| 1 | #ifndef MEMPROF_DATA_INC |
| 2 | #define MEMPROF_DATA_INC |
| 3 | /*===-- MemProfData.inc - MemProf profiling runtime structures -*- C++ -*-=== *\ |
| 4 | |* |
| 5 | |* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 6 | |* See https://llvm.org/LICENSE.txt for license information. |
| 7 | |* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 8 | |* |
| 9 | \*===----------------------------------------------------------------------===*/ |
| 10 | /* |
| 11 | * This is the main file that defines all the data structure, signature, |
| 12 | * constant literals that are shared across profiling runtime library, |
| 13 | * and host tools (reader/writer). |
| 14 | * |
| 15 | * This file has two identical copies. The primary copy lives in LLVM and |
| 16 | * the other one sits in compiler-rt/include/profile directory. To make changes |
| 17 | * in this file, first modify the primary copy and copy it over to compiler-rt. |
| 18 | * Testing of any change in this file can start only after the two copies are |
| 19 | * synced up. |
| 20 | * |
| 21 | \*===----------------------------------------------------------------------===*/ |
| 22 | #include <string.h> |
| 23 | |
| 24 | #ifdef _MSC_VER |
| 25 | #define PACKED(...) __pragma(pack(push,1)) __VA_ARGS__ __pragma(pack(pop)) |
| 26 | #else |
| 27 | #define PACKED(...) __VA_ARGS__ __attribute__((__packed__)) |
| 28 | #endif |
| 29 | |
| 30 | // A 64-bit magic number to uniquely identify the raw binary memprof profile file. |
| 31 | #define MEMPROF_RAW_MAGIC_64 \ |
| 32 | ((uint64_t)255 << 56 | (uint64_t)'m' << 48 | (uint64_t)'p' << 40 | (uint64_t)'r' << 32 | \ |
| 33 | (uint64_t)'o' << 24 | (uint64_t)'f' << 16 | (uint64_t)'r' << 8 | (uint64_t)129) |
| 34 | |
| 35 | // The version number of the raw binary format. |
| 36 | #define MEMPROF_RAW_VERSION 5ULL |
| 37 | |
| 38 | // Currently supported versions. |
| 39 | #define MEMPROF_RAW_SUPPORTED_VERSIONS {3ULL, 4ULL, 5ULL} |
| 40 | |
| 41 | #define MEMPROF_V3_MIB_SIZE 132ULL; |
| 42 | |
| 43 | #define MEMPROF_BUILDID_MAX_SIZE 32ULL |
| 44 | |
| 45 | namespace llvm { |
| 46 | namespace memprof { |
| 47 | // A struct describing the header used for the raw binary memprof profile format. |
| 48 | PACKED(struct { |
| 49 | uint64_t ; |
| 50 | uint64_t ; |
| 51 | uint64_t ; |
| 52 | uint64_t ; |
| 53 | uint64_t ; |
| 54 | uint64_t ; |
| 55 | }); |
| 56 | |
| 57 | // A struct describing the information necessary to describe a /proc/maps |
| 58 | // segment entry for a particular binary/library identified by its build id. |
| 59 | PACKED(struct SegmentEntry { |
| 60 | uint64_t Start; |
| 61 | uint64_t End; |
| 62 | uint64_t Offset; |
| 63 | uint64_t BuildIdSize; |
| 64 | uint8_t BuildId[MEMPROF_BUILDID_MAX_SIZE] = {0}; |
| 65 | |
| 66 | // This constructor is only used in tests so don't set the BuildId. |
| 67 | SegmentEntry(uint64_t S, uint64_t E, uint64_t O) |
| 68 | : Start(S), End(E), Offset(O), BuildIdSize(0) {} |
| 69 | |
| 70 | SegmentEntry(const SegmentEntry& S) { |
| 71 | Start = S.Start; |
| 72 | End = S.End; |
| 73 | Offset = S.Offset; |
| 74 | BuildIdSize = S.BuildIdSize; |
| 75 | memcpy(BuildId, S.BuildId, S.BuildIdSize); |
| 76 | } |
| 77 | |
| 78 | SegmentEntry& operator=(const SegmentEntry& S) { |
| 79 | Start = S.Start; |
| 80 | End = S.End; |
| 81 | Offset = S.Offset; |
| 82 | BuildIdSize = S.BuildIdSize; |
| 83 | memcpy(BuildId, S.BuildId, S.BuildIdSize); |
| 84 | return *this; |
| 85 | } |
| 86 | |
| 87 | bool operator==(const SegmentEntry& S) const { |
| 88 | return Start == S.Start && End == S.End && Offset == S.Offset && |
| 89 | BuildIdSize == S.BuildIdSize && |
| 90 | memcmp(BuildId, S.BuildId, S.BuildIdSize) == 0; |
| 91 | } |
| 92 | }); |
| 93 | |
| 94 | // Packed struct definition for MSVC. We can't use the PACKED macro defined in |
| 95 | // MemProfData.inc since it would mean we are embedding a directive (the |
| 96 | // #include for MIBEntryDef) into the macros which is undefined behaviour. |
| 97 | #ifdef _MSC_VER |
| 98 | __pragma(pack(push,1)) |
| 99 | #endif |
| 100 | |
| 101 | // A struct representing the heap allocation characteristics of a particular |
| 102 | // runtime context. This struct is shared between the compiler-rt runtime and |
| 103 | // the raw profile reader. The indexed format uses a separate, self-describing |
| 104 | // backwards compatible format. |
| 105 | struct MemInfoBlock{ |
| 106 | |
| 107 | #define MIBEntryDef(NameTag, Name, Type) Type Name; |
| 108 | #include "MIBEntryDef.inc" |
| 109 | #undef MIBEntryDef |
| 110 | |
| 111 | bool operator==(const MemInfoBlock& Other) const { |
| 112 | bool IsEqual = true; |
| 113 | #define MIBEntryDef(NameTag, Name, Type) \ |
| 114 | IsEqual = (IsEqual && Name == Other.Name); |
| 115 | #include "MIBEntryDef.inc" |
| 116 | #undef MIBEntryDef |
| 117 | return IsEqual; |
| 118 | } |
| 119 | |
| 120 | MemInfoBlock() { |
| 121 | #define MIBEntryDef(NameTag, Name, Type) Name = Type(); |
| 122 | #include "MIBEntryDef.inc" |
| 123 | #undef MIBEntryDef |
| 124 | } |
| 125 | |
| 126 | MemInfoBlock(uint32_t Size, uint64_t AccessCount, uint32_t AllocTs, |
| 127 | uint32_t DeallocTs, uint32_t AllocCpu, uint32_t DeallocCpu, |
| 128 | uintptr_t Histogram, uint32_t HistogramSize) |
| 129 | : MemInfoBlock() { |
| 130 | AllocCount = 1U; |
| 131 | TotalAccessCount = AccessCount; |
| 132 | MinAccessCount = AccessCount; |
| 133 | MaxAccessCount = AccessCount; |
| 134 | TotalSize = Size; |
| 135 | MinSize = Size; |
| 136 | MaxSize = Size; |
| 137 | AllocTimestamp = AllocTs; |
| 138 | DeallocTimestamp = DeallocTs; |
| 139 | TotalLifetime = DeallocTimestamp - AllocTimestamp; |
| 140 | MinLifetime = TotalLifetime; |
| 141 | MaxLifetime = TotalLifetime; |
| 142 | // Access density is accesses per byte. Multiply by 100 to include the |
| 143 | // fractional part. |
| 144 | TotalAccessDensity = AccessCount * 100 / Size; |
| 145 | MinAccessDensity = TotalAccessDensity; |
| 146 | MaxAccessDensity = TotalAccessDensity; |
| 147 | // Lifetime access density is the access density per second of lifetime. |
| 148 | // Multiply by 1000 to convert denominator lifetime to seconds (using a |
| 149 | // minimum lifetime of 1ms to avoid divide by 0. Do the multiplication first |
| 150 | // to reduce truncations to 0. |
| 151 | TotalLifetimeAccessDensity = |
| 152 | TotalAccessDensity * 1000 / (TotalLifetime ? TotalLifetime : 1); |
| 153 | MinLifetimeAccessDensity = TotalLifetimeAccessDensity; |
| 154 | MaxLifetimeAccessDensity = TotalLifetimeAccessDensity; |
| 155 | AllocCpuId = AllocCpu; |
| 156 | DeallocCpuId = DeallocCpu; |
| 157 | NumMigratedCpu = AllocCpuId != DeallocCpuId; |
| 158 | AccessHistogramSize = HistogramSize; |
| 159 | AccessHistogram = Histogram; |
| 160 | } |
| 161 | |
| 162 | void Merge(const MemInfoBlock &newMIB) { |
| 163 | AllocCount += newMIB.AllocCount; |
| 164 | |
| 165 | TotalAccessCount += newMIB.TotalAccessCount; |
| 166 | MinAccessCount = newMIB.MinAccessCount < MinAccessCount ? newMIB.MinAccessCount : MinAccessCount; |
| 167 | MaxAccessCount = newMIB.MaxAccessCount > MaxAccessCount ? newMIB.MaxAccessCount : MaxAccessCount; |
| 168 | |
| 169 | TotalSize += newMIB.TotalSize; |
| 170 | MinSize = newMIB.MinSize < MinSize ? newMIB.MinSize : MinSize; |
| 171 | MaxSize = newMIB.MaxSize > MaxSize ? newMIB.MaxSize : MaxSize; |
| 172 | |
| 173 | TotalLifetime += newMIB.TotalLifetime; |
| 174 | MinLifetime = newMIB.MinLifetime < MinLifetime ? newMIB.MinLifetime : MinLifetime; |
| 175 | MaxLifetime = newMIB.MaxLifetime > MaxLifetime ? newMIB.MaxLifetime : MaxLifetime; |
| 176 | |
| 177 | TotalAccessDensity += newMIB.TotalAccessDensity; |
| 178 | MinAccessDensity = newMIB.MinAccessDensity < MinAccessDensity |
| 179 | ? newMIB.MinAccessDensity |
| 180 | : MinAccessDensity; |
| 181 | MaxAccessDensity = newMIB.MaxAccessDensity > MaxAccessDensity |
| 182 | ? newMIB.MaxAccessDensity |
| 183 | : MaxAccessDensity; |
| 184 | |
| 185 | TotalLifetimeAccessDensity += newMIB.TotalLifetimeAccessDensity; |
| 186 | MinLifetimeAccessDensity = |
| 187 | newMIB.MinLifetimeAccessDensity < MinLifetimeAccessDensity |
| 188 | ? newMIB.MinLifetimeAccessDensity |
| 189 | : MinLifetimeAccessDensity; |
| 190 | MaxLifetimeAccessDensity = |
| 191 | newMIB.MaxLifetimeAccessDensity > MaxLifetimeAccessDensity |
| 192 | ? newMIB.MaxLifetimeAccessDensity |
| 193 | : MaxLifetimeAccessDensity; |
| 194 | |
| 195 | // We know newMIB was deallocated later, so just need to check if it was |
| 196 | // allocated before last one deallocated. |
| 197 | NumLifetimeOverlaps += newMIB.AllocTimestamp < DeallocTimestamp; |
| 198 | AllocTimestamp = newMIB.AllocTimestamp; |
| 199 | DeallocTimestamp = newMIB.DeallocTimestamp; |
| 200 | |
| 201 | NumSameAllocCpu += AllocCpuId == newMIB.AllocCpuId; |
| 202 | NumSameDeallocCpu += DeallocCpuId == newMIB.DeallocCpuId; |
| 203 | AllocCpuId = newMIB.AllocCpuId; |
| 204 | DeallocCpuId = newMIB.DeallocCpuId; |
| 205 | |
| 206 | // For merging histograms, we always keep the longer histogram, and add |
| 207 | // values of shorter histogram to larger one. |
| 208 | uintptr_t ShorterHistogram; |
| 209 | uint32_t ShorterHistogramSize; |
| 210 | if (newMIB.AccessHistogramSize > AccessHistogramSize) { |
| 211 | ShorterHistogram = AccessHistogram; |
| 212 | ShorterHistogramSize = AccessHistogramSize; |
| 213 | // Swap histogram of current to larger histogram |
| 214 | AccessHistogram = newMIB.AccessHistogram; |
| 215 | AccessHistogramSize = newMIB.AccessHistogramSize; |
| 216 | } else { |
| 217 | ShorterHistogram = newMIB.AccessHistogram; |
| 218 | ShorterHistogramSize = newMIB.AccessHistogramSize; |
| 219 | } |
| 220 | for (size_t i = 0; i < ShorterHistogramSize; ++i) { |
| 221 | ((uint64_t *)AccessHistogram)[i] += ((uint64_t *)ShorterHistogram)[i]; |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | #ifdef _MSC_VER |
| 226 | } __pragma(pack(pop)); |
| 227 | #else |
| 228 | } __attribute__((__packed__)); |
| 229 | #endif |
| 230 | |
| 231 | constexpr int MantissaBits = 12; |
| 232 | constexpr int ExponentBits = 4; |
| 233 | constexpr uint16_t MaxMantissa = (1U << MantissaBits) - 1; |
| 234 | constexpr uint16_t MaxExponent = (1U << ExponentBits) - 1; |
| 235 | constexpr uint64_t MaxRepresentableValue = static_cast<uint64_t>(MaxMantissa) |
| 236 | << MaxExponent; |
| 237 | |
| 238 | // Encodes a 64-bit unsigned integer into a 16-bit scaled integer format. |
| 239 | inline uint16_t encodeHistogramCount(uint64_t Count) { |
| 240 | if (Count == 0) |
| 241 | return 0; |
| 242 | |
| 243 | if (Count > MaxRepresentableValue) |
| 244 | Count = MaxRepresentableValue; |
| 245 | |
| 246 | if (Count <= MaxMantissa) |
| 247 | return Count; |
| 248 | |
| 249 | uint64_t M = Count; |
| 250 | uint16_t E = 0; |
| 251 | while (M > MaxMantissa) { |
| 252 | M = (M + 1) >> 1; |
| 253 | E++; |
| 254 | } |
| 255 | return (E << MantissaBits) | static_cast<uint16_t>(M); |
| 256 | } |
| 257 | |
| 258 | // Decodes a 16-bit scaled integer and returns the |
| 259 | // decoded 64-bit unsigned integer. |
| 260 | inline uint64_t decodeHistogramCount(uint16_t EncodedValue) { |
| 261 | const uint16_t E = EncodedValue >> MantissaBits; |
| 262 | const uint16_t M = EncodedValue & MaxMantissa; |
| 263 | return static_cast<uint64_t>(M) << E; |
| 264 | } |
| 265 | |
| 266 | } // namespace memprof |
| 267 | } // namespace llvm |
| 268 | |
| 269 | #endif |
| 270 | |