[Profile] Improve profile dumping with merging

[davidxl]

Summary

When profile merging is on, the name section in the profile data file won't change after merging. This patch adds support to skip over name section when possible to reduce disk IO.

Do a Clang self build with instrumented clang (IR based), the build time improves a little ~1%. With IR PGO, the name section is relatively small. The total profile size is ~15MB while the compressed name strings is only 2.4MB.

[vsk]

This sounds like a great idea.

lib/profile/InstrProfilingFile.c

228

Can you remove the extra fseek in a separate commit?

lib/profile/InstrProfilingWriter.c

281

If we aren't writing the name data, can't we also avoid writing __llvm_profile_data, and the padding?

[davidxl]

lib/profile/InstrProfilingWriter.c

281

The profile reader uses the counter begin recorded in the header and the counter ptr recorded in data to compute counter offset. Assuming the relative offset remain unchanged, I think it is possible to skip them as well. Since the major cost comes from name section, I'd like to limit the change to name skipping only.

[vsk]

Thanks, lgtm.

lib/profile/InstrProfilingWriter.c

281

Ok, I see this in RawInstrProfReader::getCounter.

lib/profile/InstrProfilingBuffer.c

 void initBufferWriter(ProfDataWriter *BufferWriter, char *Buffer) {
   BufferWriter->Write = lprofBufferWriter;
   BufferWriter->WriterCtx = Buffer;
 }
 
 COMPILER_RT_VISIBILITY int __llvm_profile_write_buffer(char *Buffer) {
   ProfDataWriter BufferWriter;
   initBufferWriter(&BufferWriter, Buffer);
-  return lprofWriteData(&BufferWriter, 0);
+  return lprofWriteData(&BufferWriter, 0, 0);
 }
 
 COMPILER_RT_VISIBILITY int __llvm_profile_write_buffer_internal(
     char *Buffer, const __llvm_profile_data *DataBegin,
     const __llvm_profile_data *DataEnd, const uint64_t *CountersBegin,
     const uint64_t *CountersEnd, const char *NamesBegin, const char *NamesEnd) {
   ProfDataWriter BufferWriter;
   initBufferWriter(&BufferWriter, Buffer);
   return lprofWriteDataImpl(&BufferWriter, DataBegin, DataEnd, CountersBegin,
-                            CountersEnd, 0, NamesBegin, NamesEnd);
+                            CountersEnd, 0, NamesBegin, NamesEnd, 0);
 }

lib/profile/InstrProfilingFile.c

 int getpid(void);
 static int getCurFilenameLength();
 static const char *getCurFilename(char *FilenameBuf);
 static unsigned doMerging() { return lprofCurFilename.MergePoolSize; }
 
 /* Return 1 if there is an error, otherwise return  0.  */
 static uint32_t fileWriter(ProfDataWriter *This, ProfDataIOVec *IOVecs,
                            uint32_t NumIOVecs) {
-
   uint32_t I;
   FILE *File = (FILE *)This->WriterCtx;
   for (I = 0; I < NumIOVecs; I++) {
+    if (IOVecs[I].Data) {
       if (fwrite(IOVecs[I].Data, IOVecs[I].ElmSize, IOVecs[I].NumElm, File) !=
           IOVecs[I].NumElm)
         return 1;
+    } else {
+      if (fseek(File, IOVecs[I].ElmSize * IOVecs[I].NumElm, SEEK_CUR) == -1)
+        return 1;
+    }
   }
   return 0;
 }
 
 static void initFileWriter(ProfDataWriter *This, FILE *File) {
   This->Write = fileWriter;
   This->WriterCtx = File;
 }
static void setupIOBuffer() {
   if (BufferSzStr && BufferSzStr[0]) {
     VPBufferSize = atoi(BufferSzStr);
     DynamicBufferIOBuffer = (uint8_t *)calloc(VPBufferSize, 1);
   }
 }
 
 /* Read profile data in \c ProfileFile and merge with in-memory
    profile counters. Returns -1 if there is fatal error, otheriwse
-   0 is returned.
+   0 is returned. Returning 0 does not mean merge is actually
+   performed. If merge is actually done, *MergeDone is set to 1.
 */
-static int doProfileMerging(FILE *ProfileFile) {
+static int doProfileMerging(FILE *ProfileFile, int *MergeDone) {
   uint64_t ProfileFileSize;
   char *ProfileBuffer;
 
   if (fseek(ProfileFile, 0L, SEEK_END) == -1) {
     PROF_ERR("Unable to merge profile data, unable to get size: %s\n",
              strerror(errno));
     return -1;
   }
PROF_WARN("Unable to merge profile data: %s\n",
               "source profile file is not compatible.");
     return 0;
   }
 
   /* Now start merging */
   __llvm_profile_merge_from_buffer(ProfileBuffer, ProfileFileSize);
   (void)munmap(ProfileBuffer, ProfileFileSize);
 
+  *MergeDone = 1;
+
   return 0;
 }
 
 /* Create the directory holding the file, if needed. */
 static void createProfileDir(const char *Filename) {
   size_t Length = strlen(Filename);
   if (lprofFindFirstDirSeparator(Filename)) {
     char *Copy = (char *)COMPILER_RT_ALLOCA(Length + 1);
     strncpy(Copy, Filename, Length + 1);
     __llvm_profile_recursive_mkdir(Copy);
   }
 }
 
 /* Open the profile data for merging. It opens the file in r+b mode with
  * file locking.  If the file has content which is compatible with the
  * current process, it also reads in the profile data in the file and merge
  * it with in-memory counters. After the profile data is merged in memory,
  * the original profile data is truncated and gets ready for the profile
  * dumper. With profile merging enabled, each executable as well as any of
  * its instrumented shared libraries dump profile data into their own data file.
 */
-static FILE *openFileForMerging(const char *ProfileFileName) {
+static FILE *openFileForMerging(const char *ProfileFileName, int *MergeDone) {
   FILE *ProfileFile;
   int rc;
 
   createProfileDir(ProfileFileName);
   ProfileFile = lprofOpenFileEx(ProfileFileName);
   if (!ProfileFile)
     return NULL;
 
-  rc = doProfileMerging(ProfileFile);
-  if (rc || COMPILER_RT_FTRUNCATE(ProfileFile, 0L) ||
+  rc = doProfileMerging(ProfileFile, MergeDone);
+  if (rc || (!*MergeDone && COMPILER_RT_FTRUNCATE(ProfileFile, 0L)) ||
       fseek(ProfileFile, 0L, SEEK_SET) == -1) {
     PROF_ERR("Profile Merging of file %s failed: %s\n", ProfileFileName,
              strerror(errno));
     fclose(ProfileFile);
     return NULL;
   }
-  fseek(ProfileFile, 0L, SEEK_SET);
   return ProfileFile;
 }
 
 /* Write profile data to file \c OutputName.  */
 static int writeFile(const char *OutputName) {
   int RetVal;
   FILE *OutputFile;
 
+  int MergeDone = 0;
   if (!doMerging())
     OutputFile = fopen(OutputName, "ab");
   else
-    OutputFile = openFileForMerging(OutputName);
+    OutputFile = openFileForMerging(OutputName, &MergeDone);
 
   if (!OutputFile)
     return -1;
 
   FreeHook = &free;
   setupIOBuffer();
   ProfDataWriter fileWriter;
   initFileWriter(&fileWriter, OutputFile);
-  RetVal = lprofWriteData(&fileWriter, lprofGetVPDataReader());
+  RetVal = lprofWriteData(&fileWriter, lprofGetVPDataReader(), MergeDone);
 
   fclose(OutputFile);
   return RetVal;
 }
 
 static void truncateCurrentFile(void) {
   const char *Filename;
   char *FilenameBuf;

lib/profile/InstrProfilingInternal.h

  */
 int lprofBufferIOFlush(ProfBufferIO *BufferIO);
 
 /* The low level interface to write data into a buffer. It is used as the
  * callback by other high level writer methods such as buffered IO writer
  * and profile data writer.  */
 uint32_t lprofBufferWriter(ProfDataWriter *This, ProfDataIOVec *IOVecs,
                            uint32_t NumIOVecs);
-
 void initBufferWriter(ProfDataWriter *BufferWriter, char *Buffer);
 
 struct ValueProfData;
 struct ValueProfRecord;
 struct InstrProfValueData;
 struct ValueProfNode;
 
 /*!

    * If there is nothing left, it returns NULL.
    */
   struct ValueProfNode *(*GetValueData)(uint32_t ValueKind, uint32_t Site,
                                         struct InstrProfValueData *Dst,
                                         struct ValueProfNode *StartNode,
                                         uint32_t N);
 } VPDataReaderType;
 
-int lprofWriteData(ProfDataWriter *Writer, VPDataReaderType *VPDataReader);
+/* Write profile data to destinitation. If SkipNameDataWrite is set to 1,
+   the name data is already in destintation, we just skip over it. */
+int lprofWriteData(ProfDataWriter *Writer, VPDataReaderType *VPDataReader,
+                   int SkipNameDataWrite);
 int lprofWriteDataImpl(ProfDataWriter *Writer,
                        const __llvm_profile_data *DataBegin,
                        const __llvm_profile_data *DataEnd,
                        const uint64_t *CountersBegin,
                        const uint64_t *CountersEnd,
                        VPDataReaderType *VPDataReader, const char *NamesBegin,
-                       const char *NamesEnd);
+                       const char *NamesEnd, int SkipNameDataWrite);
 
 /* Merge value profile data pointed to by SrcValueProfData into
  * in-memory profile counters pointed by to DstData.  */
 void lprofMergeValueProfData(struct ValueProfData *SrcValueProfData,
                              __llvm_profile_data *DstData);
 
 VPDataReaderType *lprofGetVPDataReader();
 

lib/profile/InstrProfilingWriter.c

  */
 COMPILER_RT_VISIBILITY uint32_t lprofBufferWriter(ProfDataWriter *This,
                                                   ProfDataIOVec *IOVecs,
                                                   uint32_t NumIOVecs) {
   uint32_t I;
   char **Buffer = (char **)&This->WriterCtx;
   for (I = 0; I < NumIOVecs; I++) {
     size_t Length = IOVecs[I].ElmSize * IOVecs[I].NumElm;
+    if (IOVecs[I].Data)
       memcpy(*Buffer, IOVecs[I].Data, Length);
     *Buffer += Length;
   }
   return 0;
 }
 
 static void llvmInitBufferIO(ProfBufferIO *BufferIO, ProfDataWriter *FileWriter,
                              uint8_t *Buffer, uint32_t BufferSz) {
   BufferIO->FileWriter = FileWriter;
static int writeValueProfData(ProfDataWriter *Writer,
   if (lprofBufferIOFlush(BufferIO) != 0)
     return -1;
   lprofDeleteBufferIO(BufferIO);
 
   return 0;
 }
 
 COMPILER_RT_VISIBILITY int lprofWriteData(ProfDataWriter *Writer,
-                                          VPDataReaderType *VPDataReader) {
+                                          VPDataReaderType *VPDataReader,
+                                          int SkipNameDataWrite) {
   /* Match logic in __llvm_profile_write_buffer(). */
   const __llvm_profile_data *DataBegin = __llvm_profile_begin_data();
   const __llvm_profile_data *DataEnd = __llvm_profile_end_data();
   const uint64_t *CountersBegin = __llvm_profile_begin_counters();
   const uint64_t *CountersEnd = __llvm_profile_end_counters();
   const char *NamesBegin = __llvm_profile_begin_names();
   const char *NamesEnd = __llvm_profile_end_names();
   return lprofWriteDataImpl(Writer, DataBegin, DataEnd, CountersBegin,
-                            CountersEnd, VPDataReader, NamesBegin, NamesEnd);
+                            CountersEnd, VPDataReader, NamesBegin, NamesEnd,
+                            SkipNameDataWrite);
 }
 
 COMPILER_RT_VISIBILITY int
 lprofWriteDataImpl(ProfDataWriter *Writer, const __llvm_profile_data *DataBegin,
                    const __llvm_profile_data *DataEnd,
                    const uint64_t *CountersBegin, const uint64_t *CountersEnd,
                    VPDataReaderType *VPDataReader, const char *NamesBegin,
-                   const char *NamesEnd) {
+                   const char *NamesEnd, int SkipNameDataWrite) {
 
   /* Calculate size of sections. */
   const uint64_t DataSize = __llvm_profile_get_data_size(DataBegin, DataEnd);
   const uint64_t CountersSize = CountersEnd - CountersBegin;
   const uint64_t NamesSize = NamesEnd - NamesBegin;
   const uint64_t Padding = __llvm_profile_get_num_padding_bytes(NamesSize);
 
   /* Enough zeroes for padding. */
   const char Zeroes[sizeof(uint64_t)] = {0};
 
   /* Create the header. */
   __llvm_profile_header Header;
 
   if (!DataSize)
     return 0;
 
 /* Initialize header structure.  */
 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) Header.Name = Init;
 #include "InstrProfData.inc"
 
   /* Write the data. */
-  ProfDataIOVec IOVec[] = {{&Header, sizeof(__llvm_profile_header), 1},
-                           {DataBegin, sizeof(__llvm_profile_data), DataSize},
-                           {CountersBegin, sizeof(uint64_t), CountersSize},
-                           {NamesBegin, sizeof(uint8_t), NamesSize},
-                           {Zeroes, sizeof(uint8_t), Padding}};
+  ProfDataIOVec IOVec[] = {
+      {&Header, sizeof(__llvm_profile_header), 1},
+      {SkipNameDataWrite ? NULL : NamesBegin, sizeof(uint8_t), NamesSize},
   if (Writer->Write(Writer, IOVec, sizeof(IOVec) / sizeof(*IOVec)))
     return -1;
 
   return writeValueProfData(Writer, VPDataReader, DataBegin, DataEnd);
 }