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 /// Extract the value profile data from \p Inst which is annotated with
288 /// value profile meta data. Return false if there is no value data annotated,
289 /// otherwise return true.
290 bool getValueProfDataFromInst(const Instruction &Inst,
291                               InstrProfValueKind ValueKind,
292                               uint32_t MaxNumValueData,
293                               InstrProfValueData ValueData[],
294                               uint32_t &ActualNumValueData, uint64_t &TotalC,
295                               bool GetNoICPValue = false);
296 
297 /// Extract the value profile data from \p Inst and returns them if \p Inst is
298 /// annotated with value profile data. Returns nullptr otherwise. It's similar
299 /// to `getValueProfDataFromInst` above except that an array is allocated only
300 /// after a preliminary checking that the value profiles of kind `ValueKind`
301 /// exist.
302 std::unique_ptr<InstrProfValueData[]>
303 getValueProfDataFromInst(const Instruction &Inst, InstrProfValueKind ValueKind,
304                          uint32_t MaxNumValueData, uint32_t &ActualNumValueData,
305                          uint64_t &TotalC, bool GetNoICPValue = false);
306 
getPGOFuncNameMetadataName()307 inline StringRef getPGOFuncNameMetadataName() { return "PGOFuncName"; }
308 
309 /// Return the PGOFuncName meta data associated with a function.
310 MDNode *getPGOFuncNameMetadata(const Function &F);
311 
312 std::string getPGOName(const GlobalVariable &V, bool InLTO = false);
313 
314 /// Create the PGOFuncName meta data if PGOFuncName is different from
315 /// function's raw name. This should only apply to internal linkage functions
316 /// declared by users only.
317 void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName);
318 
319 /// Check if we can use Comdat for profile variables. This will eliminate
320 /// the duplicated profile variables for Comdat functions.
321 bool needsComdatForCounter(const GlobalObject &GV, const Module &M);
322 
323 /// An enum describing the attributes of an instrumented profile.
324 enum class InstrProfKind {
325   Unknown = 0x0,
326   // A frontend clang profile, incompatible with other attrs.
327   FrontendInstrumentation = 0x1,
328   // An IR-level profile (default when -fprofile-generate is used).
329   IRInstrumentation = 0x2,
330   // A profile with entry basic block instrumentation.
331   FunctionEntryInstrumentation = 0x4,
332   // A context sensitive IR-level profile.
333   ContextSensitive = 0x8,
334   // Use single byte probes for coverage.
335   SingleByteCoverage = 0x10,
336   // Only instrument the function entry basic block.
337   FunctionEntryOnly = 0x20,
338   // A memory profile collected using -fprofile=memory.
339   MemProf = 0x40,
340   // A temporal profile.
341   TemporalProfile = 0x80,
342   LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/TemporalProfile)
343 };
344 
345 const std::error_category &instrprof_category();
346 
347 enum class instrprof_error {
348   success = 0,
349   eof,
350   unrecognized_format,
351   bad_magic,
352   bad_header,
353   unsupported_version,
354   unsupported_hash_type,
355   too_large,
356   truncated,
357   malformed,
358   missing_correlation_info,
359   unexpected_correlation_info,
360   unable_to_correlate_profile,
361   unknown_function,
362   invalid_prof,
363   hash_mismatch,
364   count_mismatch,
365   bitmap_mismatch,
366   counter_overflow,
367   value_site_count_mismatch,
368   compress_failed,
369   uncompress_failed,
370   empty_raw_profile,
371   zlib_unavailable,
372   raw_profile_version_mismatch,
373   counter_value_too_large,
374 };
375 
376 /// An ordered list of functions identified by their NameRef found in
377 /// INSTR_PROF_DATA
378 struct TemporalProfTraceTy {
379   std::vector<uint64_t> FunctionNameRefs;
380   uint64_t Weight;
381   TemporalProfTraceTy(std::initializer_list<uint64_t> Trace = {},
382                       uint64_t Weight = 1)
FunctionNameRefsTemporalProfTraceTy383       : FunctionNameRefs(Trace), Weight(Weight) {}
384 
385   /// Use a set of temporal profile traces to create a list of balanced
386   /// partitioning function nodes used by BalancedPartitioning to generate a
387   /// function order that reduces page faults during startup
388   static std::vector<BPFunctionNode>
389   createBPFunctionNodes(ArrayRef<TemporalProfTraceTy> Traces);
390 };
391 
make_error_code(instrprof_error E)392 inline std::error_code make_error_code(instrprof_error E) {
393   return std::error_code(static_cast<int>(E), instrprof_category());
394 }
395 
396 class InstrProfError : public ErrorInfo<InstrProfError> {
397 public:
398   InstrProfError(instrprof_error Err, const Twine &ErrStr = Twine())
Err(Err)399       : Err(Err), Msg(ErrStr.str()) {
400     assert(Err != instrprof_error::success && "Not an error");
401   }
402 
403   std::string message() const override;
404 
log(raw_ostream & OS)405   void log(raw_ostream &OS) const override { OS << message(); }
406 
convertToErrorCode()407   std::error_code convertToErrorCode() const override {
408     return make_error_code(Err);
409   }
410 
get()411   instrprof_error get() const { return Err; }
getMessage()412   const std::string &getMessage() const { return Msg; }
413 
414   /// Consume an Error and return the raw enum value contained within it, and
415   /// the optional error message. The Error must either be a success value, or
416   /// contain a single InstrProfError.
take(Error E)417   static std::pair<instrprof_error, std::string> take(Error E) {
418     auto Err = instrprof_error::success;
419     std::string Msg = "";
420     handleAllErrors(std::move(E), [&Err, &Msg](const InstrProfError &IPE) {
421       assert(Err == instrprof_error::success && "Multiple errors encountered");
422       Err = IPE.get();
423       Msg = IPE.getMessage();
424     });
425     return {Err, Msg};
426   }
427 
428   static char ID;
429 
430 private:
431   instrprof_error Err;
432   std::string Msg;
433 };
434 
435 namespace object {
436 
437 class SectionRef;
438 
439 } // end namespace object
440 
441 namespace IndexedInstrProf {
442 
443 uint64_t ComputeHash(StringRef K);
444 
445 } // end namespace IndexedInstrProf
446 
447 /// A symbol table used for function [IR]PGO name look-up with keys
448 /// (such as pointers, md5hash values) to the function. A function's
449 /// [IR]PGO name or name's md5hash are used in retrieving the profile
450 /// data of the function. See \c getIRPGOFuncName() and \c getPGOFuncName
451 /// methods for details how [IR]PGO name is formed.
452 class InstrProfSymtab {
453 public:
454   using AddrHashMap = std::vector<std::pair<uint64_t, uint64_t>>;
455 
456 private:
457   using AddrIntervalMap =
458       IntervalMap<uint64_t, uint64_t, 4, IntervalMapHalfOpenInfo<uint64_t>>;
459   StringRef Data;
460   uint64_t Address = 0;
461   // Unique name strings. Used to ensure entries in MD5NameMap (a vector that's
462   // going to be sorted) has unique MD5 keys in the first place.
463   StringSet<> NameTab;
464   // Records the unique virtual table names. This is used by InstrProfWriter to
465   // write out an on-disk chained hash table of virtual table names.
466   // InstrProfWriter stores per function profile data (keyed by function names)
467   // so it doesn't use a StringSet for function names.
468   StringSet<> VTableNames;
469   // A map from MD5 keys to function name strings.
470   std::vector<std::pair<uint64_t, StringRef>> MD5NameMap;
471   // A map from MD5 keys to function define. We only populate this map
472   // when build the Symtab from a Module.
473   std::vector<std::pair<uint64_t, Function *>> MD5FuncMap;
474   // A map from MD5 to the global variable. This map is only populated when
475   // building the symtab from a module. Use separate container instances for
476   // `MD5FuncMap` and `MD5VTableMap`.
477   // TODO: Unify the container type and the lambda function 'mapName' inside
478   // add{Func,VTable}WithName.
479   DenseMap<uint64_t, GlobalVariable *> MD5VTableMap;
480   // A map from function runtime address to function name MD5 hash.
481   // This map is only populated and used by raw instr profile reader.
482   AddrHashMap AddrToMD5Map;
483 
484   AddrIntervalMap::Allocator VTableAddrMapAllocator;
485   // This map is only populated and used by raw instr profile reader.
486   AddrIntervalMap VTableAddrMap;
487   bool Sorted = false;
488 
getExternalSymbol()489   static StringRef getExternalSymbol() { return "** External Symbol **"; }
490 
491   // Returns the canonial name of the given PGOName. In a canonical name, all
492   // suffixes that begins with "." except ".__uniq." are stripped.
493   // FIXME: Unify this with `FunctionSamples::getCanonicalFnName`.
494   static StringRef getCanonicalName(StringRef PGOName);
495 
496   // Add the function into the symbol table, by creating the following
497   // map entries:
498   // name-set = {PGOFuncName} union {getCanonicalName(PGOFuncName)}
499   // - In MD5NameMap: <MD5Hash(name), name> for name in name-set
500   // - In MD5FuncMap: <MD5Hash(name), &F> for name in name-set
501   Error addFuncWithName(Function &F, StringRef PGOFuncName);
502 
503   // Add the vtable into the symbol table, by creating the following
504   // map entries:
505   // name-set = {PGOName} union {getCanonicalName(PGOName)}
506   // - In MD5NameMap:  <MD5Hash(name), name> for name in name-set
507   // - In MD5VTableMap: <MD5Hash(name), name> for name in name-set
508   Error addVTableWithName(GlobalVariable &V, StringRef PGOVTableName);
509 
510   // If the symtab is created by a series of calls to \c addFuncName, \c
511   // finalizeSymtab needs to be called before looking up function names.
512   // This is required because the underlying map is a vector (for space
513   // efficiency) which needs to be sorted.
514   inline void finalizeSymtab();
515 
516 public:
InstrProfSymtab()517   InstrProfSymtab() : VTableAddrMap(VTableAddrMapAllocator) {}
518 
519   // Not copyable or movable.
520   // Consider std::unique_ptr for move.
521   InstrProfSymtab(const InstrProfSymtab &) = delete;
522   InstrProfSymtab &operator=(const InstrProfSymtab &) = delete;
523   InstrProfSymtab(InstrProfSymtab &&) = delete;
524   InstrProfSymtab &operator=(InstrProfSymtab &&) = delete;
525 
526   /// Create InstrProfSymtab from an object file section which
527   /// contains function PGO names. When section may contain raw
528   /// string data or string data in compressed form. This method
529   /// only initialize the symtab with reference to the data and
530   /// the section base address. The decompression will be delayed
531   /// until before it is used. See also \c create(StringRef) method.
532   Error create(object::SectionRef &Section);
533 
534   /// \c NameStrings is a string composed of one of more sub-strings
535   ///  encoded in the format described in \c collectPGOFuncNameStrings.
536   /// This method is a wrapper to \c readAndDecodeStrings method.
537   Error create(StringRef NameStrings);
538 
539   /// Initialize symtab states with function names and vtable names. \c
540   /// FuncNameStrings is a string composed of one or more encoded function name
541   /// strings, and \c VTableNameStrings composes of one or more encoded vtable
542   /// names. This interface is solely used by raw profile reader.
543   Error create(StringRef FuncNameStrings, StringRef VTableNameStrings);
544 
545   /// Initialize 'this' with the set of vtable names encoded in
546   /// \c CompressedVTableNames.
547   Error initVTableNamesFromCompressedStrings(StringRef CompressedVTableNames);
548 
549   /// This interface is used by reader of CoverageMapping test
550   /// format.
551   inline Error create(StringRef D, uint64_t BaseAddr);
552 
553   /// A wrapper interface to populate the PGO symtab with functions
554   /// decls from module \c M. This interface is used by transformation
555   /// passes such as indirect function call promotion. Variable \c InLTO
556   /// indicates if this is called from LTO optimization passes.
557   Error create(Module &M, bool InLTO = false);
558 
559   /// Create InstrProfSymtab from a set of names iteratable from
560   /// \p IterRange. This interface is used by IndexedProfReader.
561   template <typename NameIterRange>
562   Error create(const NameIterRange &IterRange);
563 
564   /// Create InstrProfSymtab from a set of function names and vtable
565   /// names iteratable from \p IterRange. This interface is used by
566   /// IndexedProfReader.
567   template <typename FuncNameIterRange, typename VTableNameIterRange>
568   Error create(const FuncNameIterRange &FuncIterRange,
569                const VTableNameIterRange &VTableIterRange);
570 
571   // Map the MD5 of the symbol name to the name.
addSymbolName(StringRef SymbolName)572   Error addSymbolName(StringRef SymbolName) {
573     if (SymbolName.empty())
574       return make_error<InstrProfError>(instrprof_error::malformed,
575                                         "symbol name is empty");
576 
577     // Insert into NameTab so that MD5NameMap (a vector that will be sorted)
578     // won't have duplicated entries in the first place.
579     auto Ins = NameTab.insert(SymbolName);
580     if (Ins.second) {
581       MD5NameMap.push_back(std::make_pair(
582           IndexedInstrProf::ComputeHash(SymbolName), Ins.first->getKey()));
583       Sorted = false;
584     }
585     return Error::success();
586   }
587 
588   /// The method name is kept since there are many callers.
589   /// It just forwards to 'addSymbolName'.
addFuncName(StringRef FuncName)590   Error addFuncName(StringRef FuncName) { return addSymbolName(FuncName); }
591 
592   /// Adds VTableName as a known symbol, and inserts it to a map that
593   /// tracks all vtable names.
addVTableName(StringRef VTableName)594   Error addVTableName(StringRef VTableName) {
595     if (Error E = addSymbolName(VTableName))
596       return E;
597 
598     // Record VTableName. InstrProfWriter uses this set. The comment around
599     // class member explains why.
600     VTableNames.insert(VTableName);
601     return Error::success();
602   }
603 
getVTableNames()604   const StringSet<> &getVTableNames() const { return VTableNames; }
605 
606   /// Map a function address to its name's MD5 hash. This interface
607   /// is only used by the raw profiler reader.
mapAddress(uint64_t Addr,uint64_t MD5Val)608   void mapAddress(uint64_t Addr, uint64_t MD5Val) {
609     AddrToMD5Map.push_back(std::make_pair(Addr, MD5Val));
610   }
611 
612   /// Map the address range (i.e., [start_address, end_address)) of a variable
613   /// to  its names' MD5 hash. This interface is only used by the raw profile
614   /// reader.
mapVTableAddress(uint64_t StartAddr,uint64_t EndAddr,uint64_t MD5Val)615   void mapVTableAddress(uint64_t StartAddr, uint64_t EndAddr, uint64_t MD5Val) {
616     VTableAddrMap.insert(StartAddr, EndAddr, MD5Val);
617   }
618 
619   /// Return a function's hash, or 0, if the function isn't in this SymTab.
620   uint64_t getFunctionHashFromAddress(uint64_t Address);
621 
622   /// Return a vtable's hash, or 0 if the vtable doesn't exist in this SymTab.
623   uint64_t getVTableHashFromAddress(uint64_t Address);
624 
625   /// Return function's PGO name from the function name's symbol
626   /// address in the object file. If an error occurs, return
627   /// an empty string.
628   StringRef getFuncName(uint64_t FuncNameAddress, size_t NameSize);
629 
630   /// Return name of functions or global variables from the name's md5 hash
631   /// value. If not found, return an empty string.
632   inline StringRef getFuncOrVarName(uint64_t ValMD5Hash);
633 
634   /// Just like getFuncOrVarName, except that it will return literal string
635   /// 'External Symbol' if the function or global variable is external to
636   /// this symbol table.
637   inline StringRef getFuncOrVarNameIfDefined(uint64_t ValMD5Hash);
638 
639   /// True if Symbol is the value used to represent external symbols.
isExternalSymbol(const StringRef & Symbol)640   static bool isExternalSymbol(const StringRef &Symbol) {
641     return Symbol == InstrProfSymtab::getExternalSymbol();
642   }
643 
644   /// Return function from the name's md5 hash. Return nullptr if not found.
645   inline Function *getFunction(uint64_t FuncMD5Hash);
646 
647   /// Return the global variable corresponding to md5 hash. Return nullptr if
648   /// not found.
649   inline GlobalVariable *getGlobalVariable(uint64_t MD5Hash);
650 
651   /// Return the name section data.
getNameData()652   inline StringRef getNameData() const { return Data; }
653 
654   /// Dump the symbols in this table.
655   void dumpNames(raw_ostream &OS) const;
656 };
657 
create(StringRef D,uint64_t BaseAddr)658 Error InstrProfSymtab::create(StringRef D, uint64_t BaseAddr) {
659   Data = D;
660   Address = BaseAddr;
661   return Error::success();
662 }
663 
664 template <typename NameIterRange>
create(const NameIterRange & IterRange)665 Error InstrProfSymtab::create(const NameIterRange &IterRange) {
666   for (auto Name : IterRange)
667     if (Error E = addFuncName(Name))
668       return E;
669 
670   finalizeSymtab();
671   return Error::success();
672 }
673 
674 template <typename FuncNameIterRange, typename VTableNameIterRange>
create(const FuncNameIterRange & FuncIterRange,const VTableNameIterRange & VTableIterRange)675 Error InstrProfSymtab::create(const FuncNameIterRange &FuncIterRange,
676                               const VTableNameIterRange &VTableIterRange) {
677   // Iterate elements by StringRef rather than by const reference.
678   // StringRef is small enough, so the loop is efficient whether
679   // element in the range is std::string or StringRef.
680   for (StringRef Name : FuncIterRange)
681     if (Error E = addFuncName(Name))
682       return E;
683 
684   for (StringRef VTableName : VTableIterRange)
685     if (Error E = addVTableName(VTableName))
686       return E;
687 
688   finalizeSymtab();
689   return Error::success();
690 }
691 
finalizeSymtab()692 void InstrProfSymtab::finalizeSymtab() {
693   if (Sorted)
694     return;
695   llvm::sort(MD5NameMap, less_first());
696   llvm::sort(MD5FuncMap, less_first());
697   llvm::sort(AddrToMD5Map, less_first());
698   AddrToMD5Map.erase(std::unique(AddrToMD5Map.begin(), AddrToMD5Map.end()),
699                      AddrToMD5Map.end());
700   Sorted = true;
701 }
702 
getFuncOrVarNameIfDefined(uint64_t MD5Hash)703 StringRef InstrProfSymtab::getFuncOrVarNameIfDefined(uint64_t MD5Hash) {
704   StringRef ret = getFuncOrVarName(MD5Hash);
705   if (ret.empty())
706     return InstrProfSymtab::getExternalSymbol();
707   return ret;
708 }
709 
getFuncOrVarName(uint64_t MD5Hash)710 StringRef InstrProfSymtab::getFuncOrVarName(uint64_t MD5Hash) {
711   finalizeSymtab();
712   auto Result = llvm::lower_bound(MD5NameMap, MD5Hash,
713                                   [](const std::pair<uint64_t, StringRef> &LHS,
714                                      uint64_t RHS) { return LHS.first < RHS; });
715   if (Result != MD5NameMap.end() && Result->first == MD5Hash)
716     return Result->second;
717   return StringRef();
718 }
719 
getFunction(uint64_t FuncMD5Hash)720 Function* InstrProfSymtab::getFunction(uint64_t FuncMD5Hash) {
721   finalizeSymtab();
722   auto Result = llvm::lower_bound(MD5FuncMap, FuncMD5Hash,
723                                   [](const std::pair<uint64_t, Function *> &LHS,
724                                      uint64_t RHS) { return LHS.first < RHS; });
725   if (Result != MD5FuncMap.end() && Result->first == FuncMD5Hash)
726     return Result->second;
727   return nullptr;
728 }
729 
getGlobalVariable(uint64_t MD5Hash)730 GlobalVariable *InstrProfSymtab::getGlobalVariable(uint64_t MD5Hash) {
731   if (auto Iter = MD5VTableMap.find(MD5Hash); Iter != MD5VTableMap.end())
732     return Iter->second;
733   return nullptr;
734 }
735 
736 // To store the sums of profile count values, or the percentage of
737 // the sums of the total count values.
738 struct CountSumOrPercent {
739   uint64_t NumEntries;
740   double CountSum;
741   double ValueCounts[IPVK_Last - IPVK_First + 1];
CountSumOrPercentCountSumOrPercent742   CountSumOrPercent() : NumEntries(0), CountSum(0.0f), ValueCounts() {}
resetCountSumOrPercent743   void reset() {
744     NumEntries = 0;
745     CountSum = 0.0f;
746     for (double &VC : ValueCounts)
747       VC = 0.0f;
748   }
749 };
750 
751 // Function level or program level overlap information.
752 struct OverlapStats {
753   enum OverlapStatsLevel { ProgramLevel, FunctionLevel };
754   // Sum of the total count values for the base profile.
755   CountSumOrPercent Base;
756   // Sum of the total count values for the test profile.
757   CountSumOrPercent Test;
758   // Overlap lap score. Should be in range of [0.0f to 1.0f].
759   CountSumOrPercent Overlap;
760   CountSumOrPercent Mismatch;
761   CountSumOrPercent Unique;
762   OverlapStatsLevel Level;
763   const std::string *BaseFilename;
764   const std::string *TestFilename;
765   StringRef FuncName;
766   uint64_t FuncHash;
767   bool Valid;
768 
769   OverlapStats(OverlapStatsLevel L = ProgramLevel)
LevelOverlapStats770       : Level(L), BaseFilename(nullptr), TestFilename(nullptr), FuncHash(0),
771         Valid(false) {}
772 
773   void dump(raw_fd_ostream &OS) const;
774 
setFuncInfoOverlapStats775   void setFuncInfo(StringRef Name, uint64_t Hash) {
776     FuncName = Name;
777     FuncHash = Hash;
778   }
779 
780   Error accumulateCounts(const std::string &BaseFilename,
781                          const std::string &TestFilename, bool IsCS);
782   void addOneMismatch(const CountSumOrPercent &MismatchFunc);
783   void addOneUnique(const CountSumOrPercent &UniqueFunc);
784 
scoreOverlapStats785   static inline double score(uint64_t Val1, uint64_t Val2, double Sum1,
786                              double Sum2) {
787     if (Sum1 < 1.0f || Sum2 < 1.0f)
788       return 0.0f;
789     return std::min(Val1 / Sum1, Val2 / Sum2);
790   }
791 };
792 
793 // This is used to filter the functions whose overlap information
794 // to be output.
795 struct OverlapFuncFilters {
796   uint64_t ValueCutoff;
797   const std::string NameFilter;
798 };
799 
800 struct InstrProfValueSiteRecord {
801   /// Value profiling data pairs at a given value site.
802   std::list<InstrProfValueData> ValueData;
803 
InstrProfValueSiteRecordInstrProfValueSiteRecord804   InstrProfValueSiteRecord() { ValueData.clear(); }
805   template <class InputIterator>
InstrProfValueSiteRecordInstrProfValueSiteRecord806   InstrProfValueSiteRecord(InputIterator F, InputIterator L)
807       : ValueData(F, L) {}
808 
809   /// Sort ValueData ascending by Value
sortByTargetValuesInstrProfValueSiteRecord810   void sortByTargetValues() {
811     ValueData.sort(
812         [](const InstrProfValueData &left, const InstrProfValueData &right) {
813           return left.Value < right.Value;
814         });
815   }
816   /// Sort ValueData Descending by Count
817   inline void sortByCount();
818 
819   /// Merge data from another InstrProfValueSiteRecord
820   /// Optionally scale merged counts by \p Weight.
821   void merge(InstrProfValueSiteRecord &Input, uint64_t Weight,
822              function_ref<void(instrprof_error)> Warn);
823   /// Scale up value profile data counts by N (Numerator) / D (Denominator).
824   void scale(uint64_t N, uint64_t D, function_ref<void(instrprof_error)> Warn);
825 
826   /// Compute the overlap b/w this record and Input record.
827   void overlap(InstrProfValueSiteRecord &Input, uint32_t ValueKind,
828                OverlapStats &Overlap, OverlapStats &FuncLevelOverlap);
829 };
830 
831 /// Profiling information for a single function.
832 struct InstrProfRecord {
833   std::vector<uint64_t> Counts;
834   std::vector<uint8_t> BitmapBytes;
835 
836   InstrProfRecord() = default;
InstrProfRecordInstrProfRecord837   InstrProfRecord(std::vector<uint64_t> Counts) : Counts(std::move(Counts)) {}
InstrProfRecordInstrProfRecord838   InstrProfRecord(std::vector<uint64_t> Counts,
839                   std::vector<uint8_t> BitmapBytes)
840       : Counts(std::move(Counts)), BitmapBytes(std::move(BitmapBytes)) {}
841   InstrProfRecord(InstrProfRecord &&) = default;
InstrProfRecordInstrProfRecord842   InstrProfRecord(const InstrProfRecord &RHS)
843       : Counts(RHS.Counts), BitmapBytes(RHS.BitmapBytes),
844         ValueData(RHS.ValueData
845                       ? std::make_unique<ValueProfData>(*RHS.ValueData)
846                       : nullptr) {}
847   InstrProfRecord &operator=(InstrProfRecord &&) = default;
848   InstrProfRecord &operator=(const InstrProfRecord &RHS) {
849     Counts = RHS.Counts;
850     BitmapBytes = RHS.BitmapBytes;
851     if (!RHS.ValueData) {
852       ValueData = nullptr;
853       return *this;
854     }
855     if (!ValueData)
856       ValueData = std::make_unique<ValueProfData>(*RHS.ValueData);
857     else
858       *ValueData = *RHS.ValueData;
859     return *this;
860   }
861 
862   /// Return the number of value profile kinds with non-zero number
863   /// of profile sites.
864   inline uint32_t getNumValueKinds() const;
865   /// Return the number of instrumented sites for ValueKind.
866   inline uint32_t getNumValueSites(uint32_t ValueKind) const;
867 
868   /// Return the total number of ValueData for ValueKind.
869   inline uint32_t getNumValueData(uint32_t ValueKind) const;
870 
871   /// Return the number of value data collected for ValueKind at profiling
872   /// site: Site.
873   inline uint32_t getNumValueDataForSite(uint32_t ValueKind,
874                                          uint32_t Site) const;
875 
876   /// Return the array of profiled values at \p Site. If \p TotalC
877   /// is not null, the total count of all target values at this site
878   /// will be stored in \c *TotalC.
879   inline std::unique_ptr<InstrProfValueData[]>
880   getValueForSite(uint32_t ValueKind, uint32_t Site,
881                   uint64_t *TotalC = nullptr) const;
882 
883   /// Get the target value/counts of kind \p ValueKind collected at site
884   /// \p Site and store the result in array \p Dest. Return the total
885   /// counts of all target values at this site.
886   inline uint64_t getValueForSite(InstrProfValueData Dest[], uint32_t ValueKind,
887                                   uint32_t Site) const;
888 
889   /// Reserve space for NumValueSites sites.
890   inline void reserveSites(uint32_t ValueKind, uint32_t NumValueSites);
891 
892   /// Add ValueData for ValueKind at value Site.
893   void addValueData(uint32_t ValueKind, uint32_t Site,
894                     InstrProfValueData *VData, uint32_t N,
895                     InstrProfSymtab *SymTab);
896 
897   /// Merge the counts in \p Other into this one.
898   /// Optionally scale merged counts by \p Weight.
899   void merge(InstrProfRecord &Other, uint64_t Weight,
900              function_ref<void(instrprof_error)> Warn);
901 
902   /// Scale up profile counts (including value profile data) by
903   /// a factor of (N / D).
904   void scale(uint64_t N, uint64_t D, function_ref<void(instrprof_error)> Warn);
905 
906   /// Sort value profile data (per site) by count.
sortValueDataInstrProfRecord907   void sortValueData() {
908     for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
909       for (auto &SR : getValueSitesForKind(Kind))
910         SR.sortByCount();
911   }
912 
913   /// Clear value data entries and edge counters.
ClearInstrProfRecord914   void Clear() {
915     Counts.clear();
916     clearValueData();
917   }
918 
919   /// Clear value data entries
clearValueDataInstrProfRecord920   void clearValueData() { ValueData = nullptr; }
921 
922   /// Compute the sums of all counts and store in Sum.
923   void accumulateCounts(CountSumOrPercent &Sum) const;
924 
925   /// Compute the overlap b/w this IntrprofRecord and Other.
926   void overlap(InstrProfRecord &Other, OverlapStats &Overlap,
927                OverlapStats &FuncLevelOverlap, uint64_t ValueCutoff);
928 
929   /// Compute the overlap of value profile counts.
930   void overlapValueProfData(uint32_t ValueKind, InstrProfRecord &Src,
931                             OverlapStats &Overlap,
932                             OverlapStats &FuncLevelOverlap);
933 
934   enum CountPseudoKind {
935     NotPseudo = 0,
936     PseudoHot,
937     PseudoWarm,
938   };
939   enum PseudoCountVal {
940     HotFunctionVal = -1,
941     WarmFunctionVal = -2,
942   };
getCountPseudoKindInstrProfRecord943   CountPseudoKind getCountPseudoKind() const {
944     uint64_t FirstCount = Counts[0];
945     if (FirstCount == (uint64_t)HotFunctionVal)
946       return PseudoHot;
947     if (FirstCount == (uint64_t)WarmFunctionVal)
948       return PseudoWarm;
949     return NotPseudo;
950   }
setPseudoCountInstrProfRecord951   void setPseudoCount(CountPseudoKind Kind) {
952     if (Kind == PseudoHot)
953       Counts[0] = (uint64_t)HotFunctionVal;
954     else if (Kind == PseudoWarm)
955       Counts[0] = (uint64_t)WarmFunctionVal;
956   }
957 
958 private:
959   struct ValueProfData {
960     std::vector<InstrProfValueSiteRecord> IndirectCallSites;
961     std::vector<InstrProfValueSiteRecord> MemOPSizes;
962     std::vector<InstrProfValueSiteRecord> VTableTargets;
963   };
964   std::unique_ptr<ValueProfData> ValueData;
965 
966   MutableArrayRef<InstrProfValueSiteRecord>
getValueSitesForKindInstrProfRecord967   getValueSitesForKind(uint32_t ValueKind) {
968     // Cast to /add/ const (should be an implicit_cast, ideally, if that's ever
969     // implemented in LLVM) to call the const overload of this function, then
970     // cast away the constness from the result.
971     auto AR = const_cast<const InstrProfRecord *>(this)->getValueSitesForKind(
972         ValueKind);
973     return MutableArrayRef(
974         const_cast<InstrProfValueSiteRecord *>(AR.data()), AR.size());
975   }
976   ArrayRef<InstrProfValueSiteRecord>
getValueSitesForKindInstrProfRecord977   getValueSitesForKind(uint32_t ValueKind) const {
978     if (!ValueData)
979       return std::nullopt;
980     switch (ValueKind) {
981     case IPVK_IndirectCallTarget:
982       return ValueData->IndirectCallSites;
983     case IPVK_MemOPSize:
984       return ValueData->MemOPSizes;
985     case IPVK_VTableTarget:
986       return ValueData->VTableTargets;
987     default:
988       llvm_unreachable("Unknown value kind!");
989     }
990   }
991 
992   std::vector<InstrProfValueSiteRecord> &
getOrCreateValueSitesForKindInstrProfRecord993   getOrCreateValueSitesForKind(uint32_t ValueKind) {
994     if (!ValueData)
995       ValueData = std::make_unique<ValueProfData>();
996     switch (ValueKind) {
997     case IPVK_IndirectCallTarget:
998       return ValueData->IndirectCallSites;
999     case IPVK_MemOPSize:
1000       return ValueData->MemOPSizes;
1001     case IPVK_VTableTarget:
1002       return ValueData->VTableTargets;
1003     default:
1004       llvm_unreachable("Unknown value kind!");
1005     }
1006   }
1007 
1008   // Map indirect call target name hash to name string.
1009   uint64_t remapValue(uint64_t Value, uint32_t ValueKind,
1010                       InstrProfSymtab *SymTab);
1011 
1012   // Merge Value Profile data from Src record to this record for ValueKind.
1013   // Scale merged value counts by \p Weight.
1014   void mergeValueProfData(uint32_t ValkeKind, InstrProfRecord &Src,
1015                           uint64_t Weight,
1016                           function_ref<void(instrprof_error)> Warn);
1017 
1018   // Scale up value profile data count by N (Numerator) / D (Denominator).
1019   void scaleValueProfData(uint32_t ValueKind, uint64_t N, uint64_t D,
1020                           function_ref<void(instrprof_error)> Warn);
1021 };
1022 
1023 struct NamedInstrProfRecord : InstrProfRecord {
1024   StringRef Name;
1025   uint64_t Hash;
1026 
1027   // We reserve this bit as the flag for context sensitive profile record.
1028   static const int CS_FLAG_IN_FUNC_HASH = 60;
1029 
1030   NamedInstrProfRecord() = default;
NamedInstrProfRecordNamedInstrProfRecord1031   NamedInstrProfRecord(StringRef Name, uint64_t Hash,
1032                        std::vector<uint64_t> Counts)
1033       : InstrProfRecord(std::move(Counts)), Name(Name), Hash(Hash) {}
NamedInstrProfRecordNamedInstrProfRecord1034   NamedInstrProfRecord(StringRef Name, uint64_t Hash,
1035                        std::vector<uint64_t> Counts,
1036                        std::vector<uint8_t> BitmapBytes)
1037       : InstrProfRecord(std::move(Counts), std::move(BitmapBytes)), Name(Name),
1038         Hash(Hash) {}
1039 
hasCSFlagInHashNamedInstrProfRecord1040   static bool hasCSFlagInHash(uint64_t FuncHash) {
1041     return ((FuncHash >> CS_FLAG_IN_FUNC_HASH) & 1);
1042   }
setCSFlagInHashNamedInstrProfRecord1043   static void setCSFlagInHash(uint64_t &FuncHash) {
1044     FuncHash |= ((uint64_t)1 << CS_FLAG_IN_FUNC_HASH);
1045   }
1046 };
1047 
getNumValueKinds()1048 uint32_t InstrProfRecord::getNumValueKinds() const {
1049   uint32_t NumValueKinds = 0;
1050   for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
1051     NumValueKinds += !(getValueSitesForKind(Kind).empty());
1052   return NumValueKinds;
1053 }
1054 
getNumValueData(uint32_t ValueKind)1055 uint32_t InstrProfRecord::getNumValueData(uint32_t ValueKind) const {
1056   uint32_t N = 0;
1057   for (const auto &SR : getValueSitesForKind(ValueKind))
1058     N += SR.ValueData.size();
1059   return N;
1060 }
1061 
getNumValueSites(uint32_t ValueKind)1062 uint32_t InstrProfRecord::getNumValueSites(uint32_t ValueKind) const {
1063   return getValueSitesForKind(ValueKind).size();
1064 }
1065 
getNumValueDataForSite(uint32_t ValueKind,uint32_t Site)1066 uint32_t InstrProfRecord::getNumValueDataForSite(uint32_t ValueKind,
1067                                                  uint32_t Site) const {
1068   return getValueSitesForKind(ValueKind)[Site].ValueData.size();
1069 }
1070 
1071 std::unique_ptr<InstrProfValueData[]>
getValueForSite(uint32_t ValueKind,uint32_t Site,uint64_t * TotalC)1072 InstrProfRecord::getValueForSite(uint32_t ValueKind, uint32_t Site,
1073                                  uint64_t *TotalC) const {
1074   uint64_t Dummy = 0;
1075   uint64_t &TotalCount = (TotalC == nullptr ? Dummy : *TotalC);
1076   uint32_t N = getNumValueDataForSite(ValueKind, Site);
1077   if (N == 0) {
1078     TotalCount = 0;
1079     return std::unique_ptr<InstrProfValueData[]>(nullptr);
1080   }
1081 
1082   auto VD = std::make_unique<InstrProfValueData[]>(N);
1083   TotalCount = getValueForSite(VD.get(), ValueKind, Site);
1084 
1085   return VD;
1086 }
1087 
getValueForSite(InstrProfValueData Dest[],uint32_t ValueKind,uint32_t Site)1088 uint64_t InstrProfRecord::getValueForSite(InstrProfValueData Dest[],
1089                                           uint32_t ValueKind,
1090                                           uint32_t Site) const {
1091   uint32_t I = 0;
1092   uint64_t TotalCount = 0;
1093   for (auto V : getValueSitesForKind(ValueKind)[Site].ValueData) {
1094     Dest[I].Value = V.Value;
1095     Dest[I].Count = V.Count;
1096     TotalCount = SaturatingAdd(TotalCount, V.Count);
1097     I++;
1098   }
1099   return TotalCount;
1100 }
1101 
reserveSites(uint32_t ValueKind,uint32_t NumValueSites)1102 void InstrProfRecord::reserveSites(uint32_t ValueKind, uint32_t NumValueSites) {
1103   if (!NumValueSites)
1104     return;
1105   getOrCreateValueSitesForKind(ValueKind).reserve(NumValueSites);
1106 }
1107 
1108 // Include definitions for value profile data
1109 #define INSTR_PROF_VALUE_PROF_DATA
1110 #include "llvm/ProfileData/InstrProfData.inc"
1111 
sortByCount()1112 void InstrProfValueSiteRecord::sortByCount() {
1113   ValueData.sort(
1114       [](const InstrProfValueData &left, const InstrProfValueData &right) {
1115         return left.Count > right.Count;
1116       });
1117   // Now truncate
1118   size_t max_s = INSTR_PROF_MAX_NUM_VAL_PER_SITE;
1119   if (ValueData.size() > max_s)
1120     ValueData.resize(max_s);
1121 }
1122 
1123 namespace IndexedInstrProf {
1124 
1125 enum class HashT : uint32_t {
1126   MD5,
1127   Last = MD5
1128 };
1129 
ComputeHash(HashT Type,StringRef K)1130 inline uint64_t ComputeHash(HashT Type, StringRef K) {
1131   switch (Type) {
1132   case HashT::MD5:
1133     return MD5Hash(K);
1134   }
1135   llvm_unreachable("Unhandled hash type");
1136 }
1137 
1138 const uint64_t Magic = 0x8169666f72706cff; // "\xfflprofi\x81"
1139 
1140 enum ProfVersion {
1141   // Version 1 is the first version. In this version, the value of
1142   // a key/value pair can only include profile data of a single function.
1143   // Due to this restriction, the number of block counters for a given
1144   // function is not recorded but derived from the length of the value.
1145   Version1 = 1,
1146   // The version 2 format supports recording profile data of multiple
1147   // functions which share the same key in one value field. To support this,
1148   // the number block counters is recorded as an uint64_t field right after the
1149   // function structural hash.
1150   Version2 = 2,
1151   // Version 3 supports value profile data. The value profile data is expected
1152   // to follow the block counter profile data.
1153   Version3 = 3,
1154   // In this version, profile summary data \c IndexedInstrProf::Summary is
1155   // stored after the profile header.
1156   Version4 = 4,
1157   // In this version, the frontend PGO stable hash algorithm defaults to V2.
1158   Version5 = 5,
1159   // In this version, the frontend PGO stable hash algorithm got fixed and
1160   // may produce hashes different from Version5.
1161   Version6 = 6,
1162   // An additional counter is added around logical operators.
1163   Version7 = 7,
1164   // An additional (optional) memory profile type is added.
1165   Version8 = 8,
1166   // Binary ids are added.
1167   Version9 = 9,
1168   // An additional (optional) temporal profile traces section is added.
1169   Version10 = 10,
1170   // An additional field is used for bitmap bytes.
1171   Version11 = 11,
1172   // VTable profiling,
1173   Version12 = 12,
1174   // The current version is 12.
1175   CurrentVersion = INSTR_PROF_INDEX_VERSION
1176 };
1177 const uint64_t Version = ProfVersion::CurrentVersion;
1178 
1179 const HashT HashType = HashT::MD5;
1180 
ComputeHash(StringRef K)1181 inline uint64_t ComputeHash(StringRef K) { return ComputeHash(HashType, K); }
1182 
1183 // This structure defines the file header of the LLVM profile
1184 // data file in indexed-format. Please update llvm/docs/InstrProfileFormat.rst
1185 // as appropriate when updating the indexed profile format.
1186 struct Header {
1187   uint64_t Magic;
1188   // The lower 32 bits specify the version of the indexed profile.
1189   // The most significant 32 bits are reserved to specify the variant types of
1190   // the profile.
1191   uint64_t Version;
1192   uint64_t Unused; // Becomes unused since version 4
1193   uint64_t HashType;
1194   // This field records the offset of this hash table's metadata (i.e., the
1195   // number of buckets and entries), which follows right after the payload of
1196   // the entire hash table.
1197   uint64_t HashOffset;
1198   uint64_t MemProfOffset;
1199   uint64_t BinaryIdOffset;
1200   uint64_t TemporalProfTracesOffset;
1201   uint64_t VTableNamesOffset;
1202   // New fields should only be added at the end to ensure that the size
1203   // computation is correct. The methods below need to be updated to ensure that
1204   // the new field is read correctly.
1205 
1206   // Reads a header struct from the buffer.
1207   static Expected<Header> readFromBuffer(const unsigned char *Buffer);
1208 
1209   // Returns the size of the header in bytes for all valid fields based on the
1210   // version. I.e a older version header will return a smaller size.
1211   size_t size() const;
1212 
1213   // Returns the format version in little endian. The header retains the version
1214   // in native endian of the compiler runtime.
1215   uint64_t formatVersion() const;
1216 };
1217 
1218 // Profile summary data recorded in the profile data file in indexed
1219 // format. It is introduced in version 4. The summary data follows
1220 // right after the profile file header.
1221 struct Summary {
1222   struct Entry {
1223     uint64_t Cutoff; ///< The required percentile of total execution count.
1224     uint64_t
1225         MinBlockCount;  ///< The minimum execution count for this percentile.
1226     uint64_t NumBlocks; ///< Number of blocks >= the minumum execution count.
1227   };
1228   // The field kind enumerator to assigned value mapping should remain
1229   // unchanged  when a new kind is added or an old kind gets deleted in
1230   // the future.
1231   enum SummaryFieldKind {
1232     /// The total number of functions instrumented.
1233     TotalNumFunctions = 0,
1234     /// Total number of instrumented blocks/edges.
1235     TotalNumBlocks = 1,
1236     /// The maximal execution count among all functions.
1237     /// This field does not exist for profile data from IR based
1238     /// instrumentation.
1239     MaxFunctionCount = 2,
1240     /// Max block count of the program.
1241     MaxBlockCount = 3,
1242     /// Max internal block count of the program (excluding entry blocks).
1243     MaxInternalBlockCount = 4,
1244     /// The sum of all instrumented block counts.
1245     TotalBlockCount = 5,
1246     NumKinds = TotalBlockCount + 1
1247   };
1248 
1249   // The number of summmary fields following the summary header.
1250   uint64_t NumSummaryFields;
1251   // The number of Cutoff Entries (Summary::Entry) following summary fields.
1252   uint64_t NumCutoffEntries;
1253 
1254   Summary() = delete;
SummarySummary1255   Summary(uint32_t Size) { memset(this, 0, Size); }
1256 
deleteSummary1257   void operator delete(void *ptr) { ::operator delete(ptr); }
1258 
getSizeSummary1259   static uint32_t getSize(uint32_t NumSumFields, uint32_t NumCutoffEntries) {
1260     return sizeof(Summary) + NumCutoffEntries * sizeof(Entry) +
1261            NumSumFields * sizeof(uint64_t);
1262   }
1263 
getSummaryDataBaseSummary1264   const uint64_t *getSummaryDataBase() const {
1265     return reinterpret_cast<const uint64_t *>(this + 1);
1266   }
1267 
getSummaryDataBaseSummary1268   uint64_t *getSummaryDataBase() {
1269     return reinterpret_cast<uint64_t *>(this + 1);
1270   }
1271 
getCutoffEntryBaseSummary1272   const Entry *getCutoffEntryBase() const {
1273     return reinterpret_cast<const Entry *>(
1274         &getSummaryDataBase()[NumSummaryFields]);
1275   }
1276 
getCutoffEntryBaseSummary1277   Entry *getCutoffEntryBase() {
1278     return reinterpret_cast<Entry *>(&getSummaryDataBase()[NumSummaryFields]);
1279   }
1280 
getSummary1281   uint64_t get(SummaryFieldKind K) const {
1282     return getSummaryDataBase()[K];
1283   }
1284 
setSummary1285   void set(SummaryFieldKind K, uint64_t V) {
1286     getSummaryDataBase()[K] = V;
1287   }
1288 
getEntrySummary1289   const Entry &getEntry(uint32_t I) const { return getCutoffEntryBase()[I]; }
1290 
setEntrySummary1291   void setEntry(uint32_t I, const ProfileSummaryEntry &E) {
1292     Entry &ER = getCutoffEntryBase()[I];
1293     ER.Cutoff = E.Cutoff;
1294     ER.MinBlockCount = E.MinCount;
1295     ER.NumBlocks = E.NumCounts;
1296   }
1297 };
1298 
allocSummary(uint32_t TotalSize)1299 inline std::unique_ptr<Summary> allocSummary(uint32_t TotalSize) {
1300   return std::unique_ptr<Summary>(new (::operator new(TotalSize))
1301                                       Summary(TotalSize));
1302 }
1303 
1304 } // end namespace IndexedInstrProf
1305 
1306 namespace RawInstrProf {
1307 
1308 // Version 1: First version
1309 // Version 2: Added value profile data section. Per-function control data
1310 // struct has more fields to describe value profile information.
1311 // Version 3: Compressed name section support. Function PGO name reference
1312 // from control data struct is changed from raw pointer to Name's MD5 value.
1313 // Version 4: ValueDataBegin and ValueDataSizes fields are removed from the
1314 // raw header.
1315 // Version 5: Bit 60 of FuncHash is reserved for the flag for the context
1316 // sensitive records.
1317 // Version 6: Added binary id.
1318 // Version 7: Reorder binary id and include version in signature.
1319 // Version 8: Use relative counter pointer.
1320 // Version 9: Added relative bitmap bytes pointer and count used by MC/DC.
1321 // Version 10: Added vtable, a new type of value profile data.
1322 const uint64_t Version = INSTR_PROF_RAW_VERSION;
1323 
1324 template <class IntPtrT> inline uint64_t getMagic();
1325 template <> inline uint64_t getMagic<uint64_t>() {
1326   return INSTR_PROF_RAW_MAGIC_64;
1327 }
1328 
1329 template <> inline uint64_t getMagic<uint32_t>() {
1330   return INSTR_PROF_RAW_MAGIC_32;
1331 }
1332 
1333 // Per-function profile data header/control structure.
1334 // The definition should match the structure defined in
1335 // compiler-rt/lib/profile/InstrProfiling.h.
1336 // It should also match the synthesized type in
1337 // Transforms/Instrumentation/InstrProfiling.cpp:getOrCreateRegionCounters.
1338 template <class IntPtrT> struct alignas(8) ProfileData {
1339 #define INSTR_PROF_DATA(Type, LLVMType, Name, Init) Type Name;
1340 #include "llvm/ProfileData/InstrProfData.inc"
1341 };
1342 
1343 template <class IntPtrT> struct alignas(8) VTableProfileData {
1344 #define INSTR_PROF_VTABLE_DATA(Type, LLVMType, Name, Init) Type Name;
1345 #include "llvm/ProfileData/InstrProfData.inc"
1346 };
1347 
1348 // File header structure of the LLVM profile data in raw format.
1349 // The definition should match the header referenced in
1350 // compiler-rt/lib/profile/InstrProfilingFile.c  and
1351 // InstrProfilingBuffer.c.
1352 struct Header {
1353 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
1354 #include "llvm/ProfileData/InstrProfData.inc"
1355 };
1356 
1357 } // end namespace RawInstrProf
1358 
1359 // Create the variable for the profile file name.
1360 void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput);
1361 
1362 // Whether to compress function names in profile records, and filenames in
1363 // code coverage mappings. Used by the Instrumentation library and unit tests.
1364 extern cl::opt<bool> DoInstrProfNameCompression;
1365 
1366 } // end namespace llvm
1367 #endif // LLVM_PROFILEDATA_INSTRPROF_H
1368