[TypeProf][InstrPGO] Introduce raw and instr profile format change for type profiling. (#81691)

* Raw profile format
- Header: records the byte size of compressed vtable names, and the
number of profiled vtable entries (call it `VTableProfData`). Header
also records padded bytes of each section.
- Payload: adds a section for compressed vtable names, and a section to
store `VTableProfData`. Both sections are padded so the size is a
multiple of 8.
* Indexed profile format
  - Header: records the byte offset of compressed vtable names.
- Payload: adds a section to store compressed vtable names. This section
is used by `llvm-profdata` to show the list of vtables profiled for an
instrumented site.
  
[The originally reviewed
patch](#66825) will have
profile reader/write change and llvm-profdata change.
- To ensure this PR has all the necessary profile format change along
with profile version bump, created a copy of the originally reviewed
patch in #80761. The copy
doesn't have profile format change, but it has the set of tests which
covers type profile generation, profile read and profile merge. Tests
pass there.
  
rfc in
https://discourse.llvm.org/t/rfc-dynamic-type-profiling-and-optimizations-in-llvm/74600

---------

Co-authored-by: modiking <modiking213@gmail.com>

Author: mingmingl-llvm

compiler-rt/include/profile/InstrProfData.inc

@@ -96,6 +96,25 @@ INSTR_PROF_DATA(const uint32_t, llvm::Type::getInt32Ty(Ctx), NumBitmapBytes, \
 #undef INSTR_PROF_DATA
 /* INSTR_PROF_DATA end. */
 
+/* For a virtual table object, record the name hash to associate profiled
+ * addresses with global variables, and record {starting address, size in bytes}
+ * to map the profiled virtual table (which usually have an offset from the
+ * starting address) back to a virtual table object. */
+#ifndef INSTR_PROF_VTABLE_DATA
+#define INSTR_PROF_VTABLE_DATA(Type, LLVMType, Name, Initializer)
+#else
+#define INSTR_PROF_VTABLE_DATA_DEFINED
+#endif
+INSTR_PROF_VTABLE_DATA(const uint64_t, llvm::Type::getInt64Ty(Ctx), \
+                       VTableNameHash, ConstantInt::get(llvm::Type::getInt64Ty(Ctx), \
+                       IndexedInstrProf::ComputeHash(PGOVTableName)))
+INSTR_PROF_VTABLE_DATA(const IntPtrT, llvm::PointerType::getUnqual(Ctx), \
+                       VTablePointer, VTableAddr)
+INSTR_PROF_VTABLE_DATA(const uint32_t, llvm::Type::getInt32Ty(Ctx), VTableSize, \
+                       ConstantInt::get(llvm::Type::getInt32Ty(Ctx), \
+                                        VTableSizeVal))
+#undef INSTR_PROF_VTABLE_DATA
+/* INSTR_PROF_VTABLE_DATA end. */
 
 /* This is an internal data structure used by value profiler. It
  * is defined here to allow serialization code sharing by LLVM
@@ -147,6 +166,8 @@ INSTR_PROF_RAW_HEADER(uint64_t, CountersDelta,
 INSTR_PROF_RAW_HEADER(uint64_t, BitmapDelta,
                       (uintptr_t)BitmapBegin - (uintptr_t)DataBegin)
 INSTR_PROF_RAW_HEADER(uint64_t, NamesDelta, (uintptr_t)NamesBegin)
+INSTR_PROF_RAW_HEADER(uint64_t, NumVTables, NumVTables)
+INSTR_PROF_RAW_HEADER(uint64_t, VNamesSize, VNamesSize)
 INSTR_PROF_RAW_HEADER(uint64_t, ValueKindLast, IPVK_Last)
 #undef INSTR_PROF_RAW_HEADER
 /* INSTR_PROF_RAW_HEADER  end */
@@ -188,13 +209,26 @@ VALUE_PROF_FUNC_PARAM(uint32_t, CounterIndex, Type::getInt32Ty(Ctx))
 VALUE_PROF_KIND(IPVK_IndirectCallTarget, 0, "indirect call target")
 /* For memory intrinsic functions size profiling. */
 VALUE_PROF_KIND(IPVK_MemOPSize, 1, "memory intrinsic functions size")
+/* For virtual table address profiling, the address point of the virtual table
+ * (i.e., the address contained in objects pointing to a virtual table) are
+ * profiled. Note this may not be the address of the per C++ class virtual table
+ *  object (e.g., there might be an offset).
+ *
+ * The profiled addresses are stored in raw profile, together with the following
+ * two types of information.
+ * 1. The (starting and ending) addresses of per C++ class virtual table objects.
+ * 2. The (compressed) virtual table object names.
+ * RawInstrProfReader converts profiled virtual table addresses to virtual table
+ *  objects' MD5 hash.
+ */
+VALUE_PROF_KIND(IPVK_VTableTarget, 2, "The profiled address point of the vtable")
 /* These two kinds must be the last to be
  * declared. This is to make sure the string
  * array created with the template can be
  * indexed with the kind value.
  */
 VALUE_PROF_KIND(IPVK_First, IPVK_IndirectCallTarget, "first")
-VALUE_PROF_KIND(IPVK_Last, IPVK_MemOPSize, "last")
+VALUE_PROF_KIND(IPVK_Last, IPVK_VTableTarget, "last")
 
 #undef VALUE_PROF_KIND
 /* VALUE_PROF_KIND end */
@@ -284,12 +318,18 @@ INSTR_PROF_SECT_ENTRY(IPSK_bitmap, \
 INSTR_PROF_SECT_ENTRY(IPSK_name, \
                       INSTR_PROF_QUOTE(INSTR_PROF_NAME_COMMON), \
                       INSTR_PROF_NAME_COFF, "__DATA,")
+INSTR_PROF_SECT_ENTRY(IPSK_vname, \
+                      INSTR_PROF_QUOTE(INSTR_PROF_VNAME_COMMON), \
+                      INSTR_PROF_VNAME_COFF, "__DATA,")
 INSTR_PROF_SECT_ENTRY(IPSK_vals, \
                       INSTR_PROF_QUOTE(INSTR_PROF_VALS_COMMON), \
                       INSTR_PROF_VALS_COFF, "__DATA,")
 INSTR_PROF_SECT_ENTRY(IPSK_vnodes, \
                       INSTR_PROF_QUOTE(INSTR_PROF_VNODES_COMMON), \
                       INSTR_PROF_VNODES_COFF, "__DATA,")
+INSTR_PROF_SECT_ENTRY(IPSK_vtab, \
+                      INSTR_PROF_QUOTE(INSTR_PROF_VTAB_COMMON), \
+                      INSTR_PROF_VTAB_COFF, "__DATA,")
 INSTR_PROF_SECT_ENTRY(IPSK_covmap, \
                       INSTR_PROF_QUOTE(INSTR_PROF_COVMAP_COMMON), \
                       INSTR_PROF_COVMAP_COFF, "__LLVM_COV,")
@@ -668,9 +708,9 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
         (uint64_t)'f' << 16 | (uint64_t)'R' << 8 | (uint64_t)129
 
 /* Raw profile format version (start from 1). */
-#define INSTR_PROF_RAW_VERSION 9
+#define INSTR_PROF_RAW_VERSION 10
 /* Indexed profile format version (start from 1). */
-#define INSTR_PROF_INDEX_VERSION 11
+#define INSTR_PROF_INDEX_VERSION 12
 /* Coverage mapping format version (start from 0). */
 #define INSTR_PROF_COVMAP_VERSION 6
 
@@ -708,10 +748,12 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
    than WIN32 */
 #define INSTR_PROF_DATA_COMMON __llvm_prf_data
 #define INSTR_PROF_NAME_COMMON __llvm_prf_names
+#define INSTR_PROF_VNAME_COMMON __llvm_prf_vtabnames
 #define INSTR_PROF_CNTS_COMMON __llvm_prf_cnts
 #define INSTR_PROF_BITS_COMMON __llvm_prf_bits
 #define INSTR_PROF_VALS_COMMON __llvm_prf_vals
 #define INSTR_PROF_VNODES_COMMON __llvm_prf_vnds
+#define INSTR_PROF_VTAB_COMMON __llvm_prf_vtab
 #define INSTR_PROF_COVMAP_COMMON __llvm_covmap
 #define INSTR_PROF_COVFUN_COMMON __llvm_covfun
 #define INSTR_PROF_COVDATA_COMMON __llvm_covdata
@@ -722,10 +764,12 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
  */
 #define INSTR_PROF_DATA_COFF ".lprfd$M"
 #define INSTR_PROF_NAME_COFF ".lprfn$M"
+#define INSTR_PROF_VNAME_COFF ".lprfvn$M"
 #define INSTR_PROF_CNTS_COFF ".lprfc$M"
 #define INSTR_PROF_BITS_COFF ".lprfb$M"
 #define INSTR_PROF_VALS_COFF ".lprfv$M"
 #define INSTR_PROF_VNODES_COFF ".lprfnd$M"
+#define INSTR_PROF_VTAB_COFF ".lprfvt$M"
 #define INSTR_PROF_COVMAP_COFF ".lcovmap$M"
 #define INSTR_PROF_COVFUN_COFF ".lcovfun$M"
 /* Since cov data and cov names sections are not allocated, we don't need to

compiler-rt/lib/profile/InstrProfiling.h

@@ -49,6 +49,12 @@ typedef struct ValueProfNode {
 #include "profile/InstrProfData.inc"
 } ValueProfNode;
 
+typedef void *IntPtrT;
+typedef struct COMPILER_RT_ALIGNAS(INSTR_PROF_DATA_ALIGNMENT) VTableProfData {
+#define INSTR_PROF_VTABLE_DATA(Type, LLVMType, Name, Initializer) Type Name;
+#include "profile/InstrProfData.inc"
+} VTableProfData;
+
 /*!
  * \brief Return 1 if profile counters are continuously synced to the raw
  * profile via an mmap(). This is in contrast to the default mode, in which
@@ -103,12 +109,16 @@ const __llvm_profile_data *__llvm_profile_begin_data(void);
 const __llvm_profile_data *__llvm_profile_end_data(void);
 const char *__llvm_profile_begin_names(void);
 const char *__llvm_profile_end_names(void);
+const char *__llvm_profile_begin_vtabnames(void);
+const char *__llvm_profile_end_vtabnames(void);
 char *__llvm_profile_begin_counters(void);
 char *__llvm_profile_end_counters(void);
 char *__llvm_profile_begin_bitmap(void);
 char *__llvm_profile_end_bitmap(void);
 ValueProfNode *__llvm_profile_begin_vnodes();
 ValueProfNode *__llvm_profile_end_vnodes();
+VTableProfData *__llvm_profile_begin_vtables();
+VTableProfData *__llvm_profile_end_vtables();
 uint32_t *__llvm_profile_begin_orderfile();
 
 /*!
@@ -252,20 +262,31 @@ uint64_t __llvm_profile_get_num_bitmap_bytes(const char *Begin,
 /*! \brief Get the size of the profile name section in bytes. */
 uint64_t __llvm_profile_get_name_size(const char *Begin, const char *End);
 
-/* ! \brief Given the sizes of the data and counter information, return the
- * number of padding bytes before and after the counters, and after the names,
- * in the raw profile.
+/*! \brief Get the number of virtual table profile data entries */
+uint64_t __llvm_profile_get_num_vtable(const VTableProfData *Begin,
+                                       const VTableProfData *End);
+
+/*! \brief Get the size of virtual table profile data in bytes. */
+uint64_t __llvm_profile_get_vtable_section_size(const VTableProfData *Begin,
+                                                const VTableProfData *End);
+
+/* ! \brief Given the sizes of the data and counter information, computes the
+ * number of padding bytes before and after the counter section, as well as the
+ * number of padding bytes after other setions in the raw profile.
+ * Returns -1 upon errors and 0 upon success. Output parameters should be used
+ * iff return value is 0.
  *
  * Note: When mmap() mode is disabled, no padding bytes before/after counters
  * are needed. However, in mmap() mode, the counter section in the raw profile
  * must be page-aligned: this API computes the number of padding bytes
  * needed to achieve that.
  */
-void __llvm_profile_get_padding_sizes_for_counters(
+int __llvm_profile_get_padding_sizes_for_counters(
     uint64_t DataSize, uint64_t CountersSize, uint64_t NumBitmapBytes,
-    uint64_t NamesSize, uint64_t *PaddingBytesBeforeCounters,
-    uint64_t *PaddingBytesAfterCounters, uint64_t *PaddingBytesAfterBitmap,
-    uint64_t *PaddingBytesAfterNames);
+    uint64_t NamesSize, uint64_t VTableSize, uint64_t VNameSize,
+    uint64_t *PaddingBytesBeforeCounters, uint64_t *PaddingBytesAfterCounters,
+    uint64_t *PaddingBytesAfterBitmap, uint64_t *PaddingBytesAfterNames,
+    uint64_t *PaddingBytesAfterVTable, uint64_t *PaddingBytesAfterVNames);
 
 /*!
  * \brief Set the flag that profile data has been dumped to the file.

compiler-rt/lib/profile/InstrProfilingBuffer.c

@@ -51,16 +51,29 @@ uint64_t __llvm_profile_get_size_for_buffer(void) {
   const char *BitmapEnd = __llvm_profile_end_bitmap();
   const char *NamesBegin = __llvm_profile_begin_names();
   const char *NamesEnd = __llvm_profile_end_names();
+  const VTableProfData *VTableBegin = __llvm_profile_begin_vtables();
+  const VTableProfData *VTableEnd = __llvm_profile_end_vtables();
+  const char *VNamesBegin = __llvm_profile_begin_vtabnames();
+  const char *VNamesEnd = __llvm_profile_end_vtabnames();
 
   return __llvm_profile_get_size_for_buffer_internal(
       DataBegin, DataEnd, CountersBegin, CountersEnd, BitmapBegin, BitmapEnd,
-      NamesBegin, NamesEnd);
+      NamesBegin, NamesEnd, VTableBegin, VTableEnd, VNamesBegin, VNamesEnd);
 }
 
 COMPILER_RT_VISIBILITY
 uint64_t __llvm_profile_get_num_data(const __llvm_profile_data *Begin,
                                      const __llvm_profile_data *End) {
   intptr_t BeginI = (intptr_t)Begin, EndI = (intptr_t)End;
+  // `sizeof(__llvm_profile_data) - 1` is required in the numerator when
+  // [Begin, End] represents an inclusive range.
+  // For ELF, [Begin, End) represents the address of linker-inserted
+  // symbols  `__start__<elf-section>` and `__stop_<elf-section>`.
+  // Thereby, `End` is one byte past the inclusive range, and
+  // `sizeof(__llvm_profile_data) - 1` is not necessary in the numerator to get
+  // the correct number of profile data.
+  // FIXME: Consider removing `sizeof(__llvm_profile_data) - 1` if this is true
+  // across platforms.
   return ((EndI + sizeof(__llvm_profile_data) - 1) - BeginI) /
          sizeof(__llvm_profile_data);
 }
@@ -71,6 +84,26 @@ uint64_t __llvm_profile_get_data_size(const __llvm_profile_data *Begin,
   return __llvm_profile_get_num_data(Begin, End) * sizeof(__llvm_profile_data);
 }
 
+// Counts the number of `VTableProfData` elements within the range of [Begin,
+// End). Caller should guarantee that End points to one byte past the inclusive
+// range.
+// FIXME: Add a compiler-rt test to make sure the number of vtables in the
+// raw profile is the same as the number of vtable elements in the instrumented
+// binary.
+COMPILER_RT_VISIBILITY
+uint64_t __llvm_profile_get_num_vtable(const VTableProfData *Begin,
+                                       const VTableProfData *End) {
+  // Convert pointers to intptr_t to use integer arithmetic.
+  intptr_t EndI = (intptr_t)End, BeginI = (intptr_t)Begin;
+  return (EndI - BeginI) / sizeof(VTableProfData);
+}
+
+COMPILER_RT_VISIBILITY
+uint64_t __llvm_profile_get_vtable_section_size(const VTableProfData *Begin,
+                                                const VTableProfData *End) {
+  return (intptr_t)(End) - (intptr_t)(Begin);
+}
+
 COMPILER_RT_VISIBILITY size_t __llvm_profile_counter_entry_size(void) {
   if (__llvm_profile_get_version() & VARIANT_MASK_BYTE_COVERAGE)
     return sizeof(uint8_t);
@@ -119,21 +152,33 @@ static int needsCounterPadding(void) {
 }
 
 COMPILER_RT_VISIBILITY
-void __llvm_profile_get_padding_sizes_for_counters(
+int __llvm_profile_get_padding_sizes_for_counters(
     uint64_t DataSize, uint64_t CountersSize, uint64_t NumBitmapBytes,
-    uint64_t NamesSize, uint64_t *PaddingBytesBeforeCounters,
-    uint64_t *PaddingBytesAfterCounters, uint64_t *PaddingBytesAfterBitmapBytes,
-    uint64_t *PaddingBytesAfterNames) {
+    uint64_t NamesSize, uint64_t VTableSize, uint64_t VNameSize,
+    uint64_t *PaddingBytesBeforeCounters, uint64_t *PaddingBytesAfterCounters,
+    uint64_t *PaddingBytesAfterBitmapBytes, uint64_t *PaddingBytesAfterNames,
+    uint64_t *PaddingBytesAfterVTable, uint64_t *PaddingBytesAfterVName) {
+  // Counter padding is needed only if continuous mode is enabled.
   if (!needsCounterPadding()) {
     *PaddingBytesBeforeCounters = 0;
     *PaddingBytesAfterCounters =
         __llvm_profile_get_num_padding_bytes(CountersSize);
     *PaddingBytesAfterBitmapBytes =
         __llvm_profile_get_num_padding_bytes(NumBitmapBytes);
     *PaddingBytesAfterNames = __llvm_profile_get_num_padding_bytes(NamesSize);
-    return;
+    if (PaddingBytesAfterVTable != NULL)
+      *PaddingBytesAfterVTable =
+          __llvm_profile_get_num_padding_bytes(VTableSize);
+    if (PaddingBytesAfterVName != NULL)
+      *PaddingBytesAfterVName = __llvm_profile_get_num_padding_bytes(VNameSize);
+    return 0;
   }
 
+  // Value profiling not supported in continuous mode at profile-write time.
+  // Return -1 to alert the incompatibility.
+  if (VTableSize != 0 || VNameSize != 0)
+    return -1;
+
   // In continuous mode, the file offsets for headers and for the start of
   // counter sections need to be page-aligned.
   *PaddingBytesBeforeCounters =
@@ -142,34 +187,52 @@ void __llvm_profile_get_padding_sizes_for_counters(
   *PaddingBytesAfterBitmapBytes =
       calculateBytesNeededToPageAlign(NumBitmapBytes);
   *PaddingBytesAfterNames = calculateBytesNeededToPageAlign(NamesSize);
+  // Set these two variables to zero to avoid uninitialized variables
+  // even if VTableSize and VNameSize are known to be zero.
+  if (PaddingBytesAfterVTable != NULL)
+    *PaddingBytesAfterVTable = 0;
+  if (PaddingBytesAfterVName != NULL)
+    *PaddingBytesAfterVName = 0;
+  return 0;
 }
 
 COMPILER_RT_VISIBILITY
 uint64_t __llvm_profile_get_size_for_buffer_internal(
     const __llvm_profile_data *DataBegin, const __llvm_profile_data *DataEnd,
     const char *CountersBegin, const char *CountersEnd, const char *BitmapBegin,
-    const char *BitmapEnd, const char *NamesBegin, const char *NamesEnd) {
+    const char *BitmapEnd, const char *NamesBegin, const char *NamesEnd,
+    const VTableProfData *VTableBegin, const VTableProfData *VTableEnd,
+    const char *VNamesBegin, const char *VNamesEnd) {
   /* Match logic in __llvm_profile_write_buffer(). */
   const uint64_t NamesSize = (NamesEnd - NamesBegin) * sizeof(char);
   uint64_t DataSize = __llvm_profile_get_data_size(DataBegin, DataEnd);
   uint64_t CountersSize =
       __llvm_profile_get_counters_size(CountersBegin, CountersEnd);
   const uint64_t NumBitmapBytes =
       __llvm_profile_get_num_bitmap_bytes(BitmapBegin, BitmapEnd);
+  const uint64_t VTableSize =
+      __llvm_profile_get_vtable_section_size(VTableBegin, VTableEnd);
+  const uint64_t VNameSize =
+      __llvm_profile_get_name_size(VNamesBegin, VNamesEnd);
 
   /* Determine how much padding is needed before/after the counters and after
    * the names. */
   uint64_t PaddingBytesBeforeCounters, PaddingBytesAfterCounters,
-      PaddingBytesAfterNames, PaddingBytesAfterBitmapBytes;
+      PaddingBytesAfterNames, PaddingBytesAfterBitmapBytes,
+      PaddingBytesAfterVTable, PaddingBytesAfterVNames;
   __llvm_profile_get_padding_sizes_for_counters(
-      DataSize, CountersSize, NumBitmapBytes, NamesSize,
-      &PaddingBytesBeforeCounters, &PaddingBytesAfterCounters,
-      &PaddingBytesAfterBitmapBytes, &PaddingBytesAfterNames);
+      DataSize, CountersSize, NumBitmapBytes, NamesSize, 0 /* VTableSize */,
+      0 /* VNameSize */, &PaddingBytesBeforeCounters,
+      &PaddingBytesAfterCounters, &PaddingBytesAfterBitmapBytes,
+      &PaddingBytesAfterNames, &PaddingBytesAfterVTable,
+      &PaddingBytesAfterVNames);
 
   return sizeof(__llvm_profile_header) + __llvm_write_binary_ids(NULL) +
          DataSize + PaddingBytesBeforeCounters + CountersSize +
          PaddingBytesAfterCounters + NumBitmapBytes +
-         PaddingBytesAfterBitmapBytes + NamesSize + PaddingBytesAfterNames;
+         PaddingBytesAfterBitmapBytes + NamesSize + PaddingBytesAfterNames +
+         VTableSize + PaddingBytesAfterVTable + VNameSize +
+         PaddingBytesAfterVNames;
 }
 
 COMPILER_RT_VISIBILITY
@@ -191,7 +254,10 @@ COMPILER_RT_VISIBILITY int __llvm_profile_write_buffer_internal(
     const char *NamesBegin, const char *NamesEnd) {
   ProfDataWriter BufferWriter;
   initBufferWriter(&BufferWriter, Buffer);
-  return lprofWriteDataImpl(&BufferWriter, DataBegin, DataEnd, CountersBegin,
-                            CountersEnd, BitmapBegin, BitmapEnd, 0, NamesBegin,
-                            NamesEnd, 0);
+  // Set virtual table arguments to NULL since they are not supported yet.
+  return lprofWriteDataImpl(
+      &BufferWriter, DataBegin, DataEnd, CountersBegin, CountersEnd,
+      BitmapBegin, BitmapEnd, /*VPDataReader=*/0, NamesBegin, NamesEnd,
+      /*VTableBegin=*/NULL, /*VTableEnd=*/NULL, /*VNamesBegin=*/NULL,
+      /*VNamesEnd=*/NULL, /*SkipNameDataWrite=*/0);
 }

compiler-rt/lib/profile/InstrProfilingInternal.h

@@ -22,7 +22,9 @@
 uint64_t __llvm_profile_get_size_for_buffer_internal(
     const __llvm_profile_data *DataBegin, const __llvm_profile_data *DataEnd,
     const char *CountersBegin, const char *CountersEnd, const char *BitmapBegin,
-    const char *BitmapEnd, const char *NamesBegin, const char *NamesEnd);
+    const char *BitmapEnd, const char *NamesBegin, const char *NamesEnd,
+    const VTableProfData *VTableBegin, const VTableProfData *VTableEnd,
+    const char *VNamesBegin, const char *VNamesEnd);
 
 /*!
  * \brief Write instrumentation data to the given buffer, given explicit
@@ -156,7 +158,9 @@ int lprofWriteDataImpl(ProfDataWriter *Writer,
                        const char *CountersBegin, const char *CountersEnd,
                        const char *BitmapBegin, const char *BitmapEnd,
                        VPDataReaderType *VPDataReader, const char *NamesBegin,
-                       const char *NamesEnd, int SkipNameDataWrite);
+                       const char *NamesEnd, const VTableProfData *VTableBegin,
+                       const VTableProfData *VTableEnd, const char *VNamesBegin,
+                       const char *VNamesEnd, int SkipNameDataWrite);
 
 /* Merge value profile data pointed to by SrcValueProfData into
  * in-memory profile counters pointed by to DstData.  */