuilder(ProfileSummaryBuilder::DefaultCutoffs); |
1107 | |
1108 | auto checkSampleProfileHasFUnique = [&Reader]() { |
1109 | for (const auto &PD : Reader->getProfiles()) { |
1110 | auto &FContext = PD.second.getContext(); |
1111 | if (FContext.toString().find(s: FunctionSamples::UniqSuffix) != |
1112 | std::string::npos) { |
1113 | return true; |
1114 | } |
1115 | } |
1116 | return false; |
1117 | }; |
1118 | |
1119 | bool SampleProfileHasFUnique = checkSampleProfileHasFUnique(); |
1120 | |
1121 | auto buildStaticFuncMap = [&StaticFuncMap, |
1122 | SampleProfileHasFUnique](const StringRef Name) { |
1123 | std::string FilePrefixes[] = {".cpp" , "cc" , ".c" , ".hpp" , ".h" }; |
1124 | size_t PrefixPos = StringRef::npos; |
1125 | for (auto &FilePrefix : FilePrefixes) { |
1126 | std::string NamePrefix = FilePrefix + GlobalIdentifierDelimiter; |
1127 | PrefixPos = Name.find_insensitive(Str: NamePrefix); |
1128 | if (PrefixPos == StringRef::npos) |
1129 | continue; |
1130 | PrefixPos += NamePrefix.size(); |
1131 | break; |
1132 | } |
1133 | |
1134 | if (PrefixPos == StringRef::npos) { |
1135 | return; |
1136 | } |
1137 | |
1138 | StringRef NewName = Name.drop_front(N: PrefixPos); |
1139 | StringRef FName = Name.substr(Start: 0, N: PrefixPos - 1); |
1140 | if (NewName.size() == 0) { |
1141 | return; |
1142 | } |
1143 | |
1144 | // This name should have a static linkage. |
1145 | size_t PostfixPos = NewName.find(Str: FunctionSamples::UniqSuffix); |
1146 | bool ProfileHasFUnique = (PostfixPos != StringRef::npos); |
1147 | |
1148 | // If sample profile and instrumented profile do not agree on symbol |
1149 | // uniqification. |
1150 | if (SampleProfileHasFUnique != ProfileHasFUnique) { |
1151 | // If instrumented profile uses -funique-internal-linkage-symbols, |
1152 | // we need to trim the name. |
1153 | if (ProfileHasFUnique) { |
1154 | NewName = NewName.substr(Start: 0, N: PostfixPos); |
1155 | } else { |
1156 | // If sample profile uses -funique-internal-linkage-symbols, |
1157 | // we build the map. |
1158 | std::string NStr = |
1159 | NewName.str() + getUniqueInternalLinkagePostfix(FName); |
1160 | NewName = StringRef(NStr); |
1161 | StaticFuncMap[NewName] = Name; |
1162 | return; |
1163 | } |
1164 | } |
1165 | |
1166 | if (!StaticFuncMap.contains(Key: NewName)) { |
1167 | StaticFuncMap[NewName] = Name; |
1168 | } else { |
1169 | StaticFuncMap[NewName] = DuplicateNameStr; |
1170 | } |
1171 | }; |
1172 | |
1173 | // We need to flatten the SampleFDO profile as the InstrFDO |
1174 | // profile does not have inlined callsite profiles. |
1175 | // One caveat is the pre-inlined function -- their samples |
1176 | // should be collapsed into the caller function. |
1177 | // Here we do a DFS traversal to get the flatten profile |
1178 | // info: the sum of entrycount and the max of maxcount. |
1179 | // Here is the algorithm: |
1180 | // recursive (FS, root_name) { |
1181 | // name = FS->getName(); |
1182 | // get samples for FS; |
1183 | // if (InstrProf.find(name) { |
1184 | // root_name = name; |
1185 | // } else { |
1186 | // if (name is in static_func map) { |
1187 | // root_name = static_name; |
1188 | // } |
1189 | // } |
1190 | // update the Map entry for root_name; |
1191 | // for (subfs: FS) { |
1192 | // recursive(subfs, root_name); |
1193 | // } |
1194 | // } |
1195 | // |
1196 | // Here is an example. |
1197 | // |
1198 | // SampleProfile: |
1199 | // foo:12345:1000 |
1200 | // 1: 1000 |
1201 | // 2.1: 1000 |
1202 | // 15: 5000 |
1203 | // 4: bar:1000 |
1204 | // 1: 1000 |
1205 | // 2: goo:3000 |
1206 | // 1: 3000 |
1207 | // 8: bar:40000 |
1208 | // 1: 10000 |
1209 | // 2: goo:30000 |
1210 | // 1: 30000 |
1211 | // |
1212 | // InstrProfile has two entries: |
1213 | // foo |
1214 | // bar.cc;bar |
1215 | // |
1216 | // After BuildMaxSampleMap, we should have the following in FlattenSampleMap: |
1217 | // {"foo", {1000, 5000}} |
1218 | // {"bar.cc;bar", {11000, 30000}} |
1219 | // |
1220 | // foo's has an entry count of 1000, and max body count of 5000. |
1221 | // bar.cc;bar has an entry count of 11000 (sum two callsites of 1000 and |
1222 | // 10000), and max count of 30000 (from the callsite in line 8). |
1223 | // |
1224 | // Note that goo's count will remain in bar.cc;bar() as it does not have an |
1225 | // entry in InstrProfile. |
1226 | llvm::StringMap<std::pair<uint64_t, uint64_t>> FlattenSampleMap; |
1227 | auto BuildMaxSampleMap = [&FlattenSampleMap, &StaticFuncMap, |
1228 | &InstrProfileMap](const FunctionSamples &FS, |
1229 | const StringRef &RootName) { |
1230 | auto BuildMaxSampleMapImpl = [&](const FunctionSamples &FS, |
1231 | const StringRef &RootName, |
1232 | auto &BuildImpl) -> void { |
1233 | std::string NameStr = FS.getFunction().str(); |
1234 | const StringRef Name = NameStr; |
1235 | const StringRef *NewRootName = &RootName; |
1236 | uint64_t EntrySample = FS.getHeadSamplesEstimate(); |
1237 | uint64_t MaxBodySample = FS.getMaxCountInside(/* SkipCallSite*/ true); |
1238 | |
1239 | auto It = InstrProfileMap.find(Key: Name); |
1240 | if (It != InstrProfileMap.end()) { |
1241 | NewRootName = &Name; |
1242 | } else { |
1243 | auto NewName = StaticFuncMap.find(Key: Name); |
1244 | if (NewName != StaticFuncMap.end()) { |
1245 | It = InstrProfileMap.find(Key: NewName->second.str()); |
1246 | if (NewName->second != DuplicateNameStr) { |
1247 | NewRootName = &NewName->second; |
1248 | } |
1249 | } else { |
1250 | // Here the EntrySample is of an inlined function, so we should not |
1251 | // update the EntrySample in the map. |
1252 | EntrySample = 0; |
1253 | } |
1254 | } |
1255 | EntrySample += FlattenSampleMap[*NewRootName].first; |
1256 | MaxBodySample = |
1257 | std::max(a: FlattenSampleMap[*NewRootName].second, b: MaxBodySample); |
1258 | FlattenSampleMap[*NewRootName] = |
1259 | std::make_pair(x&: EntrySample, y&: MaxBodySample); |
1260 | |
1261 | for (const auto &C : FS.getCallsiteSamples()) |
1262 | for (const auto &F : C.second) |
1263 | BuildImpl(F.second, *NewRootName, BuildImpl); |
1264 | }; |
1265 | BuildMaxSampleMapImpl(FS, RootName, BuildMaxSampleMapImpl); |
1266 | }; |
1267 | |
1268 | for (auto &PD : WC->Writer.getProfileData()) { |
1269 | // Populate IPBuilder. |
1270 | for (const auto &PDV : PD.getValue()) { |
1271 | InstrProfRecord Record = PDV.second; |
1272 | IPBuilder.addRecord(Record); |
1273 | } |
1274 | |
1275 | // If a function has multiple entries in instr profile, skip it. |
1276 | if (PD.getValue().size() != 1) |
1277 | continue; |
1278 | |
1279 | // Initialize InstrProfileMap. |
1280 | InstrProfRecord *R = &PD.getValue().begin()->second; |
1281 | StringRef FullName = PD.getKey(); |
1282 | InstrProfileMap[FullName] = InstrProfileEntry(R); |
1283 | buildStaticFuncMap(FullName); |
1284 | } |
1285 | |
1286 | for (auto &PD : Reader->getProfiles()) { |
1287 | sampleprof::FunctionSamples &FS = PD.second; |
1288 | std::string Name = FS.getFunction().str(); |
1289 | BuildMaxSampleMap(FS, Name); |
1290 | } |
1291 | |
1292 | ProfileSummary InstrPS = *IPBuilder.getSummary(); |
1293 | ProfileSummary SamplePS = Reader->getSummary(); |
1294 | |
1295 | // Compute cold thresholds for instr profile and sample profile. |
1296 | uint64_t HotSampleThreshold = |
1297 | ProfileSummaryBuilder::getEntryForPercentile( |
1298 | DS: SamplePS.getDetailedSummary(), |
1299 | Percentile: ProfileSummaryBuilder::DefaultCutoffs[HotPercentileIdx]) |
1300 | .MinCount; |
1301 | uint64_t ColdSampleThreshold = |
1302 | ProfileSummaryBuilder::getEntryForPercentile( |
1303 | DS: SamplePS.getDetailedSummary(), |
1304 | Percentile: ProfileSummaryBuilder::DefaultCutoffs[ColdPercentileIdx]) |
1305 | .MinCount; |
1306 | uint64_t HotInstrThreshold = |
1307 | ProfileSummaryBuilder::getEntryForPercentile( |
1308 | DS: InstrPS.getDetailedSummary(), |
1309 | Percentile: ProfileSummaryBuilder::DefaultCutoffs[HotPercentileIdx]) |
1310 | .MinCount; |
1311 | uint64_t ColdInstrThreshold = |
1312 | InstrProfColdThreshold |
1313 | ? InstrProfColdThreshold |
1314 | : ProfileSummaryBuilder::getEntryForPercentile( |
1315 | DS: InstrPS.getDetailedSummary(), |
1316 | Percentile: ProfileSummaryBuilder::DefaultCutoffs[ColdPercentileIdx]) |
1317 | .MinCount; |
1318 | |
1319 | // Find hot/warm functions in sample profile which is cold in instr profile |
1320 | // and adjust the profiles of those functions in the instr profile. |
1321 | for (const auto &E : FlattenSampleMap) { |
1322 | uint64_t SampleMaxCount = std::max(a: E.second.first, b: E.second.second); |
1323 | if (SampleMaxCount < ColdSampleThreshold) |
1324 | continue; |
1325 | StringRef Name = E.first(); |
1326 | auto It = InstrProfileMap.find(Key: Name); |
1327 | if (It == InstrProfileMap.end()) { |
1328 | auto NewName = StaticFuncMap.find(Key: Name); |
1329 | if (NewName != StaticFuncMap.end()) { |
1330 | It = InstrProfileMap.find(Key: NewName->second.str()); |
1331 | if (NewName->second == DuplicateNameStr) { |
1332 | WithColor::warning() |
1333 | << "Static function " << Name |
1334 | << " has multiple promoted names, cannot adjust profile.\n" ; |
1335 | } |
1336 | } |
1337 | } |
1338 | if (It == InstrProfileMap.end() || |
1339 | It->second.MaxCount > ColdInstrThreshold || |
1340 | It->second.NumEdgeCounters < SupplMinSizeThreshold) |
1341 | continue; |
1342 | bool SetToHot = SampleMaxCount >= HotSampleThreshold; |
1343 | updateInstrProfileEntry(IFE&: It->second, SetToHot, HotInstrThreshold, |
1344 | ColdInstrThreshold, ZeroCounterThreshold); |
1345 | } |
1346 | } |
1347 | |
1348 | /// The main function to supplement instr profile with sample profile. |
1349 | /// \Inputs contains the instr profile. \p SampleFilename specifies the |
1350 | /// sample profile. \p OutputFilename specifies the output profile name. |
1351 | /// \p OutputFormat specifies the output profile format. \p OutputSparse |
1352 | /// specifies whether to generate sparse profile. \p SupplMinSizeThreshold |
1353 | /// specifies the minimal size for the functions whose profile will be |
1354 | /// adjusted. \p ZeroCounterThreshold is the threshold to check whether |
1355 | /// a function contains too many zero counters and whether its profile |
1356 | /// should be dropped. \p InstrProfColdThreshold is the user specified |
1357 | /// cold threshold which will override the cold threshold got from the |
1358 | /// instr profile summary. |
1359 | static void supplementInstrProfile(const WeightedFileVector &Inputs, |
1360 | StringRef SampleFilename, bool OutputSparse, |
1361 | unsigned SupplMinSizeThreshold, |
1362 | float ZeroCounterThreshold, |
1363 | unsigned InstrProfColdThreshold) { |
1364 | if (OutputFilename == "-" ) |
1365 | exitWithError(Message: "cannot write indexed profdata format to stdout" ); |
1366 | if (Inputs.size() != 1) |
1367 | exitWithError(Message: "expect one input to be an instr profile" ); |
1368 | if (Inputs[0].Weight != 1) |
1369 | exitWithError(Message: "expect instr profile doesn't have weight" ); |
1370 | |
1371 | StringRef InstrFilename = Inputs[0].Filename; |
1372 | |
1373 | // Read sample profile. |
1374 | LLVMContext Context; |
1375 | auto FS = vfs::getRealFileSystem(); |
1376 | auto ReaderOrErr = sampleprof::SampleProfileReader::create( |
1377 | Filename: SampleFilename.str(), C&: Context, FS&: *FS, P: FSDiscriminatorPassOption); |
1378 | if (std::error_code EC = ReaderOrErr.getError()) |
1379 | exitWithErrorCode(EC, Whence: SampleFilename); |
1380 | auto Reader = std::move(ReaderOrErr.get()); |
1381 | if (std::error_code EC = Reader->read()) |
1382 | exitWithErrorCode(EC, Whence: SampleFilename); |
1383 | |
1384 | // Read instr profile. |
1385 | std::mutex ErrorLock; |
1386 | SmallSet<instrprof_error, 4> WriterErrorCodes; |
1387 | auto WC = std::make_unique<WriterContext>(args&: OutputSparse, args&: ErrorLock, |
1388 | args&: WriterErrorCodes); |
1389 | loadInput(Input: Inputs[0], Remapper: nullptr, Correlator: nullptr, /*ProfiledBinary=*/"" , WC: WC.get()); |
1390 | if (WC->Errors.size() > 0) |
1391 | exitWithError(E: std::move(WC->Errors[0].first), Whence: InstrFilename); |
1392 | |
1393 | adjustInstrProfile(WC, Reader, SupplMinSizeThreshold, ZeroCounterThreshold, |
1394 | InstrProfColdThreshold); |
1395 | writeInstrProfile(OutputFilename, OutputFormat, Writer&: WC->Writer); |
1396 | } |
1397 | |
1398 | /// Make a copy of the given function samples with all symbol names remapped |
1399 | /// by the provided symbol remapper. |
1400 | static sampleprof::FunctionSamples |
1401 | remapSamples(const sampleprof::FunctionSamples &Samples, |
1402 | SymbolRemapper &Remapper, sampleprof_error &Error) { |
1403 | sampleprof::FunctionSamples Result; |
1404 | Result.setFunction(Remapper(Samples.getFunction())); |
1405 | Result.addTotalSamples(Num: Samples.getTotalSamples()); |
1406 | Result.addHeadSamples(Num: Samples.getHeadSamples()); |
1407 | for (const auto &BodySample : Samples.getBodySamples()) { |
1408 | uint32_t MaskedDiscriminator = |
1409 | BodySample.first.Discriminator & getDiscriminatorMask(); |
1410 | Result.addBodySamples(LineOffset: BodySample.first.LineOffset, Discriminator: MaskedDiscriminator, |
1411 | Num: BodySample.second.getSamples()); |
1412 | for (const auto &Target : BodySample.second.getCallTargets()) { |
1413 | Result.addCalledTargetSamples(LineOffset: BodySample.first.LineOffset, |
1414 | Discriminator: MaskedDiscriminator, |
1415 | Func: Remapper(Target.first), Num: Target.second); |
1416 | } |
1417 | } |
1418 | for (const auto &CallsiteSamples : Samples.getCallsiteSamples()) { |
1419 | sampleprof::FunctionSamplesMap &Target = |
1420 | Result.functionSamplesAt(Loc: CallsiteSamples.first); |
1421 | for (const auto &Callsite : CallsiteSamples.second) { |
1422 | sampleprof::FunctionSamples Remapped = |
1423 | remapSamples(Samples: Callsite.second, Remapper, Error); |
1424 | mergeSampleProfErrors(Accumulator&: Error, |
1425 | Result: Target[Remapped.getFunction()].merge(Other: Remapped)); |
1426 | } |
1427 | } |
1428 | return Result; |
1429 | } |
1430 | |
1431 | static sampleprof::SampleProfileFormat FormatMap[] = { |
1432 | sampleprof::SPF_None, |
1433 | sampleprof::SPF_Text, |
1434 | sampleprof::SPF_None, |
1435 | sampleprof::SPF_Ext_Binary, |
1436 | sampleprof::SPF_GCC, |
1437 | sampleprof::SPF_Binary}; |
1438 | |
1439 | static std::unique_ptr<MemoryBuffer> |
1440 | getInputFileBuf(const StringRef &InputFile) { |
1441 | if (InputFile == "" ) |
1442 | return {}; |
1443 | |
1444 | auto BufOrError = MemoryBuffer::getFileOrSTDIN(Filename: InputFile); |
1445 | if (!BufOrError) |
1446 | exitWithErrorCode(EC: BufOrError.getError(), Whence: InputFile); |
1447 | |
1448 | return std::move(*BufOrError); |
1449 | } |
1450 | |
1451 | static void populateProfileSymbolList(MemoryBuffer *Buffer, |
1452 | sampleprof::ProfileSymbolList &PSL) { |
1453 | if (!Buffer) |
1454 | return; |
1455 | |
1456 | SmallVector<StringRef, 32> SymbolVec; |
1457 | StringRef Data = Buffer->getBuffer(); |
1458 | Data.split(A&: SymbolVec, Separator: '\n', /*MaxSplit=*/-1, /*KeepEmpty=*/false); |
1459 | |
1460 | for (StringRef SymbolStr : SymbolVec) |
1461 | PSL.add(Name: SymbolStr.trim()); |
1462 | } |
1463 | |
1464 | static void handleExtBinaryWriter(sampleprof::SampleProfileWriter &Writer, |
1465 | ProfileFormat OutputFormat, |
1466 | MemoryBuffer *Buffer, |
1467 | sampleprof::ProfileSymbolList &WriterList, |
1468 | bool CompressAllSections, bool UseMD5, |
1469 | bool GenPartialProfile) { |
1470 | populateProfileSymbolList(Buffer, PSL&: WriterList); |
1471 | if (WriterList.size() > 0 && OutputFormat != PF_Ext_Binary) |
1472 | warn(Message: "Profile Symbol list is not empty but the output format is not " |
1473 | "ExtBinary format. The list will be lost in the output. " ); |
1474 | |
1475 | Writer.setProfileSymbolList(&WriterList); |
1476 | |
1477 | if (CompressAllSections) { |
1478 | if (OutputFormat != PF_Ext_Binary) |
1479 | warn(Message: "-compress-all-section is ignored. Specify -extbinary to enable it" ); |
1480 | else |
1481 | Writer.setToCompressAllSections(); |
1482 | } |
1483 | if (UseMD5) { |
1484 | if (OutputFormat != PF_Ext_Binary) |
1485 | warn(Message: "-use-md5 is ignored. Specify -extbinary to enable it" ); |
1486 | else |
1487 | Writer.setUseMD5(); |
1488 | } |
1489 | if (GenPartialProfile) { |
1490 | if (OutputFormat != PF_Ext_Binary) |
1491 | warn(Message: "-gen-partial-profile is ignored. Specify -extbinary to enable it" ); |
1492 | else |
1493 | Writer.setPartialProfile(); |
1494 | } |
1495 | } |
1496 | |
1497 | static void mergeSampleProfile(const WeightedFileVector &Inputs, |
1498 | SymbolRemapper *Remapper, |
1499 | StringRef ProfileSymbolListFile, |
1500 | size_t OutputSizeLimit) { |
1501 | using namespace sampleprof; |
1502 | SampleProfileMap ProfileMap; |
1503 | SmallVector<std::unique_ptr<sampleprof::SampleProfileReader>, 5> Readers; |
1504 | LLVMContext Context; |
1505 | sampleprof::ProfileSymbolList WriterList; |
1506 | std::optional<bool> ProfileIsProbeBased; |
1507 | std::optional<bool> ProfileIsCS; |
1508 | for (const auto &Input : Inputs) { |
1509 | auto FS = vfs::getRealFileSystem(); |
1510 | auto ReaderOrErr = SampleProfileReader::create(Filename: Input.Filename, C&: Context, FS&: *FS, |
1511 | P: FSDiscriminatorPassOption); |
1512 | if (std::error_code EC = ReaderOrErr.getError()) { |
1513 | warnOrExitGivenError(FailMode, EC, Whence: Input.Filename); |
1514 | continue; |
1515 | } |
1516 | |
1517 | // We need to keep the readers around until after all the files are |
1518 | // read so that we do not lose the function names stored in each |
1519 | // reader's memory. The function names are needed to write out the |
1520 | // merged profile map. |
1521 | Readers.push_back(Elt: std::move(ReaderOrErr.get())); |
1522 | const auto Reader = Readers.back().get(); |
1523 | if (std::error_code EC = Reader->read()) { |
1524 | warnOrExitGivenError(FailMode, EC, Whence: Input.Filename); |
1525 | Readers.pop_back(); |
1526 | continue; |
1527 | } |
1528 | |
1529 | SampleProfileMap &Profiles = Reader->getProfiles(); |
1530 | if (ProfileIsProbeBased && |
1531 | ProfileIsProbeBased != FunctionSamples::ProfileIsProbeBased) |
1532 | exitWithError( |
1533 | Message: "cannot merge probe-based profile with non-probe-based profile" ); |
1534 | ProfileIsProbeBased = FunctionSamples::ProfileIsProbeBased; |
1535 | if (ProfileIsCS && ProfileIsCS != FunctionSamples::ProfileIsCS) |
1536 | exitWithError(Message: "cannot merge CS profile with non-CS profile" ); |
1537 | ProfileIsCS = FunctionSamples::ProfileIsCS; |
1538 | for (SampleProfileMap::iterator I = Profiles.begin(), E = Profiles.end(); |
1539 | I != E; ++I) { |
1540 | sampleprof_error Result = sampleprof_error::success; |
1541 | FunctionSamples Remapped = |
1542 | Remapper ? remapSamples(Samples: I->second, Remapper&: *Remapper, Error&: Result) |
1543 | : FunctionSamples(); |
1544 | FunctionSamples &Samples = Remapper ? Remapped : I->second; |
1545 | SampleContext FContext = Samples.getContext(); |
1546 | mergeSampleProfErrors(Accumulator&: Result, |
1547 | Result: ProfileMap[FContext].merge(Other: Samples, Weight: Input.Weight)); |
1548 | if (Result != sampleprof_error::success) { |
1549 | std::error_code EC = make_error_code(E: Result); |
1550 | handleMergeWriterError(E: errorCodeToError(EC), WhenceFile: Input.Filename, |
1551 | WhenceFunction: FContext.toString()); |
1552 | } |
1553 | } |
1554 | |
1555 | if (!DropProfileSymbolList) { |
1556 | std::unique_ptr<sampleprof::ProfileSymbolList> ReaderList = |
1557 | Reader->getProfileSymbolList(); |
1558 | if (ReaderList) |
1559 | WriterList.merge(List: *ReaderList); |
1560 | } |
1561 | } |
1562 | |
1563 | if (ProfileIsCS && (SampleMergeColdContext || SampleTrimColdContext)) { |
1564 | // Use threshold calculated from profile summary unless specified. |
1565 | SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs); |
1566 | auto Summary = Builder.computeSummaryForProfiles(Profiles: ProfileMap); |
1567 | uint64_t SampleProfColdThreshold = |
1568 | ProfileSummaryBuilder::getColdCountThreshold( |
1569 | DS: (Summary->getDetailedSummary())); |
1570 | |
1571 | // Trim and merge cold context profile using cold threshold above; |
1572 | SampleContextTrimmer(ProfileMap) |
1573 | .trimAndMergeColdContextProfiles( |
1574 | ColdCountThreshold: SampleProfColdThreshold, TrimColdContext: SampleTrimColdContext, |
1575 | MergeColdContext: SampleMergeColdContext, ColdContextFrameLength: SampleColdContextFrameDepth, TrimBaseProfileOnly: false); |
1576 | } |
1577 | |
1578 | if (ProfileLayout == llvm::sampleprof::SPL_Flat) { |
1579 | ProfileConverter::flattenProfile(ProfileMap, ProfileIsCS: FunctionSamples::ProfileIsCS); |
1580 | ProfileIsCS = FunctionSamples::ProfileIsCS = false; |
1581 | } else if (ProfileIsCS && ProfileLayout == llvm::sampleprof::SPL_Nest) { |
1582 | ProfileConverter CSConverter(ProfileMap); |
1583 | CSConverter.convertCSProfiles(); |
1584 | ProfileIsCS = FunctionSamples::ProfileIsCS = false; |
1585 | } |
1586 | |
1587 | filterFunctions(ProfileMap); |
1588 | |
1589 | auto WriterOrErr = |
1590 | SampleProfileWriter::create(Filename: OutputFilename, Format: FormatMap[OutputFormat]); |
1591 | if (std::error_code EC = WriterOrErr.getError()) |
1592 | exitWithErrorCode(EC, Whence: OutputFilename); |
1593 | |
1594 | auto Writer = std::move(WriterOrErr.get()); |
1595 | // WriterList will have StringRef refering to string in Buffer. |
1596 | // Make sure Buffer lives as long as WriterList. |
1597 | auto Buffer = getInputFileBuf(InputFile: ProfileSymbolListFile); |
1598 | handleExtBinaryWriter(Writer&: *Writer, OutputFormat, Buffer: Buffer.get(), WriterList, |
1599 | CompressAllSections, UseMD5, GenPartialProfile); |
1600 | |
1601 | // If OutputSizeLimit is 0 (default), it is the same as write(). |
1602 | if (std::error_code EC = |
1603 | Writer->writeWithSizeLimit(ProfileMap, OutputSizeLimit)) |
1604 | exitWithErrorCode(EC); |
1605 | } |
1606 | |
1607 | static WeightedFile parseWeightedFile(const StringRef &WeightedFilename) { |
1608 | StringRef WeightStr, FileName; |
1609 | std::tie(args&: WeightStr, args&: FileName) = WeightedFilename.split(Separator: ','); |
1610 | |
1611 | uint64_t Weight; |
1612 | if (WeightStr.getAsInteger(Radix: 10, Result&: Weight) || Weight < 1) |
1613 | exitWithError(Message: "input weight must be a positive integer" ); |
1614 | |
1615 | return {.Filename: std::string(FileName), .Weight: Weight}; |
1616 | } |
1617 | |
1618 | static void addWeightedInput(WeightedFileVector &WNI, const WeightedFile &WF) { |
1619 | StringRef Filename = WF.Filename; |
1620 | uint64_t Weight = WF.Weight; |
1621 | |
1622 | // If it's STDIN just pass it on. |
1623 | if (Filename == "-" ) { |
1624 | WNI.push_back(Elt: {.Filename: std::string(Filename), .Weight: Weight}); |
1625 | return; |
1626 | } |
1627 | |
1628 | llvm::sys::fs::file_status Status; |
1629 | llvm::sys::fs::status(path: Filename, result&: Status); |
1630 | if (!llvm::sys::fs::exists(status: Status)) |
1631 | exitWithErrorCode(EC: make_error_code(E: errc::no_such_file_or_directory), |
1632 | Whence: Filename); |
1633 | // If it's a source file, collect it. |
1634 | if (llvm::sys::fs::is_regular_file(status: Status)) { |
1635 | WNI.push_back(Elt: {.Filename: std::string(Filename), .Weight: Weight}); |
1636 | return; |
1637 | } |
1638 | |
1639 | if (llvm::sys::fs::is_directory(status: Status)) { |
1640 | std::error_code EC; |
1641 | for (llvm::sys::fs::recursive_directory_iterator F(Filename, EC), E; |
1642 | F != E && !EC; F.increment(ec&: EC)) { |
1643 | if (llvm::sys::fs::is_regular_file(Path: F->path())) { |
1644 | addWeightedInput(WNI, WF: {.Filename: F->path(), .Weight: Weight}); |
1645 | } |
1646 | } |
1647 | if (EC) |
1648 | exitWithErrorCode(EC, Whence: Filename); |
1649 | } |
1650 | } |
1651 | |
1652 | static void parseInputFilenamesFile(MemoryBuffer *Buffer, |
1653 | WeightedFileVector &WFV) { |
1654 | if (!Buffer) |
1655 | return; |
1656 | |
1657 | SmallVector<StringRef, 8> Entries; |
1658 | StringRef Data = Buffer->getBuffer(); |
1659 | Data.split(A&: Entries, Separator: '\n', /*MaxSplit=*/-1, /*KeepEmpty=*/false); |
1660 | for (const StringRef &FileWeightEntry : Entries) { |
1661 | StringRef SanitizedEntry = FileWeightEntry.trim(Chars: " \t\v\f\r" ); |
1662 | // Skip comments. |
1663 | if (SanitizedEntry.starts_with(Prefix: "#" )) |
1664 | continue; |
1665 | // If there's no comma, it's an unweighted profile. |
1666 | else if (!SanitizedEntry.contains(C: ',')) |
1667 | addWeightedInput(WNI&: WFV, WF: {.Filename: std::string(SanitizedEntry), .Weight: 1}); |
1668 | else |
1669 | addWeightedInput(WNI&: WFV, WF: parseWeightedFile(WeightedFilename: SanitizedEntry)); |
1670 | } |
1671 | } |
1672 | |
1673 | static int merge_main(StringRef ProgName) { |
1674 | WeightedFileVector WeightedInputs; |
1675 | for (StringRef Filename : InputFilenames) |
1676 | addWeightedInput(WNI&: WeightedInputs, WF: {.Filename: std::string(Filename), .Weight: 1}); |
1677 | for (StringRef WeightedFilename : WeightedInputFilenames) |
1678 | addWeightedInput(WNI&: WeightedInputs, WF: parseWeightedFile(WeightedFilename)); |
1679 | |
1680 | // Make sure that the file buffer stays alive for the duration of the |
1681 | // weighted input vector's lifetime. |
1682 | auto Buffer = getInputFileBuf(InputFile: InputFilenamesFile); |
1683 | parseInputFilenamesFile(Buffer: Buffer.get(), WFV&: WeightedInputs); |
1684 | |
1685 | if (WeightedInputs.empty()) |
1686 | exitWithError(Message: "no input files specified. See " + ProgName + " merge -help" ); |
1687 | |
1688 | if (DumpInputFileList) { |
1689 | for (auto &WF : WeightedInputs) |
1690 | outs() << WF.Weight << "," << WF.Filename << "\n" ; |
1691 | return 0; |
1692 | } |
1693 | |
1694 | std::unique_ptr<SymbolRemapper> Remapper; |
1695 | if (!RemappingFile.empty()) |
1696 | Remapper = SymbolRemapper::create(InputFile: RemappingFile); |
1697 | |
1698 | if (!SupplInstrWithSample.empty()) { |
1699 | if (ProfileKind != instr) |
1700 | exitWithError( |
1701 | Message: "-supplement-instr-with-sample can only work with -instr. " ); |
1702 | |
1703 | supplementInstrProfile(Inputs: WeightedInputs, SampleFilename: SupplInstrWithSample, OutputSparse, |
1704 | SupplMinSizeThreshold, ZeroCounterThreshold, |
1705 | InstrProfColdThreshold); |
1706 | return 0; |
1707 | } |
1708 | |
1709 | if (ProfileKind == instr) |
1710 | mergeInstrProfile(Inputs: WeightedInputs, Remapper: Remapper.get(), MaxDbgCorrelationWarnings, |
1711 | ProfiledBinary); |
1712 | else |
1713 | mergeSampleProfile(Inputs: WeightedInputs, Remapper: Remapper.get(), ProfileSymbolListFile, |
1714 | OutputSizeLimit); |
1715 | return 0; |
1716 | } |
1717 | |
1718 | /// Computer the overlap b/w profile BaseFilename and profile TestFilename. |
1719 | static void overlapInstrProfile(const std::string &BaseFilename, |
1720 | const std::string &TestFilename, |
1721 | const OverlapFuncFilters &FuncFilter, |
1722 | raw_fd_ostream &OS, bool IsCS) { |
1723 | std::mutex ErrorLock; |
1724 | SmallSet<instrprof_error, 4> WriterErrorCodes; |
1725 | WriterContext Context(false, ErrorLock, WriterErrorCodes); |
1726 | WeightedFile WeightedInput{.Filename: BaseFilename, .Weight: 1}; |
1727 | OverlapStats Overlap; |
1728 | Error E = Overlap.accumulateCounts(BaseFilename, TestFilename, IsCS); |
1729 | if (E) |
1730 | exitWithError(E: std::move(E), Whence: "error in getting profile count sums" ); |
1731 | if (Overlap.Base.CountSum < 1.0f) { |
1732 | OS << "Sum of edge counts for profile " << BaseFilename << " is 0.\n" ; |
1733 | exit(status: 0); |
1734 | } |
1735 | if (Overlap.Test.CountSum < 1.0f) { |
1736 | OS << "Sum of edge counts for profile " << TestFilename << " is 0.\n" ; |
1737 | exit(status: 0); |
1738 | } |
1739 | loadInput(Input: WeightedInput, Remapper: nullptr, Correlator: nullptr, /*ProfiledBinary=*/"" , WC: &Context); |
1740 | overlapInput(BaseFilename, TestFilename, WC: &Context, Overlap, FuncFilter, OS, |
1741 | IsCS); |
1742 | Overlap.dump(OS); |
1743 | } |
1744 | |
1745 | namespace { |
1746 | struct SampleOverlapStats { |
1747 | SampleContext BaseName; |
1748 | SampleContext TestName; |
1749 | // Number of overlap units |
1750 | uint64_t OverlapCount = 0; |
1751 | // Total samples of overlap units |
1752 | uint64_t OverlapSample = 0; |
1753 | // Number of and total samples of units that only present in base or test |
1754 | // profile |
1755 | uint64_t BaseUniqueCount = 0; |
1756 | uint64_t BaseUniqueSample = 0; |
1757 | uint64_t TestUniqueCount = 0; |
1758 | uint64_t TestUniqueSample = 0; |
1759 | // Number of units and total samples in base or test profile |
1760 | uint64_t BaseCount = 0; |
1761 | uint64_t BaseSample = 0; |
1762 | uint64_t TestCount = 0; |
1763 | uint64_t TestSample = 0; |
1764 | // Number of and total samples of units that present in at least one profile |
1765 | uint64_t UnionCount = 0; |
1766 | uint64_t UnionSample = 0; |
1767 | // Weighted similarity |
1768 | double Similarity = 0.0; |
1769 | // For SampleOverlapStats instances representing functions, weights of the |
1770 | // function in base and test profiles |
1771 | double BaseWeight = 0.0; |
1772 | double TestWeight = 0.0; |
1773 | |
1774 | SampleOverlapStats() = default; |
1775 | }; |
1776 | } // end anonymous namespace |
1777 | |
1778 | namespace { |
1779 | struct FuncSampleStats { |
1780 | uint64_t SampleSum = 0; |
1781 | uint64_t MaxSample = 0; |
1782 | uint64_t HotBlockCount = 0; |
1783 | FuncSampleStats() = default; |
1784 | FuncSampleStats(uint64_t SampleSum, uint64_t MaxSample, |
1785 | uint64_t HotBlockCount) |
1786 | : SampleSum(SampleSum), MaxSample(MaxSample), |
1787 | HotBlockCount(HotBlockCount) {} |
1788 | }; |
1789 | } // end anonymous namespace |
1790 | |
1791 | namespace { |
1792 | enum MatchStatus { MS_Match, MS_FirstUnique, MS_SecondUnique, MS_None }; |
1793 | |
1794 | // Class for updating merging steps for two sorted maps. The class should be |
1795 | // instantiated with a map iterator type. |
1796 | template <class T> class MatchStep { |
1797 | public: |
1798 | MatchStep() = delete; |
1799 | |
1800 | MatchStep(T FirstIter, T FirstEnd, T SecondIter, T SecondEnd) |
1801 | : FirstIter(FirstIter), FirstEnd(FirstEnd), SecondIter(SecondIter), |
1802 | SecondEnd(SecondEnd), Status(MS_None) {} |
1803 | |
1804 | bool areBothFinished() const { |
1805 | return (FirstIter == FirstEnd && SecondIter == SecondEnd); |
1806 | } |
1807 | |
1808 | bool isFirstFinished() const { return FirstIter == FirstEnd; } |
1809 | |
1810 | bool isSecondFinished() const { return SecondIter == SecondEnd; } |
1811 | |
1812 | /// Advance one step based on the previous match status unless the previous |
1813 | /// status is MS_None. Then update Status based on the comparison between two |
1814 | /// container iterators at the current step. If the previous status is |
1815 | /// MS_None, it means two iterators are at the beginning and no comparison has |
1816 | /// been made, so we simply update Status without advancing the iterators. |
1817 | void updateOneStep(); |
1818 | |
1819 | T getFirstIter() const { return FirstIter; } |
1820 | |
1821 | T getSecondIter() const { return SecondIter; } |
1822 | |
1823 | MatchStatus getMatchStatus() const { return Status; } |
1824 | |
1825 | private: |
1826 | // Current iterator and end iterator of the first container. |
1827 | T FirstIter; |
1828 | T FirstEnd; |
1829 | // Current iterator and end iterator of the second container. |
1830 | T SecondIter; |
1831 | T SecondEnd; |
1832 | // Match status of the current step. |
1833 | MatchStatus Status; |
1834 | }; |
1835 | } // end anonymous namespace |
1836 | |
1837 | template <class T> void MatchStep<T>::updateOneStep() { |
1838 | switch (Status) { |
1839 | case MS_Match: |
1840 | ++FirstIter; |
1841 | ++SecondIter; |
1842 | break; |
1843 | case MS_FirstUnique: |
1844 | ++FirstIter; |
1845 | break; |
1846 | case MS_SecondUnique: |
1847 | ++SecondIter; |
1848 | break; |
1849 | case MS_None: |
1850 | break; |
1851 | } |
1852 | |
1853 | // Update Status according to iterators at the current step. |
1854 | if (areBothFinished()) |
1855 | return; |
1856 | if (FirstIter != FirstEnd && |
1857 | (SecondIter == SecondEnd || FirstIter->first < SecondIter->first)) |
1858 | Status = MS_FirstUnique; |
1859 | else if (SecondIter != SecondEnd && |
1860 | (FirstIter == FirstEnd || SecondIter->first < FirstIter->first)) |
1861 | Status = MS_SecondUnique; |
1862 | else |
1863 | Status = MS_Match; |
1864 | } |
1865 | |
1866 | // Return the sum of line/block samples, the max line/block sample, and the |
1867 | // number of line/block samples above the given threshold in a function |
1868 | // including its inlinees. |
1869 | static void getFuncSampleStats(const sampleprof::FunctionSamples &Func, |
1870 | FuncSampleStats &FuncStats, |
1871 | uint64_t HotThreshold) { |
1872 | for (const auto &L : Func.getBodySamples()) { |
1873 | uint64_t Sample = L.second.getSamples(); |
1874 | FuncStats.SampleSum += Sample; |
1875 | FuncStats.MaxSample = std::max(a: FuncStats.MaxSample, b: Sample); |
1876 | if (Sample >= HotThreshold) |
1877 | ++FuncStats.HotBlockCount; |
1878 | } |
1879 | |
1880 | for (const auto &C : Func.getCallsiteSamples()) { |
1881 | for (const auto &F : C.second) |
1882 | getFuncSampleStats(Func: F.second, FuncStats, HotThreshold); |
1883 | } |
1884 | } |
1885 | |
1886 | /// Predicate that determines if a function is hot with a given threshold. We |
1887 | /// keep it separate from its callsites for possible extension in the future. |
1888 | static bool isFunctionHot(const FuncSampleStats &FuncStats, |
1889 | uint64_t HotThreshold) { |
1890 | // We intentionally compare the maximum sample count in a function with the |
1891 | // HotThreshold to get an approximate determination on hot functions. |
1892 | return (FuncStats.MaxSample >= HotThreshold); |
1893 | } |
1894 | |
1895 | namespace { |
1896 | class SampleOverlapAggregator { |
1897 | public: |
1898 | SampleOverlapAggregator(const std::string &BaseFilename, |
1899 | const std::string &TestFilename, |
1900 | double LowSimilarityThreshold, double Epsilon, |
1901 | const OverlapFuncFilters &FuncFilter) |
1902 | : BaseFilename(BaseFilename), TestFilename(TestFilename), |
1903 | LowSimilarityThreshold(LowSimilarityThreshold), Epsilon(Epsilon), |
1904 | FuncFilter(FuncFilter) {} |
1905 | |
1906 | /// Detect 0-sample input profile and report to output stream. This interface |
1907 | /// should be called after loadProfiles(). |
1908 | bool detectZeroSampleProfile(raw_fd_ostream &OS) const; |
1909 | |
1910 | /// Write out function-level similarity statistics for functions specified by |
1911 | /// options --function, --value-cutoff, and --similarity-cutoff. |
1912 | void dumpFuncSimilarity(raw_fd_ostream &OS) const; |
1913 | |
1914 | /// Write out program-level similarity and overlap statistics. |
1915 | void dumpProgramSummary(raw_fd_ostream &OS) const; |
1916 | |
1917 | /// Write out hot-function and hot-block statistics for base_profile, |
1918 | /// test_profile, and their overlap. For both cases, the overlap HO is |
1919 | /// calculated as follows: |
1920 | /// Given the number of functions (or blocks) that are hot in both profiles |
1921 | /// HCommon and the number of functions (or blocks) that are hot in at |
1922 | /// least one profile HUnion, HO = HCommon / HUnion. |
1923 | void dumpHotFuncAndBlockOverlap(raw_fd_ostream &OS) const; |
1924 | |
1925 | /// This function tries matching functions in base and test profiles. For each |
1926 | /// pair of matched functions, it aggregates the function-level |
1927 | /// similarity into a profile-level similarity. It also dump function-level |
1928 | /// similarity information of functions specified by --function, |
1929 | /// --value-cutoff, and --similarity-cutoff options. The program-level |
1930 | /// similarity PS is computed as follows: |
1931 | /// Given function-level similarity FS(A) for all function A, the |
1932 | /// weight of function A in base profile WB(A), and the weight of function |
1933 | /// A in test profile WT(A), compute PS(base_profile, test_profile) = |
1934 | /// sum_A(FS(A) * avg(WB(A), WT(A))) ranging in [0.0f to 1.0f] with 0.0 |
1935 | /// meaning no-overlap. |
1936 | void computeSampleProfileOverlap(raw_fd_ostream &OS); |
1937 | |
1938 | /// Initialize ProfOverlap with the sum of samples in base and test |
1939 | /// profiles. This function also computes and keeps the sum of samples and |
1940 | /// max sample counts of each function in BaseStats and TestStats for later |
1941 | /// use to avoid re-computations. |
1942 | void initializeSampleProfileOverlap(); |
1943 | |
1944 | /// Load profiles specified by BaseFilename and TestFilename. |
1945 | std::error_code loadProfiles(); |
1946 | |
1947 | using FuncSampleStatsMap = |
1948 | std::unordered_map<SampleContext, FuncSampleStats, SampleContext::Hash>; |
1949 | |
1950 | private: |
1951 | SampleOverlapStats ProfOverlap; |
1952 | SampleOverlapStats HotFuncOverlap; |
1953 | SampleOverlapStats HotBlockOverlap; |
1954 | std::string BaseFilename; |
1955 | std::string TestFilename; |
1956 | std::unique_ptr<sampleprof::SampleProfileReader> BaseReader; |
1957 | std::unique_ptr<sampleprof::SampleProfileReader> TestReader; |
1958 | // BaseStats and TestStats hold FuncSampleStats for each function, with |
1959 | // function name as the key. |
1960 | FuncSampleStatsMap BaseStats; |
1961 | FuncSampleStatsMap TestStats; |
1962 | // Low similarity threshold in floating point number |
1963 | double LowSimilarityThreshold; |
1964 | // Block samples above BaseHotThreshold or TestHotThreshold are considered hot |
1965 | // for tracking hot blocks. |
1966 | uint64_t BaseHotThreshold; |
1967 | uint64_t TestHotThreshold; |
1968 | // A small threshold used to round the results of floating point accumulations |
1969 | // to resolve imprecision. |
1970 | const double Epsilon; |
1971 | std::multimap<double, SampleOverlapStats, std::greater<double>> |
1972 | FuncSimilarityDump; |
1973 | // FuncFilter carries specifications in options --value-cutoff and |
1974 | // --function. |
1975 | OverlapFuncFilters FuncFilter; |
1976 | // Column offsets for printing the function-level details table. |
1977 | static const unsigned int TestWeightCol = 15; |
1978 | static const unsigned int SimilarityCol = 30; |
1979 | static const unsigned int OverlapCol = 43; |
1980 | static const unsigned int BaseUniqueCol = 53; |
1981 | static const unsigned int TestUniqueCol = 67; |
1982 | static const unsigned int BaseSampleCol = 81; |
1983 | static const unsigned int TestSampleCol = 96; |
1984 | static const unsigned int FuncNameCol = 111; |
1985 | |
1986 | /// Return a similarity of two line/block sample counters in the same |
1987 | /// function in base and test profiles. The line/block-similarity BS(i) is |
1988 | /// computed as follows: |
1989 | /// For an offsets i, given the sample count at i in base profile BB(i), |
1990 | /// the sample count at i in test profile BT(i), the sum of sample counts |
1991 | /// in this function in base profile SB, and the sum of sample counts in |
1992 | /// this function in test profile ST, compute BS(i) = 1.0 - fabs(BB(i)/SB - |
1993 | /// BT(i)/ST), ranging in [0.0f to 1.0f] with 0.0 meaning no-overlap. |
1994 | double computeBlockSimilarity(uint64_t BaseSample, uint64_t TestSample, |
1995 | const SampleOverlapStats &FuncOverlap) const; |
1996 | |
1997 | void updateHotBlockOverlap(uint64_t BaseSample, uint64_t TestSample, |
1998 | uint64_t HotBlockCount); |
1999 | |
2000 | void getHotFunctions(const FuncSampleStatsMap &ProfStats, |
2001 | FuncSampleStatsMap &HotFunc, |
2002 | uint64_t HotThreshold) const; |
2003 | |
2004 | void computeHotFuncOverlap(); |
2005 | |
2006 | /// This function updates statistics in FuncOverlap, HotBlockOverlap, and |
2007 | /// Difference for two sample units in a matched function according to the |
2008 | /// given match status. |
2009 | void updateOverlapStatsForFunction(uint64_t BaseSample, uint64_t TestSample, |
2010 | uint64_t HotBlockCount, |
2011 | SampleOverlapStats &FuncOverlap, |
2012 | double &Difference, MatchStatus Status); |
2013 | |
2014 | /// This function updates statistics in FuncOverlap, HotBlockOverlap, and |
2015 | /// Difference for unmatched callees that only present in one profile in a |
2016 | /// matched caller function. |
2017 | void updateForUnmatchedCallee(const sampleprof::FunctionSamples &Func, |
2018 | SampleOverlapStats &FuncOverlap, |
2019 | double &Difference, MatchStatus Status); |
2020 | |
2021 | /// This function updates sample overlap statistics of an overlap function in |
2022 | /// base and test profile. It also calculates a function-internal similarity |
2023 | /// FIS as follows: |
2024 | /// For offsets i that have samples in at least one profile in this |
2025 | /// function A, given BS(i) returned by computeBlockSimilarity(), compute |
2026 | /// FIS(A) = (2.0 - sum_i(1.0 - BS(i))) / 2, ranging in [0.0f to 1.0f] with |
2027 | /// 0.0 meaning no overlap. |
2028 | double computeSampleFunctionInternalOverlap( |
2029 | const sampleprof::FunctionSamples &BaseFunc, |
2030 | const sampleprof::FunctionSamples &TestFunc, |
2031 | SampleOverlapStats &FuncOverlap); |
2032 | |
2033 | /// Function-level similarity (FS) is a weighted value over function internal |
2034 | /// similarity (FIS). This function computes a function's FS from its FIS by |
2035 | /// applying the weight. |
2036 | double weightForFuncSimilarity(double FuncSimilarity, uint64_t BaseFuncSample, |
2037 | uint64_t TestFuncSample) const; |
2038 | |
2039 | /// The function-level similarity FS(A) for a function A is computed as |
2040 | /// follows: |
2041 | /// Compute a function-internal similarity FIS(A) by |
2042 | /// computeSampleFunctionInternalOverlap(). Then, with the weight of |
2043 | /// function A in base profile WB(A), and the weight of function A in test |
2044 | /// profile WT(A), compute FS(A) = FIS(A) * (1.0 - fabs(WB(A) - WT(A))) |
2045 | /// ranging in [0.0f to 1.0f] with 0.0 meaning no overlap. |
2046 | double |
2047 | computeSampleFunctionOverlap(const sampleprof::FunctionSamples *BaseFunc, |
2048 | const sampleprof::FunctionSamples *TestFunc, |
2049 | SampleOverlapStats *FuncOverlap, |
2050 | uint64_t BaseFuncSample, |
2051 | uint64_t TestFuncSample); |
2052 | |
2053 | /// Profile-level similarity (PS) is a weighted aggregate over function-level |
2054 | /// similarities (FS). This method weights the FS value by the function |
2055 | /// weights in the base and test profiles for the aggregation. |
2056 | double weightByImportance(double FuncSimilarity, uint64_t BaseFuncSample, |
2057 | uint64_t TestFuncSample) const; |
2058 | }; |
2059 | } // end anonymous namespace |
2060 | |
2061 | bool SampleOverlapAggregator::detectZeroSampleProfile( |
2062 | raw_fd_ostream &OS) const { |
2063 | bool HaveZeroSample = false; |
2064 | if (ProfOverlap.BaseSample == 0) { |
2065 | OS << "Sum of sample counts for profile " << BaseFilename << " is 0.\n" ; |
2066 | HaveZeroSample = true; |
2067 | } |
2068 | if (ProfOverlap.TestSample == 0) { |
2069 | OS << "Sum of sample counts for profile " << TestFilename << " is 0.\n" ; |
2070 | HaveZeroSample = true; |
2071 | } |
2072 | return HaveZeroSample; |
2073 | } |
2074 | |
2075 | double SampleOverlapAggregator::computeBlockSimilarity( |
2076 | uint64_t BaseSample, uint64_t TestSample, |
2077 | const SampleOverlapStats &FuncOverlap) const { |
2078 | double BaseFrac = 0.0; |
2079 | double TestFrac = 0.0; |
2080 | if (FuncOverlap.BaseSample > 0) |
2081 | BaseFrac = static_cast<double>(BaseSample) / FuncOverlap.BaseSample; |
2082 | if (FuncOverlap.TestSample > 0) |
2083 | TestFrac = static_cast<double>(TestSample) / FuncOverlap.TestSample; |
2084 | return 1.0 - std::fabs(x: BaseFrac - TestFrac); |
2085 | } |
2086 | |
2087 | void SampleOverlapAggregator::updateHotBlockOverlap(uint64_t BaseSample, |
2088 | uint64_t TestSample, |
2089 | uint64_t HotBlockCount) { |
2090 | bool IsBaseHot = (BaseSample >= BaseHotThreshold); |
2091 | bool IsTestHot = (TestSample >= TestHotThreshold); |
2092 | if (!IsBaseHot && !IsTestHot) |
2093 | return; |
2094 | |
2095 | HotBlockOverlap.UnionCount += HotBlockCount; |
2096 | if (IsBaseHot) |
2097 | HotBlockOverlap.BaseCount += HotBlockCount; |
2098 | if (IsTestHot) |
2099 | HotBlockOverlap.TestCount += HotBlockCount; |
2100 | if (IsBaseHot && IsTestHot) |
2101 | HotBlockOverlap.OverlapCount += HotBlockCount; |
2102 | } |
2103 | |
2104 | void SampleOverlapAggregator::getHotFunctions( |
2105 | const FuncSampleStatsMap &ProfStats, FuncSampleStatsMap &HotFunc, |
2106 | uint64_t HotThreshold) const { |
2107 | for (const auto &F : ProfStats) { |
2108 | if (isFunctionHot(FuncStats: F.second, HotThreshold)) |
2109 | HotFunc.emplace(args: F.first, args: F.second); |
2110 | } |
2111 | } |
2112 | |
2113 | void SampleOverlapAggregator::computeHotFuncOverlap() { |
2114 | FuncSampleStatsMap BaseHotFunc; |
2115 | getHotFunctions(ProfStats: BaseStats, HotFunc&: BaseHotFunc, HotThreshold: BaseHotThreshold); |
2116 | HotFuncOverlap.BaseCount = BaseHotFunc.size(); |
2117 | |
2118 | FuncSampleStatsMap TestHotFunc; |
2119 | getHotFunctions(ProfStats: TestStats, HotFunc&: TestHotFunc, HotThreshold: TestHotThreshold); |
2120 | HotFuncOverlap.TestCount = TestHotFunc.size(); |
2121 | HotFuncOverlap.UnionCount = HotFuncOverlap.TestCount; |
2122 | |
2123 | for (const auto &F : BaseHotFunc) { |
2124 | if (TestHotFunc.count(x: F.first)) |
2125 | ++HotFuncOverlap.OverlapCount; |
2126 | else |
2127 | ++HotFuncOverlap.UnionCount; |
2128 | } |
2129 | } |
2130 | |
2131 | void SampleOverlapAggregator::updateOverlapStatsForFunction( |
2132 | uint64_t BaseSample, uint64_t TestSample, uint64_t HotBlockCount, |
2133 | SampleOverlapStats &FuncOverlap, double &Difference, MatchStatus Status) { |
2134 | assert(Status != MS_None && |
2135 | "Match status should be updated before updating overlap statistics" ); |
2136 | if (Status == MS_FirstUnique) { |
2137 | TestSample = 0; |
2138 | FuncOverlap.BaseUniqueSample += BaseSample; |
2139 | } else if (Status == MS_SecondUnique) { |
2140 | BaseSample = 0; |
2141 | FuncOverlap.TestUniqueSample += TestSample; |
2142 | } else { |
2143 | ++FuncOverlap.OverlapCount; |
2144 | } |
2145 | |
2146 | FuncOverlap.UnionSample += std::max(a: BaseSample, b: TestSample); |
2147 | FuncOverlap.OverlapSample += std::min(a: BaseSample, b: TestSample); |
2148 | Difference += |
2149 | 1.0 - computeBlockSimilarity(BaseSample, TestSample, FuncOverlap); |
2150 | updateHotBlockOverlap(BaseSample, TestSample, HotBlockCount); |
2151 | } |
2152 | |
2153 | void SampleOverlapAggregator::updateForUnmatchedCallee( |
2154 | const sampleprof::FunctionSamples &Func, SampleOverlapStats &FuncOverlap, |
2155 | double &Difference, MatchStatus Status) { |
2156 | assert((Status == MS_FirstUnique || Status == MS_SecondUnique) && |
2157 | "Status must be either of the two unmatched cases" ); |
2158 | FuncSampleStats FuncStats; |
2159 | if (Status == MS_FirstUnique) { |
2160 | getFuncSampleStats(Func, FuncStats, HotThreshold: BaseHotThreshold); |
2161 | updateOverlapStatsForFunction(BaseSample: FuncStats.SampleSum, TestSample: 0, |
2162 | HotBlockCount: FuncStats.HotBlockCount, FuncOverlap, |
2163 | Difference, Status); |
2164 | } else { |
2165 | getFuncSampleStats(Func, FuncStats, HotThreshold: TestHotThreshold); |
2166 | updateOverlapStatsForFunction(BaseSample: 0, TestSample: FuncStats.SampleSum, |
2167 | HotBlockCount: FuncStats.HotBlockCount, FuncOverlap, |
2168 | Difference, Status); |
2169 | } |
2170 | } |
2171 | |
2172 | double SampleOverlapAggregator::computeSampleFunctionInternalOverlap( |
2173 | const sampleprof::FunctionSamples &BaseFunc, |
2174 | const sampleprof::FunctionSamples &TestFunc, |
2175 | SampleOverlapStats &FuncOverlap) { |
2176 | |
2177 | using namespace sampleprof; |
2178 | |
2179 | double Difference = 0; |
2180 | |
2181 | // Accumulate Difference for regular line/block samples in the function. |
2182 | // We match them through sort-merge join algorithm because |
2183 | // FunctionSamples::getBodySamples() returns a map of sample counters ordered |
2184 | // by their offsets. |
2185 | MatchStep<BodySampleMap::const_iterator> BlockIterStep( |
2186 | BaseFunc.getBodySamples().cbegin(), BaseFunc.getBodySamples().cend(), |
2187 | TestFunc.getBodySamples().cbegin(), TestFunc.getBodySamples().cend()); |
2188 | BlockIterStep.updateOneStep(); |
2189 | while (!BlockIterStep.areBothFinished()) { |
2190 | uint64_t BaseSample = |
2191 | BlockIterStep.isFirstFinished() |
2192 | ? 0 |
2193 | : BlockIterStep.getFirstIter()->second.getSamples(); |
2194 | uint64_t TestSample = |
2195 | BlockIterStep.isSecondFinished() |
2196 | ? 0 |
2197 | : BlockIterStep.getSecondIter()->second.getSamples(); |
2198 | updateOverlapStatsForFunction(BaseSample, TestSample, HotBlockCount: 1, FuncOverlap, |
2199 | Difference, Status: BlockIterStep.getMatchStatus()); |
2200 | |
2201 | BlockIterStep.updateOneStep(); |
2202 | } |
2203 | |
2204 | // Accumulate Difference for callsite lines in the function. We match |
2205 | // them through sort-merge algorithm because |
2206 | // FunctionSamples::getCallsiteSamples() returns a map of callsite records |
2207 | // ordered by their offsets. |
2208 | MatchStep<CallsiteSampleMap::const_iterator> CallsiteIterStep( |
2209 | BaseFunc.getCallsiteSamples().cbegin(), |
2210 | BaseFunc.getCallsiteSamples().cend(), |
2211 | TestFunc.getCallsiteSamples().cbegin(), |
2212 | TestFunc.getCallsiteSamples().cend()); |
2213 | CallsiteIterStep.updateOneStep(); |
2214 | while (!CallsiteIterStep.areBothFinished()) { |
2215 | MatchStatus CallsiteStepStatus = CallsiteIterStep.getMatchStatus(); |
2216 | assert(CallsiteStepStatus != MS_None && |
2217 | "Match status should be updated before entering loop body" ); |
2218 | |
2219 | if (CallsiteStepStatus != MS_Match) { |
2220 | auto Callsite = (CallsiteStepStatus == MS_FirstUnique) |
2221 | ? CallsiteIterStep.getFirstIter() |
2222 | : CallsiteIterStep.getSecondIter(); |
2223 | for (const auto &F : Callsite->second) |
2224 | updateForUnmatchedCallee(Func: F.second, FuncOverlap, Difference, |
2225 | Status: CallsiteStepStatus); |
2226 | } else { |
2227 | // There may be multiple inlinees at the same offset, so we need to try |
2228 | // matching all of them. This match is implemented through sort-merge |
2229 | // algorithm because callsite records at the same offset are ordered by |
2230 | // function names. |
2231 | MatchStep<FunctionSamplesMap::const_iterator> CalleeIterStep( |
2232 | CallsiteIterStep.getFirstIter()->second.cbegin(), |
2233 | CallsiteIterStep.getFirstIter()->second.cend(), |
2234 | CallsiteIterStep.getSecondIter()->second.cbegin(), |
2235 | CallsiteIterStep.getSecondIter()->second.cend()); |
2236 | CalleeIterStep.updateOneStep(); |
2237 | while (!CalleeIterStep.areBothFinished()) { |
2238 | MatchStatus CalleeStepStatus = CalleeIterStep.getMatchStatus(); |
2239 | if (CalleeStepStatus != MS_Match) { |
2240 | auto Callee = (CalleeStepStatus == MS_FirstUnique) |
2241 | ? CalleeIterStep.getFirstIter() |
2242 | : CalleeIterStep.getSecondIter(); |
2243 | updateForUnmatchedCallee(Func: Callee->second, FuncOverlap, Difference, |
2244 | Status: CalleeStepStatus); |
2245 | } else { |
2246 | // An inlined function can contain other inlinees inside, so compute |
2247 | // the Difference recursively. |
2248 | Difference += 2.0 - 2 * computeSampleFunctionInternalOverlap( |
2249 | BaseFunc: CalleeIterStep.getFirstIter()->second, |
2250 | TestFunc: CalleeIterStep.getSecondIter()->second, |
2251 | FuncOverlap); |
2252 | } |
2253 | CalleeIterStep.updateOneStep(); |
2254 | } |
2255 | } |
2256 | CallsiteIterStep.updateOneStep(); |
2257 | } |
2258 | |
2259 | // Difference reflects the total differences of line/block samples in this |
2260 | // function and ranges in [0.0f to 2.0f]. Take (2.0 - Difference) / 2 to |
2261 | // reflect the similarity between function profiles in [0.0f to 1.0f]. |
2262 | return (2.0 - Difference) / 2; |
2263 | } |
2264 | |
2265 | double SampleOverlapAggregator::weightForFuncSimilarity( |
2266 | double FuncInternalSimilarity, uint64_t BaseFuncSample, |
2267 | uint64_t TestFuncSample) const { |
2268 | // Compute the weight as the distance between the function weights in two |
2269 | // profiles. |
2270 | double BaseFrac = 0.0; |
2271 | double TestFrac = 0.0; |
2272 | assert(ProfOverlap.BaseSample > 0 && |
2273 | "Total samples in base profile should be greater than 0" ); |
2274 | BaseFrac = static_cast<double>(BaseFuncSample) / ProfOverlap.BaseSample; |
2275 | assert(ProfOverlap.TestSample > 0 && |
2276 | "Total samples in test profile should be greater than 0" ); |
2277 | TestFrac = static_cast<double>(TestFuncSample) / ProfOverlap.TestSample; |
2278 | double WeightDistance = std::fabs(x: BaseFrac - TestFrac); |
2279 | |
2280 | // Take WeightDistance into the similarity. |
2281 | return FuncInternalSimilarity * (1 - WeightDistance); |
2282 | } |
2283 | |
2284 | double |
2285 | SampleOverlapAggregator::weightByImportance(double FuncSimilarity, |
2286 | uint64_t BaseFuncSample, |
2287 | uint64_t TestFuncSample) const { |
2288 | |
2289 | double BaseFrac = 0.0; |
2290 | double TestFrac = 0.0; |
2291 | assert(ProfOverlap.BaseSample > 0 && |
2292 | "Total samples in base profile should be greater than 0" ); |
2293 | BaseFrac = static_cast<double>(BaseFuncSample) / ProfOverlap.BaseSample / 2.0; |
2294 | assert(ProfOverlap.TestSample > 0 && |
2295 | "Total samples in test profile should be greater than 0" ); |
2296 | TestFrac = static_cast<double>(TestFuncSample) / ProfOverlap.TestSample / 2.0; |
2297 | return FuncSimilarity * (BaseFrac + TestFrac); |
2298 | } |
2299 | |
2300 | double SampleOverlapAggregator::computeSampleFunctionOverlap( |
2301 | const sampleprof::FunctionSamples *BaseFunc, |
2302 | const sampleprof::FunctionSamples *TestFunc, |
2303 | SampleOverlapStats *FuncOverlap, uint64_t BaseFuncSample, |
2304 | uint64_t TestFuncSample) { |
2305 | // Default function internal similarity before weighted, meaning two functions |
2306 | // has no overlap. |
2307 | const double DefaultFuncInternalSimilarity = 0; |
2308 | double FuncSimilarity; |
2309 | double FuncInternalSimilarity; |
2310 | |
2311 | // If BaseFunc or TestFunc is nullptr, it means the functions do not overlap. |
2312 | // In this case, we use DefaultFuncInternalSimilarity as the function internal |
2313 | // similarity. |
2314 | if (!BaseFunc || !TestFunc) { |
2315 | FuncInternalSimilarity = DefaultFuncInternalSimilarity; |
2316 | } else { |
2317 | assert(FuncOverlap != nullptr && |
2318 | "FuncOverlap should be provided in this case" ); |
2319 | FuncInternalSimilarity = computeSampleFunctionInternalOverlap( |
2320 | BaseFunc: *BaseFunc, TestFunc: *TestFunc, FuncOverlap&: *FuncOverlap); |
2321 | // Now, FuncInternalSimilarity may be a little less than 0 due to |
2322 | // imprecision of floating point accumulations. Make it zero if the |
2323 | // difference is below Epsilon. |
2324 | FuncInternalSimilarity = (std::fabs(x: FuncInternalSimilarity - 0) < Epsilon) |
2325 | ? 0 |
2326 | : FuncInternalSimilarity; |
2327 | } |
2328 | FuncSimilarity = weightForFuncSimilarity(FuncInternalSimilarity, |
2329 | BaseFuncSample, TestFuncSample); |
2330 | return FuncSimilarity; |
2331 | } |
2332 | |
2333 | void SampleOverlapAggregator::computeSampleProfileOverlap(raw_fd_ostream &OS) { |
2334 | using namespace sampleprof; |
2335 | |
2336 | std::unordered_map<SampleContext, const FunctionSamples *, |
2337 | SampleContext::Hash> |
2338 | BaseFuncProf; |
2339 | const auto &BaseProfiles = BaseReader->getProfiles(); |
2340 | for (const auto &BaseFunc : BaseProfiles) { |
2341 | BaseFuncProf.emplace(args&: BaseFunc.second.getContext(), args: &(BaseFunc.second)); |
2342 | } |
2343 | ProfOverlap.UnionCount = BaseFuncProf.size(); |
2344 | |
2345 | const auto &TestProfiles = TestReader->getProfiles(); |
2346 | for (const auto &TestFunc : TestProfiles) { |
2347 | SampleOverlapStats FuncOverlap; |
2348 | FuncOverlap.TestName = TestFunc.second.getContext(); |
2349 | assert(TestStats.count(FuncOverlap.TestName) && |
2350 | "TestStats should have records for all functions in test profile " |
2351 | "except inlinees" ); |
2352 | FuncOverlap.TestSample = TestStats[FuncOverlap.TestName].SampleSum; |
2353 | |
2354 | bool Matched = false; |
2355 | const auto Match = BaseFuncProf.find(x: FuncOverlap.TestName); |
2356 | if (Match == BaseFuncProf.end()) { |
2357 | const FuncSampleStats &FuncStats = TestStats[FuncOverlap.TestName]; |
2358 | ++ProfOverlap.TestUniqueCount; |
2359 | ProfOverlap.TestUniqueSample += FuncStats.SampleSum; |
2360 | FuncOverlap.TestUniqueSample = FuncStats.SampleSum; |
2361 | |
2362 | updateHotBlockOverlap(BaseSample: 0, TestSample: FuncStats.SampleSum, HotBlockCount: FuncStats.HotBlockCount); |
2363 | |
2364 | double FuncSimilarity = computeSampleFunctionOverlap( |
2365 | BaseFunc: nullptr, TestFunc: nullptr, FuncOverlap: nullptr, BaseFuncSample: 0, TestFuncSample: FuncStats.SampleSum); |
2366 | ProfOverlap.Similarity += |
2367 | weightByImportance(FuncSimilarity, BaseFuncSample: 0, TestFuncSample: FuncStats.SampleSum); |
2368 | |
2369 | ++ProfOverlap.UnionCount; |
2370 | ProfOverlap.UnionSample += FuncStats.SampleSum; |
2371 | } else { |
2372 | ++ProfOverlap.OverlapCount; |
2373 | |
2374 | // Two functions match with each other. Compute function-level overlap and |
2375 | // aggregate them into profile-level overlap. |
2376 | FuncOverlap.BaseName = Match->second->getContext(); |
2377 | assert(BaseStats.count(FuncOverlap.BaseName) && |
2378 | "BaseStats should have records for all functions in base profile " |
2379 | "except inlinees" ); |
2380 | FuncOverlap.BaseSample = BaseStats[FuncOverlap.BaseName].SampleSum; |
2381 | |
2382 | FuncOverlap.Similarity = computeSampleFunctionOverlap( |
2383 | BaseFunc: Match->second, TestFunc: &TestFunc.second, FuncOverlap: &FuncOverlap, BaseFuncSample: FuncOverlap.BaseSample, |
2384 | TestFuncSample: FuncOverlap.TestSample); |
2385 | ProfOverlap.Similarity += |
2386 | weightByImportance(FuncSimilarity: FuncOverlap.Similarity, BaseFuncSample: FuncOverlap.BaseSample, |
2387 | TestFuncSample: FuncOverlap.TestSample); |
2388 | ProfOverlap.OverlapSample += FuncOverlap.OverlapSample; |
2389 | ProfOverlap.UnionSample += FuncOverlap.UnionSample; |
2390 | |
2391 | // Accumulate the percentage of base unique and test unique samples into |
2392 | // ProfOverlap. |
2393 | ProfOverlap.BaseUniqueSample += FuncOverlap.BaseUniqueSample; |
2394 | ProfOverlap.TestUniqueSample += FuncOverlap.TestUniqueSample; |
2395 | |
2396 | // Remove matched base functions for later reporting functions not found |
2397 | // in test profile. |
2398 | BaseFuncProf.erase(position: Match); |
2399 | Matched = true; |
2400 | } |
2401 | |
2402 | // Print function-level similarity information if specified by options. |
2403 | assert(TestStats.count(FuncOverlap.TestName) && |
2404 | "TestStats should have records for all functions in test profile " |
2405 | "except inlinees" ); |
2406 | if (TestStats[FuncOverlap.TestName].MaxSample >= FuncFilter.ValueCutoff || |
2407 | (Matched && FuncOverlap.Similarity < LowSimilarityThreshold) || |
2408 | (Matched && !FuncFilter.NameFilter.empty() && |
2409 | FuncOverlap.BaseName.toString().find(str: FuncFilter.NameFilter) != |
2410 | std::string::npos)) { |
2411 | assert(ProfOverlap.BaseSample > 0 && |
2412 | "Total samples in base profile should be greater than 0" ); |
2413 | FuncOverlap.BaseWeight = |
2414 | static_cast<double>(FuncOverlap.BaseSample) / ProfOverlap.BaseSample; |
2415 | assert(ProfOverlap.TestSample > 0 && |
2416 | "Total samples in test profile should be greater than 0" ); |
2417 | FuncOverlap.TestWeight = |
2418 | static_cast<double>(FuncOverlap.TestSample) / ProfOverlap.TestSample; |
2419 | FuncSimilarityDump.emplace(args&: FuncOverlap.BaseWeight, args&: FuncOverlap); |
2420 | } |
2421 | } |
2422 | |
2423 | // Traverse through functions in base profile but not in test profile. |
2424 | for (const auto &F : BaseFuncProf) { |
2425 | assert(BaseStats.count(F.second->getContext()) && |
2426 | "BaseStats should have records for all functions in base profile " |
2427 | "except inlinees" ); |
2428 | const FuncSampleStats &FuncStats = BaseStats[F.second->getContext()]; |
2429 | ++ProfOverlap.BaseUniqueCount; |
2430 | ProfOverlap.BaseUniqueSample += FuncStats.SampleSum; |
2431 | |
2432 | updateHotBlockOverlap(BaseSample: FuncStats.SampleSum, TestSample: 0, HotBlockCount: FuncStats.HotBlockCount); |
2433 | |
2434 | double FuncSimilarity = computeSampleFunctionOverlap( |
2435 | BaseFunc: nullptr, TestFunc: nullptr, FuncOverlap: nullptr, BaseFuncSample: FuncStats.SampleSum, TestFuncSample: 0); |
2436 | ProfOverlap.Similarity += |
2437 | weightByImportance(FuncSimilarity, BaseFuncSample: FuncStats.SampleSum, TestFuncSample: 0); |
2438 | |
2439 | ProfOverlap.UnionSample += FuncStats.SampleSum; |
2440 | } |
2441 | |
2442 | // Now, ProfSimilarity may be a little greater than 1 due to imprecision |
2443 | // of floating point accumulations. Make it 1.0 if the difference is below |
2444 | // Epsilon. |
2445 | ProfOverlap.Similarity = (std::fabs(x: ProfOverlap.Similarity - 1) < Epsilon) |
2446 | ? 1 |
2447 | : ProfOverlap.Similarity; |
2448 | |
2449 | computeHotFuncOverlap(); |
2450 | } |
2451 | |
2452 | void SampleOverlapAggregator::initializeSampleProfileOverlap() { |
2453 | const auto &BaseProf = BaseReader->getProfiles(); |
2454 | for (const auto &I : BaseProf) { |
2455 | ++ProfOverlap.BaseCount; |
2456 | FuncSampleStats FuncStats; |
2457 | getFuncSampleStats(Func: I.second, FuncStats, HotThreshold: BaseHotThreshold); |
2458 | ProfOverlap.BaseSample += FuncStats.SampleSum; |
2459 | BaseStats.emplace(args&: I.second.getContext(), args&: FuncStats); |
2460 | } |
2461 | |
2462 | const auto &TestProf = TestReader->getProfiles(); |
2463 | for (const auto &I : TestProf) { |
2464 | ++ProfOverlap.TestCount; |
2465 | FuncSampleStats FuncStats; |
2466 | getFuncSampleStats(Func: I.second, FuncStats, HotThreshold: TestHotThreshold); |
2467 | ProfOverlap.TestSample += FuncStats.SampleSum; |
2468 | TestStats.emplace(args&: I.second.getContext(), args&: FuncStats); |
2469 | } |
2470 | |
2471 | ProfOverlap.BaseName = StringRef(BaseFilename); |
2472 | ProfOverlap.TestName = StringRef(TestFilename); |
2473 | } |
2474 | |
2475 | void SampleOverlapAggregator::dumpFuncSimilarity(raw_fd_ostream &OS) const { |
2476 | using namespace sampleprof; |
2477 | |
2478 | if (FuncSimilarityDump.empty()) |
2479 | return; |
2480 | |
2481 | formatted_raw_ostream FOS(OS); |
2482 | FOS << "Function-level details:\n" ; |
2483 | FOS << "Base weight" ; |
2484 | FOS.PadToColumn(NewCol: TestWeightCol); |
2485 | FOS << "Test weight" ; |
2486 | FOS.PadToColumn(NewCol: SimilarityCol); |
2487 | FOS << "Similarity" ; |
2488 | FOS.PadToColumn(NewCol: OverlapCol); |
2489 | FOS << "Overlap" ; |
2490 | FOS.PadToColumn(NewCol: BaseUniqueCol); |
2491 | FOS << "Base unique" ; |
2492 | FOS.PadToColumn(NewCol: TestUniqueCol); |
2493 | FOS << "Test unique" ; |
2494 | FOS.PadToColumn(NewCol: BaseSampleCol); |
2495 | FOS << "Base samples" ; |
2496 | FOS.PadToColumn(NewCol: TestSampleCol); |
2497 | FOS << "Test samples" ; |
2498 | FOS.PadToColumn(NewCol: FuncNameCol); |
2499 | FOS << "Function name\n" ; |
2500 | for (const auto &F : FuncSimilarityDump) { |
2501 | double OverlapPercent = |
2502 | F.second.UnionSample > 0 |
2503 | ? static_cast<double>(F.second.OverlapSample) / F.second.UnionSample |
2504 | : 0; |
2505 | double BaseUniquePercent = |
2506 | F.second.BaseSample > 0 |
2507 | ? static_cast<double>(F.second.BaseUniqueSample) / |
2508 | F.second.BaseSample |
2509 | : 0; |
2510 | double TestUniquePercent = |
2511 | F.second.TestSample > 0 |
2512 | ? static_cast<double>(F.second.TestUniqueSample) / |
2513 | F.second.TestSample |
2514 | : 0; |
2515 | |
2516 | FOS << format(Fmt: "%.2f%%" , Vals: F.second.BaseWeight * 100); |
2517 | FOS.PadToColumn(NewCol: TestWeightCol); |
2518 | FOS << format(Fmt: "%.2f%%" , Vals: F.second.TestWeight * 100); |
2519 | FOS.PadToColumn(NewCol: SimilarityCol); |
2520 | FOS << format(Fmt: "%.2f%%" , Vals: F.second.Similarity * 100); |
2521 | FOS.PadToColumn(NewCol: OverlapCol); |
2522 | FOS << format(Fmt: "%.2f%%" , Vals: OverlapPercent * 100); |
2523 | FOS.PadToColumn(NewCol: BaseUniqueCol); |
2524 | FOS << format(Fmt: "%.2f%%" , Vals: BaseUniquePercent * 100); |
2525 | FOS.PadToColumn(NewCol: TestUniqueCol); |
2526 | FOS << format(Fmt: "%.2f%%" , Vals: TestUniquePercent * 100); |
2527 | FOS.PadToColumn(NewCol: BaseSampleCol); |
2528 | FOS << F.second.BaseSample; |
2529 | FOS.PadToColumn(NewCol: TestSampleCol); |
2530 | FOS << F.second.TestSample; |
2531 | FOS.PadToColumn(NewCol: FuncNameCol); |
2532 | FOS << F.second.TestName.toString() << "\n" ; |
2533 | } |
2534 | } |
2535 | |
2536 | void SampleOverlapAggregator::dumpProgramSummary(raw_fd_ostream &OS) const { |
2537 | OS << "Profile overlap infomation for base_profile: " |
2538 | << ProfOverlap.BaseName.toString() |
2539 | << " and test_profile: " << ProfOverlap.TestName.toString() |
2540 | << "\nProgram level:\n" ; |
2541 | |
2542 | OS << " Whole program profile similarity: " |
2543 | << format(Fmt: "%.3f%%" , Vals: ProfOverlap.Similarity * 100) << "\n" ; |
2544 | |
2545 | assert(ProfOverlap.UnionSample > 0 && |
2546 | "Total samples in two profile should be greater than 0" ); |
2547 | double OverlapPercent = |
2548 | static_cast<double>(ProfOverlap.OverlapSample) / ProfOverlap.UnionSample; |
2549 | assert(ProfOverlap.BaseSample > 0 && |
2550 | "Total samples in base profile should be greater than 0" ); |
2551 | double BaseUniquePercent = static_cast<double>(ProfOverlap.BaseUniqueSample) / |
2552 | ProfOverlap.BaseSample; |
2553 | assert(ProfOverlap.TestSample > 0 && |
2554 | "Total samples in test profile should be greater than 0" ); |
2555 | double TestUniquePercent = static_cast<double>(ProfOverlap.TestUniqueSample) / |
2556 | ProfOverlap.TestSample; |
2557 | |
2558 | OS << " Whole program sample overlap: " |
2559 | << format(Fmt: "%.3f%%" , Vals: OverlapPercent * 100) << "\n" ; |
2560 | OS << " percentage of samples unique in base profile: " |
2561 | << format(Fmt: "%.3f%%" , Vals: BaseUniquePercent * 100) << "\n" ; |
2562 | OS << " percentage of samples unique in test profile: " |
2563 | << format(Fmt: "%.3f%%" , Vals: TestUniquePercent * 100) << "\n" ; |
2564 | OS << " total samples in base profile: " << ProfOverlap.BaseSample << "\n" |
2565 | << " total samples in test profile: " << ProfOverlap.TestSample << "\n" ; |
2566 | |
2567 | assert(ProfOverlap.UnionCount > 0 && |
2568 | "There should be at least one function in two input profiles" ); |
2569 | double FuncOverlapPercent = |
2570 | static_cast<double>(ProfOverlap.OverlapCount) / ProfOverlap.UnionCount; |
2571 | OS << " Function overlap: " << format(Fmt: "%.3f%%" , Vals: FuncOverlapPercent * 100) |
2572 | << "\n" ; |
2573 | OS << " overlap functions: " << ProfOverlap.OverlapCount << "\n" ; |
2574 | OS << " functions unique in base profile: " << ProfOverlap.BaseUniqueCount |
2575 | << "\n" ; |
2576 | OS << " functions unique in test profile: " << ProfOverlap.TestUniqueCount |
2577 | << "\n" ; |
2578 | } |
2579 | |
2580 | void SampleOverlapAggregator::dumpHotFuncAndBlockOverlap( |
2581 | raw_fd_ostream &OS) const { |
2582 | assert(HotFuncOverlap.UnionCount > 0 && |
2583 | "There should be at least one hot function in two input profiles" ); |
2584 | OS << " Hot-function overlap: " |
2585 | << format(Fmt: "%.3f%%" , Vals: static_cast<double>(HotFuncOverlap.OverlapCount) / |
2586 | HotFuncOverlap.UnionCount * 100) |
2587 | << "\n" ; |
2588 | OS << " overlap hot functions: " << HotFuncOverlap.OverlapCount << "\n" ; |
2589 | OS << " hot functions unique in base profile: " |
2590 | << HotFuncOverlap.BaseCount - HotFuncOverlap.OverlapCount << "\n" ; |
2591 | OS << " hot functions unique in test profile: " |
2592 | << HotFuncOverlap.TestCount - HotFuncOverlap.OverlapCount << "\n" ; |
2593 | |
2594 | assert(HotBlockOverlap.UnionCount > 0 && |
2595 | "There should be at least one hot block in two input profiles" ); |
2596 | OS << " Hot-block overlap: " |
2597 | << format(Fmt: "%.3f%%" , Vals: static_cast<double>(HotBlockOverlap.OverlapCount) / |
2598 | HotBlockOverlap.UnionCount * 100) |
2599 | << "\n" ; |
2600 | OS << " overlap hot blocks: " << HotBlockOverlap.OverlapCount << "\n" ; |
2601 | OS << " hot blocks unique in base profile: " |
2602 | << HotBlockOverlap.BaseCount - HotBlockOverlap.OverlapCount << "\n" ; |
2603 | OS << " hot blocks unique in test profile: " |
2604 | << HotBlockOverlap.TestCount - HotBlockOverlap.OverlapCount << "\n" ; |
2605 | } |
2606 | |
2607 | std::error_code SampleOverlapAggregator::loadProfiles() { |
2608 | using namespace sampleprof; |
2609 | |
2610 | LLVMContext Context; |
2611 | auto FS = vfs::getRealFileSystem(); |
2612 | auto BaseReaderOrErr = SampleProfileReader::create(Filename: BaseFilename, C&: Context, FS&: *FS, |
2613 | P: FSDiscriminatorPassOption); |
2614 | if (std::error_code EC = BaseReaderOrErr.getError()) |
2615 | exitWithErrorCode(EC, Whence: BaseFilename); |
2616 | |
2617 | auto TestReaderOrErr = SampleProfileReader::create(Filename: TestFilename, C&: Context, FS&: *FS, |
2618 | P: FSDiscriminatorPassOption); |
2619 | if (std::error_code EC = TestReaderOrErr.getError()) |
2620 | exitWithErrorCode(EC, Whence: TestFilename); |
2621 | |
2622 | BaseReader = std::move(BaseReaderOrErr.get()); |
2623 | TestReader = std::move(TestReaderOrErr.get()); |
2624 | |
2625 | if (std::error_code EC = BaseReader->read()) |
2626 | exitWithErrorCode(EC, Whence: BaseFilename); |
2627 | if (std::error_code EC = TestReader->read()) |
2628 | exitWithErrorCode(EC, Whence: TestFilename); |
2629 | if (BaseReader->profileIsProbeBased() != TestReader->profileIsProbeBased()) |
2630 | exitWithError( |
2631 | Message: "cannot compare probe-based profile with non-probe-based profile" ); |
2632 | if (BaseReader->profileIsCS() != TestReader->profileIsCS()) |
2633 | exitWithError(Message: "cannot compare CS profile with non-CS profile" ); |
2634 | |
2635 | // Load BaseHotThreshold and TestHotThreshold as 99-percentile threshold in |
2636 | // profile summary. |
2637 | ProfileSummary &BasePS = BaseReader->getSummary(); |
2638 | ProfileSummary &TestPS = TestReader->getSummary(); |
2639 | BaseHotThreshold = |
2640 | ProfileSummaryBuilder::getHotCountThreshold(DS: BasePS.getDetailedSummary()); |
2641 | TestHotThreshold = |
2642 | ProfileSummaryBuilder::getHotCountThreshold(DS: TestPS.getDetailedSummary()); |
2643 | |
2644 | return std::error_code(); |
2645 | } |
2646 | |
2647 | void overlapSampleProfile(const std::string &BaseFilename, |
2648 | const std::string &TestFilename, |
2649 | const OverlapFuncFilters &FuncFilter, |
2650 | uint64_t SimilarityCutoff, raw_fd_ostream &OS) { |
2651 | using namespace sampleprof; |
2652 | |
2653 | // We use 0.000005 to initialize OverlapAggr.Epsilon because the final metrics |
2654 | // report 2--3 places after decimal point in percentage numbers. |
2655 | SampleOverlapAggregator OverlapAggr( |
2656 | BaseFilename, TestFilename, |
2657 | static_cast<double>(SimilarityCutoff) / 1000000, 0.000005, FuncFilter); |
2658 | if (std::error_code EC = OverlapAggr.loadProfiles()) |
2659 | exitWithErrorCode(EC); |
2660 | |
2661 | OverlapAggr.initializeSampleProfileOverlap(); |
2662 | if (OverlapAggr.detectZeroSampleProfile(OS)) |
2663 | return; |
2664 | |
2665 | OverlapAggr.computeSampleProfileOverlap(OS); |
2666 | |
2667 | OverlapAggr.dumpProgramSummary(OS); |
2668 | OverlapAggr.dumpHotFuncAndBlockOverlap(OS); |
2669 | OverlapAggr.dumpFuncSimilarity(OS); |
2670 | } |
2671 | |
2672 | static int overlap_main() { |
2673 | std::error_code EC; |
2674 | raw_fd_ostream OS(OutputFilename.data(), EC, sys::fs::OF_TextWithCRLF); |
2675 | if (EC) |
2676 | exitWithErrorCode(EC, Whence: OutputFilename); |
2677 | |
2678 | if (ProfileKind == instr) |
2679 | overlapInstrProfile(BaseFilename, TestFilename, |
2680 | FuncFilter: OverlapFuncFilters{.ValueCutoff: OverlapValueCutoff, .NameFilter: FuncNameFilter}, |
2681 | OS, IsCS); |
2682 | else |
2683 | overlapSampleProfile(BaseFilename, TestFilename, |
2684 | FuncFilter: OverlapFuncFilters{.ValueCutoff: OverlapValueCutoff, .NameFilter: FuncNameFilter}, |
2685 | SimilarityCutoff, OS); |
2686 | |
2687 | return 0; |
2688 | } |
2689 | |
2690 | namespace { |
2691 | struct ValueSitesStats { |
2692 | ValueSitesStats() = default; |
2693 | uint64_t TotalNumValueSites = 0; |
2694 | uint64_t TotalNumValueSitesWithValueProfile = 0; |
2695 | uint64_t TotalNumValues = 0; |
2696 | std::vector<unsigned> ValueSitesHistogram; |
2697 | }; |
2698 | } // namespace |
2699 | |
2700 | static void traverseAllValueSites(const InstrProfRecord &Func, uint32_t VK, |
2701 | ValueSitesStats &Stats, raw_fd_ostream &OS, |
2702 | InstrProfSymtab *Symtab) { |
2703 | uint32_t NS = Func.getNumValueSites(ValueKind: VK); |
2704 | Stats.TotalNumValueSites += NS; |
2705 | for (size_t I = 0; I < NS; ++I) { |
2706 | auto VD = Func.getValueArrayForSite(ValueKind: VK, Site: I); |
2707 | uint32_t NV = VD.size(); |
2708 | if (NV == 0) |
2709 | continue; |
2710 | Stats.TotalNumValues += NV; |
2711 | Stats.TotalNumValueSitesWithValueProfile++; |
2712 | if (NV > Stats.ValueSitesHistogram.size()) |
2713 | Stats.ValueSitesHistogram.resize(new_size: NV, x: 0); |
2714 | Stats.ValueSitesHistogram[NV - 1]++; |
2715 | |
2716 | uint64_t SiteSum = 0; |
2717 | for (const auto &V : VD) |
2718 | SiteSum += V.Count; |
2719 | if (SiteSum == 0) |
2720 | SiteSum = 1; |
2721 | |
2722 | for (const auto &V : VD) { |
2723 | OS << "\t[ " << format(Fmt: "%2u" , Vals: I) << ", " ; |
2724 | if (Symtab == nullptr) |
2725 | OS << format(Fmt: "%4" PRIu64, Vals: V.Value); |
2726 | else |
2727 | OS << Symtab->getFuncOrVarName(MD5Hash: V.Value); |
2728 | OS << ", " << format(Fmt: "%10" PRId64, Vals: V.Count) << " ] (" |
2729 | << format(Fmt: "%.2f%%" , Vals: (V.Count * 100.0 / SiteSum)) << ")\n" ; |
2730 | } |
2731 | } |
2732 | } |
2733 | |
2734 | static void showValueSitesStats(raw_fd_ostream &OS, uint32_t VK, |
2735 | ValueSitesStats &Stats) { |
2736 | OS << " Total number of sites: " << Stats.TotalNumValueSites << "\n" ; |
2737 | OS << " Total number of sites with values: " |
2738 | << Stats.TotalNumValueSitesWithValueProfile << "\n" ; |
2739 | OS << " Total number of profiled values: " << Stats.TotalNumValues << "\n" ; |
2740 | |
2741 | OS << " Value sites histogram:\n\tNumTargets, SiteCount\n" ; |
2742 | for (unsigned I = 0; I < Stats.ValueSitesHistogram.size(); I++) { |
2743 | if (Stats.ValueSitesHistogram[I] > 0) |
2744 | OS << "\t" << I + 1 << ", " << Stats.ValueSitesHistogram[I] << "\n" ; |
2745 | } |
2746 | } |
2747 | |
2748 | static int showInstrProfile(ShowFormat SFormat, raw_fd_ostream &OS) { |
2749 | if (SFormat == ShowFormat::Json) |
2750 | exitWithError(Message: "JSON output is not supported for instr profiles" ); |
2751 | if (SFormat == ShowFormat::Yaml) |
2752 | exitWithError(Message: "YAML output is not supported for instr profiles" ); |
2753 | auto FS = vfs::getRealFileSystem(); |
2754 | auto ReaderOrErr = InstrProfReader::create(Path: Filename, FS&: *FS); |
2755 | std::vector<uint32_t> Cutoffs = std::move(DetailedSummaryCutoffs); |
2756 | if (ShowDetailedSummary && Cutoffs.empty()) { |
2757 | Cutoffs = ProfileSummaryBuilder::DefaultCutoffs; |
2758 | } |
2759 | InstrProfSummaryBuilder Builder(std::move(Cutoffs)); |
2760 | if (Error E = ReaderOrErr.takeError()) |
2761 | exitWithError(E: std::move(E), Whence: Filename); |
2762 | |
2763 | auto Reader = std::move(ReaderOrErr.get()); |
2764 | bool IsIRInstr = Reader->isIRLevelProfile(); |
2765 | size_t ShownFunctions = 0; |
2766 | size_t BelowCutoffFunctions = 0; |
2767 | int NumVPKind = IPVK_Last - IPVK_First + 1; |
2768 | std::vector<ValueSitesStats> VPStats(NumVPKind); |
2769 | |
2770 | auto MinCmp = [](const std::pair<std::string, uint64_t> &v1, |
2771 | const std::pair<std::string, uint64_t> &v2) { |
2772 | return v1.second > v2.second; |
2773 | }; |
2774 | |
2775 | std::priority_queue<std::pair<std::string, uint64_t>, |
2776 | std::vector<std::pair<std::string, uint64_t>>, |
2777 | decltype(MinCmp)> |
2778 | HottestFuncs(MinCmp); |
2779 | |
2780 | if (!TextFormat && OnlyListBelow) { |
2781 | OS << "The list of functions with the maximum counter less than " |
2782 | << ShowValueCutoff << ":\n" ; |
2783 | } |
2784 | |
2785 | // Add marker so that IR-level instrumentation round-trips properly. |
2786 | if (TextFormat && IsIRInstr) |
2787 | OS << ":ir\n" ; |
2788 | |
2789 | for (const auto &Func : *Reader) { |
2790 | if (Reader->isIRLevelProfile()) { |
2791 | bool FuncIsCS = NamedInstrProfRecord::hasCSFlagInHash(FuncHash: Func.Hash); |
2792 | if (FuncIsCS != ShowCS) |
2793 | continue; |
2794 | } |
2795 | bool Show = ShowAllFunctions || |
2796 | (!FuncNameFilter.empty() && Func.Name.contains(Other: FuncNameFilter)); |
2797 | |
2798 | bool doTextFormatDump = (Show && TextFormat); |
2799 | |
2800 | if (doTextFormatDump) { |
2801 | InstrProfSymtab &Symtab = Reader->getSymtab(); |
2802 | InstrProfWriter::writeRecordInText(Name: Func.Name, Hash: Func.Hash, Counters: Func, Symtab, |
2803 | OS); |
2804 | continue; |
2805 | } |
2806 | |
2807 | assert(Func.Counts.size() > 0 && "function missing entry counter" ); |
2808 | Builder.addRecord(Func); |
2809 | |
2810 | if (ShowCovered) { |
2811 | if (llvm::any_of(Range: Func.Counts, P: [](uint64_t C) { return C; })) |
2812 | OS << Func.Name << "\n" ; |
2813 | continue; |
2814 | } |
2815 | |
2816 | uint64_t FuncMax = 0; |
2817 | uint64_t FuncSum = 0; |
2818 | |
2819 | auto PseudoKind = Func.getCountPseudoKind(); |
2820 | if (PseudoKind != InstrProfRecord::NotPseudo) { |
2821 | if (Show) { |
2822 | if (!ShownFunctions) |
2823 | OS << "Counters:\n" ; |
2824 | ++ShownFunctions; |
2825 | OS << " " << Func.Name << ":\n" |
2826 | << " Hash: " << format(Fmt: "0x%016" PRIx64, Vals: Func.Hash) << "\n" |
2827 | << " Counters: " << Func.Counts.size(); |
2828 | if (PseudoKind == InstrProfRecord::PseudoHot) |
2829 | OS << " <PseudoHot>\n" ; |
2830 | else if (PseudoKind == InstrProfRecord::PseudoWarm) |
2831 | OS << " <PseudoWarm>\n" ; |
2832 | else |
2833 | llvm_unreachable("Unknown PseudoKind" ); |
2834 | } |
2835 | continue; |
2836 | } |
2837 | |
2838 | for (size_t I = 0, E = Func.Counts.size(); I < E; ++I) { |
2839 | FuncMax = std::max(a: FuncMax, b: Func.Counts[I]); |
2840 | FuncSum += Func.Counts[I]; |
2841 | } |
2842 | |
2843 | if (FuncMax < ShowValueCutoff) { |
2844 | ++BelowCutoffFunctions; |
2845 | if (OnlyListBelow) { |
2846 | OS << " " << Func.Name << ": (Max = " << FuncMax |
2847 | << " Sum = " << FuncSum << ")\n" ; |
2848 | } |
2849 | continue; |
2850 | } else if (OnlyListBelow) |
2851 | continue; |
2852 | |
2853 | if (TopNFunctions) { |
2854 | if (HottestFuncs.size() == TopNFunctions) { |
2855 | if (HottestFuncs.top().second < FuncMax) { |
2856 | HottestFuncs.pop(); |
2857 | HottestFuncs.emplace(args: std::make_pair(x: std::string(Func.Name), y&: FuncMax)); |
2858 | } |
2859 | } else |
2860 | HottestFuncs.emplace(args: std::make_pair(x: std::string(Func.Name), y&: FuncMax)); |
2861 | } |
2862 | |
2863 | if (Show) { |
2864 | if (!ShownFunctions) |
2865 | OS << "Counters:\n" ; |
2866 | |
2867 | ++ShownFunctions; |
2868 | |
2869 | OS << " " << Func.Name << ":\n" |
2870 | << " Hash: " << format(Fmt: "0x%016" PRIx64, Vals: Func.Hash) << "\n" |
2871 | << " Counters: " << Func.Counts.size() << "\n" ; |
2872 | if (!IsIRInstr) |
2873 | OS << " Function count: " << Func.Counts[0] << "\n" ; |
2874 | |
2875 | if (ShowIndirectCallTargets) |
2876 | OS << " Indirect Call Site Count: " |
2877 | << Func.getNumValueSites(ValueKind: IPVK_IndirectCallTarget) << "\n" ; |
2878 | |
2879 | if (ShowVTables) |
2880 | OS << " Number of instrumented vtables: " |
2881 | << Func.getNumValueSites(ValueKind: IPVK_VTableTarget) << "\n" ; |
2882 | |
2883 | uint32_t NumMemOPCalls = Func.getNumValueSites(ValueKind: IPVK_MemOPSize); |
2884 | if (ShowMemOPSizes && NumMemOPCalls > 0) |
2885 | OS << " Number of Memory Intrinsics Calls: " << NumMemOPCalls |
2886 | << "\n" ; |
2887 | |
2888 | if (ShowCounts) { |
2889 | OS << " Block counts: [" ; |
2890 | size_t Start = (IsIRInstr ? 0 : 1); |
2891 | for (size_t I = Start, E = Func.Counts.size(); I < E; ++I) { |
2892 | OS << (I == Start ? "" : ", " ) << Func.Counts[I]; |
2893 | } |
2894 | OS << "]\n" ; |
2895 | } |
2896 | |
2897 | if (ShowIndirectCallTargets) { |
2898 | OS << " Indirect Target Results:\n" ; |
2899 | traverseAllValueSites(Func, VK: IPVK_IndirectCallTarget, |
2900 | Stats&: VPStats[IPVK_IndirectCallTarget], OS, |
2901 | Symtab: &(Reader->getSymtab())); |
2902 | } |
2903 | |
2904 | if (ShowVTables) { |
2905 | OS << " VTable Results:\n" ; |
2906 | traverseAllValueSites(Func, VK: IPVK_VTableTarget, |
2907 | Stats&: VPStats[IPVK_VTableTarget], OS, |
2908 | Symtab: &(Reader->getSymtab())); |
2909 | } |
2910 | |
2911 | if (ShowMemOPSizes && NumMemOPCalls > 0) { |
2912 | OS << " Memory Intrinsic Size Results:\n" ; |
2913 | traverseAllValueSites(Func, VK: IPVK_MemOPSize, Stats&: VPStats[IPVK_MemOPSize], OS, |
2914 | Symtab: nullptr); |
2915 | } |
2916 | } |
2917 | } |
2918 | if (Reader->hasError()) |
2919 | exitWithError(E: Reader->getError(), Whence: Filename); |
2920 | |
2921 | if (TextFormat || ShowCovered) |
2922 | return 0; |
2923 | std::unique_ptr<ProfileSummary> PS(Builder.getSummary()); |
2924 | bool IsIR = Reader->isIRLevelProfile(); |
2925 | OS << "Instrumentation level: " << (IsIR ? "IR" : "Front-end" ); |
2926 | if (IsIR) |
2927 | OS << " entry_first = " << Reader->instrEntryBBEnabled(); |
2928 | OS << "\n" ; |
2929 | if (ShowAllFunctions || !FuncNameFilter.empty()) |
2930 | OS << "Functions shown: " << ShownFunctions << "\n" ; |
2931 | OS << "Total functions: " << PS->getNumFunctions() << "\n" ; |
2932 | if (ShowValueCutoff > 0) { |
2933 | OS << "Number of functions with maximum count (< " << ShowValueCutoff |
2934 | << "): " << BelowCutoffFunctions << "\n" ; |
2935 | OS << "Number of functions with maximum count (>= " << ShowValueCutoff |
2936 | << "): " << PS->getNumFunctions() - BelowCutoffFunctions << "\n" ; |
2937 | } |
2938 | OS << "Maximum function count: " << PS->getMaxFunctionCount() << "\n" ; |
2939 | OS << "Maximum internal block count: " << PS->getMaxInternalCount() << "\n" ; |
2940 | |
2941 | if (TopNFunctions) { |
2942 | std::vector<std::pair<std::string, uint64_t>> SortedHottestFuncs; |
2943 | while (!HottestFuncs.empty()) { |
2944 | SortedHottestFuncs.emplace_back(args: HottestFuncs.top()); |
2945 | HottestFuncs.pop(); |
2946 | } |
2947 | OS << "Top " << TopNFunctions |
2948 | << " functions with the largest internal block counts: \n" ; |
2949 | for (auto &hotfunc : llvm::reverse(C&: SortedHottestFuncs)) |
2950 | OS << " " << hotfunc.first << ", max count = " << hotfunc.second << "\n" ; |
2951 | } |
2952 | |
2953 | if (ShownFunctions && ShowIndirectCallTargets) { |
2954 | OS << "Statistics for indirect call sites profile:\n" ; |
2955 | showValueSitesStats(OS, VK: IPVK_IndirectCallTarget, |
2956 | Stats&: VPStats[IPVK_IndirectCallTarget]); |
2957 | } |
2958 | |
2959 | if (ShownFunctions && ShowVTables) { |
2960 | OS << "Statistics for vtable profile:\n" ; |
2961 | showValueSitesStats(OS, VK: IPVK_VTableTarget, Stats&: VPStats[IPVK_VTableTarget]); |
2962 | } |
2963 | |
2964 | if (ShownFunctions && ShowMemOPSizes) { |
2965 | OS << "Statistics for memory intrinsic calls sizes profile:\n" ; |
2966 | showValueSitesStats(OS, VK: IPVK_MemOPSize, Stats&: VPStats[IPVK_MemOPSize]); |
2967 | } |
2968 | |
2969 | if (ShowDetailedSummary) { |
2970 | OS << "Total number of blocks: " << PS->getNumCounts() << "\n" ; |
2971 | OS << "Total count: " << PS->getTotalCount() << "\n" ; |
2972 | PS->printDetailedSummary(OS); |
2973 | } |
2974 | |
2975 | if (ShowBinaryIds) |
2976 | if (Error E = Reader->printBinaryIds(OS)) |
2977 | exitWithError(E: std::move(E), Whence: Filename); |
2978 | |
2979 | if (ShowProfileVersion) |
2980 | OS << "Profile version: " << Reader->getVersion() << "\n" ; |
2981 | |
2982 | if (ShowTemporalProfTraces) { |
2983 | auto &Traces = Reader->getTemporalProfTraces(); |
2984 | OS << "Temporal Profile Traces (samples=" << Traces.size() |
2985 | << " seen=" << Reader->getTemporalProfTraceStreamSize() << "):\n" ; |
2986 | for (unsigned i = 0; i < Traces.size(); i++) { |
2987 | OS << " Temporal Profile Trace " << i << " (weight=" << Traces[i].Weight |
2988 | << " count=" << Traces[i].FunctionNameRefs.size() << "):\n" ; |
2989 | for (auto &NameRef : Traces[i].FunctionNameRefs) |
2990 | OS << " " << Reader->getSymtab().getFuncOrVarName(MD5Hash: NameRef) << "\n" ; |
2991 | } |
2992 | } |
2993 | |
2994 | return 0; |
2995 | } |
2996 | |
2997 | static void showSectionInfo(sampleprof::SampleProfileReader *Reader, |
2998 | raw_fd_ostream &OS) { |
2999 | if (!Reader->dumpSectionInfo(OS)) { |
3000 | WithColor::warning() << "-show-sec-info-only is only supported for " |
3001 | << "sample profile in extbinary format and is " |
3002 | << "ignored for other formats.\n" ; |
3003 | return; |
3004 | } |
3005 | } |
3006 | |
3007 | namespace { |
3008 | struct HotFuncInfo { |
3009 | std::string FuncName; |
3010 | uint64_t TotalCount = 0; |
3011 | double TotalCountPercent = 0.0f; |
3012 | uint64_t MaxCount = 0; |
3013 | uint64_t EntryCount = 0; |
3014 | |
3015 | HotFuncInfo() = default; |
3016 | |
3017 | HotFuncInfo(StringRef FN, uint64_t TS, double TSP, uint64_t MS, uint64_t ES) |
3018 | : FuncName(FN.begin(), FN.end()), TotalCount(TS), TotalCountPercent(TSP), |
3019 | MaxCount(MS), EntryCount(ES) {} |
3020 | }; |
3021 | } // namespace |
3022 | |
3023 | // Print out detailed information about hot functions in PrintValues vector. |
3024 | // Users specify titles and offset of every columns through ColumnTitle and |
3025 | // ColumnOffset. The size of ColumnTitle and ColumnOffset need to be the same |
3026 | // and at least 4. Besides, users can optionally give a HotFuncMetric string to |
3027 | // print out or let it be an empty string. |
3028 | static void dumpHotFunctionList(const std::vector<std::string> &ColumnTitle, |
3029 | const std::vector<int> &ColumnOffset, |
3030 | const std::vector<HotFuncInfo> &PrintValues, |
3031 | uint64_t HotFuncCount, uint64_t TotalFuncCount, |
3032 | uint64_t HotProfCount, uint64_t TotalProfCount, |
3033 | const std::string &HotFuncMetric, |
3034 | uint32_t TopNFunctions, raw_fd_ostream &OS) { |
3035 | assert(ColumnOffset.size() == ColumnTitle.size() && |
3036 | "ColumnOffset and ColumnTitle should have the same size" ); |
3037 | assert(ColumnTitle.size() >= 4 && |
3038 | "ColumnTitle should have at least 4 elements" ); |
3039 | assert(TotalFuncCount > 0 && |
3040 | "There should be at least one function in the profile" ); |
3041 | double TotalProfPercent = 0; |
3042 | if (TotalProfCount > 0) |
3043 | TotalProfPercent = static_cast<double>(HotProfCount) / TotalProfCount * 100; |
3044 | |
3045 | formatted_raw_ostream FOS(OS); |
3046 | FOS << HotFuncCount << " out of " << TotalFuncCount |
3047 | << " functions with profile (" |
3048 | << format(Fmt: "%.2f%%" , |
3049 | Vals: (static_cast<double>(HotFuncCount) / TotalFuncCount * 100)) |
3050 | << ") are considered hot functions" ; |
3051 | if (!HotFuncMetric.empty()) |
3052 | FOS << " (" << HotFuncMetric << ")" ; |
3053 | FOS << ".\n" ; |
3054 | FOS << HotProfCount << " out of " << TotalProfCount << " profile counts (" |
3055 | << format(Fmt: "%.2f%%" , Vals: TotalProfPercent) << ") are from hot functions.\n" ; |
3056 | |
3057 | for (size_t I = 0; I < ColumnTitle.size(); ++I) { |
3058 | FOS.PadToColumn(NewCol: ColumnOffset[I]); |
3059 | FOS << ColumnTitle[I]; |
3060 | } |
3061 | FOS << "\n" ; |
3062 | |
3063 | uint32_t Count = 0; |
3064 | for (const auto &R : PrintValues) { |
3065 | if (TopNFunctions && (Count++ == TopNFunctions)) |
3066 | break; |
3067 | FOS.PadToColumn(NewCol: ColumnOffset[0]); |
3068 | FOS << R.TotalCount << " (" << format(Fmt: "%.2f%%" , Vals: R.TotalCountPercent) << ")" ; |
3069 | FOS.PadToColumn(NewCol: ColumnOffset[1]); |
3070 | FOS << R.MaxCount; |
3071 | FOS.PadToColumn(NewCol: ColumnOffset[2]); |
3072 | FOS << R.EntryCount; |
3073 | FOS.PadToColumn(NewCol: ColumnOffset[3]); |
3074 | FOS << R.FuncName << "\n" ; |
3075 | } |
3076 | } |
3077 | |
3078 | static int showHotFunctionList(const sampleprof::SampleProfileMap &Profiles, |
3079 | ProfileSummary &PS, uint32_t TopN, |
3080 | raw_fd_ostream &OS) { |
3081 | using namespace sampleprof; |
3082 | |
3083 | const uint32_t HotFuncCutoff = 990000; |
3084 | auto &SummaryVector = PS.getDetailedSummary(); |
3085 | uint64_t MinCountThreshold = 0; |
3086 | for (const ProfileSummaryEntry &SummaryEntry : SummaryVector) { |
3087 | if (SummaryEntry.Cutoff == HotFuncCutoff) { |
3088 | MinCountThreshold = SummaryEntry.MinCount; |
3089 | break; |
3090 | } |
3091 | } |
3092 | |
3093 | // Traverse all functions in the profile and keep only hot functions. |
3094 | // The following loop also calculates the sum of total samples of all |
3095 | // functions. |
3096 | std::multimap<uint64_t, std::pair<const FunctionSamples *, const uint64_t>, |
3097 | std::greater<uint64_t>> |
3098 | HotFunc; |
3099 | uint64_t ProfileTotalSample = 0; |
3100 | uint64_t HotFuncSample = 0; |
3101 | uint64_t HotFuncCount = 0; |
3102 | |
3103 | for (const auto &I : Profiles) { |
3104 | FuncSampleStats FuncStats; |
3105 | const FunctionSamples &FuncProf = I.second; |
3106 | ProfileTotalSample += FuncProf.getTotalSamples(); |
3107 | getFuncSampleStats(Func: FuncProf, FuncStats, HotThreshold: MinCountThreshold); |
3108 | |
3109 | if (isFunctionHot(FuncStats, HotThreshold: MinCountThreshold)) { |
3110 | HotFunc.emplace(args: FuncProf.getTotalSamples(), |
3111 | args: std::make_pair(x: &(I.second), y&: FuncStats.MaxSample)); |
3112 | HotFuncSample += FuncProf.getTotalSamples(); |
3113 | ++HotFuncCount; |
3114 | } |
3115 | } |
3116 | |
3117 | std::vector<std::string> ColumnTitle{"Total sample (%)" , "Max sample" , |
3118 | "Entry sample" , "Function name" }; |
3119 | std::vector<int> ColumnOffset{0, 24, 42, 58}; |
3120 | std::string Metric = |
3121 | std::string("max sample >= " ) + std::to_string(val: MinCountThreshold); |
3122 | std::vector<HotFuncInfo> PrintValues; |
3123 | for (const auto &FuncPair : HotFunc) { |
3124 | const FunctionSamples &Func = *FuncPair.second.first; |
3125 | double TotalSamplePercent = |
3126 | (ProfileTotalSample > 0) |
3127 | ? (Func.getTotalSamples() * 100.0) / ProfileTotalSample |
3128 | : 0; |
3129 | PrintValues.emplace_back( |
3130 | args: HotFuncInfo(Func.getContext().toString(), Func.getTotalSamples(), |
3131 | TotalSamplePercent, FuncPair.second.second, |
3132 | Func.getHeadSamplesEstimate())); |
3133 | } |
3134 | dumpHotFunctionList(ColumnTitle, ColumnOffset, PrintValues, HotFuncCount, |
3135 | TotalFuncCount: Profiles.size(), HotProfCount: HotFuncSample, TotalProfCount: ProfileTotalSample, |
3136 | HotFuncMetric: Metric, TopNFunctions: TopN, OS); |
3137 | |
3138 | return 0; |
3139 | } |
3140 | |
3141 | static int showSampleProfile(ShowFormat SFormat, raw_fd_ostream &OS) { |
3142 | if (SFormat == ShowFormat::Yaml) |
3143 | exitWithError(Message: "YAML output is not supported for sample profiles" ); |
3144 | using namespace sampleprof; |
3145 | LLVMContext Context; |
3146 | auto FS = vfs::getRealFileSystem(); |
3147 | auto ReaderOrErr = SampleProfileReader::create(Filename, C&: Context, FS&: *FS, |
3148 | P: FSDiscriminatorPassOption); |
3149 | if (std::error_code EC = ReaderOrErr.getError()) |
3150 | exitWithErrorCode(EC, Whence: Filename); |
3151 | |
3152 | auto Reader = std::move(ReaderOrErr.get()); |
3153 | if (ShowSectionInfoOnly) { |
3154 | showSectionInfo(Reader: Reader.get(), OS); |
3155 | return 0; |
3156 | } |
3157 | |
3158 | if (std::error_code EC = Reader->read()) |
3159 | exitWithErrorCode(EC, Whence: Filename); |
3160 | |
3161 | if (ShowAllFunctions || FuncNameFilter.empty()) { |
3162 | if (SFormat == ShowFormat::Json) |
3163 | Reader->dumpJson(OS); |
3164 | else |
3165 | Reader->dump(OS); |
3166 | } else { |
3167 | if (SFormat == ShowFormat::Json) |
3168 | exitWithError( |
3169 | Message: "the JSON format is supported only when all functions are to " |
3170 | "be printed" ); |
3171 | |
3172 | // TODO: parse context string to support filtering by contexts. |
3173 | FunctionSamples *FS = Reader->getSamplesFor(Fname: StringRef(FuncNameFilter)); |
3174 | Reader->dumpFunctionProfile(FS: FS ? *FS : FunctionSamples(), OS); |
3175 | } |
3176 | |
3177 | if (ShowProfileSymbolList) { |
3178 | std::unique_ptr<sampleprof::ProfileSymbolList> ReaderList = |
3179 | Reader->getProfileSymbolList(); |
3180 | ReaderList->dump(OS); |
3181 | } |
3182 | |
3183 | if (ShowDetailedSummary) { |
3184 | auto &PS = Reader->getSummary(); |
3185 | PS.printSummary(OS); |
3186 | PS.printDetailedSummary(OS); |
3187 | } |
3188 | |
3189 | if (ShowHotFuncList || TopNFunctions) |
3190 | showHotFunctionList(Profiles: Reader->getProfiles(), PS&: Reader->getSummary(), |
3191 | TopN: TopNFunctions, OS); |
3192 | |
3193 | return 0; |
3194 | } |
3195 | |
3196 | static int showMemProfProfile(ShowFormat SFormat, raw_fd_ostream &OS) { |
3197 | if (SFormat == ShowFormat::Json) |
3198 | exitWithError(Message: "JSON output is not supported for MemProf" ); |
3199 | auto ReaderOr = llvm::memprof::RawMemProfReader::create( |
3200 | Path: Filename, ProfiledBinary, /*KeepNames=*/KeepName: true); |
3201 | if (Error E = ReaderOr.takeError()) |
3202 | // Since the error can be related to the profile or the binary we do not |
3203 | // pass whence. Instead additional context is provided where necessary in |
3204 | // the error message. |
3205 | exitWithError(E: std::move(E), /*Whence*/ "" ); |
3206 | |
3207 | std::unique_ptr<llvm::memprof::RawMemProfReader> Reader( |
3208 | ReaderOr.get().release()); |
3209 | |
3210 | Reader->printYAML(OS); |
3211 | return 0; |
3212 | } |
3213 | |
3214 | static int showDebugInfoCorrelation(const std::string &Filename, |
3215 | ShowFormat SFormat, raw_fd_ostream &OS) { |
3216 | if (SFormat == ShowFormat::Json) |
3217 | exitWithError(Message: "JSON output is not supported for debug info correlation" ); |
3218 | std::unique_ptr<InstrProfCorrelator> Correlator; |
3219 | if (auto Err = |
3220 | InstrProfCorrelator::get(Filename, FileKind: InstrProfCorrelator::DEBUG_INFO) |
3221 | .moveInto(Value&: Correlator)) |
3222 | exitWithError(E: std::move(Err), Whence: Filename); |
3223 | if (SFormat == ShowFormat::Yaml) { |
3224 | if (auto Err = Correlator->dumpYaml(MaxWarnings: MaxDbgCorrelationWarnings, OS)) |
3225 | exitWithError(E: std::move(Err), Whence: Filename); |
3226 | return 0; |
3227 | } |
3228 | |
3229 | if (auto Err = Correlator->correlateProfileData(MaxWarnings: MaxDbgCorrelationWarnings)) |
3230 | exitWithError(E: std::move(Err), Whence: Filename); |
3231 | |
3232 | InstrProfSymtab Symtab; |
3233 | if (auto Err = Symtab.create( |
3234 | NameStrings: StringRef(Correlator->getNamesPointer(), Correlator->getNamesSize()))) |
3235 | exitWithError(E: std::move(Err), Whence: Filename); |
3236 | |
3237 | if (ShowProfileSymbolList) |
3238 | Symtab.dumpNames(OS); |
3239 | // TODO: Read "Profile Data Type" from debug info to compute and show how many |
3240 | // counters the section holds. |
3241 | if (ShowDetailedSummary) |
3242 | OS << "Counters section size: 0x" |
3243 | << Twine::utohexstr(Val: Correlator->getCountersSectionSize()) << " bytes\n" ; |
3244 | OS << "Found " << Correlator->getDataSize() << " functions\n" ; |
3245 | |
3246 | return 0; |
3247 | } |
3248 | |
3249 | static int show_main(StringRef ProgName) { |
3250 | if (Filename.empty() && DebugInfoFilename.empty()) |
3251 | exitWithError( |
3252 | Message: "the positional argument '<profdata-file>' is required unless '--" + |
3253 | DebugInfoFilename.ArgStr + "' is provided" ); |
3254 | |
3255 | if (Filename == OutputFilename) { |
3256 | errs() << ProgName |
3257 | << " show: Input file name cannot be the same as the output file " |
3258 | "name!\n" ; |
3259 | return 1; |
3260 | } |
3261 | if (JsonFormat) |
3262 | SFormat = ShowFormat::Json; |
3263 | |
3264 | std::error_code EC; |
3265 | raw_fd_ostream OS(OutputFilename.data(), EC, sys::fs::OF_TextWithCRLF); |
3266 | if (EC) |
3267 | exitWithErrorCode(EC, Whence: OutputFilename); |
3268 | |
3269 | if (ShowAllFunctions && !FuncNameFilter.empty()) |
3270 | WithColor::warning() << "-function argument ignored: showing all functions\n" ; |
3271 | |
3272 | if (!DebugInfoFilename.empty()) |
3273 | return showDebugInfoCorrelation(Filename: DebugInfoFilename, SFormat, OS); |
3274 | |
3275 | if (ShowProfileKind == instr) |
3276 | return showInstrProfile(SFormat, OS); |
3277 | if (ShowProfileKind == sample) |
3278 | return showSampleProfile(SFormat, OS); |
3279 | return showMemProfProfile(SFormat, OS); |
3280 | } |
3281 | |
3282 | static int order_main() { |
3283 | std::error_code EC; |
3284 | raw_fd_ostream OS(OutputFilename.data(), EC, sys::fs::OF_TextWithCRLF); |
3285 | if (EC) |
3286 | exitWithErrorCode(EC, Whence: OutputFilename); |
3287 | auto FS = vfs::getRealFileSystem(); |
3288 | auto ReaderOrErr = InstrProfReader::create(Path: Filename, FS&: *FS); |
3289 | if (Error E = ReaderOrErr.takeError()) |
3290 | exitWithError(E: std::move(E), Whence: Filename); |
3291 | |
3292 | auto Reader = std::move(ReaderOrErr.get()); |
3293 | for (auto &I : *Reader) { |
3294 | // Read all entries |
3295 | (void)I; |
3296 | } |
3297 | ArrayRef Traces = Reader->getTemporalProfTraces(); |
3298 | if (NumTestTraces && NumTestTraces >= Traces.size()) |
3299 | exitWithError( |
3300 | Message: "--" + NumTestTraces.ArgStr + |
3301 | " must be smaller than the total number of traces: expected: < " + |
3302 | Twine(Traces.size()) + ", actual: " + Twine(NumTestTraces)); |
3303 | ArrayRef TestTraces = Traces.take_back(N: NumTestTraces); |
3304 | Traces = Traces.drop_back(N: NumTestTraces); |
3305 | |
3306 | std::vector<BPFunctionNode> Nodes; |
3307 | TemporalProfTraceTy::createBPFunctionNodes(Traces, Nodes); |
3308 | BalancedPartitioningConfig Config; |
3309 | BalancedPartitioning BP(Config); |
3310 | BP.run(Nodes); |
3311 | |
3312 | OS << "# Ordered " << Nodes.size() << " functions\n" ; |
3313 | if (!TestTraces.empty()) { |
3314 | // Since we don't know the symbol sizes, we assume 32 functions per page. |
3315 | DenseMap<BPFunctionNode::IDT, unsigned> IdToPageNumber; |
3316 | for (auto &Node : Nodes) |
3317 | IdToPageNumber[Node.Id] = IdToPageNumber.size() / 32; |
3318 | |
3319 | SmallSet<unsigned, 0> TouchedPages; |
3320 | unsigned Area = 0; |
3321 | for (auto &Trace : TestTraces) { |
3322 | for (auto Id : Trace.FunctionNameRefs) { |
3323 | auto It = IdToPageNumber.find(Val: Id); |
3324 | if (It == IdToPageNumber.end()) |
3325 | continue; |
3326 | TouchedPages.insert(V: It->getSecond()); |
3327 | Area += TouchedPages.size(); |
3328 | } |
3329 | TouchedPages.clear(); |
3330 | } |
3331 | OS << "# Total area under the page fault curve: " << (float)Area << "\n" ; |
3332 | } |
3333 | OS << "# Warning: Mach-O may prefix symbols with \"_\" depending on the " |
3334 | "linkage and this output does not take that into account. Some " |
3335 | "post-processing may be required before passing to the linker via " |
3336 | "-order_file.\n" ; |
3337 | for (auto &N : Nodes) { |
3338 | auto [Filename, ParsedFuncName] = |
3339 | getParsedIRPGOName(IRPGOName: Reader->getSymtab().getFuncOrVarName(MD5Hash: N.Id)); |
3340 | if (!Filename.empty()) |
3341 | OS << "# " << Filename << "\n" ; |
3342 | OS << ParsedFuncName << "\n" ; |
3343 | } |
3344 | return 0; |
3345 | } |
3346 | |
3347 | int llvm_profdata_main(int argc, char **argvNonConst, |
3348 | const llvm::ToolContext &) { |
3349 | const char **argv = const_cast<const char **>(argvNonConst); |
3350 | |
3351 | StringRef ProgName(sys::path::filename(path: argv[0])); |
3352 | |
3353 | if (argc < 2) { |
3354 | errs() << ProgName |
3355 | << ": No subcommand specified! Run llvm-profata --help for usage.\n" ; |
3356 | return 1; |
3357 | } |
3358 | |
3359 | cl::ParseCommandLineOptions(argc, argv, Overview: "LLVM profile data\n" ); |
3360 | |
3361 | if (ShowSubcommand) |
3362 | return show_main(ProgName); |
3363 | |
3364 | if (OrderSubcommand) |
3365 | return order_main(); |
3366 | |
3367 | if (OverlapSubcommand) |
3368 | return overlap_main(); |
3369 | |
3370 | if (MergeSubcommand) |
3371 | return merge_main(ProgName); |
3372 | |
3373 | errs() << ProgName |
3374 | << ": Unknown command. Run llvm-profdata --help for usage.\n" ; |
3375 | return 1; |
3376 | } |
3377 | |