xref: /freebsd/contrib/llvm-project/llvm/include/llvm/ProfileData/InstrProf.h (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- InstrProf.h - Instrumented profiling format support ------*- C++ -*-===//
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 // Instrumentation-based profiling data is generated by instrumented
10 // binaries through library functions in compiler-rt, and read by the clang
11 // frontend to feed PGO.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_PROFILEDATA_INSTRPROF_H
16 #define LLVM_PROFILEDATA_INSTRPROF_H
17 
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/BitmaskEnum.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/IntervalMap.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/ADT/StringSet.h"
25 #include "llvm/IR/GlobalValue.h"
26 #include "llvm/IR/ProfileSummary.h"
27 #include "llvm/ProfileData/InstrProfData.inc"
28 #include "llvm/Support/BalancedPartitioning.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/Support/Error.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MD5.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/TargetParser/Host.h"
37 #include "llvm/TargetParser/Triple.h"
38 #include <algorithm>
39 #include <cassert>
40 #include <cstddef>
41 #include <cstdint>
42 #include <cstring>
43 #include <list>
44 #include <memory>
45 #include <string>
46 #include <system_error>
47 #include <utility>
48 #include <vector>
49 
50 namespace llvm {
51 
52 class Function;
53 class GlobalVariable;
54 struct InstrProfRecord;
55 class InstrProfSymtab;
56 class Instruction;
57 class MDNode;
58 class Module;
59 
60 enum InstrProfSectKind {
61 #define INSTR_PROF_SECT_ENTRY(Kind, SectNameCommon, SectNameCoff, Prefix) Kind,
62 #include "llvm/ProfileData/InstrProfData.inc"
63 };
64 
65 /// Return the max count value. We reserver a few large values for special use.
getInstrMaxCountValue()66 inline uint64_t getInstrMaxCountValue() {
67   return std::numeric_limits<uint64_t>::max() - 2;
68 }
69 
70 /// Return the name of the profile section corresponding to \p IPSK.
71 ///
72 /// The name of the section depends on the object format type \p OF. If
73 /// \p AddSegmentInfo is true, a segment prefix and additional linker hints may
74 /// be added to the section name (this is the default).
75 std::string getInstrProfSectionName(InstrProfSectKind IPSK,
76                                     Triple::ObjectFormatType OF,
77                                     bool AddSegmentInfo = true);
78 
79 /// Return the name profile runtime entry point to do value profiling
80 /// for a given site.
getInstrProfValueProfFuncName()81 inline StringRef getInstrProfValueProfFuncName() {
82   return INSTR_PROF_VALUE_PROF_FUNC_STR;
83 }
84 
85 /// Return the name profile runtime entry point to do memop size value
86 /// profiling.
getInstrProfValueProfMemOpFuncName()87 inline StringRef getInstrProfValueProfMemOpFuncName() {
88   return INSTR_PROF_VALUE_PROF_MEMOP_FUNC_STR;
89 }
90 
91 /// Return the name prefix of variables containing instrumented function names.
getInstrProfNameVarPrefix()92 inline StringRef getInstrProfNameVarPrefix() { return "__profn_"; }
93 
94 /// Return the name prefix of variables containing virtual table profile data.
getInstrProfVTableVarPrefix()95 inline StringRef getInstrProfVTableVarPrefix() { return "__profvt_"; }
96 
97 /// Return the name prefix of variables containing per-function control data.
getInstrProfDataVarPrefix()98 inline StringRef getInstrProfDataVarPrefix() { return "__profd_"; }
99 
100 /// Return the name prefix of profile counter variables.
getInstrProfCountersVarPrefix()101 inline StringRef getInstrProfCountersVarPrefix() { return "__profc_"; }
102 
103 /// Return the name prefix of profile bitmap variables.
getInstrProfBitmapVarPrefix()104 inline StringRef getInstrProfBitmapVarPrefix() { return "__profbm_"; }
105 
106 /// Return the name prefix of value profile variables.
getInstrProfValuesVarPrefix()107 inline StringRef getInstrProfValuesVarPrefix() { return "__profvp_"; }
108 
109 /// Return the name of value profile node array variables:
getInstrProfVNodesVarName()110 inline StringRef getInstrProfVNodesVarName() { return "__llvm_prf_vnodes"; }
111 
112 /// Return the name of the variable holding the strings (possibly compressed)
113 /// of all function's PGO names.
getInstrProfNamesVarName()114 inline StringRef getInstrProfNamesVarName() { return "__llvm_prf_nm"; }
115 
getInstrProfVTableNamesVarName()116 inline StringRef getInstrProfVTableNamesVarName() { return "__llvm_prf_vnm"; }
117 
118 /// Return the name of a covarage mapping variable (internal linkage)
119 /// for each instrumented source module. Such variables are allocated
120 /// in the __llvm_covmap section.
getCoverageMappingVarName()121 inline StringRef getCoverageMappingVarName() {
122   return "__llvm_coverage_mapping";
123 }
124 
125 /// Return the name of the internal variable recording the array
126 /// of PGO name vars referenced by the coverage mapping. The owning
127 /// functions of those names are not emitted by FE (e.g, unused inline
128 /// functions.)
getCoverageUnusedNamesVarName()129 inline StringRef getCoverageUnusedNamesVarName() {
130   return "__llvm_coverage_names";
131 }
132 
133 /// Return the name of function that registers all the per-function control
134 /// data at program startup time by calling __llvm_register_function. This
135 /// function has internal linkage and is called by  __llvm_profile_init
136 /// runtime method. This function is not generated for these platforms:
137 /// Darwin, Linux, and FreeBSD.
getInstrProfRegFuncsName()138 inline StringRef getInstrProfRegFuncsName() {
139   return "__llvm_profile_register_functions";
140 }
141 
142 /// Return the name of the runtime interface that registers per-function control
143 /// data for one instrumented function.
getInstrProfRegFuncName()144 inline StringRef getInstrProfRegFuncName() {
145   return "__llvm_profile_register_function";
146 }
147 
148 /// Return the name of the runtime interface that registers the PGO name
149 /// strings.
getInstrProfNamesRegFuncName()150 inline StringRef getInstrProfNamesRegFuncName() {
151   return "__llvm_profile_register_names_function";
152 }
153 
154 /// Return the name of the runtime initialization method that is generated by
155 /// the compiler. The function calls __llvm_profile_register_functions and
156 /// __llvm_profile_override_default_filename functions if needed. This function
157 /// has internal linkage and invoked at startup time via init_array.
getInstrProfInitFuncName()158 inline StringRef getInstrProfInitFuncName() { return "__llvm_profile_init"; }
159 
160 /// Return the name of the hook variable defined in profile runtime library.
161 /// A reference to the variable causes the linker to link in the runtime
162 /// initialization module (which defines the hook variable).
getInstrProfRuntimeHookVarName()163 inline StringRef getInstrProfRuntimeHookVarName() {
164   return INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_RUNTIME_VAR);
165 }
166 
167 /// Return the name of the compiler generated function that references the
168 /// runtime hook variable. The function is a weak global.
getInstrProfRuntimeHookVarUseFuncName()169 inline StringRef getInstrProfRuntimeHookVarUseFuncName() {
170   return "__llvm_profile_runtime_user";
171 }
172 
getInstrProfCounterBiasVarName()173 inline StringRef getInstrProfCounterBiasVarName() {
174   return INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_COUNTER_BIAS_VAR);
175 }
176 
177 /// Return the marker used to separate PGO names during serialization.
getInstrProfNameSeparator()178 inline StringRef getInstrProfNameSeparator() { return "\01"; }
179 
180 /// Please use getIRPGOFuncName for LLVM IR instrumentation. This function is
181 /// for front-end (Clang, etc) instrumentation.
182 /// Return the modified name for function \c F suitable to be
183 /// used the key for profile lookup. Variable \c InLTO indicates if this
184 /// is called in LTO optimization passes.
185 std::string getPGOFuncName(const Function &F, bool InLTO = false,
186                            uint64_t Version = INSTR_PROF_INDEX_VERSION);
187 
188 /// Return the modified name for a function suitable to be
189 /// used the key for profile lookup. The function's original
190 /// name is \c RawFuncName and has linkage of type \c Linkage.
191 /// The function is defined in module \c FileName.
192 std::string getPGOFuncName(StringRef RawFuncName,
193                            GlobalValue::LinkageTypes Linkage,
194                            StringRef FileName,
195                            uint64_t Version = INSTR_PROF_INDEX_VERSION);
196 
197 /// \return the modified name for function \c F suitable to be
198 /// used as the key for IRPGO profile lookup. \c InLTO indicates if this is
199 /// called from LTO optimization passes.
200 std::string getIRPGOFuncName(const Function &F, bool InLTO = false);
201 
202 /// \return the filename and the function name parsed from the output of
203 /// \c getIRPGOFuncName()
204 std::pair<StringRef, StringRef> getParsedIRPGOName(StringRef IRPGOName);
205 
206 /// Return the name of the global variable used to store a function
207 /// name in PGO instrumentation. \c FuncName is the IRPGO function name
208 /// (returned by \c getIRPGOFuncName) for LLVM IR instrumentation and PGO
209 /// function name (returned by \c getPGOFuncName) for front-end instrumentation.
210 std::string getPGOFuncNameVarName(StringRef FuncName,
211                                   GlobalValue::LinkageTypes Linkage);
212 
213 /// Create and return the global variable for function name used in PGO
214 /// instrumentation. \c FuncName is the IRPGO function name (returned by
215 /// \c getIRPGOFuncName) for LLVM IR instrumentation and PGO function name
216 /// (returned by \c getPGOFuncName) for front-end instrumentation.
217 GlobalVariable *createPGOFuncNameVar(Function &F, StringRef PGOFuncName);
218 
219 /// Create and return the global variable for function name used in PGO
220 /// instrumentation. \c FuncName is the IRPGO function name (returned by
221 /// \c getIRPGOFuncName) for LLVM IR instrumentation and PGO function name
222 /// (returned by \c getPGOFuncName) for front-end instrumentation.
223 GlobalVariable *createPGOFuncNameVar(Module &M,
224                                      GlobalValue::LinkageTypes Linkage,
225                                      StringRef PGOFuncName);
226 
227 /// Return the initializer in string of the PGO name var \c NameVar.
228 StringRef getPGOFuncNameVarInitializer(GlobalVariable *NameVar);
229 
230 /// Given a PGO function name, remove the filename prefix and return
231 /// the original (static) function name.
232 StringRef getFuncNameWithoutPrefix(StringRef PGOFuncName,
233                                    StringRef FileName = "<unknown>");
234 
235 /// Given a vector of strings (names of global objects like functions or,
236 /// virtual tables) \c NameStrs, the method generates a combined string \c
237 /// Result that is ready to be serialized.  The \c Result string is comprised of
238 /// three fields: The first field is the length of the uncompressed strings, and
239 /// the the second field is the length of the zlib-compressed string. Both
240 /// fields are encoded in ULEB128.  If \c doCompress is false, the
241 ///  third field is the uncompressed strings; otherwise it is the
242 /// compressed string. When the string compression is off, the
243 /// second field will have value zero.
244 Error collectGlobalObjectNameStrings(ArrayRef<std::string> NameStrs,
245                                      bool doCompression, std::string &Result);
246 
247 /// Produce \c Result string with the same format described above. The input
248 /// is vector of PGO function name variables that are referenced.
249 /// The global variable element in 'NameVars' is a string containing the pgo
250 /// name of a function. See `createPGOFuncNameVar` that creates these global
251 /// variables.
252 Error collectPGOFuncNameStrings(ArrayRef<GlobalVariable *> NameVars,
253                                 std::string &Result, bool doCompression = true);
254 
255 Error collectVTableStrings(ArrayRef<GlobalVariable *> VTables,
256                            std::string &Result, bool doCompression);
257 
258 /// Check if INSTR_PROF_RAW_VERSION_VAR is defined. This global is only being
259 /// set in IR PGO compilation.
260 bool isIRPGOFlagSet(const Module *M);
261 
262 /// Check if we can safely rename this Comdat function. Instances of the same
263 /// comdat function may have different control flows thus can not share the
264 /// same counter variable.
265 bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken = false);
266 
267 enum InstrProfValueKind : uint32_t {
268 #define VALUE_PROF_KIND(Enumerator, Value, Descr) Enumerator = Value,
269 #include "llvm/ProfileData/InstrProfData.inc"
270 };
271 
272 /// Get the value profile data for value site \p SiteIdx from \p InstrProfR
273 /// and annotate the instruction \p Inst with the value profile meta data.
274 /// Annotate up to \p MaxMDCount (default 3) number of records per value site.
275 void annotateValueSite(Module &M, Instruction &Inst,
276                        const InstrProfRecord &InstrProfR,
277                        InstrProfValueKind ValueKind, uint32_t SiteIndx,
278                        uint32_t MaxMDCount = 3);
279 
280 /// Same as the above interface but using an ArrayRef, as well as \p Sum.
281 /// This function will not annotate !prof metadata on the instruction if the
282 /// referenced array is empty.
283 void annotateValueSite(Module &M, Instruction &Inst,
284                        ArrayRef<InstrProfValueData> VDs, uint64_t Sum,
285                        InstrProfValueKind ValueKind, uint32_t MaxMDCount);
286 
287 // TODO: Unify metadata name 'PGOFuncName' and 'PGOName', by supporting read
288 // of this metadata for backward compatibility and generating 'PGOName' only.
289 /// Extract the value profile data from \p Inst and returns them if \p Inst is
290 /// annotated with value profile data. Returns an empty vector otherwise.
291 SmallVector<InstrProfValueData, 4>
292 getValueProfDataFromInst(const Instruction &Inst, InstrProfValueKind ValueKind,
293                          uint32_t MaxNumValueData, uint64_t &TotalC,
294                          bool GetNoICPValue = false);
295 
getPGOFuncNameMetadataName()296 inline StringRef getPGOFuncNameMetadataName() { return "PGOFuncName"; }
297 
getPGONameMetadataName()298 inline StringRef getPGONameMetadataName() { return "PGOName"; }
299 
300 /// Return the PGOFuncName meta data associated with a function.
301 MDNode *getPGOFuncNameMetadata(const Function &F);
302 
303 std::string getPGOName(const GlobalVariable &V, bool InLTO = false);
304 
305 /// Create the PGOFuncName meta data if PGOFuncName is different from
306 /// function's raw name. This should only apply to internal linkage functions
307 /// declared by users only.
308 /// TODO: Update all callers to 'createPGONameMetadata' and deprecate this
309 /// function.
310 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName);
311 
312 /// Create the PGOName metadata if a global object's PGO name is different from
313 /// its mangled name. This should apply to local-linkage global objects only.
314 void createPGONameMetadata(GlobalObject &GO, StringRef PGOName);
315 
316 /// Check if we can use Comdat for profile variables. This will eliminate
317 /// the duplicated profile variables for Comdat functions.
318 bool needsComdatForCounter(const GlobalObject &GV, const Module &M);
319 
320 /// An enum describing the attributes of an instrumented profile.
321 enum class InstrProfKind {
322   Unknown = 0x0,
323   // A frontend clang profile, incompatible with other attrs.
324   FrontendInstrumentation = 0x1,
325   // An IR-level profile (default when -fprofile-generate is used).
326   IRInstrumentation = 0x2,
327   // A profile with entry basic block instrumentation.
328   FunctionEntryInstrumentation = 0x4,
329   // A context sensitive IR-level profile.
330   ContextSensitive = 0x8,
331   // Use single byte probes for coverage.
332   SingleByteCoverage = 0x10,
333   // Only instrument the function entry basic block.
334   FunctionEntryOnly = 0x20,
335   // A memory profile collected using -fprofile=memory.
336   MemProf = 0x40,
337   // A temporal profile.
338   TemporalProfile = 0x80,
339   LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/TemporalProfile)
340 };
341 
342 const std::error_category &instrprof_category();
343 
344 enum class instrprof_error {
345   success = 0,
346   eof,
347   unrecognized_format,
348   bad_magic,
349   bad_header,
350   unsupported_version,
351   unsupported_hash_type,
352   too_large,
353   truncated,
354   malformed,
355   missing_correlation_info,
356   unexpected_correlation_info,
357   unable_to_correlate_profile,
358   unknown_function,
359   invalid_prof,
360   hash_mismatch,
361   count_mismatch,
362   bitmap_mismatch,
363   counter_overflow,
364   value_site_count_mismatch,
365   compress_failed,
366   uncompress_failed,
367   empty_raw_profile,
368   zlib_unavailable,
369   raw_profile_version_mismatch,
370   counter_value_too_large,
371 };
372 
373 /// An ordered list of functions identified by their NameRef found in
374 /// INSTR_PROF_DATA
375 struct TemporalProfTraceTy {
376   std::vector<uint64_t> FunctionNameRefs;
377   uint64_t Weight;
378   TemporalProfTraceTy(std::initializer_list<uint64_t> Trace = {},
379                       uint64_t Weight = 1)
FunctionNameRefsTemporalProfTraceTy380       : FunctionNameRefs(Trace), Weight(Weight) {}
381 
382   /// Use a set of temporal profile traces to create a list of balanced
383   /// partitioning function nodes used by BalancedPartitioning to generate a
384   /// function order that reduces page faults during startup
385   static void createBPFunctionNodes(ArrayRef<TemporalProfTraceTy> Traces,
386                                     std::vector<BPFunctionNode> &Nodes,
387                                     bool RemoveOutlierUNs = true);
388 };
389 
make_error_code(instrprof_error E)390 inline std::error_code make_error_code(instrprof_error E) {
391   return std::error_code(static_cast<int>(E), instrprof_category());
392 }
393 
394 class InstrProfError : public ErrorInfo<InstrProfError> {
395 public:
396   InstrProfError(instrprof_error Err, const Twine &ErrStr = Twine())
Err(Err)397       : Err(Err), Msg(ErrStr.str()) {
398     assert(Err != instrprof_error::success && "Not an error");
399   }
400 
401   std::string message() const override;
402 
log(raw_ostream & OS)403   void log(raw_ostream &OS) const override { OS << message(); }
404 
convertToErrorCode()405   std::error_code convertToErrorCode() const override {
406     return make_error_code(Err);
407   }
408 
get()409   instrprof_error get() const { return Err; }
getMessage()410   const std::string &getMessage() const { return Msg; }
411 
412   /// Consume an Error and return the raw enum value contained within it, and
413   /// the optional error message. The Error must either be a success value, or
414   /// contain a single InstrProfError.
take(Error E)415   static std::pair<instrprof_error, std::string> take(Error E) {
416     auto Err = instrprof_error::success;
417     std::string Msg = "";
418     handleAllErrors(std::move(E), [&Err, &Msg](const InstrProfError &IPE) {
419       assert(Err == instrprof_error::success && "Multiple errors encountered");
420       Err = IPE.get();
421       Msg = IPE.getMessage();
422     });
423     return {Err, Msg};
424   }
425 
426   static char ID;
427 
428 private:
429   instrprof_error Err;
430   std::string Msg;
431 };
432 
433 namespace object {
434 
435 class SectionRef;
436 
437 } // end namespace object
438 
439 namespace IndexedInstrProf {
440 
441 uint64_t ComputeHash(StringRef K);
442 
443 } // end namespace IndexedInstrProf
444 
445 /// A symbol table used for function [IR]PGO name look-up with keys
446 /// (such as pointers, md5hash values) to the function. A function's
447 /// [IR]PGO name or name's md5hash are used in retrieving the profile
448 /// data of the function. See \c getIRPGOFuncName() and \c getPGOFuncName
449 /// methods for details how [IR]PGO name is formed.
450 class InstrProfSymtab {
451 public:
452   using AddrHashMap = std::vector<std::pair<uint64_t, uint64_t>>;
453 
454 private:
455   using AddrIntervalMap =
456       IntervalMap<uint64_t, uint64_t, 4, IntervalMapHalfOpenInfo<uint64_t>>;
457   StringRef Data;
458   uint64_t Address = 0;
459   // Unique name strings. Used to ensure entries in MD5NameMap (a vector that's
460   // going to be sorted) has unique MD5 keys in the first place.
461   StringSet<> NameTab;
462   // Records the unique virtual table names. This is used by InstrProfWriter to
463   // write out an on-disk chained hash table of virtual table names.
464   // InstrProfWriter stores per function profile data (keyed by function names)
465   // so it doesn't use a StringSet for function names.
466   StringSet<> VTableNames;
467   // A map from MD5 keys to function name strings.
468   std::vector<std::pair<uint64_t, StringRef>> MD5NameMap;
469   // A map from MD5 keys to function define. We only populate this map
470   // when build the Symtab from a Module.
471   std::vector<std::pair<uint64_t, Function *>> MD5FuncMap;
472   // A map from MD5 to the global variable. This map is only populated when
473   // building the symtab from a module. Use separate container instances for
474   // `MD5FuncMap` and `MD5VTableMap`.
475   // TODO: Unify the container type and the lambda function 'mapName' inside
476   // add{Func,VTable}WithName.
477   DenseMap<uint64_t, GlobalVariable *> MD5VTableMap;
478   // A map from function runtime address to function name MD5 hash.
479   // This map is only populated and used by raw instr profile reader.
480   AddrHashMap AddrToMD5Map;
481 
482   AddrIntervalMap::Allocator VTableAddrMapAllocator;
483   // This map is only populated and used by raw instr profile reader.
484   AddrIntervalMap VTableAddrMap;
485   bool Sorted = false;
486 
getExternalSymbol()487   static StringRef getExternalSymbol() { return "** External Symbol **"; }
488 
489   // Returns the canonial name of the given PGOName. In a canonical name, all
490   // suffixes that begins with "." except ".__uniq." are stripped.
491   // FIXME: Unify this with `FunctionSamples::getCanonicalFnName`.
492   static StringRef getCanonicalName(StringRef PGOName);
493 
494   // Add the function into the symbol table, by creating the following
495   // map entries:
496   // name-set = {PGOFuncName} union {getCanonicalName(PGOFuncName)}
497   // - In MD5NameMap: <MD5Hash(name), name> for name in name-set
498   // - In MD5FuncMap: <MD5Hash(name), &F> for name in name-set
499   Error addFuncWithName(Function &F, StringRef PGOFuncName);
500 
501   // Add the vtable into the symbol table, by creating the following
502   // map entries:
503   // name-set = {PGOName} union {getCanonicalName(PGOName)}
504   // - In MD5NameMap:  <MD5Hash(name), name> for name in name-set
505   // - In MD5VTableMap: <MD5Hash(name), name> for name in name-set
506   Error addVTableWithName(GlobalVariable &V, StringRef PGOVTableName);
507 
508   // If the symtab is created by a series of calls to \c addFuncName, \c
509   // finalizeSymtab needs to be called before looking up function names.
510   // This is required because the underlying map is a vector (for space
511   // efficiency) which needs to be sorted.
512   inline void finalizeSymtab();
513 
514 public:
InstrProfSymtab()515   InstrProfSymtab() : VTableAddrMap(VTableAddrMapAllocator) {}
516 
517   // Not copyable or movable.
518   // Consider std::unique_ptr for move.
519   InstrProfSymtab(const InstrProfSymtab &) = delete;
520   InstrProfSymtab &operator=(const InstrProfSymtab &) = delete;
521   InstrProfSymtab(InstrProfSymtab &&) = delete;
522   InstrProfSymtab &operator=(InstrProfSymtab &&) = delete;
523 
524   /// Create InstrProfSymtab from an object file section which
525   /// contains function PGO names. When section may contain raw
526   /// string data or string data in compressed form. This method
527   /// only initialize the symtab with reference to the data and
528   /// the section base address. The decompression will be delayed
529   /// until before it is used. See also \c create(StringRef) method.
530   Error create(object::SectionRef &Section);
531 
532   /// \c NameStrings is a string composed of one of more sub-strings
533   ///  encoded in the format described in \c collectPGOFuncNameStrings.
534   /// This method is a wrapper to \c readAndDecodeStrings method.
535   Error create(StringRef NameStrings);
536 
537   /// Initialize symtab states with function names and vtable names. \c
538   /// FuncNameStrings is a string composed of one or more encoded function name
539   /// strings, and \c VTableNameStrings composes of one or more encoded vtable
540   /// names. This interface is solely used by raw profile reader.
541   Error create(StringRef FuncNameStrings, StringRef VTableNameStrings);
542 
543   /// Initialize 'this' with the set of vtable names encoded in
544   /// \c CompressedVTableNames.
545   Error initVTableNamesFromCompressedStrings(StringRef CompressedVTableNames);
546 
547   /// This interface is used by reader of CoverageMapping test
548   /// format.
549   inline Error create(StringRef D, uint64_t BaseAddr);
550 
551   /// A wrapper interface to populate the PGO symtab with functions
552   /// decls from module \c M. This interface is used by transformation
553   /// passes such as indirect function call promotion. Variable \c InLTO
554   /// indicates if this is called from LTO optimization passes.
555   Error create(Module &M, bool InLTO = false);
556 
557   /// Create InstrProfSymtab from a set of names iteratable from
558   /// \p IterRange. This interface is used by IndexedProfReader.
559   template <typename NameIterRange>
560   Error create(const NameIterRange &IterRange);
561 
562   /// Create InstrProfSymtab from a set of function names and vtable
563   /// names iteratable from \p IterRange. This interface is used by
564   /// IndexedProfReader.
565   template <typename FuncNameIterRange, typename VTableNameIterRange>
566   Error create(const FuncNameIterRange &FuncIterRange,
567                const VTableNameIterRange &VTableIterRange);
568 
569   // Map the MD5 of the symbol name to the name.
addSymbolName(StringRef SymbolName)570   Error addSymbolName(StringRef SymbolName) {
571     if (SymbolName.empty())
572       return make_error<InstrProfError>(instrprof_error::malformed,
573                                         "symbol name is empty");
574 
575     // Insert into NameTab so that MD5NameMap (a vector that will be sorted)
576     // won't have duplicated entries in the first place.
577     auto Ins = NameTab.insert(SymbolName);
578     if (Ins.second) {
579       MD5NameMap.push_back(std::make_pair(
580           IndexedInstrProf::ComputeHash(SymbolName), Ins.first->getKey()));
581       Sorted = false;
582     }
583     return Error::success();
584   }
585 
586   /// The method name is kept since there are many callers.
587   /// It just forwards to 'addSymbolName'.
addFuncName(StringRef FuncName)588   Error addFuncName(StringRef FuncName) { return addSymbolName(FuncName); }
589 
590   /// Adds VTableName as a known symbol, and inserts it to a map that
591   /// tracks all vtable names.
addVTableName(StringRef VTableName)592   Error addVTableName(StringRef VTableName) {
593     if (Error E = addSymbolName(VTableName))
594       return E;
595 
596     // Record VTableName. InstrProfWriter uses this set. The comment around
597     // class member explains why.
598     VTableNames.insert(VTableName);
599     return Error::success();
600   }
601 
getVTableNames()602   const StringSet<> &getVTableNames() const { return VTableNames; }
603 
604   /// Map a function address to its name's MD5 hash. This interface
605   /// is only used by the raw profiler reader.
mapAddress(uint64_t Addr,uint64_t MD5Val)606   void mapAddress(uint64_t Addr, uint64_t MD5Val) {
607     AddrToMD5Map.push_back(std::make_pair(Addr, MD5Val));
608   }
609 
610   /// Map the address range (i.e., [start_address, end_address)) of a variable
611   /// to  its names' MD5 hash. This interface is only used by the raw profile
612   /// reader.
mapVTableAddress(uint64_t StartAddr,uint64_t EndAddr,uint64_t MD5Val)613   void mapVTableAddress(uint64_t StartAddr, uint64_t EndAddr, uint64_t MD5Val) {
614     VTableAddrMap.insert(StartAddr, EndAddr, MD5Val);
615   }
616 
617   /// Return a function's hash, or 0, if the function isn't in this SymTab.
618   uint64_t getFunctionHashFromAddress(uint64_t Address);
619 
620   /// Return a vtable's hash, or 0 if the vtable doesn't exist in this SymTab.
621   uint64_t getVTableHashFromAddress(uint64_t Address);
622 
623   /// Return function's PGO name from the function name's symbol
624   /// address in the object file. If an error occurs, return
625   /// an empty string.
626   StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
627 
628   /// Return name of functions or global variables from the name's md5 hash
629   /// value. If not found, return an empty string.
630   inline StringRef getFuncOrVarName(uint64_t ValMD5Hash);
631 
632   /// Just like getFuncOrVarName, except that it will return literal string
633   /// 'External Symbol' if the function or global variable is external to
634   /// this symbol table.
635   inline StringRef getFuncOrVarNameIfDefined(uint64_t ValMD5Hash);
636 
637   /// True if Symbol is the value used to represent external symbols.
isExternalSymbol(const StringRef & Symbol)638   static bool isExternalSymbol(const StringRef &Symbol) {
639     return Symbol == InstrProfSymtab::getExternalSymbol();
640   }
641 
642   /// Return function from the name's md5 hash. Return nullptr if not found.
643   inline Function *getFunction(uint64_t FuncMD5Hash);
644 
645   /// Return the global variable corresponding to md5 hash. Return nullptr if
646   /// not found.
647   inline GlobalVariable *getGlobalVariable(uint64_t MD5Hash);
648 
649   /// Return the name section data.
getNameData()650   inline StringRef getNameData() const { return Data; }
651 
652   /// Dump the symbols in this table.
653   void dumpNames(raw_ostream &OS) const;
654 };
655 
create(StringRef D,uint64_t BaseAddr)656 Error InstrProfSymtab::create(StringRef D, uint64_t BaseAddr) {
657   Data = D;
658   Address = BaseAddr;
659   return Error::success();
660 }
661 
662 template <typename NameIterRange>
create(const NameIterRange & IterRange)663 Error InstrProfSymtab::create(const NameIterRange &IterRange) {
664   for (auto Name : IterRange)
665     if (Error E = addFuncName(Name))
666       return E;
667 
668   finalizeSymtab();
669   return Error::success();
670 }
671 
672 template <typename FuncNameIterRange, typename VTableNameIterRange>
create(const FuncNameIterRange & FuncIterRange,const VTableNameIterRange & VTableIterRange)673 Error InstrProfSymtab::create(const FuncNameIterRange &FuncIterRange,
674                               const VTableNameIterRange &VTableIterRange) {
675   // Iterate elements by StringRef rather than by const reference.
676   // StringRef is small enough, so the loop is efficient whether
677   // element in the range is std::string or StringRef.
678   for (StringRef Name : FuncIterRange)
679     if (Error E = addFuncName(Name))
680       return E;
681 
682   for (StringRef VTableName : VTableIterRange)
683     if (Error E = addVTableName(VTableName))
684       return E;
685 
686   finalizeSymtab();
687   return Error::success();
688 }
689 
finalizeSymtab()690 void InstrProfSymtab::finalizeSymtab() {
691   if (Sorted)
692     return;
693   llvm::sort(MD5NameMap, less_first());
694   llvm::sort(MD5FuncMap, less_first());
695   llvm::sort(AddrToMD5Map, less_first());
696   AddrToMD5Map.erase(llvm::unique(AddrToMD5Map), AddrToMD5Map.end());
697   Sorted = true;
698 }
699 
getFuncOrVarNameIfDefined(uint64_t MD5Hash)700 StringRef InstrProfSymtab::getFuncOrVarNameIfDefined(uint64_t MD5Hash) {
701   StringRef ret = getFuncOrVarName(MD5Hash);
702   if (ret.empty())
703     return InstrProfSymtab::getExternalSymbol();
704   return ret;
705 }
706 
getFuncOrVarName(uint64_t MD5Hash)707 StringRef InstrProfSymtab::getFuncOrVarName(uint64_t MD5Hash) {
708   finalizeSymtab();
709   auto Result = llvm::lower_bound(MD5NameMap, MD5Hash,
710                                   [](const std::pair<uint64_t, StringRef> &LHS,
711                                      uint64_t RHS) { return LHS.first < RHS; });
712   if (Result != MD5NameMap.end() && Result->first == MD5Hash)
713     return Result->second;
714   return StringRef();
715 }
716 
getFunction(uint64_t FuncMD5Hash)717 Function* InstrProfSymtab::getFunction(uint64_t FuncMD5Hash) {
718   finalizeSymtab();
719   auto Result = llvm::lower_bound(MD5FuncMap, FuncMD5Hash,
720                                   [](const std::pair<uint64_t, Function *> &LHS,
721                                      uint64_t RHS) { return LHS.first < RHS; });
722   if (Result != MD5FuncMap.end() && Result->first == FuncMD5Hash)
723     return Result->second;
724   return nullptr;
725 }
726 
getGlobalVariable(uint64_t MD5Hash)727 GlobalVariable *InstrProfSymtab::getGlobalVariable(uint64_t MD5Hash) {
728   return MD5VTableMap.lookup(MD5Hash);
729 }
730 
731 // To store the sums of profile count values, or the percentage of
732 // the sums of the total count values.
733 struct CountSumOrPercent {
734   uint64_t NumEntries = 0;
735   double CountSum = 0.0f;
736   std::array<double, IPVK_Last - IPVK_First + 1> ValueCounts = {};
737   CountSumOrPercent() = default;
resetCountSumOrPercent738   void reset() {
739     NumEntries = 0;
740     CountSum = 0.0f;
741     ValueCounts.fill(0.0f);
742   }
743 };
744 
745 // Function level or program level overlap information.
746 struct OverlapStats {
747   enum OverlapStatsLevel { ProgramLevel, FunctionLevel };
748   // Sum of the total count values for the base profile.
749   CountSumOrPercent Base;
750   // Sum of the total count values for the test profile.
751   CountSumOrPercent Test;
752   // Overlap lap score. Should be in range of [0.0f to 1.0f].
753   CountSumOrPercent Overlap;
754   CountSumOrPercent Mismatch;
755   CountSumOrPercent Unique;
756   OverlapStatsLevel Level;
757   const std::string *BaseFilename = nullptr;
758   const std::string *TestFilename = nullptr;
759   StringRef FuncName;
760   uint64_t FuncHash = 0;
761   bool Valid = false;
762 
LevelOverlapStats763   OverlapStats(OverlapStatsLevel L = ProgramLevel) : Level(L) {}
764 
765   void dump(raw_fd_ostream &OS) const;
766 
setFuncInfoOverlapStats767   void setFuncInfo(StringRef Name, uint64_t Hash) {
768     FuncName = Name;
769     FuncHash = Hash;
770   }
771 
772   Error accumulateCounts(const std::string &BaseFilename,
773                          const std::string &TestFilename, bool IsCS);
774   void addOneMismatch(const CountSumOrPercent &MismatchFunc);
775   void addOneUnique(const CountSumOrPercent &UniqueFunc);
776 
scoreOverlapStats777   static inline double score(uint64_t Val1, uint64_t Val2, double Sum1,
778                              double Sum2) {
779     if (Sum1 < 1.0f || Sum2 < 1.0f)
780       return 0.0f;
781     return std::min(Val1 / Sum1, Val2 / Sum2);
782   }
783 };
784 
785 // This is used to filter the functions whose overlap information
786 // to be output.
787 struct OverlapFuncFilters {
788   uint64_t ValueCutoff;
789   const std::string NameFilter;
790 };
791 
792 struct InstrProfValueSiteRecord {
793   /// Value profiling data pairs at a given value site.
794   std::vector<InstrProfValueData> ValueData;
795 
796   InstrProfValueSiteRecord() = default;
InstrProfValueSiteRecordInstrProfValueSiteRecord797   InstrProfValueSiteRecord(std::vector<InstrProfValueData> &&VD)
798       : ValueData(VD) {}
799 
800   /// Sort ValueData ascending by Value
sortByTargetValuesInstrProfValueSiteRecord801   void sortByTargetValues() {
802     llvm::sort(ValueData,
803                [](const InstrProfValueData &L, const InstrProfValueData &R) {
804                  return L.Value < R.Value;
805                });
806   }
807   /// Sort ValueData Descending by Count
808   inline void sortByCount();
809 
810   /// Merge data from another InstrProfValueSiteRecord
811   /// Optionally scale merged counts by \p Weight.
812   void merge(InstrProfValueSiteRecord &Input, uint64_t Weight,
813              function_ref<void(instrprof_error)> Warn);
814   /// Scale up value profile data counts by N (Numerator) / D (Denominator).
815   void scale(uint64_t N, uint64_t D, function_ref<void(instrprof_error)> Warn);
816 
817   /// Compute the overlap b/w this record and Input record.
818   void overlap(InstrProfValueSiteRecord &Input, uint32_t ValueKind,
819                OverlapStats &Overlap, OverlapStats &FuncLevelOverlap);
820 };
821 
822 /// Profiling information for a single function.
823 struct InstrProfRecord {
824   std::vector<uint64_t> Counts;
825   std::vector<uint8_t> BitmapBytes;
826 
827   InstrProfRecord() = default;
InstrProfRecordInstrProfRecord828   InstrProfRecord(std::vector<uint64_t> Counts) : Counts(std::move(Counts)) {}
InstrProfRecordInstrProfRecord829   InstrProfRecord(std::vector<uint64_t> Counts,
830                   std::vector<uint8_t> BitmapBytes)
831       : Counts(std::move(Counts)), BitmapBytes(std::move(BitmapBytes)) {}
832   InstrProfRecord(InstrProfRecord &&) = default;
InstrProfRecordInstrProfRecord833   InstrProfRecord(const InstrProfRecord &RHS)
834       : Counts(RHS.Counts), BitmapBytes(RHS.BitmapBytes),
835         ValueData(RHS.ValueData
836                       ? std::make_unique<ValueProfData>(*RHS.ValueData)
837                       : nullptr) {}
838   InstrProfRecord &operator=(InstrProfRecord &&) = default;
839   InstrProfRecord &operator=(const InstrProfRecord &RHS) {
840     Counts = RHS.Counts;
841     BitmapBytes = RHS.BitmapBytes;
842     if (!RHS.ValueData) {
843       ValueData = nullptr;
844       return *this;
845     }
846     if (!ValueData)
847       ValueData = std::make_unique<ValueProfData>(*RHS.ValueData);
848     else
849       *ValueData = *RHS.ValueData;
850     return *this;
851   }
852 
853   /// Return the number of value profile kinds with non-zero number
854   /// of profile sites.
855   inline uint32_t getNumValueKinds() const;
856   /// Return the number of instrumented sites for ValueKind.
857   inline uint32_t getNumValueSites(uint32_t ValueKind) const;
858 
859   /// Return the total number of ValueData for ValueKind.
860   inline uint32_t getNumValueData(uint32_t ValueKind) const;
861 
862   /// Return the array of profiled values at \p Site.
863   inline ArrayRef<InstrProfValueData> getValueArrayForSite(uint32_t ValueKind,
864                                                            uint32_t Site) const;
865 
866   /// Reserve space for NumValueSites sites.
867   inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
868 
869   /// Add ValueData for ValueKind at value Site.  We do not support adding sites
870   /// out of order.  Site must go up from 0 one by one.
871   void addValueData(uint32_t ValueKind, uint32_t Site,
872                     ArrayRef<InstrProfValueData> VData,
873                     InstrProfSymtab *SymTab);
874 
875   /// Merge the counts in \p Other into this one.
876   /// Optionally scale merged counts by \p Weight.
877   void merge(InstrProfRecord &Other, uint64_t Weight,
878              function_ref<void(instrprof_error)> Warn);
879 
880   /// Scale up profile counts (including value profile data) by
881   /// a factor of (N / D).
882   void scale(uint64_t N, uint64_t D, function_ref<void(instrprof_error)> Warn);
883 
884   /// Sort value profile data (per site) by count.
sortValueDataInstrProfRecord885   void sortValueData() {
886     for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
887       for (auto &SR : getValueSitesForKind(Kind))
888         SR.sortByCount();
889   }
890 
891   /// Clear value data entries and edge counters.
ClearInstrProfRecord892   void Clear() {
893     Counts.clear();
894     clearValueData();
895   }
896 
897   /// Clear value data entries
clearValueDataInstrProfRecord898   void clearValueData() { ValueData = nullptr; }
899 
900   /// Compute the sums of all counts and store in Sum.
901   void accumulateCounts(CountSumOrPercent &Sum) const;
902 
903   /// Compute the overlap b/w this IntrprofRecord and Other.
904   void overlap(InstrProfRecord &Other, OverlapStats &Overlap,
905                OverlapStats &FuncLevelOverlap, uint64_t ValueCutoff);
906 
907   /// Compute the overlap of value profile counts.
908   void overlapValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
909                             OverlapStats &Overlap,
910                             OverlapStats &FuncLevelOverlap);
911 
912   enum CountPseudoKind {
913     NotPseudo = 0,
914     PseudoHot,
915     PseudoWarm,
916   };
917   enum PseudoCountVal {
918     HotFunctionVal = -1,
919     WarmFunctionVal = -2,
920   };
getCountPseudoKindInstrProfRecord921   CountPseudoKind getCountPseudoKind() const {
922     uint64_t FirstCount = Counts[0];
923     if (FirstCount == (uint64_t)HotFunctionVal)
924       return PseudoHot;
925     if (FirstCount == (uint64_t)WarmFunctionVal)
926       return PseudoWarm;
927     return NotPseudo;
928   }
setPseudoCountInstrProfRecord929   void setPseudoCount(CountPseudoKind Kind) {
930     if (Kind == PseudoHot)
931       Counts[0] = (uint64_t)HotFunctionVal;
932     else if (Kind == PseudoWarm)
933       Counts[0] = (uint64_t)WarmFunctionVal;
934   }
935 
936 private:
937   using ValueProfData = std::array<std::vector<InstrProfValueSiteRecord>,
938                                    IPVK_Last - IPVK_First + 1>;
939   std::unique_ptr<ValueProfData> ValueData;
940 
941   MutableArrayRef<InstrProfValueSiteRecord>
getValueSitesForKindInstrProfRecord942   getValueSitesForKind(uint32_t ValueKind) {
943     // Cast to /add/ const (should be an implicit_cast, ideally, if that's ever
944     // implemented in LLVM) to call the const overload of this function, then
945     // cast away the constness from the result.
946     auto AR = const_cast<const InstrProfRecord *>(this)->getValueSitesForKind(
947         ValueKind);
948     return MutableArrayRef(
949         const_cast<InstrProfValueSiteRecord *>(AR.data()), AR.size());
950   }
951   ArrayRef<InstrProfValueSiteRecord>
getValueSitesForKindInstrProfRecord952   getValueSitesForKind(uint32_t ValueKind) const {
953     if (!ValueData)
954       return std::nullopt;
955     assert(IPVK_First <= ValueKind && ValueKind <= IPVK_Last &&
956            "Unknown value kind!");
957     return (*ValueData)[ValueKind - IPVK_First];
958   }
959 
960   std::vector<InstrProfValueSiteRecord> &
getOrCreateValueSitesForKindInstrProfRecord961   getOrCreateValueSitesForKind(uint32_t ValueKind) {
962     if (!ValueData)
963       ValueData = std::make_unique<ValueProfData>();
964     assert(IPVK_First <= ValueKind && ValueKind <= IPVK_Last &&
965            "Unknown value kind!");
966     return (*ValueData)[ValueKind - IPVK_First];
967   }
968 
969   // Map indirect call target name hash to name string.
970   uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
971                       InstrProfSymtab *SymTab);
972 
973   // Merge Value Profile data from Src record to this record for ValueKind.
974   // Scale merged value counts by \p Weight.
975   void mergeValueProfData(uint32_t ValkeKind, InstrProfRecord &Src,
976                           uint64_t Weight,
977                           function_ref<void(instrprof_error)> Warn);
978 
979   // Scale up value profile data count by N (Numerator) / D (Denominator).
980   void scaleValueProfData(uint32_t ValueKind, uint64_t N, uint64_t D,
981                           function_ref<void(instrprof_error)> Warn);
982 };
983 
984 struct NamedInstrProfRecord : InstrProfRecord {
985   StringRef Name;
986   uint64_t Hash;
987 
988   // We reserve this bit as the flag for context sensitive profile record.
989   static const int CS_FLAG_IN_FUNC_HASH = 60;
990 
991   NamedInstrProfRecord() = default;
NamedInstrProfRecordNamedInstrProfRecord992   NamedInstrProfRecord(StringRef Name, uint64_t Hash,
993                        std::vector<uint64_t> Counts)
994       : InstrProfRecord(std::move(Counts)), Name(Name), Hash(Hash) {}
NamedInstrProfRecordNamedInstrProfRecord995   NamedInstrProfRecord(StringRef Name, uint64_t Hash,
996                        std::vector<uint64_t> Counts,
997                        std::vector<uint8_t> BitmapBytes)
998       : InstrProfRecord(std::move(Counts), std::move(BitmapBytes)), Name(Name),
999         Hash(Hash) {}
1000 
hasCSFlagInHashNamedInstrProfRecord1001   static bool hasCSFlagInHash(uint64_t FuncHash) {
1002     return ((FuncHash >> CS_FLAG_IN_FUNC_HASH) & 1);
1003   }
setCSFlagInHashNamedInstrProfRecord1004   static void setCSFlagInHash(uint64_t &FuncHash) {
1005     FuncHash |= ((uint64_t)1 << CS_FLAG_IN_FUNC_HASH);
1006   }
1007 };
1008 
getNumValueKinds()1009 uint32_t InstrProfRecord::getNumValueKinds() const {
1010   uint32_t NumValueKinds = 0;
1011   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
1012     NumValueKinds += !(getValueSitesForKind(Kind).empty());
1013   return NumValueKinds;
1014 }
1015 
getNumValueData(uint32_t ValueKind)1016 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
1017   uint32_t N = 0;
1018   for (const auto &SR : getValueSitesForKind(ValueKind))
1019     N += SR.ValueData.size();
1020   return N;
1021 }
1022 
getNumValueSites(uint32_t ValueKind)1023 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
1024   return getValueSitesForKind(ValueKind).size();
1025 }
1026 
1027 ArrayRef<InstrProfValueData>
getValueArrayForSite(uint32_t ValueKind,uint32_t Site)1028 InstrProfRecord::getValueArrayForSite(uint32_t ValueKind, uint32_t Site) const {
1029   return getValueSitesForKind(ValueKind)[Site].ValueData;
1030 }
1031 
reserveSites(uint32_t ValueKind,uint32_t NumValueSites)1032 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
1033   if (!NumValueSites)
1034     return;
1035   getOrCreateValueSitesForKind(ValueKind).reserve(NumValueSites);
1036 }
1037 
1038 // Include definitions for value profile data
1039 #define INSTR_PROF_VALUE_PROF_DATA
1040 #include "llvm/ProfileData/InstrProfData.inc"
1041 
sortByCount()1042 void InstrProfValueSiteRecord::sortByCount() {
1043   llvm::stable_sort(
1044       ValueData, [](const InstrProfValueData &L, const InstrProfValueData &R) {
1045         return L.Count > R.Count;
1046       });
1047   // Now truncate
1048   size_t max_s = INSTR_PROF_MAX_NUM_VAL_PER_SITE;
1049   if (ValueData.size() > max_s)
1050     ValueData.resize(max_s);
1051 }
1052 
1053 namespace IndexedInstrProf {
1054 
1055 enum class HashT : uint32_t {
1056   MD5,
1057   Last = MD5
1058 };
1059 
ComputeHash(HashT Type,StringRef K)1060 inline uint64_t ComputeHash(HashT Type, StringRef K) {
1061   switch (Type) {
1062   case HashT::MD5:
1063     return MD5Hash(K);
1064   }
1065   llvm_unreachable("Unhandled hash type");
1066 }
1067 
1068 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
1069 
1070 enum ProfVersion {
1071   // Version 1 is the first version. In this version, the value of
1072   // a key/value pair can only include profile data of a single function.
1073   // Due to this restriction, the number of block counters for a given
1074   // function is not recorded but derived from the length of the value.
1075   Version1 = 1,
1076   // The version 2 format supports recording profile data of multiple
1077   // functions which share the same key in one value field. To support this,
1078   // the number block counters is recorded as an uint64_t field right after the
1079   // function structural hash.
1080   Version2 = 2,
1081   // Version 3 supports value profile data. The value profile data is expected
1082   // to follow the block counter profile data.
1083   Version3 = 3,
1084   // In this version, profile summary data \c IndexedInstrProf::Summary is
1085   // stored after the profile header.
1086   Version4 = 4,
1087   // In this version, the frontend PGO stable hash algorithm defaults to V2.
1088   Version5 = 5,
1089   // In this version, the frontend PGO stable hash algorithm got fixed and
1090   // may produce hashes different from Version5.
1091   Version6 = 6,
1092   // An additional counter is added around logical operators.
1093   Version7 = 7,
1094   // An additional (optional) memory profile type is added.
1095   Version8 = 8,
1096   // Binary ids are added.
1097   Version9 = 9,
1098   // An additional (optional) temporal profile traces section is added.
1099   Version10 = 10,
1100   // An additional field is used for bitmap bytes.
1101   Version11 = 11,
1102   // VTable profiling, decision record and bitmap are modified for mcdc.
1103   Version12 = 12,
1104   // The current version is 12.
1105   CurrentVersion = INSTR_PROF_INDEX_VERSION
1106 };
1107 const uint64_t Version = ProfVersion::CurrentVersion;
1108 
1109 const HashT HashType = HashT::MD5;
1110 
ComputeHash(StringRef K)1111 inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); }
1112 
1113 // This structure defines the file header of the LLVM profile
1114 // data file in indexed-format. Please update llvm/docs/InstrProfileFormat.rst
1115 // as appropriate when updating the indexed profile format.
1116 struct Header {
1117   uint64_t Magic = IndexedInstrProf::Magic;
1118   // The lower 32 bits specify the version of the indexed profile.
1119   // The most significant 32 bits are reserved to specify the variant types of
1120   // the profile.
1121   uint64_t Version = 0;
1122   uint64_t Unused = 0; // Becomes unused since version 4
1123   uint64_t HashType = static_cast<uint64_t>(IndexedInstrProf::HashType);
1124   // This field records the offset of this hash table's metadata (i.e., the
1125   // number of buckets and entries), which follows right after the payload of
1126   // the entire hash table.
1127   uint64_t HashOffset = 0;
1128   uint64_t MemProfOffset = 0;
1129   uint64_t BinaryIdOffset = 0;
1130   uint64_t TemporalProfTracesOffset = 0;
1131   uint64_t VTableNamesOffset = 0;
1132   // New fields should only be added at the end to ensure that the size
1133   // computation is correct. The methods below need to be updated to ensure that
1134   // the new field is read correctly.
1135 
1136   // Reads a header struct from the buffer. Header fields are in machine native
1137   // endianness.
1138   static Expected<Header> readFromBuffer(const unsigned char *Buffer);
1139 
1140   // Returns the size of the header in bytes for all valid fields based on the
1141   // version. I.e a older version header will return a smaller size.
1142   size_t size() const;
1143 
1144   // Return the indexed profile version, i.e., the least significant 32 bits
1145   // in Header.Version.
1146   uint64_t getIndexedProfileVersion() const;
1147 };
1148 
1149 // Profile summary data recorded in the profile data file in indexed
1150 // format. It is introduced in version 4. The summary data follows
1151 // right after the profile file header.
1152 struct Summary {
1153   struct Entry {
1154     uint64_t Cutoff; ///< The required percentile of total execution count.
1155     uint64_t
1156         MinBlockCount;  ///< The minimum execution count for this percentile.
1157     uint64_t NumBlocks; ///< Number of blocks >= the minumum execution count.
1158   };
1159   // The field kind enumerator to assigned value mapping should remain
1160   // unchanged  when a new kind is added or an old kind gets deleted in
1161   // the future.
1162   enum SummaryFieldKind {
1163     /// The total number of functions instrumented.
1164     TotalNumFunctions = 0,
1165     /// Total number of instrumented blocks/edges.
1166     TotalNumBlocks = 1,
1167     /// The maximal execution count among all functions.
1168     /// This field does not exist for profile data from IR based
1169     /// instrumentation.
1170     MaxFunctionCount = 2,
1171     /// Max block count of the program.
1172     MaxBlockCount = 3,
1173     /// Max internal block count of the program (excluding entry blocks).
1174     MaxInternalBlockCount = 4,
1175     /// The sum of all instrumented block counts.
1176     TotalBlockCount = 5,
1177     NumKinds = TotalBlockCount + 1
1178   };
1179 
1180   // The number of summmary fields following the summary header.
1181   uint64_t NumSummaryFields;
1182   // The number of Cutoff Entries (Summary::Entry) following summary fields.
1183   uint64_t NumCutoffEntries;
1184 
1185   Summary() = delete;
SummarySummary1186   Summary(uint32_t Size) { memset(this, 0, Size); }
1187 
deleteSummary1188   void operator delete(void *ptr) { ::operator delete(ptr); }
1189 
getSizeSummary1190   static uint32_t getSize(uint32_t NumSumFields, uint32_t NumCutoffEntries) {
1191     return sizeof(Summary) + NumCutoffEntries * sizeof(Entry) +
1192            NumSumFields * sizeof(uint64_t);
1193   }
1194 
getSummaryDataBaseSummary1195   const uint64_t *getSummaryDataBase() const {
1196     return reinterpret_cast<const uint64_t *>(this + 1);
1197   }
1198 
getSummaryDataBaseSummary1199   uint64_t *getSummaryDataBase() {
1200     return reinterpret_cast<uint64_t *>(this + 1);
1201   }
1202 
getCutoffEntryBaseSummary1203   const Entry *getCutoffEntryBase() const {
1204     return reinterpret_cast<const Entry *>(
1205         &getSummaryDataBase()[NumSummaryFields]);
1206   }
1207 
getCutoffEntryBaseSummary1208   Entry *getCutoffEntryBase() {
1209     return reinterpret_cast<Entry *>(&getSummaryDataBase()[NumSummaryFields]);
1210   }
1211 
getSummary1212   uint64_t get(SummaryFieldKind K) const {
1213     return getSummaryDataBase()[K];
1214   }
1215 
setSummary1216   void set(SummaryFieldKind K, uint64_t V) {
1217     getSummaryDataBase()[K] = V;
1218   }
1219 
getEntrySummary1220   const Entry &getEntry(uint32_t I) const { return getCutoffEntryBase()[I]; }
1221 
setEntrySummary1222   void setEntry(uint32_t I, const ProfileSummaryEntry &E) {
1223     Entry &ER = getCutoffEntryBase()[I];
1224     ER.Cutoff = E.Cutoff;
1225     ER.MinBlockCount = E.MinCount;
1226     ER.NumBlocks = E.NumCounts;
1227   }
1228 };
1229 
allocSummary(uint32_t TotalSize)1230 inline std::unique_ptr<Summary> allocSummary(uint32_t TotalSize) {
1231   return std::unique_ptr<Summary>(new (::operator new(TotalSize))
1232                                       Summary(TotalSize));
1233 }
1234 
1235 } // end namespace IndexedInstrProf
1236 
1237 namespace RawInstrProf {
1238 
1239 // Version 1: First version
1240 // Version 2: Added value profile data section. Per-function control data
1241 // struct has more fields to describe value profile information.
1242 // Version 3: Compressed name section support. Function PGO name reference
1243 // from control data struct is changed from raw pointer to Name's MD5 value.
1244 // Version 4: ValueDataBegin and ValueDataSizes fields are removed from the
1245 // raw header.
1246 // Version 5: Bit 60 of FuncHash is reserved for the flag for the context
1247 // sensitive records.
1248 // Version 6: Added binary id.
1249 // Version 7: Reorder binary id and include version in signature.
1250 // Version 8: Use relative counter pointer.
1251 // Version 9: Added relative bitmap bytes pointer and count used by MC/DC.
1252 // Version 10: Added vtable, a new type of value profile data.
1253 const uint64_t Version = INSTR_PROF_RAW_VERSION;
1254 
1255 template <class IntPtrT> inline uint64_t getMagic();
1256 template <> inline uint64_t getMagic<uint64_t>() {
1257   return INSTR_PROF_RAW_MAGIC_64;
1258 }
1259 
1260 template <> inline uint64_t getMagic<uint32_t>() {
1261   return INSTR_PROF_RAW_MAGIC_32;
1262 }
1263 
1264 // Per-function profile data header/control structure.
1265 // The definition should match the structure defined in
1266 // compiler-rt/lib/profile/InstrProfiling.h.
1267 // It should also match the synthesized type in
1268 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
1269 template <class IntPtrT> struct alignas(8) ProfileData {
1270 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
1271 #include "llvm/ProfileData/InstrProfData.inc"
1272 };
1273 
1274 template <class IntPtrT> struct alignas(8) VTableProfileData {
1275 #define INSTR_PROF_VTABLE_DATA(Type, LLVMType, Name, Init) Type Name;
1276 #include "llvm/ProfileData/InstrProfData.inc"
1277 };
1278 
1279 // File header structure of the LLVM profile data in raw format.
1280 // The definition should match the header referenced in
1281 // compiler-rt/lib/profile/InstrProfilingFile.c  and
1282 // InstrProfilingBuffer.c.
1283 struct Header {
1284 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
1285 #include "llvm/ProfileData/InstrProfData.inc"
1286 };
1287 
1288 } // end namespace RawInstrProf
1289 
1290 // Create the variable for the profile file name.
1291 void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput);
1292 
1293 // Whether to compress function names in profile records, and filenames in
1294 // code coverage mappings. Used by the Instrumentation library and unit tests.
1295 extern cl::opt<bool> DoInstrProfNameCompression;
1296 
1297 } // end namespace llvm
1298 #endif // LLVM_PROFILEDATA_INSTRPROF_H
1299