diff --git a/src/external/tracy/COMMIT b/src/external/tracy/COMMIT index 490a0cdcc..1bd5ffe10 100644 --- a/src/external/tracy/COMMIT +++ b/src/external/tracy/COMMIT @@ -1 +1 @@ -v0.9 +v0.9.1 diff --git a/src/external/tracy/TracyClient.cpp b/src/external/tracy/TracyClient.cpp index 4aa4647a5..77f81a4a7 100644 --- a/src/external/tracy/TracyClient.cpp +++ b/src/external/tracy/TracyClient.cpp @@ -28,6 +28,7 @@ #include "client/tracy_rpmalloc.cpp" #include "client/TracyDxt1.cpp" #include "client/TracyAlloc.cpp" +#include "client/TracyOverride.cpp" #if TRACY_HAS_CALLSTACK == 2 || TRACY_HAS_CALLSTACK == 3 || TRACY_HAS_CALLSTACK == 4 || TRACY_HAS_CALLSTACK == 6 # include "libbacktrace/alloc.cpp" diff --git a/src/external/tracy/client/TracyCallstack.cpp b/src/external/tracy/client/TracyCallstack.cpp index 1cfe9e795..a874446c2 100644 --- a/src/external/tracy/client/TracyCallstack.cpp +++ b/src/external/tracy/client/TracyCallstack.cpp @@ -227,6 +227,10 @@ void InitCallstack() const auto res = GetModuleFileNameA( mod[i], name, 1021 ); if( res > 0 ) { + // This may be a new module loaded since our call to SymInitialize. + // Just in case, force DbgHelp to load its pdb ! + SymLoadModuleEx(proc, NULL, name, NULL, (DWORD64)info.lpBaseOfDll, info.SizeOfImage, NULL, 0); + auto ptr = name + res; while( ptr > name && *ptr != '\\' && *ptr != '/' ) ptr--; if( ptr > name ) ptr++; diff --git a/src/external/tracy/client/TracyCallstack.hpp b/src/external/tracy/client/TracyCallstack.hpp index 5c24e7db2..0b522b730 100644 --- a/src/external/tracy/client/TracyCallstack.hpp +++ b/src/external/tracy/client/TracyCallstack.hpp @@ -2,6 +2,7 @@ #define __TRACYCALLSTACK_HPP__ #include "../common/TracyApi.h" +#include "../common/TracyForceInline.hpp" #include "TracyCallstack.h" #if TRACY_HAS_CALLSTACK == 2 || TRACY_HAS_CALLSTACK == 5 @@ -11,7 +12,14 @@ #endif -#ifdef TRACY_HAS_CALLSTACK +#ifndef TRACY_HAS_CALLSTACK + +namespace tracy +{ +static tracy_force_inline void* Callstack( int depth ) { return nullptr; } +} + +#else #ifdef TRACY_DEBUGINFOD # include @@ -21,7 +29,6 @@ #include #include "../common/TracyAlloc.hpp" -#include "../common/TracyForceInline.hpp" namespace tracy { diff --git a/src/external/tracy/client/TracyOverride.cpp b/src/external/tracy/client/TracyOverride.cpp new file mode 100644 index 000000000..591508a7f --- /dev/null +++ b/src/external/tracy/client/TracyOverride.cpp @@ -0,0 +1,26 @@ +#ifdef TRACY_ENABLE +# ifdef __linux__ +# include "TracyDebug.hpp" +# ifdef TRACY_VERBOSE +# include +# include +# endif + +extern "C" int dlclose( void* hnd ) +{ +#ifdef TRACY_VERBOSE + struct link_map* lm; + if( dlinfo( hnd, RTLD_DI_LINKMAP, &lm ) == 0 ) + { + TracyDebug( "Overriding dlclose for %s\n", lm->l_name ); + } + else + { + TracyDebug( "Overriding dlclose for unknown object (%s)\n", dlerror() ); + } +#endif + return 0; +} + +# endif +#endif diff --git a/src/external/tracy/client/TracyProfiler.cpp b/src/external/tracy/client/TracyProfiler.cpp index ea168e4fa..6104a7edd 100644 --- a/src/external/tracy/client/TracyProfiler.cpp +++ b/src/external/tracy/client/TracyProfiler.cpp @@ -1415,7 +1415,9 @@ void Profiler::SpawnWorkerThreads() #if defined _WIN32 && !defined TRACY_UWP && !defined TRACY_NO_CRASH_HANDLER s_profilerThreadId = GetThreadId( s_thread->Handle() ); +# ifdef TRACY_HAS_CALLSTACK s_symbolThreadId = GetThreadId( s_symbolThread->Handle() ); +# endif m_exceptionHandler = AddVectoredExceptionHandler( 1, CrashFilter ); #endif @@ -1454,7 +1456,7 @@ Profiler::~Profiler() if( m_crashHandlerInstalled ) RemoveVectoredExceptionHandler( m_exceptionHandler ); #endif -#ifdef __linux__ +#if defined __linux__ && !defined TRACY_NO_CRASH_HANDLER if( m_crashHandlerInstalled ) { sigaction( TRACY_CRASH_SIGNAL, &m_prevSignal.pwr, nullptr ); @@ -1520,7 +1522,7 @@ bool Profiler::ShouldExit() void Profiler::Worker() { -#ifdef __linux__ +#if defined __linux__ && !defined TRACY_NO_CRASH_HANDLER s_profilerTid = syscall( SYS_gettid ); #endif @@ -3024,9 +3026,9 @@ void Profiler::SendSourceLocation( uint64_t ptr ) MemWrite( &item.srcloc.file, (uint64_t)srcloc->file ); MemWrite( &item.srcloc.function, (uint64_t)srcloc->function ); MemWrite( &item.srcloc.line, srcloc->line ); - MemWrite( &item.srcloc.r, uint8_t( ( srcloc->color ) & 0xFF ) ); + MemWrite( &item.srcloc.b, uint8_t( ( srcloc->color ) & 0xFF ) ); MemWrite( &item.srcloc.g, uint8_t( ( srcloc->color >> 8 ) & 0xFF ) ); - MemWrite( &item.srcloc.b, uint8_t( ( srcloc->color >> 16 ) & 0xFF ) ); + MemWrite( &item.srcloc.r, uint8_t( ( srcloc->color >> 16 ) & 0xFF ) ); AppendData( &item, QueueDataSize[(int)QueueType::SourceLocation] ); } @@ -3329,10 +3331,8 @@ bool Profiler::HandleServerQuery() uint8_t type; uint64_t ptr; - uint32_t extra; memcpy( &type, &payload.type, sizeof( payload.type ) ); memcpy( &ptr, &payload.ptr, sizeof( payload.ptr ) ); - memcpy( &extra, &payload.extra, sizeof( payload.extra ) ); switch( type ) { @@ -3379,7 +3379,7 @@ bool Profiler::HandleServerQuery() break; #ifndef TRACY_NO_CODE_TRANSFER case ServerQuerySymbolCode: - HandleSymbolCodeQuery( ptr, extra ); + HandleSymbolCodeQuery( ptr, payload.extra ); break; #endif case ServerQuerySourceCode: @@ -3396,7 +3396,7 @@ bool Profiler::HandleServerQuery() break; case ServerQueryDataTransferPart: memcpy( m_queryDataPtr, &ptr, 8 ); - memcpy( m_queryDataPtr+8, &extra, 4 ); + memcpy( m_queryDataPtr+8, &payload.extra, 4 ); m_queryDataPtr += 12; AckServerQuery(); break; @@ -3546,7 +3546,7 @@ void Profiler::CalibrateDelay() constexpr int Events = Iterations * 2; // start + end static_assert( Events < QueuePrealloc, "Delay calibration loop will allocate memory in queue" ); - static const tracy::SourceLocationData __tracy_source_location { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; + static const tracy::SourceLocationData __tracy_source_location { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; const auto t0 = GetTime(); for( int i=0; izoneColor.r, uint8_t( ( color ) & 0xFF ) ); + tracy::MemWrite( &item->zoneColor.b, uint8_t( ( color ) & 0xFF ) ); tracy::MemWrite( &item->zoneColor.g, uint8_t( ( color >> 8 ) & 0xFF ) ); - tracy::MemWrite( &item->zoneColor.b, uint8_t( ( color >> 16 ) & 0xFF ) ); + tracy::MemWrite( &item->zoneColor.r, uint8_t( ( color >> 16 ) & 0xFF ) ); TracyQueueCommitC( zoneColorThread ); } } @@ -4147,6 +4147,8 @@ TRACY_API void ___tracy_emit_frame_mark_start( const char* name ) { tracy::Profi TRACY_API void ___tracy_emit_frame_mark_end( const char* name ) { tracy::Profiler::SendFrameMark( name, tracy::QueueType::FrameMarkMsgEnd ); } TRACY_API void ___tracy_emit_frame_image( const void* image, uint16_t w, uint16_t h, uint8_t offset, int flip ) { tracy::Profiler::SendFrameImage( image, w, h, offset, flip ); } TRACY_API void ___tracy_emit_plot( const char* name, double val ) { tracy::Profiler::PlotData( name, val ); } +TRACY_API void ___tracy_emit_plot_float( const char* name, float val ) { tracy::Profiler::PlotData( name, val ); } +TRACY_API void ___tracy_emit_plot_int( const char* name, int64_t val ) { tracy::Profiler::PlotData( name, val ); } TRACY_API void ___tracy_emit_message( const char* txt, size_t size, int callstack ) { tracy::Profiler::Message( txt, size, callstack ); } TRACY_API void ___tracy_emit_messageL( const char* txt, int callstack ) { tracy::Profiler::Message( txt, callstack ); } TRACY_API void ___tracy_emit_messageC( const char* txt, size_t size, uint32_t color, int callstack ) { tracy::Profiler::MessageColor( txt, size, color, callstack ); } diff --git a/src/external/tracy/client/TracyProfiler.hpp b/src/external/tracy/client/TracyProfiler.hpp index 23603535e..1ed66f666 100644 --- a/src/external/tracy/client/TracyProfiler.hpp +++ b/src/external/tracy/client/TracyProfiler.hpp @@ -414,9 +414,9 @@ public: TracyQueuePrepare( callstack == 0 ? QueueType::MessageColor : QueueType::MessageColorCallstack ); MemWrite( &item->messageColorFat.time, GetTime() ); MemWrite( &item->messageColorFat.text, (uint64_t)ptr ); - MemWrite( &item->messageColorFat.r, uint8_t( ( color ) & 0xFF ) ); + MemWrite( &item->messageColorFat.b, uint8_t( ( color ) & 0xFF ) ); MemWrite( &item->messageColorFat.g, uint8_t( ( color >> 8 ) & 0xFF ) ); - MemWrite( &item->messageColorFat.b, uint8_t( ( color >> 16 ) & 0xFF ) ); + MemWrite( &item->messageColorFat.r, uint8_t( ( color >> 16 ) & 0xFF ) ); MemWrite( &item->messageColorFat.size, (uint16_t)size ); TracyQueueCommit( messageColorFatThread ); } @@ -434,9 +434,9 @@ public: TracyQueuePrepare( callstack == 0 ? QueueType::MessageLiteralColor : QueueType::MessageLiteralColorCallstack ); MemWrite( &item->messageColorLiteral.time, GetTime() ); MemWrite( &item->messageColorLiteral.text, (uint64_t)txt ); - MemWrite( &item->messageColorLiteral.r, uint8_t( ( color ) & 0xFF ) ); + MemWrite( &item->messageColorLiteral.b, uint8_t( ( color ) & 0xFF ) ); MemWrite( &item->messageColorLiteral.g, uint8_t( ( color >> 8 ) & 0xFF ) ); - MemWrite( &item->messageColorLiteral.b, uint8_t( ( color >> 16 ) & 0xFF ) ); + MemWrite( &item->messageColorLiteral.r, uint8_t( ( color >> 16 ) & 0xFF ) ); TracyQueueCommit( messageColorLiteralThread ); } diff --git a/src/external/tracy/client/TracyScoped.hpp b/src/external/tracy/client/TracyScoped.hpp index 3c8105e6d..bc1307916 100644 --- a/src/external/tracy/client/TracyScoped.hpp +++ b/src/external/tracy/client/TracyScoped.hpp @@ -143,9 +143,9 @@ public: if( GetProfiler().ConnectionId() != m_connectionId ) return; #endif TracyQueuePrepare( QueueType::ZoneColor ); - MemWrite( &item->zoneColor.r, uint8_t( ( color ) & 0xFF ) ); + MemWrite( &item->zoneColor.b, uint8_t( ( color ) & 0xFF ) ); MemWrite( &item->zoneColor.g, uint8_t( ( color >> 8 ) & 0xFF ) ); - MemWrite( &item->zoneColor.b, uint8_t( ( color >> 16 ) & 0xFF ) ); + MemWrite( &item->zoneColor.r, uint8_t( ( color >> 16 ) & 0xFF ) ); TracyQueueCommit( zoneColorThread ); } @@ -166,7 +166,7 @@ private: const bool m_active; #ifdef TRACY_ON_DEMAND - uint64_t m_connectionId; + uint64_t m_connectionId = 0; #endif }; diff --git a/src/external/tracy/client/TracySysTrace.cpp b/src/external/tracy/client/TracySysTrace.cpp index 23b1020a5..4a562eaae 100644 --- a/src/external/tracy/client/TracySysTrace.cpp +++ b/src/external/tracy/client/TracySysTrace.cpp @@ -409,6 +409,7 @@ bool SysTraceStart( int64_t& samplingPeriod ) return false; } +#ifndef TRACY_NO_SAMPLING if( isOs64Bit ) { CLASSIC_EVENT_ID stackId[2] = {}; @@ -423,6 +424,7 @@ bool SysTraceStart( int64_t& samplingPeriod ) return false; } } +#endif #ifdef UNICODE WCHAR KernelLoggerName[sizeof( KERNEL_LOGGER_NAME )]; @@ -1263,8 +1265,7 @@ void SysTraceWorker( void* ptr ) type = QueueType::HwSampleBranchMiss; break; default: - assert( false ); - break; + abort(); } TracyLfqPrepare( type ); diff --git a/src/external/tracy/client/tracy_concurrentqueue.h b/src/external/tracy/client/tracy_concurrentqueue.h index 3149deb71..4178d39ea 100644 --- a/src/external/tracy/client/tracy_concurrentqueue.h +++ b/src/external/tracy/client/tracy_concurrentqueue.h @@ -210,19 +210,14 @@ namespace details } }; -#ifdef _MSC_VER -#pragma warning(push) -#pragma warning(disable: 4554) -#endif template static inline bool circular_less_than(T a, T b) { static_assert(std::is_integral::value && !std::numeric_limits::is_signed, "circular_less_than is intended to be used only with unsigned integer types"); - return static_cast(a - b) > (static_cast(static_cast(1) << static_cast(sizeof(T) * CHAR_BIT - 1))); + return static_cast(a - b) > static_cast(static_cast(1) << (static_cast(sizeof(T) * CHAR_BIT - 1))); + // Note: extra parens around rhs of operator<< is MSVC bug: https://developercommunity2.visualstudio.com/t/C4554-triggers-when-both-lhs-and-rhs-is/10034931 + // silencing the bug requires #pragma warning(disable: 4554) around the calling code and has no effect when done here. } -#ifdef _MSC_VER -#pragma warning(pop) -#endif template static inline char* align_for(char* ptr) diff --git a/src/external/tracy/common/TracyQueue.hpp b/src/external/tracy/common/TracyQueue.hpp index 431fa484a..8443193af 100644 --- a/src/external/tracy/common/TracyQueue.hpp +++ b/src/external/tracy/common/TracyQueue.hpp @@ -166,9 +166,9 @@ struct QueueZoneValidationThread : public QueueZoneValidation struct QueueZoneColor { - uint8_t r; - uint8_t g; uint8_t b; + uint8_t g; + uint8_t r; }; struct QueueZoneColorThread : public QueueZoneColor @@ -222,9 +222,9 @@ struct QueueSourceLocation uint64_t function; // ptr uint64_t file; // ptr uint32_t line; - uint8_t r; - uint8_t g; uint8_t b; + uint8_t g; + uint8_t r; }; struct QueueZoneTextFat @@ -342,9 +342,9 @@ struct QueueMessage struct QueueMessageColor : public QueueMessage { - uint8_t r; - uint8_t g; uint8_t b; + uint8_t g; + uint8_t r; }; struct QueueMessageLiteral : public QueueMessage diff --git a/src/external/tracy/common/TracySystem.cpp b/src/external/tracy/common/TracySystem.cpp index f14079531..2a7d997e4 100644 --- a/src/external/tracy/common/TracySystem.cpp +++ b/src/external/tracy/common/TracySystem.cpp @@ -205,61 +205,57 @@ TRACY_API const char* GetThreadName( uint32_t id ) } ptr = ptr->next; } -#else -# if defined _WIN32 -# ifdef TRACY_UWP - static auto _GetThreadDescription = &::GetThreadDescription; -# else - static auto _GetThreadDescription = (t_GetThreadDescription)GetProcAddress( GetModuleHandleA( "kernel32.dll" ), "GetThreadDescription" ); -# endif - if( _GetThreadDescription ) - { - auto hnd = OpenThread( THREAD_QUERY_LIMITED_INFORMATION, FALSE, (DWORD)id ); - if( hnd != 0 ) - { - PWSTR tmp; - _GetThreadDescription( hnd, &tmp ); - auto ret = wcstombs( buf, tmp, 256 ); - CloseHandle( hnd ); - if( ret != 0 ) - { - return buf; - } - } - } -# elif defined __linux__ - int cs, fd; - char path[32]; -# ifdef __ANDROID__ - int tid = gettid(); -# else - int tid = (int) syscall( SYS_gettid ); -# endif - snprintf( path, sizeof( path ), "/proc/self/task/%d/comm", tid ); - sprintf( buf, "%" PRIu32, id ); -# ifndef __ANDROID__ - pthread_setcancelstate( PTHREAD_CANCEL_DISABLE, &cs ); -# endif - if ( ( fd = open( path, O_RDONLY ) ) > 0) { - int len = read( fd, buf, 255 ); - if( len > 0 ) - { - buf[len] = 0; - if( len > 1 && buf[len-1] == '\n' ) - { - buf[len-1] = 0; - } - } - close( fd ); - } -# ifndef __ANDROID__ - pthread_setcancelstate( cs, 0 ); -# endif - return buf; -# endif #endif - sprintf( buf, "%" PRIu32, id ); - return buf; + +#if defined _WIN32 +# ifdef TRACY_UWP + static auto _GetThreadDescription = &::GetThreadDescription; +# else + static auto _GetThreadDescription = (t_GetThreadDescription)GetProcAddress( GetModuleHandleA( "kernel32.dll" ), "GetThreadDescription" ); +# endif + if( _GetThreadDescription ) + { + auto hnd = OpenThread( THREAD_QUERY_LIMITED_INFORMATION, FALSE, (DWORD)id ); + if( hnd != 0 ) + { + PWSTR tmp; + _GetThreadDescription( hnd, &tmp ); + auto ret = wcstombs( buf, tmp, 256 ); + CloseHandle( hnd ); + if( ret != 0 ) + { + return buf; + } + } + } +#elif defined __linux__ + int cs, fd; + char path[32]; + snprintf( path, sizeof( path ), "/proc/self/task/%d/comm", id ); + sprintf( buf, "%" PRIu32, id ); +# ifndef __ANDROID__ + pthread_setcancelstate( PTHREAD_CANCEL_DISABLE, &cs ); +# endif + if ( ( fd = open( path, O_RDONLY ) ) > 0) { + int len = read( fd, buf, 255 ); + if( len > 0 ) + { + buf[len] = 0; + if( len > 1 && buf[len-1] == '\n' ) + { + buf[len-1] = 0; + } + } + close( fd ); + } +# ifndef __ANDROID__ + pthread_setcancelstate( cs, 0 ); +# endif + return buf; +#endif + + sprintf( buf, "%" PRIu32, id ); + return buf; } TRACY_API const char* GetEnvVar( const char* name ) diff --git a/src/external/tracy/common/TracyVersion.hpp b/src/external/tracy/common/TracyVersion.hpp index 983d1c51f..c82edf93d 100644 --- a/src/external/tracy/common/TracyVersion.hpp +++ b/src/external/tracy/common/TracyVersion.hpp @@ -7,7 +7,7 @@ namespace Version { enum { Major = 0 }; enum { Minor = 9 }; -enum { Patch = 0 }; +enum { Patch = 1 }; } } diff --git a/src/external/tracy/libbacktrace/elf.cpp b/src/external/tracy/libbacktrace/elf.cpp index 9e62f090d..e6d6c4aee 100644 --- a/src/external/tracy/libbacktrace/elf.cpp +++ b/src/external/tracy/libbacktrace/elf.cpp @@ -193,6 +193,7 @@ dl_iterate_phdr (int (*callback) (struct dl_phdr_info *, #undef STT_FUNC #undef NT_GNU_BUILD_ID #undef ELFCOMPRESS_ZLIB +#undef ELFCOMPRESS_ZSTD /* Basic types. */ @@ -350,6 +351,7 @@ typedef struct #endif /* BACKTRACE_ELF_SIZE != 32 */ #define ELFCOMPRESS_ZLIB 1 +#define ELFCOMPRESS_ZSTD 2 /* Names of sections, indexed by enum dwarf_section in internal.h. */ @@ -1130,7 +1132,7 @@ elf_uncompress_failed(void) on error. */ static int -elf_zlib_fetch (const unsigned char **ppin, const unsigned char *pinend, +elf_fetch_bits (const unsigned char **ppin, const unsigned char *pinend, uint64_t *pval, unsigned int *pbits) { unsigned int bits; @@ -1177,6 +1179,118 @@ elf_zlib_fetch (const unsigned char **ppin, const unsigned char *pinend, return 1; } +/* This is like elf_fetch_bits, but it fetchs the bits backward, and ensures at + least 16 bits. This is for zstd. */ + +static int +elf_fetch_bits_backward (const unsigned char **ppin, + const unsigned char *pinend, + uint64_t *pval, unsigned int *pbits) +{ + unsigned int bits; + const unsigned char *pin; + uint64_t val; + uint32_t next; + + bits = *pbits; + if (bits >= 16) + return 1; + pin = *ppin; + val = *pval; + + if (unlikely (pin <= pinend)) + { + if (bits == 0) + { + elf_uncompress_failed (); + return 0; + } + return 1; + } + + pin -= 4; + +#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) \ + && defined(__ORDER_BIG_ENDIAN__) \ + && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ \ + || __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) + /* We've ensured that PIN is aligned. */ + next = *(const uint32_t *)pin; + +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + next = __builtin_bswap32 (next); +#endif +#else + next = pin[0] | (pin[1] << 8) | (pin[2] << 16) | (pin[3] << 24); +#endif + + val <<= 32; + val |= next; + bits += 32; + + if (unlikely (pin < pinend)) + { + val >>= (pinend - pin) * 8; + bits -= (pinend - pin) * 8; + } + + *ppin = pin; + *pval = val; + *pbits = bits; + return 1; +} + +/* Initialize backward fetching when the bitstream starts with a 1 bit in the + last byte in memory (which is the first one that we read). This is used by + zstd decompression. Returns 1 on success, 0 on error. */ + +static int +elf_fetch_backward_init (const unsigned char **ppin, + const unsigned char *pinend, + uint64_t *pval, unsigned int *pbits) +{ + const unsigned char *pin; + unsigned int stream_start; + uint64_t val; + unsigned int bits; + + pin = *ppin; + stream_start = (unsigned int)*pin; + if (unlikely (stream_start == 0)) + { + elf_uncompress_failed (); + return 0; + } + val = 0; + bits = 0; + + /* Align to a 32-bit boundary. */ + while ((((uintptr_t)pin) & 3) != 0) + { + val <<= 8; + val |= (uint64_t)*pin; + bits += 8; + --pin; + } + + val <<= 8; + val |= (uint64_t)*pin; + bits += 8; + + *ppin = pin; + *pval = val; + *pbits = bits; + if (!elf_fetch_bits_backward (ppin, pinend, pval, pbits)) + return 0; + + *pbits -= __builtin_clz (stream_start) - (sizeof (unsigned int) - 1) * 8 + 1; + + if (!elf_fetch_bits_backward (ppin, pinend, pval, pbits)) + return 0; + + return 1; +} + /* Huffman code tables, like the rest of the zlib format, are defined by RFC 1951. We store a Huffman code table as a series of tables stored sequentially in memory. Each entry in a table is 16 bits. @@ -1211,14 +1325,14 @@ elf_zlib_fetch (const unsigned char **ppin, const unsigned char *pinend, /* Number of entries we allocate to for one code table. We get a page for the two code tables we need. */ -#define HUFFMAN_TABLE_SIZE (1024) +#define ZLIB_HUFFMAN_TABLE_SIZE (1024) /* Bit masks and shifts for the values in the table. */ -#define HUFFMAN_VALUE_MASK 0x01ff -#define HUFFMAN_BITS_SHIFT 9 -#define HUFFMAN_BITS_MASK 0x7 -#define HUFFMAN_SECONDARY_SHIFT 12 +#define ZLIB_HUFFMAN_VALUE_MASK 0x01ff +#define ZLIB_HUFFMAN_BITS_SHIFT 9 +#define ZLIB_HUFFMAN_BITS_MASK 0x7 +#define ZLIB_HUFFMAN_SECONDARY_SHIFT 12 /* For working memory while inflating we need two code tables, we need an array of code lengths (max value 15, so we use unsigned char), @@ -1226,17 +1340,17 @@ elf_zlib_fetch (const unsigned char **ppin, const unsigned char *pinend, latter two arrays must be large enough to hold the maximum number of code lengths, which RFC 1951 defines as 286 + 30. */ -#define ZDEBUG_TABLE_SIZE \ - (2 * HUFFMAN_TABLE_SIZE * sizeof (uint16_t) \ +#define ZLIB_TABLE_SIZE \ + (2 * ZLIB_HUFFMAN_TABLE_SIZE * sizeof (uint16_t) \ + (286 + 30) * sizeof (uint16_t) \ + (286 + 30) * sizeof (unsigned char)) -#define ZDEBUG_TABLE_CODELEN_OFFSET \ - (2 * HUFFMAN_TABLE_SIZE * sizeof (uint16_t) \ +#define ZLIB_TABLE_CODELEN_OFFSET \ + (2 * ZLIB_HUFFMAN_TABLE_SIZE * sizeof (uint16_t) \ + (286 + 30) * sizeof (uint16_t)) -#define ZDEBUG_TABLE_WORK_OFFSET \ - (2 * HUFFMAN_TABLE_SIZE * sizeof (uint16_t)) +#define ZLIB_TABLE_WORK_OFFSET \ + (2 * ZLIB_HUFFMAN_TABLE_SIZE * sizeof (uint16_t)) #ifdef BACKTRACE_GENERATE_FIXED_HUFFMAN_TABLE @@ -1269,7 +1383,7 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, next value after VAL with the same bit length. */ next = (uint16_t *) (((unsigned char *) zdebug_table) - + ZDEBUG_TABLE_WORK_OFFSET); + + ZLIB_TABLE_WORK_OFFSET); memset (&count[0], 0, 16 * sizeof (uint16_t)); for (i = 0; i < codes_len; ++i) @@ -1297,7 +1411,7 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, /* For each length, fill in the table for the codes of that length. */ - memset (table, 0, HUFFMAN_TABLE_SIZE * sizeof (uint16_t)); + memset (table, 0, ZLIB_HUFFMAN_TABLE_SIZE * sizeof (uint16_t)); /* Handle the values that do not require a secondary table. */ @@ -1331,13 +1445,13 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, /* In the compressed bit stream, the value VAL is encoded as J bits with the value C. */ - if (unlikely ((val & ~HUFFMAN_VALUE_MASK) != 0)) + if (unlikely ((val & ~ZLIB_HUFFMAN_VALUE_MASK) != 0)) { elf_uncompress_failed (); return 0; } - tval = val | ((j - 1) << HUFFMAN_BITS_SHIFT); + tval = val | ((j - 1) << ZLIB_HUFFMAN_BITS_SHIFT); /* The table lookup uses 8 bits. If J is less than 8, we don't know what the other bits will be. We need to fill @@ -1487,7 +1601,7 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, { /* Start a new secondary table. */ - if (unlikely ((next_secondary & HUFFMAN_VALUE_MASK) + if (unlikely ((next_secondary & ZLIB_HUFFMAN_VALUE_MASK) != next_secondary)) { elf_uncompress_failed (); @@ -1498,22 +1612,23 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, secondary_bits = j - 8; next_secondary += 1 << secondary_bits; table[primary] = (secondary - + ((j - 8) << HUFFMAN_BITS_SHIFT) - + (1U << HUFFMAN_SECONDARY_SHIFT)); + + ((j - 8) << ZLIB_HUFFMAN_BITS_SHIFT) + + (1U << ZLIB_HUFFMAN_SECONDARY_SHIFT)); } else { /* There is an existing entry. It had better be a secondary table with enough bits. */ - if (unlikely ((tprimary & (1U << HUFFMAN_SECONDARY_SHIFT)) + if (unlikely ((tprimary + & (1U << ZLIB_HUFFMAN_SECONDARY_SHIFT)) == 0)) { elf_uncompress_failed (); return 0; } - secondary = tprimary & HUFFMAN_VALUE_MASK; - secondary_bits = ((tprimary >> HUFFMAN_BITS_SHIFT) - & HUFFMAN_BITS_MASK); + secondary = tprimary & ZLIB_HUFFMAN_VALUE_MASK; + secondary_bits = ((tprimary >> ZLIB_HUFFMAN_BITS_SHIFT) + & ZLIB_HUFFMAN_BITS_MASK); if (unlikely (secondary_bits < j - 8)) { elf_uncompress_failed (); @@ -1524,7 +1639,7 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, /* Fill in secondary table entries. */ - tval = val | ((j - 8) << HUFFMAN_BITS_SHIFT); + tval = val | ((j - 8) << ZLIB_HUFFMAN_BITS_SHIFT); for (ind = code >> 8; ind < (1U << secondary_bits); @@ -1567,7 +1682,7 @@ elf_zlib_inflate_table (unsigned char *codes, size_t codes_len, #include -static uint16_t table[ZDEBUG_TABLE_SIZE]; +static uint16_t table[ZLIB_TABLE_SIZE]; static unsigned char codes[288]; int @@ -1795,7 +1910,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, const uint16_t *tlit; const uint16_t *tdist; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; last = val & 1; @@ -1883,7 +1998,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, /* Read a Huffman encoding table. The various magic numbers here are from RFC 1951. */ - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; nlit = (val & 0x1f) + 257; @@ -1908,7 +2023,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, /* There are always at least 4 elements in the table. */ - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; codebits[16] = val & 7; @@ -1928,7 +2043,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, if (nclen == 5) goto codebitsdone; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; codebits[7] = val & 7; @@ -1966,7 +2081,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, if (nclen == 10) goto codebitsdone; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; codebits[11] = val & 7; @@ -2004,7 +2119,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, if (nclen == 15) goto codebitsdone; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; codebits[2] = val & 7; @@ -2043,7 +2158,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, at the end of zdebug_table to hold them. */ plenbase = (((unsigned char *) zdebug_table) - + ZDEBUG_TABLE_CODELEN_OFFSET); + + ZLIB_TABLE_CODELEN_OFFSET); plen = plenbase; plenend = plen + nlit + ndist; while (plen < plenend) @@ -2052,24 +2167,25 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, unsigned int b; uint16_t v; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; t = zdebug_table[val & 0xff]; /* The compression here uses bit lengths up to 7, so a secondary table is never necessary. */ - if (unlikely ((t & (1U << HUFFMAN_SECONDARY_SHIFT)) != 0)) + if (unlikely ((t & (1U << ZLIB_HUFFMAN_SECONDARY_SHIFT)) + != 0)) { elf_uncompress_failed (); return 0; } - b = (t >> HUFFMAN_BITS_SHIFT) & HUFFMAN_BITS_MASK; + b = (t >> ZLIB_HUFFMAN_BITS_SHIFT) & ZLIB_HUFFMAN_BITS_MASK; val >>= b + 1; bits -= b + 1; - v = t & HUFFMAN_VALUE_MASK; + v = t & ZLIB_HUFFMAN_VALUE_MASK; if (v < 16) *plen++ = v; else if (v == 16) @@ -2086,7 +2202,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, } /* We used up to 7 bits since the last - elf_zlib_fetch, so we have at least 8 bits + elf_fetch_bits, so we have at least 8 bits available here. */ c = 3 + (val & 0x3); @@ -2121,7 +2237,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, /* Store zero 3 to 10 times. */ /* We used up to 7 bits since the last - elf_zlib_fetch, so we have at least 8 bits + elf_fetch_bits, so we have at least 8 bits available here. */ c = 3 + (val & 0x7); @@ -2167,7 +2283,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, /* Store zero 11 to 138 times. */ /* We used up to 7 bits since the last - elf_zlib_fetch, so we have at least 8 bits + elf_fetch_bits, so we have at least 8 bits available here. */ c = 11 + (val & 0x7f); @@ -2204,10 +2320,11 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, zdebug_table)) return 0; if (!elf_zlib_inflate_table (plen + nlit, ndist, zdebug_table, - zdebug_table + HUFFMAN_TABLE_SIZE)) + (zdebug_table + + ZLIB_HUFFMAN_TABLE_SIZE))) return 0; tlit = zdebug_table; - tdist = zdebug_table + HUFFMAN_TABLE_SIZE; + tdist = zdebug_table + ZLIB_HUFFMAN_TABLE_SIZE; } /* Inflate values until the end of the block. This is the @@ -2220,14 +2337,14 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, uint16_t v; unsigned int lit; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; t = tlit[val & 0xff]; - b = (t >> HUFFMAN_BITS_SHIFT) & HUFFMAN_BITS_MASK; - v = t & HUFFMAN_VALUE_MASK; + b = (t >> ZLIB_HUFFMAN_BITS_SHIFT) & ZLIB_HUFFMAN_BITS_MASK; + v = t & ZLIB_HUFFMAN_VALUE_MASK; - if ((t & (1U << HUFFMAN_SECONDARY_SHIFT)) == 0) + if ((t & (1U << ZLIB_HUFFMAN_SECONDARY_SHIFT)) == 0) { lit = v; val >>= b + 1; @@ -2236,8 +2353,8 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, else { t = tlit[v + 0x100 + ((val >> 8) & ((1U << b) - 1))]; - b = (t >> HUFFMAN_BITS_SHIFT) & HUFFMAN_BITS_MASK; - lit = t & HUFFMAN_VALUE_MASK; + b = (t >> ZLIB_HUFFMAN_BITS_SHIFT) & ZLIB_HUFFMAN_BITS_MASK; + lit = t & ZLIB_HUFFMAN_VALUE_MASK; val >>= b + 8; bits -= b + 8; } @@ -2282,7 +2399,7 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, { unsigned int extra; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; /* This is an expression for the table of length @@ -2297,14 +2414,14 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, bits -= extra; } - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) return 0; t = tdist[val & 0xff]; - b = (t >> HUFFMAN_BITS_SHIFT) & HUFFMAN_BITS_MASK; - v = t & HUFFMAN_VALUE_MASK; + b = (t >> ZLIB_HUFFMAN_BITS_SHIFT) & ZLIB_HUFFMAN_BITS_MASK; + v = t & ZLIB_HUFFMAN_VALUE_MASK; - if ((t & (1U << HUFFMAN_SECONDARY_SHIFT)) == 0) + if ((t & (1U << ZLIB_HUFFMAN_SECONDARY_SHIFT)) == 0) { dist = v; val >>= b + 1; @@ -2313,8 +2430,9 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, else { t = tdist[v + 0x100 + ((val >> 8) & ((1U << b) - 1))]; - b = (t >> HUFFMAN_BITS_SHIFT) & HUFFMAN_BITS_MASK; - dist = t & HUFFMAN_VALUE_MASK; + b = ((t >> ZLIB_HUFFMAN_BITS_SHIFT) + & ZLIB_HUFFMAN_BITS_MASK); + dist = t & ZLIB_HUFFMAN_VALUE_MASK; val >>= b + 8; bits -= b + 8; } @@ -2326,216 +2444,2576 @@ elf_zlib_inflate (const unsigned char *pin, size_t sin, uint16_t *zdebug_table, /* A distance of 1. A common case, meaning repeat the last character LEN times. */ - if (unlikely (pout == porigout)) - { - elf_uncompress_failed (); - return 0; - } + if (unlikely (pout == porigout)) + { + elf_uncompress_failed (); + return 0; + } + + if (unlikely ((unsigned int) (poutend - pout) < len)) + { + elf_uncompress_failed (); + return 0; + } + + memset (pout, pout[-1], len); + pout += len; + } + else if (unlikely (dist > 29)) + { + elf_uncompress_failed (); + return 0; + } + else + { + if (dist < 4) + dist = dist + 1; + else + { + unsigned int extra; + + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) + return 0; + + /* This is an expression for the table of + distance codes in RFC 1951 3.2.5. */ + dist -= 4; + extra = (dist >> 1) + 1; + dist = (dist & 1) << extra; + dist += 5; + dist += ((1U << (extra - 1)) - 1) << 2; + dist += val & ((1U << extra) - 1); + val >>= extra; + bits -= extra; + } + + /* Go back dist bytes, and copy len bytes from + there. */ + + if (unlikely ((unsigned int) (pout - porigout) < dist)) + { + elf_uncompress_failed (); + return 0; + } + + if (unlikely ((unsigned int) (poutend - pout) < len)) + { + elf_uncompress_failed (); + return 0; + } + + if (dist >= len) + { + memcpy (pout, pout - dist, len); + pout += len; + } + else + { + while (len > 0) + { + unsigned int copy; + + copy = len < dist ? len : dist; + memcpy (pout, pout - dist, copy); + len -= copy; + pout += copy; + } + } + } + } + } + } + } + + /* We should have filled the output buffer. */ + if (unlikely (pout != poutend)) + { + elf_uncompress_failed (); + return 0; + } + + return 1; +} + +/* Verify the zlib checksum. The checksum is in the 4 bytes at + CHECKBYTES, and the uncompressed data is at UNCOMPRESSED / + UNCOMPRESSED_SIZE. Returns 1 on success, 0 on failure. */ + +static int +elf_zlib_verify_checksum (const unsigned char *checkbytes, + const unsigned char *uncompressed, + size_t uncompressed_size) +{ + unsigned int i; + unsigned int cksum; + const unsigned char *p; + uint32_t s1; + uint32_t s2; + size_t hsz; + + cksum = 0; + for (i = 0; i < 4; i++) + cksum = (cksum << 8) | checkbytes[i]; + + s1 = 1; + s2 = 0; + + /* Minimize modulo operations. */ + + p = uncompressed; + hsz = uncompressed_size; + while (hsz >= 5552) + { + for (i = 0; i < 5552; i += 16) + { + /* Manually unroll loop 16 times. */ + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + } + hsz -= 5552; + s1 %= 65521; + s2 %= 65521; + } + + while (hsz >= 16) + { + /* Manually unroll loop 16 times. */ + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + s1 = s1 + *p++; + s2 = s2 + s1; + + hsz -= 16; + } + + for (i = 0; i < hsz; ++i) + { + s1 = s1 + *p++; + s2 = s2 + s1; + } + + s1 %= 65521; + s2 %= 65521; + + if (unlikely ((s2 << 16) + s1 != cksum)) + { + elf_uncompress_failed (); + return 0; + } + + return 1; +} + +/* Inflate a zlib stream from PIN/SIN to POUT/SOUT, and verify the + checksum. Return 1 on success, 0 on error. */ + +static int +elf_zlib_inflate_and_verify (const unsigned char *pin, size_t sin, + uint16_t *zdebug_table, unsigned char *pout, + size_t sout) +{ + if (!elf_zlib_inflate (pin, sin, zdebug_table, pout, sout)) + return 0; + if (!elf_zlib_verify_checksum (pin + sin - 4, pout, sout)) + return 0; + return 1; +} + +/* For working memory during zstd compression, we need + - a literal length FSE table: 512 64-bit values == 4096 bytes + - a match length FSE table: 512 64-bit values == 4096 bytes + - a offset FSE table: 256 64-bit values == 2048 bytes + - a Huffman tree: 2048 uint16_t values == 4096 bytes + - scratch space, one of + - to build an FSE table: 512 uint16_t values == 1024 bytes + - to build a Huffman tree: 512 uint16_t + 256 uint32_t == 2048 bytes +*/ + +#define ZSTD_TABLE_SIZE \ + (2 * 512 * sizeof (struct elf_zstd_fse_baseline_entry) \ + + 256 * sizeof (struct elf_zstd_fse_baseline_entry) \ + + 2048 * sizeof (uint16_t) \ + + 512 * sizeof (uint16_t) + 256 * sizeof (uint32_t)) + +#define ZSTD_TABLE_LITERAL_FSE_OFFSET (0) + +#define ZSTD_TABLE_MATCH_FSE_OFFSET \ + (512 * sizeof (struct elf_zstd_fse_baseline_entry)) + +#define ZSTD_TABLE_OFFSET_FSE_OFFSET \ + (ZSTD_TABLE_MATCH_FSE_OFFSET \ + + 512 * sizeof (struct elf_zstd_fse_baseline_entry)) + +#define ZSTD_TABLE_HUFFMAN_OFFSET \ + (ZSTD_TABLE_OFFSET_FSE_OFFSET \ + + 256 * sizeof (struct elf_zstd_fse_baseline_entry)) + +#define ZSTD_TABLE_WORK_OFFSET \ + (ZSTD_TABLE_HUFFMAN_OFFSET + 2048 * sizeof (uint16_t)) + +/* An entry in a zstd FSE table. */ + +struct elf_zstd_fse_entry +{ + /* The value that this FSE entry represents. */ + unsigned char symbol; + /* The number of bits to read to determine the next state. */ + unsigned char bits; + /* Add the bits to this base to get the next state. */ + uint16_t base; +}; + +static int +elf_zstd_build_fse (const int16_t *, int, uint16_t *, int, + struct elf_zstd_fse_entry *); + +/* Read a zstd FSE table and build the decoding table in *TABLE, updating *PPIN + as it reads. ZDEBUG_TABLE is scratch space; it must be enough for 512 + uint16_t values (1024 bytes). MAXIDX is the maximum number of symbols + permitted. *TABLE_BITS is the maximum number of bits for symbols in the + table: the size of *TABLE is at least 1 << *TABLE_BITS. This updates + *TABLE_BITS to the actual number of bits. Returns 1 on success, 0 on + error. */ + +static int +elf_zstd_read_fse (const unsigned char **ppin, const unsigned char *pinend, + uint16_t *zdebug_table, int maxidx, + struct elf_zstd_fse_entry *table, int *table_bits) +{ + const unsigned char *pin; + int16_t *norm; + uint16_t *next; + uint64_t val; + unsigned int bits; + int accuracy_log; + uint32_t remaining; + uint32_t threshold; + int bits_needed; + int idx; + int prev0; + + pin = *ppin; + + norm = (int16_t *) zdebug_table; + next = zdebug_table + 256; + + if (unlikely (pin + 3 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + + /* Align PIN to a 32-bit boundary. */ + + val = 0; + bits = 0; + while ((((uintptr_t) pin) & 3) != 0) + { + val |= (uint64_t)*pin << bits; + bits += 8; + ++pin; + } + + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) + return 0; + + accuracy_log = (val & 0xf) + 5; + if (accuracy_log > *table_bits) + { + elf_uncompress_failed (); + return 0; + } + *table_bits = accuracy_log; + val >>= 4; + bits -= 4; + + /* This code is mostly copied from the reference implementation. */ + + /* The number of remaining probabilities, plus 1. This sets the number of + bits that need to be read for the next value. */ + remaining = (1 << accuracy_log) + 1; + + /* The current difference between small and large values, which depends on + the number of remaining values. Small values use one less bit. */ + threshold = 1 << accuracy_log; + + /* The number of bits used to compute threshold. */ + bits_needed = accuracy_log + 1; + + /* The next character value. */ + idx = 0; + + /* Whether the last count was 0. */ + prev0 = 0; + + while (remaining > 1 && idx <= maxidx) + { + uint32_t max; + int32_t count; + + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) + return 0; + + if (prev0) + { + int zidx; + + /* Previous count was 0, so there is a 2-bit repeat flag. If the + 2-bit flag is 0b11, it adds 3 and then there is another repeat + flag. */ + zidx = idx; + while ((val & 0xfff) == 0xfff) + { + zidx += 3 * 6; + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) + return 0; + val >>= 12; + bits -= 12; + } + while ((val & 3) == 3) + { + zidx += 3; + if (!elf_fetch_bits (&pin, pinend, &val, &bits)) + return 0; + val >>= 2; + bits -= 2; + } + /* We have at least 13 bits here, don't need to fetch. */ + zidx += val & 3; + val >>= 2; + bits -= 2; + + if (unlikely (zidx > maxidx)) + { + elf_uncompress_failed (); + return 0; + } + + for (; idx < zidx; idx++) + norm[idx] = 0; + + prev0 = 0; + continue; + } + + max = (2 * threshold - 1) - remaining; + if ((val & (threshold - 1)) < max) + { + /* A small value. */ + count = (int32_t) ((uint32_t) val & (threshold - 1)); + val >>= bits_needed - 1; + bits -= bits_needed - 1; + } + else + { + /* A large value. */ + count = (int32_t) ((uint32_t) val & (2 * threshold - 1)); + if (count >= (int32_t) threshold) + count -= (int32_t) max; + val >>= bits_needed; + bits -= bits_needed; + } + + count--; + if (count >= 0) + remaining -= count; + else + remaining--; + if (unlikely (idx >= 256)) + { + elf_uncompress_failed (); + return 0; + } + norm[idx] = (int16_t) count; + ++idx; + + prev0 = count == 0; + + while (remaining < threshold) + { + bits_needed--; + threshold >>= 1; + } + } + + if (unlikely (remaining != 1)) + { + elf_uncompress_failed (); + return 0; + } + + /* If we've read ahead more than a byte, back up. */ + while (bits >= 8) + { + --pin; + bits -= 8; + } + + *ppin = pin; + + for (; idx <= maxidx; idx++) + norm[idx] = 0; + + return elf_zstd_build_fse (norm, idx, next, *table_bits, table); +} + +/* Build the FSE decoding table from a list of probabilities. This reads from + NORM of length IDX, uses NEXT as scratch space, and writes to *TABLE, whose + size is TABLE_BITS. */ + +static int +elf_zstd_build_fse (const int16_t *norm, int idx, uint16_t *next, + int table_bits, struct elf_zstd_fse_entry *table) +{ + int table_size; + int high_threshold; + int i; + int pos; + int step; + int mask; + + table_size = 1 << table_bits; + high_threshold = table_size - 1; + for (i = 0; i < idx; i++) + { + int16_t n; + + n = norm[i]; + if (n >= 0) + next[i] = (uint16_t) n; + else + { + table[high_threshold].symbol = (unsigned char) i; + high_threshold--; + next[i] = 1; + } + } + + pos = 0; + step = (table_size >> 1) + (table_size >> 3) + 3; + mask = table_size - 1; + for (i = 0; i < idx; i++) + { + int n; + int j; + + n = (int) norm[i]; + for (j = 0; j < n; j++) + { + table[pos].symbol = (unsigned char) i; + pos = (pos + step) & mask; + while (unlikely (pos > high_threshold)) + pos = (pos + step) & mask; + } + } + if (pos != 0) + { + elf_uncompress_failed (); + return 0; + } + + for (i = 0; i < table_size; i++) + { + unsigned char sym; + uint16_t next_state; + int high_bit; + int bits; + + sym = table[i].symbol; + next_state = next[sym]; + ++next[sym]; + + if (next_state == 0) + { + elf_uncompress_failed (); + return 0; + } + high_bit = 31 - __builtin_clz (next_state); + + bits = table_bits - high_bit; + table[i].bits = (unsigned char) bits; + table[i].base = (uint16_t) ((next_state << bits) - table_size); + } + + return 1; +} + +/* Encode the baseline and bits into a single 32-bit value. */ + +#define ZSTD_ENCODE_BASELINE_BITS(baseline, basebits) \ + ((uint32_t)(baseline) | ((uint32_t)(basebits) << 24)) + +#define ZSTD_DECODE_BASELINE(baseline_basebits) \ + ((uint32_t)(baseline_basebits) & 0xffffff) + +#define ZSTD_DECODE_BASEBITS(baseline_basebits) \ + ((uint32_t)(baseline_basebits) >> 24) + +/* Given a literal length code, we need to read a number of bits and add that + to a baseline. For states 0 to 15 the baseline is the state and the number + of bits is zero. */ + +#define ZSTD_LITERAL_LENGTH_BASELINE_OFFSET (16) + +static const uint32_t elf_zstd_literal_length_base[] = +{ + ZSTD_ENCODE_BASELINE_BITS(16, 1), + ZSTD_ENCODE_BASELINE_BITS(18, 1), + ZSTD_ENCODE_BASELINE_BITS(20, 1), + ZSTD_ENCODE_BASELINE_BITS(22, 1), + ZSTD_ENCODE_BASELINE_BITS(24, 2), + ZSTD_ENCODE_BASELINE_BITS(28, 2), + ZSTD_ENCODE_BASELINE_BITS(32, 3), + ZSTD_ENCODE_BASELINE_BITS(40, 3), + ZSTD_ENCODE_BASELINE_BITS(48, 4), + ZSTD_ENCODE_BASELINE_BITS(64, 6), + ZSTD_ENCODE_BASELINE_BITS(128, 7), + ZSTD_ENCODE_BASELINE_BITS(256, 8), + ZSTD_ENCODE_BASELINE_BITS(512, 9), + ZSTD_ENCODE_BASELINE_BITS(1024, 10), + ZSTD_ENCODE_BASELINE_BITS(2048, 11), + ZSTD_ENCODE_BASELINE_BITS(4096, 12), + ZSTD_ENCODE_BASELINE_BITS(8192, 13), + ZSTD_ENCODE_BASELINE_BITS(16384, 14), + ZSTD_ENCODE_BASELINE_BITS(32768, 15), + ZSTD_ENCODE_BASELINE_BITS(65536, 16) +}; + +/* The same applies to match length codes. For states 0 to 31 the baseline is + the state + 3 and the number of bits is zero. */ + +#define ZSTD_MATCH_LENGTH_BASELINE_OFFSET (32) + +static const uint32_t elf_zstd_match_length_base[] = +{ + ZSTD_ENCODE_BASELINE_BITS(35, 1), + ZSTD_ENCODE_BASELINE_BITS(37, 1), + ZSTD_ENCODE_BASELINE_BITS(39, 1), + ZSTD_ENCODE_BASELINE_BITS(41, 1), + ZSTD_ENCODE_BASELINE_BITS(43, 2), + ZSTD_ENCODE_BASELINE_BITS(47, 2), + ZSTD_ENCODE_BASELINE_BITS(51, 3), + ZSTD_ENCODE_BASELINE_BITS(59, 3), + ZSTD_ENCODE_BASELINE_BITS(67, 4), + ZSTD_ENCODE_BASELINE_BITS(83, 4), + ZSTD_ENCODE_BASELINE_BITS(99, 5), + ZSTD_ENCODE_BASELINE_BITS(131, 7), + ZSTD_ENCODE_BASELINE_BITS(259, 8), + ZSTD_ENCODE_BASELINE_BITS(515, 9), + ZSTD_ENCODE_BASELINE_BITS(1027, 10), + ZSTD_ENCODE_BASELINE_BITS(2051, 11), + ZSTD_ENCODE_BASELINE_BITS(4099, 12), + ZSTD_ENCODE_BASELINE_BITS(8195, 13), + ZSTD_ENCODE_BASELINE_BITS(16387, 14), + ZSTD_ENCODE_BASELINE_BITS(32771, 15), + ZSTD_ENCODE_BASELINE_BITS(65539, 16) +}; + +/* An entry in an FSE table used for literal/match/length values. For these we + have to map the symbol to a baseline value, and we have to read zero or more + bits and add that value to the baseline value. Rather than look the values + up in a separate table, we grow the FSE table so that we get better memory + caching. */ + +struct elf_zstd_fse_baseline_entry +{ + /* The baseline for the value that this FSE entry represents.. */ + uint32_t baseline; + /* The number of bits to read to add to the baseline. */ + unsigned char basebits; + /* The number of bits to read to determine the next state. */ + unsigned char bits; + /* Add the bits to this base to get the next state. */ + uint16_t base; +}; + +/* Convert the literal length FSE table FSE_TABLE to an FSE baseline table at + BASELINE_TABLE. Note that FSE_TABLE and BASELINE_TABLE will overlap. */ + +static int +elf_zstd_make_literal_baseline_fse ( + const struct elf_zstd_fse_entry *fse_table, + int table_bits, + struct elf_zstd_fse_baseline_entry *baseline_table) +{ + size_t count; + const struct elf_zstd_fse_entry *pfse; + struct elf_zstd_fse_baseline_entry *pbaseline; + + /* Convert backward to avoid overlap. */ + + count = 1U << table_bits; + pfse = fse_table + count; + pbaseline = baseline_table + count; + while (pfse > fse_table) + { + unsigned char symbol; + unsigned char bits; + uint16_t base; + + --pfse; + --pbaseline; + symbol = pfse->symbol; + bits = pfse->bits; + base = pfse->base; + if (symbol < ZSTD_LITERAL_LENGTH_BASELINE_OFFSET) + { + pbaseline->baseline = (uint32_t)symbol; + pbaseline->basebits = 0; + } + else + { + unsigned int idx; + uint32_t basebits; + + if (unlikely (symbol > 35)) + { + elf_uncompress_failed (); + return 0; + } + idx = symbol - ZSTD_LITERAL_LENGTH_BASELINE_OFFSET; + basebits = elf_zstd_literal_length_base[idx]; + pbaseline->baseline = ZSTD_DECODE_BASELINE(basebits); + pbaseline->basebits = ZSTD_DECODE_BASEBITS(basebits); + } + pbaseline->bits = bits; + pbaseline->base = base; + } + + return 1; +} + +/* Convert the offset length FSE table FSE_TABLE to an FSE baseline table at + BASELINE_TABLE. Note that FSE_TABLE and BASELINE_TABLE will overlap. */ + +static int +elf_zstd_make_offset_baseline_fse ( + const struct elf_zstd_fse_entry *fse_table, + int table_bits, + struct elf_zstd_fse_baseline_entry *baseline_table) +{ + size_t count; + const struct elf_zstd_fse_entry *pfse; + struct elf_zstd_fse_baseline_entry *pbaseline; + + /* Convert backward to avoid overlap. */ + + count = 1U << table_bits; + pfse = fse_table + count; + pbaseline = baseline_table + count; + while (pfse > fse_table) + { + unsigned char symbol; + unsigned char bits; + uint16_t base; + + --pfse; + --pbaseline; + symbol = pfse->symbol; + bits = pfse->bits; + base = pfse->base; + if (unlikely (symbol > 31)) + { + elf_uncompress_failed (); + return 0; + } + + /* The simple way to write this is + + pbaseline->baseline = (uint32_t)1 << symbol; + pbaseline->basebits = symbol; + + That will give us an offset value that corresponds to the one + described in the RFC. However, for offset values > 3, we have to + subtract 3. And for offset values 1, 2, 3 we use a repeated offset. + The baseline is always a power of 2, and is never 0, so for these low + values we will see one entry that is baseline 1, basebits 0, and one + entry that is baseline 2, basebits 1. All other entries will have + baseline >= 4 and basebits >= 2. + + So we can check for RFC offset <= 3 by checking for basebits <= 1. + And that means that we can subtract 3 here and not worry about doing + it in the hot loop. */ + + pbaseline->baseline = (uint32_t)1 << symbol; + if (symbol >= 2) + pbaseline->baseline -= 3; + pbaseline->basebits = symbol; + pbaseline->bits = bits; + pbaseline->base = base; + } + + return 1; +} + +/* Convert the match length FSE table FSE_TABLE to an FSE baseline table at + BASELINE_TABLE. Note that FSE_TABLE and BASELINE_TABLE will overlap. */ + +static int +elf_zstd_make_match_baseline_fse ( + const struct elf_zstd_fse_entry *fse_table, + int table_bits, + struct elf_zstd_fse_baseline_entry *baseline_table) +{ + size_t count; + const struct elf_zstd_fse_entry *pfse; + struct elf_zstd_fse_baseline_entry *pbaseline; + + /* Convert backward to avoid overlap. */ + + count = 1U << table_bits; + pfse = fse_table + count; + pbaseline = baseline_table + count; + while (pfse > fse_table) + { + unsigned char symbol; + unsigned char bits; + uint16_t base; + + --pfse; + --pbaseline; + symbol = pfse->symbol; + bits = pfse->bits; + base = pfse->base; + if (symbol < ZSTD_MATCH_LENGTH_BASELINE_OFFSET) + { + pbaseline->baseline = (uint32_t)symbol + 3; + pbaseline->basebits = 0; + } + else + { + unsigned int idx; + uint32_t basebits; + + if (unlikely (symbol > 52)) + { + elf_uncompress_failed (); + return 0; + } + idx = symbol - ZSTD_MATCH_LENGTH_BASELINE_OFFSET; + basebits = elf_zstd_match_length_base[idx]; + pbaseline->baseline = ZSTD_DECODE_BASELINE(basebits); + pbaseline->basebits = ZSTD_DECODE_BASEBITS(basebits); + } + pbaseline->bits = bits; + pbaseline->base = base; + } + + return 1; +} + +#ifdef BACKTRACE_GENERATE_ZSTD_FSE_TABLES + +/* Used to generate the predefined FSE decoding tables for zstd. */ + +#include + +/* These values are straight from RFC 8878. */ + +static int16_t lit[36] = +{ + 4, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, + 2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 2, 1, 1, 1, 1, 1, + -1,-1,-1,-1 +}; + +static int16_t match[53] = +{ + 1, 4, 3, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,-1,-1, + -1,-1,-1,-1,-1 +}; + +static int16_t offset[29] = +{ + 1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1,-1,-1,-1,-1,-1 +}; + +static uint16_t next[256]; + +static void +print_table (const struct elf_zstd_fse_baseline_entry *table, size_t size) +{ + size_t i; + + printf ("{\n"); + for (i = 0; i < size; i += 3) + { + int j; + + printf (" "); + for (j = 0; j < 3 && i + j < size; ++j) + printf (" { %u, %d, %d, %d },", table[i + j].baseline, + table[i + j].basebits, table[i + j].bits, + table[i + j].base); + printf ("\n"); + } + printf ("};\n"); +} + +int +main () +{ + struct elf_zstd_fse_entry lit_table[64]; + struct elf_zstd_fse_baseline_entry lit_baseline[64]; + struct elf_zstd_fse_entry match_table[64]; + struct elf_zstd_fse_baseline_entry match_baseline[64]; + struct elf_zstd_fse_entry offset_table[32]; + struct elf_zstd_fse_baseline_entry offset_baseline[32]; + + if (!elf_zstd_build_fse (lit, sizeof lit / sizeof lit[0], next, + 6, lit_table)) + { + fprintf (stderr, "elf_zstd_build_fse failed\n"); + exit (EXIT_FAILURE); + } + + if (!elf_zstd_make_literal_baseline_fse (lit_table, 6, lit_baseline)) + { + fprintf (stderr, "elf_zstd_make_literal_baseline_fse failed\n"); + exit (EXIT_FAILURE); + } + + printf ("static const struct elf_zstd_fse_baseline_entry " + "elf_zstd_lit_table[64] =\n"); + print_table (lit_baseline, + sizeof lit_baseline / sizeof lit_baseline[0]); + printf ("\n"); + + if (!elf_zstd_build_fse (match, sizeof match / sizeof match[0], next, + 6, match_table)) + { + fprintf (stderr, "elf_zstd_build_fse failed\n"); + exit (EXIT_FAILURE); + } + + if (!elf_zstd_make_match_baseline_fse (match_table, 6, match_baseline)) + { + fprintf (stderr, "elf_zstd_make_match_baseline_fse failed\n"); + exit (EXIT_FAILURE); + } + + printf ("static const struct elf_zstd_fse_baseline_entry " + "elf_zstd_match_table[64] =\n"); + print_table (match_baseline, + sizeof match_baseline / sizeof match_baseline[0]); + printf ("\n"); + + if (!elf_zstd_build_fse (offset, sizeof offset / sizeof offset[0], next, + 5, offset_table)) + { + fprintf (stderr, "elf_zstd_build_fse failed\n"); + exit (EXIT_FAILURE); + } + + if (!elf_zstd_make_offset_baseline_fse (offset_table, 5, offset_baseline)) + { + fprintf (stderr, "elf_zstd_make_offset_baseline_fse failed\n"); + exit (EXIT_FAILURE); + } + + printf ("static const struct elf_zstd_fse_baseline_entry " + "elf_zstd_offset_table[32] =\n"); + print_table (offset_baseline, + sizeof offset_baseline / sizeof offset_baseline[0]); + printf ("\n"); + + return 0; +} + +#endif + +/* The fixed tables generated by the #ifdef'ed out main function + above. */ + +static const struct elf_zstd_fse_baseline_entry elf_zstd_lit_table[64] = +{ + { 0, 0, 4, 0 }, { 0, 0, 4, 16 }, { 1, 0, 5, 32 }, + { 3, 0, 5, 0 }, { 4, 0, 5, 0 }, { 6, 0, 5, 0 }, + { 7, 0, 5, 0 }, { 9, 0, 5, 0 }, { 10, 0, 5, 0 }, + { 12, 0, 5, 0 }, { 14, 0, 6, 0 }, { 16, 1, 5, 0 }, + { 20, 1, 5, 0 }, { 22, 1, 5, 0 }, { 28, 2, 5, 0 }, + { 32, 3, 5, 0 }, { 48, 4, 5, 0 }, { 64, 6, 5, 32 }, + { 128, 7, 5, 0 }, { 256, 8, 6, 0 }, { 1024, 10, 6, 0 }, + { 4096, 12, 6, 0 }, { 0, 0, 4, 32 }, { 1, 0, 4, 0 }, + { 2, 0, 5, 0 }, { 4, 0, 5, 32 }, { 5, 0, 5, 0 }, + { 7, 0, 5, 32 }, { 8, 0, 5, 0 }, { 10, 0, 5, 32 }, + { 11, 0, 5, 0 }, { 13, 0, 6, 0 }, { 16, 1, 5, 32 }, + { 18, 1, 5, 0 }, { 22, 1, 5, 32 }, { 24, 2, 5, 0 }, + { 32, 3, 5, 32 }, { 40, 3, 5, 0 }, { 64, 6, 4, 0 }, + { 64, 6, 4, 16 }, { 128, 7, 5, 32 }, { 512, 9, 6, 0 }, + { 2048, 11, 6, 0 }, { 0, 0, 4, 48 }, { 1, 0, 4, 16 }, + { 2, 0, 5, 32 }, { 3, 0, 5, 32 }, { 5, 0, 5, 32 }, + { 6, 0, 5, 32 }, { 8, 0, 5, 32 }, { 9, 0, 5, 32 }, + { 11, 0, 5, 32 }, { 12, 0, 5, 32 }, { 15, 0, 6, 0 }, + { 18, 1, 5, 32 }, { 20, 1, 5, 32 }, { 24, 2, 5, 32 }, + { 28, 2, 5, 32 }, { 40, 3, 5, 32 }, { 48, 4, 5, 32 }, + { 65536, 16, 6, 0 }, { 32768, 15, 6, 0 }, { 16384, 14, 6, 0 }, + { 8192, 13, 6, 0 }, +}; + +static const struct elf_zstd_fse_baseline_entry elf_zstd_match_table[64] = +{ + { 3, 0, 6, 0 }, { 4, 0, 4, 0 }, { 5, 0, 5, 32 }, + { 6, 0, 5, 0 }, { 8, 0, 5, 0 }, { 9, 0, 5, 0 }, + { 11, 0, 5, 0 }, { 13, 0, 6, 0 }, { 16, 0, 6, 0 }, + { 19, 0, 6, 0 }, { 22, 0, 6, 0 }, { 25, 0, 6, 0 }, + { 28, 0, 6, 0 }, { 31, 0, 6, 0 }, { 34, 0, 6, 0 }, + { 37, 1, 6, 0 }, { 41, 1, 6, 0 }, { 47, 2, 6, 0 }, + { 59, 3, 6, 0 }, { 83, 4, 6, 0 }, { 131, 7, 6, 0 }, + { 515, 9, 6, 0 }, { 4, 0, 4, 16 }, { 5, 0, 4, 0 }, + { 6, 0, 5, 32 }, { 7, 0, 5, 0 }, { 9, 0, 5, 32 }, + { 10, 0, 5, 0 }, { 12, 0, 6, 0 }, { 15, 0, 6, 0 }, + { 18, 0, 6, 0 }, { 21, 0, 6, 0 }, { 24, 0, 6, 0 }, + { 27, 0, 6, 0 }, { 30, 0, 6, 0 }, { 33, 0, 6, 0 }, + { 35, 1, 6, 0 }, { 39, 1, 6, 0 }, { 43, 2, 6, 0 }, + { 51, 3, 6, 0 }, { 67, 4, 6, 0 }, { 99, 5, 6, 0 }, + { 259, 8, 6, 0 }, { 4, 0, 4, 32 }, { 4, 0, 4, 48 }, + { 5, 0, 4, 16 }, { 7, 0, 5, 32 }, { 8, 0, 5, 32 }, + { 10, 0, 5, 32 }, { 11, 0, 5, 32 }, { 14, 0, 6, 0 }, + { 17, 0, 6, 0 }, { 20, 0, 6, 0 }, { 23, 0, 6, 0 }, + { 26, 0, 6, 0 }, { 29, 0, 6, 0 }, { 32, 0, 6, 0 }, + { 65539, 16, 6, 0 }, { 32771, 15, 6, 0 }, { 16387, 14, 6, 0 }, + { 8195, 13, 6, 0 }, { 4099, 12, 6, 0 }, { 2051, 11, 6, 0 }, + { 1027, 10, 6, 0 }, +}; + +static const struct elf_zstd_fse_baseline_entry elf_zstd_offset_table[32] = +{ + { 1, 0, 5, 0 }, { 64, 6, 4, 0 }, { 512, 9, 5, 0 }, + { 32768, 15, 5, 0 }, { 2097152, 21, 5, 0 }, { 8, 3, 5, 0 }, + { 128, 7, 4, 0 }, { 4096, 12, 5, 0 }, { 262144, 18, 5, 0 }, + { 8388608, 23, 5, 0 }, { 32, 5, 5, 0 }, { 256, 8, 4, 0 }, + { 16384, 14, 5, 0 }, { 1048576, 20, 5, 0 }, { 4, 2, 5, 0 }, + { 128, 7, 4, 16 }, { 2048, 11, 5, 0 }, { 131072, 17, 5, 0 }, + { 4194304, 22, 5, 0 }, { 16, 4, 5, 0 }, { 256, 8, 4, 16 }, + { 8192, 13, 5, 0 }, { 524288, 19, 5, 0 }, { 2, 1, 5, 0 }, + { 64, 6, 4, 16 }, { 1024, 10, 5, 0 }, { 65536, 16, 5, 0 }, + { 268435456, 28, 5, 0 }, { 134217728, 27, 5, 0 }, { 67108864, 26, 5, 0 }, + { 33554432, 25, 5, 0 }, { 16777216, 24, 5, 0 }, +}; + +/* Read a zstd Huffman table and build the decoding table in *TABLE, reading + and updating *PPIN. This sets *PTABLE_BITS to the number of bits of the + table, such that the table length is 1 << *TABLE_BITS. ZDEBUG_TABLE is + scratch space; it must be enough for 512 uint16_t values + 256 32-bit values + (2048 bytes). Returns 1 on success, 0 on error. */ + +static int +elf_zstd_read_huff (const unsigned char **ppin, const unsigned char *pinend, + uint16_t *zdebug_table, uint16_t *table, int *ptable_bits) +{ + const unsigned char *pin; + unsigned char hdr; + unsigned char *weights; + size_t count; + uint32_t *weight_mark; + size_t i; + uint32_t weight_mask; + size_t table_bits; + + pin = *ppin; + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + hdr = *pin; + ++pin; + + weights = (unsigned char *) zdebug_table; + + if (hdr < 128) + { + /* Table is compressed using FSE. */ + + struct elf_zstd_fse_entry *fse_table; + int fse_table_bits; + uint16_t *scratch; + const unsigned char *pfse; + const unsigned char *pback; + uint64_t val; + unsigned int bits; + unsigned int state1, state2; + + /* SCRATCH is used temporarily by elf_zstd_read_fse. It overlaps + WEIGHTS. */ + scratch = zdebug_table; + fse_table = (struct elf_zstd_fse_entry *) (scratch + 512); + fse_table_bits = 6; + + pfse = pin; + if (!elf_zstd_read_fse (&pfse, pinend, scratch, 255, fse_table, + &fse_table_bits)) + return 0; + + if (unlikely (pin + hdr > pinend)) + { + elf_uncompress_failed (); + return 0; + } + + /* We no longer need SCRATCH. Start recording weights. We need up to + 256 bytes of weights and 64 bytes of rank counts, so it won't overlap + FSE_TABLE. */ + + pback = pin + hdr - 1; + + if (!elf_fetch_backward_init (&pback, pfse, &val, &bits)) + return 0; + + bits -= fse_table_bits; + state1 = (val >> bits) & ((1U << fse_table_bits) - 1); + bits -= fse_table_bits; + state2 = (val >> bits) & ((1U << fse_table_bits) - 1); + + /* There are two independent FSE streams, tracked by STATE1 and STATE2. + We decode them alternately. */ + + count = 0; + while (1) + { + struct elf_zstd_fse_entry *pt; + uint64_t v; + + pt = &fse_table[state1]; + + if (unlikely (pin < pinend) && bits < pt->bits) + { + if (unlikely (count >= 254)) + { + elf_uncompress_failed (); + return 0; + } + weights[count] = (unsigned char) pt->symbol; + weights[count + 1] = (unsigned char) fse_table[state2].symbol; + count += 2; + break; + } + + if (unlikely (pt->bits == 0)) + v = 0; + else + { + if (!elf_fetch_bits_backward (&pback, pfse, &val, &bits)) + return 0; + + bits -= pt->bits; + v = (val >> bits) & (((uint64_t)1 << pt->bits) - 1); + } + + state1 = pt->base + v; + + if (unlikely (count >= 255)) + { + elf_uncompress_failed (); + return 0; + } + + weights[count] = pt->symbol; + ++count; + + pt = &fse_table[state2]; + + if (unlikely (pin < pinend && bits < pt->bits)) + { + if (unlikely (count >= 254)) + { + elf_uncompress_failed (); + return 0; + } + weights[count] = (unsigned char) pt->symbol; + weights[count + 1] = (unsigned char) fse_table[state1].symbol; + count += 2; + break; + } + + if (unlikely (pt->bits == 0)) + v = 0; + else + { + if (!elf_fetch_bits_backward (&pback, pfse, &val, &bits)) + return 0; + + bits -= pt->bits; + v = (val >> bits) & (((uint64_t)1 << pt->bits) - 1); + } + + state2 = pt->base + v; + + if (unlikely (count >= 255)) + { + elf_uncompress_failed (); + return 0; + } + + weights[count] = pt->symbol; + ++count; + } + + pin += hdr; + } + else + { + /* Table is not compressed. Each weight is 4 bits. */ + + count = hdr - 127; + if (unlikely (pin + ((count + 1) / 2) >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + for (i = 0; i < count; i += 2) + { + unsigned char b; + + b = *pin; + ++pin; + weights[i] = b >> 4; + weights[i + 1] = b & 0xf; + } + } + + weight_mark = (uint32_t *) (weights + 256); + memset (weight_mark, 0, 12 * sizeof (uint32_t)); + weight_mask = 0; + for (i = 0; i < count; ++i) + { + unsigned char w; + + w = weights[i]; + if (unlikely (w > 12)) + { + elf_uncompress_failed (); + return 0; + } + ++weight_mark[w]; + if (w > 0) + weight_mask += 1U << (w - 1); + } + if (unlikely (weight_mask == 0)) + { + elf_uncompress_failed (); + return 0; + } + + table_bits = 32 - __builtin_clz (weight_mask); + if (unlikely (table_bits > 11)) + { + elf_uncompress_failed (); + return 0; + } + + /* Work out the last weight value, which is omitted because the weights must + sum to a power of two. */ + { + uint32_t left; + uint32_t high_bit; + + left = ((uint32_t)1 << table_bits) - weight_mask; + if (left == 0) + { + elf_uncompress_failed (); + return 0; + } + high_bit = 31 - __builtin_clz (left); + if (((uint32_t)1 << high_bit) != left) + { + elf_uncompress_failed (); + return 0; + } + + if (unlikely (count >= 256)) + { + elf_uncompress_failed (); + return 0; + } + + weights[count] = high_bit + 1; + ++count; + ++weight_mark[high_bit + 1]; + } + + if (weight_mark[1] < 2 || (weight_mark[1] & 1) != 0) + { + elf_uncompress_failed (); + return 0; + } + + /* Change WEIGHT_MARK from a count of weights to the index of the first + symbol for that weight. We shift the indexes to also store how many we + hae seen so far, below. */ + { + uint32_t next; + + next = 0; + for (i = 0; i < table_bits; ++i) + { + uint32_t cur; + + cur = next; + next += weight_mark[i + 1] << i; + weight_mark[i + 1] = cur; + } + } + + for (i = 0; i < count; ++i) + { + unsigned char weight; + uint32_t length; + uint16_t tval; + size_t start; + uint32_t j; + + weight = weights[i]; + if (weight == 0) + continue; + + length = 1U << (weight - 1); + tval = (i << 8) | (table_bits + 1 - weight); + start = weight_mark[weight]; + for (j = 0; j < length; ++j) + table[start + j] = tval; + weight_mark[weight] += length; + } + + *ppin = pin; + *ptable_bits = (int)table_bits; + + return 1; +} + +/* Read and decompress the literals and store them ending at POUTEND. This + works because we are going to use all the literals in the output, so they + must fit into the output buffer. HUFFMAN_TABLE, and PHUFFMAN_TABLE_BITS + store the Huffman table across calls. SCRATCH is used to read a Huffman + table. Store the start of the decompressed literals in *PPLIT. Update + *PPIN. Return 1 on success, 0 on error. */ + +static int +elf_zstd_read_literals (const unsigned char **ppin, + const unsigned char *pinend, + unsigned char *pout, + unsigned char *poutend, + uint16_t *scratch, + uint16_t *huffman_table, + int *phuffman_table_bits, + unsigned char **pplit) +{ + const unsigned char *pin; + unsigned char *plit; + unsigned char hdr; + uint32_t regenerated_size; + uint32_t compressed_size; + int streams; + uint32_t total_streams_size; + unsigned int huffman_table_bits; + uint64_t huffman_mask; + + pin = *ppin; + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + hdr = *pin; + ++pin; + + if ((hdr & 3) == 0 || (hdr & 3) == 1) + { + int raw; + + /* Raw_literals_Block or RLE_Literals_Block */ + + raw = (hdr & 3) == 0; + + switch ((hdr >> 2) & 3) + { + case 0: case 2: + regenerated_size = hdr >> 3; + break; + case 1: + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + regenerated_size = (hdr >> 4) + ((uint32_t)(*pin) << 4); + ++pin; + break; + case 3: + if (unlikely (pin + 1 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + regenerated_size = ((hdr >> 4) + + ((uint32_t)*pin << 4) + + ((uint32_t)pin[1] << 12)); + pin += 2; + break; + default: + elf_uncompress_failed (); + return 0; + } + + if (unlikely ((size_t)(poutend - pout) < regenerated_size)) + { + elf_uncompress_failed (); + return 0; + } + + plit = poutend - regenerated_size; + + if (raw) + { + if (unlikely (pin + regenerated_size >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + memcpy (plit, pin, regenerated_size); + pin += regenerated_size; + } + else + { + if (pin >= pinend) + { + elf_uncompress_failed (); + return 0; + } + memset (plit, *pin, regenerated_size); + ++pin; + } + + *ppin = pin; + *pplit = plit; + + return 1; + } + + /* Compressed_Literals_Block or Treeless_Literals_Block */ + + switch ((hdr >> 2) & 3) + { + case 0: case 1: + if (unlikely (pin + 1 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + regenerated_size = (hdr >> 4) | ((uint32_t)(*pin & 0x3f) << 4); + compressed_size = (uint32_t)*pin >> 6 | ((uint32_t)pin[1] << 2); + pin += 2; + streams = ((hdr >> 2) & 3) == 0 ? 1 : 4; + break; + case 2: + if (unlikely (pin + 2 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + regenerated_size = (((uint32_t)hdr >> 4) + | ((uint32_t)*pin << 4) + | (((uint32_t)pin[1] & 3) << 12)); + compressed_size = (((uint32_t)pin[1] >> 2) + | ((uint32_t)pin[2] << 6)); + pin += 3; + streams = 4; + break; + case 3: + if (unlikely (pin + 3 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + regenerated_size = (((uint32_t)hdr >> 4) + | ((uint32_t)*pin << 4) + | (((uint32_t)pin[1] & 0x3f) << 12)); + compressed_size = (((uint32_t)pin[1] >> 6) + | ((uint32_t)pin[2] << 2) + | ((uint32_t)pin[3] << 10)); + pin += 4; + streams = 4; + break; + default: + elf_uncompress_failed (); + return 0; + } + + if (unlikely (pin + compressed_size > pinend)) + { + elf_uncompress_failed (); + return 0; + } + + pinend = pin + compressed_size; + *ppin = pinend; + + if (unlikely ((size_t)(poutend - pout) < regenerated_size)) + { + elf_uncompress_failed (); + return 0; + } + + plit = poutend - regenerated_size; + + *pplit = plit; + + total_streams_size = compressed_size; + if ((hdr & 3) == 2) + { + const unsigned char *ptable; + + /* Compressed_Literals_Block. Read Huffman tree. */ + + ptable = pin; + if (!elf_zstd_read_huff (&ptable, pinend, scratch, huffman_table, + phuffman_table_bits)) + return 0; + + if (unlikely (total_streams_size < (size_t)(ptable - pin))) + { + elf_uncompress_failed (); + return 0; + } + + total_streams_size -= ptable - pin; + pin = ptable; + } + else + { + /* Treeless_Literals_Block. Reuse previous Huffman tree. */ + if (unlikely (*phuffman_table_bits == 0)) + { + elf_uncompress_failed (); + return 0; + } + } + + /* Decompress COMPRESSED_SIZE bytes of data at PIN using the huffman table, + storing REGENERATED_SIZE bytes of decompressed data at PLIT. */ + + huffman_table_bits = (unsigned int)*phuffman_table_bits; + huffman_mask = ((uint64_t)1 << huffman_table_bits) - 1; + + if (streams == 1) + { + const unsigned char *pback; + const unsigned char *pbackend; + uint64_t val; + unsigned int bits; + uint32_t i; + + pback = pin + compressed_size - 1; + pbackend = pin; + if (!elf_fetch_backward_init (&pback, pbackend, &val, &bits)) + return 0; + + /* This is one of the inner loops of the decompression algorithm, so we + put some effort into optimization. We can't get more than 64 bytes + from a single call to elf_fetch_bits_backward, and we can't subtract + more than 11 bits at a time. */ + + if (regenerated_size >= 64) + { + unsigned char *plitstart; + unsigned char *plitstop; + + plitstart = plit; + plitstop = plit + regenerated_size - 64; + while (plit < plitstop) + { + uint16_t t; + + if (!elf_fetch_bits_backward (&pback, pbackend, &val, &bits)) + return 0; + + if (bits < 16) + break; + + while (bits >= 33) + { + t = huffman_table[(val >> (bits - huffman_table_bits)) + & huffman_mask]; + *plit = t >> 8; + ++plit; + bits -= t & 0xff; + + t = huffman_table[(val >> (bits - huffman_table_bits)) + & huffman_mask]; + *plit = t >> 8; + ++plit; + bits -= t & 0xff; + + t = huffman_table[(val >> (bits - huffman_table_bits)) + & huffman_mask]; + *plit = t >> 8; + ++plit; + bits -= t & 0xff; + } + + while (bits > 11) + { + t = huffman_table[(val >> (bits - huffman_table_bits)) + & huffman_mask]; + *plit = t >> 8; + ++plit; + bits -= t & 0xff; + } + } + + regenerated_size -= plit - plitstart; + } + + for (i = 0; i < regenerated_size; ++i) + { + uint16_t t; + + if (!elf_fetch_bits_backward (&pback, pbackend, &val, &bits)) + return 0; + + if (unlikely (bits < huffman_table_bits)) + { + t = huffman_table[(val << (huffman_table_bits - bits)) + & huffman_mask]; + if (unlikely (bits < (t & 0xff))) + { + elf_uncompress_failed (); + return 0; + } + } + else + t = huffman_table[(val >> (bits - huffman_table_bits)) + & huffman_mask]; + + *plit = t >> 8; + ++plit; + bits -= t & 0xff; + } + + return 1; + } + + { + uint32_t stream_size1, stream_size2, stream_size3, stream_size4; + uint32_t tot; + const unsigned char *pback1, *pback2, *pback3, *pback4; + const unsigned char *pbackend1, *pbackend2, *pbackend3, *pbackend4; + uint64_t val1, val2, val3, val4; + unsigned int bits1, bits2, bits3, bits4; + unsigned char *plit1, *plit2, *plit3, *plit4; + uint32_t regenerated_stream_size; + uint32_t regenerated_stream_size4; + uint16_t t1, t2, t3, t4; + uint32_t i; + uint32_t limit; + + /* Read jump table. */ + if (unlikely (pin + 5 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + stream_size1 = (uint32_t)*pin | ((uint32_t)pin[1] << 8); + pin += 2; + stream_size2 = (uint32_t)*pin | ((uint32_t)pin[1] << 8); + pin += 2; + stream_size3 = (uint32_t)*pin | ((uint32_t)pin[1] << 8); + pin += 2; + tot = stream_size1 + stream_size2 + stream_size3; + if (unlikely (tot > total_streams_size - 6)) + { + elf_uncompress_failed (); + return 0; + } + stream_size4 = total_streams_size - 6 - tot; + + pback1 = pin + stream_size1 - 1; + pbackend1 = pin; + + pback2 = pback1 + stream_size2; + pbackend2 = pback1 + 1; + + pback3 = pback2 + stream_size3; + pbackend3 = pback2 + 1; + + pback4 = pback3 + stream_size4; + pbackend4 = pback3 + 1; + + if (!elf_fetch_backward_init (&pback1, pbackend1, &val1, &bits1)) + return 0; + if (!elf_fetch_backward_init (&pback2, pbackend2, &val2, &bits2)) + return 0; + if (!elf_fetch_backward_init (&pback3, pbackend3, &val3, &bits3)) + return 0; + if (!elf_fetch_backward_init (&pback4, pbackend4, &val4, &bits4)) + return 0; + + regenerated_stream_size = (regenerated_size + 3) / 4; + + plit1 = plit; + plit2 = plit1 + regenerated_stream_size; + plit3 = plit2 + regenerated_stream_size; + plit4 = plit3 + regenerated_stream_size; + + regenerated_stream_size4 = regenerated_size - regenerated_stream_size * 3; + + /* We can't get more than 64 literal bytes from a single call to + elf_fetch_bits_backward. The fourth stream can be up to 3 bytes less, + so use as the limit. */ + + limit = regenerated_stream_size4 <= 64 ? 0 : regenerated_stream_size4 - 64; + i = 0; + while (i < limit) + { + if (!elf_fetch_bits_backward (&pback1, pbackend1, &val1, &bits1)) + return 0; + if (!elf_fetch_bits_backward (&pback2, pbackend2, &val2, &bits2)) + return 0; + if (!elf_fetch_bits_backward (&pback3, pbackend3, &val3, &bits3)) + return 0; + if (!elf_fetch_bits_backward (&pback4, pbackend4, &val4, &bits4)) + return 0; + + /* We can't subtract more than 11 bits at a time. */ + + do + { + t1 = huffman_table[(val1 >> (bits1 - huffman_table_bits)) + & huffman_mask]; + t2 = huffman_table[(val2 >> (bits2 - huffman_table_bits)) + & huffman_mask]; + t3 = huffman_table[(val3 >> (bits3 - huffman_table_bits)) + & huffman_mask]; + t4 = huffman_table[(val4 >> (bits4 - huffman_table_bits)) + & huffman_mask]; + + *plit1 = t1 >> 8; + ++plit1; + bits1 -= t1 & 0xff; + + *plit2 = t2 >> 8; + ++plit2; + bits2 -= t2 & 0xff; + + *plit3 = t3 >> 8; + ++plit3; + bits3 -= t3 & 0xff; + + *plit4 = t4 >> 8; + ++plit4; + bits4 -= t4 & 0xff; + + ++i; + } + while (bits1 > 11 && bits2 > 11 && bits3 > 11 && bits4 > 11); + } + + while (i < regenerated_stream_size) + { + int use4; + + use4 = i < regenerated_stream_size4; + + if (!elf_fetch_bits_backward (&pback1, pbackend1, &val1, &bits1)) + return 0; + if (!elf_fetch_bits_backward (&pback2, pbackend2, &val2, &bits2)) + return 0; + if (!elf_fetch_bits_backward (&pback3, pbackend3, &val3, &bits3)) + return 0; + if (use4) + { + if (!elf_fetch_bits_backward (&pback4, pbackend4, &val4, &bits4)) + return 0; + } + + if (unlikely (bits1 < huffman_table_bits)) + { + t1 = huffman_table[(val1 << (huffman_table_bits - bits1)) + & huffman_mask]; + if (unlikely (bits1 < (t1 & 0xff))) + { + elf_uncompress_failed (); + return 0; + } + } + else + t1 = huffman_table[(val1 >> (bits1 - huffman_table_bits)) + & huffman_mask]; + + if (unlikely (bits2 < huffman_table_bits)) + { + t2 = huffman_table[(val2 << (huffman_table_bits - bits2)) + & huffman_mask]; + if (unlikely (bits2 < (t2 & 0xff))) + { + elf_uncompress_failed (); + return 0; + } + } + else + t2 = huffman_table[(val2 >> (bits2 - huffman_table_bits)) + & huffman_mask]; + + if (unlikely (bits3 < huffman_table_bits)) + { + t3 = huffman_table[(val3 << (huffman_table_bits - bits3)) + & huffman_mask]; + if (unlikely (bits3 < (t3 & 0xff))) + { + elf_uncompress_failed (); + return 0; + } + } + else + t3 = huffman_table[(val3 >> (bits3 - huffman_table_bits)) + & huffman_mask]; + + if (use4) + { + if (unlikely (bits4 < huffman_table_bits)) + { + t4 = huffman_table[(val4 << (huffman_table_bits - bits4)) + & huffman_mask]; + if (unlikely (bits4 < (t4 & 0xff))) + { + elf_uncompress_failed (); + return 0; + } + } + else + t4 = huffman_table[(val4 >> (bits4 - huffman_table_bits)) + & huffman_mask]; + + *plit4 = t4 >> 8; + ++plit4; + bits4 -= t4 & 0xff; + } + + *plit1 = t1 >> 8; + ++plit1; + bits1 -= t1 & 0xff; + + *plit2 = t2 >> 8; + ++plit2; + bits2 -= t2 & 0xff; + + *plit3 = t3 >> 8; + ++plit3; + bits3 -= t3 & 0xff; + + ++i; + } + } + + return 1; +} + +/* The information used to decompress a sequence code, which can be a literal + length, an offset, or a match length. */ + +struct elf_zstd_seq_decode +{ + const struct elf_zstd_fse_baseline_entry *table; + int table_bits; +}; + +/* Unpack a sequence code compression mode. */ + +static int +elf_zstd_unpack_seq_decode (int mode, + const unsigned char **ppin, + const unsigned char *pinend, + const struct elf_zstd_fse_baseline_entry *predef, + int predef_bits, + uint16_t *scratch, + int maxidx, + struct elf_zstd_fse_baseline_entry *table, + int table_bits, + int (*conv)(const struct elf_zstd_fse_entry *, + int, + struct elf_zstd_fse_baseline_entry *), + struct elf_zstd_seq_decode *decode) +{ + switch (mode) + { + case 0: + decode->table = predef; + decode->table_bits = predef_bits; + break; + + case 1: + { + struct elf_zstd_fse_entry entry; + + if (unlikely (*ppin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + entry.symbol = **ppin; + ++*ppin; + entry.bits = 0; + entry.base = 0; + decode->table_bits = 0; + if (!conv (&entry, 0, table)) + return 0; + } + break; + + case 2: + { + struct elf_zstd_fse_entry *fse_table; + + /* We use the same space for the simple FSE table and the baseline + table. */ + fse_table = (struct elf_zstd_fse_entry *)table; + decode->table_bits = table_bits; + if (!elf_zstd_read_fse (ppin, pinend, scratch, maxidx, fse_table, + &decode->table_bits)) + return 0; + if (!conv (fse_table, decode->table_bits, table)) + return 0; + decode->table = table; + } + break; + + case 3: + if (unlikely (decode->table_bits == -1)) + { + elf_uncompress_failed (); + return 0; + } + break; + + default: + elf_uncompress_failed (); + return 0; + } + + return 1; +} + +/* Decompress a zstd stream from PIN/SIN to POUT/SOUT. Code based on RFC 8878. + Return 1 on success, 0 on error. */ + +static int +elf_zstd_decompress (const unsigned char *pin, size_t sin, + unsigned char *zdebug_table, unsigned char *pout, + size_t sout) +{ + const unsigned char *pinend; + unsigned char *poutstart; + unsigned char *poutend; + struct elf_zstd_seq_decode literal_decode; + struct elf_zstd_fse_baseline_entry *literal_fse_table; + struct elf_zstd_seq_decode match_decode; + struct elf_zstd_fse_baseline_entry *match_fse_table; + struct elf_zstd_seq_decode offset_decode; + struct elf_zstd_fse_baseline_entry *offset_fse_table; + uint16_t *huffman_table; + int huffman_table_bits; + uint32_t repeated_offset1; + uint32_t repeated_offset2; + uint32_t repeated_offset3; + uint16_t *scratch; + unsigned char hdr; + int has_checksum; + uint64_t content_size; + int last_block; + + pinend = pin + sin; + poutstart = pout; + poutend = pout + sout; + + literal_decode.table = NULL; + literal_decode.table_bits = -1; + literal_fse_table = ((struct elf_zstd_fse_baseline_entry *) + (zdebug_table + ZSTD_TABLE_LITERAL_FSE_OFFSET)); + + match_decode.table = NULL; + match_decode.table_bits = -1; + match_fse_table = ((struct elf_zstd_fse_baseline_entry *) + (zdebug_table + ZSTD_TABLE_MATCH_FSE_OFFSET)); + + offset_decode.table = NULL; + offset_decode.table_bits = -1; + offset_fse_table = ((struct elf_zstd_fse_baseline_entry *) + (zdebug_table + ZSTD_TABLE_OFFSET_FSE_OFFSET)); + huffman_table = ((uint16_t *) + (zdebug_table + ZSTD_TABLE_HUFFMAN_OFFSET)); + huffman_table_bits = 0; + scratch = ((uint16_t *) + (zdebug_table + ZSTD_TABLE_WORK_OFFSET)); + + repeated_offset1 = 1; + repeated_offset2 = 4; + repeated_offset3 = 8; + + if (unlikely (sin < 4)) + { + elf_uncompress_failed (); + return 0; + } + + /* These values are the zstd magic number. */ + if (unlikely (pin[0] != 0x28 + || pin[1] != 0xb5 + || pin[2] != 0x2f + || pin[3] != 0xfd)) + { + elf_uncompress_failed (); + return 0; + } + + pin += 4; + + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + + hdr = *pin++; + + /* We expect a single frame. */ + if (unlikely ((hdr & (1 << 5)) == 0)) + { + elf_uncompress_failed (); + return 0; + } + /* Reserved bit must be zero. */ + if (unlikely ((hdr & (1 << 3)) != 0)) + { + elf_uncompress_failed (); + return 0; + } + /* We do not expect a dictionary. */ + if (unlikely ((hdr & 3) != 0)) + { + elf_uncompress_failed (); + return 0; + } + has_checksum = (hdr & (1 << 2)) != 0; + switch (hdr >> 6) + { + case 0: + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + content_size = (uint64_t) *pin++; + break; + case 1: + if (unlikely (pin + 1 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + content_size = (((uint64_t) pin[0]) | (((uint64_t) pin[1]) << 8)) + 256; + pin += 2; + break; + case 2: + if (unlikely (pin + 3 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + content_size = ((uint64_t) pin[0] + | (((uint64_t) pin[1]) << 8) + | (((uint64_t) pin[2]) << 16) + | (((uint64_t) pin[3]) << 24)); + pin += 4; + break; + case 3: + if (unlikely (pin + 7 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + content_size = ((uint64_t) pin[0] + | (((uint64_t) pin[1]) << 8) + | (((uint64_t) pin[2]) << 16) + | (((uint64_t) pin[3]) << 24) + | (((uint64_t) pin[4]) << 32) + | (((uint64_t) pin[5]) << 40) + | (((uint64_t) pin[6]) << 48) + | (((uint64_t) pin[7]) << 56)); + pin += 8; + break; + default: + elf_uncompress_failed (); + return 0; + } + + if (unlikely (content_size != (size_t) content_size + || (size_t) content_size != sout)) + { + elf_uncompress_failed (); + return 0; + } + + last_block = 0; + while (!last_block) + { + uint32_t block_hdr; + int block_type; + uint32_t block_size; + + if (unlikely (pin + 2 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + block_hdr = ((uint32_t) pin[0] + | (((uint32_t) pin[1]) << 8) + | (((uint32_t) pin[2]) << 16)); + pin += 3; + + last_block = block_hdr & 1; + block_type = (block_hdr >> 1) & 3; + block_size = block_hdr >> 3; + + switch (block_type) + { + case 0: + /* Raw_Block */ + if (unlikely ((size_t) block_size > (size_t) (pinend - pin))) + { + elf_uncompress_failed (); + return 0; + } + if (unlikely ((size_t) block_size > (size_t) (poutend - pout))) + { + elf_uncompress_failed (); + return 0; + } + memcpy (pout, pin, block_size); + pout += block_size; + pin += block_size; + break; + + case 1: + /* RLE_Block */ + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + if (unlikely ((size_t) block_size > (size_t) (poutend - pout))) + { + elf_uncompress_failed (); + return 0; + } + memset (pout, *pin, block_size); + pout += block_size; + pin++; + break; + + case 2: + { + const unsigned char *pblockend; + unsigned char *plitstack; + unsigned char *plit; + uint32_t literal_count; + unsigned char seq_hdr; + size_t seq_count; + size_t seq; + const unsigned char *pback; + uint64_t val; + unsigned int bits; + unsigned int literal_state; + unsigned int offset_state; + unsigned int match_state; + + /* Compressed_Block */ + if (unlikely ((size_t) block_size > (size_t) (pinend - pin))) + { + elf_uncompress_failed (); + return 0; + } + + pblockend = pin + block_size; + + /* Read the literals into the end of the output space, and leave + PLIT pointing at them. */ + + if (!elf_zstd_read_literals (&pin, pblockend, pout, poutend, + scratch, huffman_table, + &huffman_table_bits, + &plitstack)) + return 0; + plit = plitstack; + literal_count = poutend - plit; + + seq_hdr = *pin; + pin++; + if (seq_hdr < 128) + seq_count = seq_hdr; + else if (seq_hdr < 255) + { + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + seq_count = ((seq_hdr - 128) << 8) + *pin; + pin++; + } + else + { + if (unlikely (pin + 1 >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + seq_count = *pin + (pin[1] << 8) + 0x7f00; + pin += 2; + } + + if (seq_count > 0) + { + int (*pfn)(const struct elf_zstd_fse_entry *, + int, struct elf_zstd_fse_baseline_entry *); + + if (unlikely (pin >= pinend)) + { + elf_uncompress_failed (); + return 0; + } + seq_hdr = *pin; + ++pin; + + pfn = elf_zstd_make_literal_baseline_fse; + if (!elf_zstd_unpack_seq_decode ((seq_hdr >> 6) & 3, + &pin, pinend, + &elf_zstd_lit_table[0], 6, + scratch, 35, + literal_fse_table, 9, pfn, + &literal_decode)) + return 0; + + pfn = elf_zstd_make_offset_baseline_fse; + if (!elf_zstd_unpack_seq_decode ((seq_hdr >> 4) & 3, + &pin, pinend, + &elf_zstd_offset_table[0], 5, + scratch, 31, + offset_fse_table, 8, pfn, + &offset_decode)) + return 0; + + pfn = elf_zstd_make_match_baseline_fse; + if (!elf_zstd_unpack_seq_decode ((seq_hdr >> 2) & 3, + &pin, pinend, + &elf_zstd_match_table[0], 6, + scratch, 52, + match_fse_table, 9, pfn, + &match_decode)) + return 0; + } + + pback = pblockend - 1; + if (!elf_fetch_backward_init (&pback, pin, &val, &bits)) + return 0; + + bits -= literal_decode.table_bits; + literal_state = ((val >> bits) + & ((1U << literal_decode.table_bits) - 1)); + + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + bits -= offset_decode.table_bits; + offset_state = ((val >> bits) + & ((1U << offset_decode.table_bits) - 1)); + + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + bits -= match_decode.table_bits; + match_state = ((val >> bits) + & ((1U << match_decode.table_bits) - 1)); + + seq = 0; + while (1) + { + const struct elf_zstd_fse_baseline_entry *pt; + uint32_t offset_basebits; + uint32_t offset_baseline; + uint32_t offset_bits; + uint32_t offset_base; + uint32_t offset; + uint32_t match_baseline; + uint32_t match_bits; + uint32_t match_base; + uint32_t match; + uint32_t literal_baseline; + uint32_t literal_bits; + uint32_t literal_base; + uint32_t literal; + uint32_t need; + uint32_t add; + + pt = &offset_decode.table[offset_state]; + offset_basebits = pt->basebits; + offset_baseline = pt->baseline; + offset_bits = pt->bits; + offset_base = pt->base; + + /* This case can be more than 16 bits, which is all that + elf_fetch_bits_backward promises. */ + need = offset_basebits; + add = 0; + if (unlikely (need > 16)) + { + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + bits -= 16; + add = (val >> bits) & ((1U << 16) - 1); + need -= 16; + add <<= need; + } + if (need > 0) + { + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + bits -= need; + add += (val >> bits) & ((1U << need) - 1); + } + + offset = offset_baseline + add; + + pt = &match_decode.table[match_state]; + need = pt->basebits; + match_baseline = pt->baseline; + match_bits = pt->bits; + match_base = pt->base; + + add = 0; + if (need > 0) + { + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + bits -= need; + add = (val >> bits) & ((1U << need) - 1); + } + + match = match_baseline + add; + + pt = &literal_decode.table[literal_state]; + need = pt->basebits; + literal_baseline = pt->baseline; + literal_bits = pt->bits; + literal_base = pt->base; + + add = 0; + if (need > 0) + { + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + bits -= need; + add = (val >> bits) & ((1U << need) - 1); + } + + literal = literal_baseline + add; + + /* See the comment in elf_zstd_make_offset_baseline_fse. */ + if (offset_basebits > 1) + { + repeated_offset3 = repeated_offset2; + repeated_offset2 = repeated_offset1; + repeated_offset1 = offset; + } + else + { + if (unlikely (literal == 0)) + ++offset; + switch (offset) + { + case 1: + offset = repeated_offset1; + break; + case 2: + offset = repeated_offset2; + repeated_offset2 = repeated_offset1; + repeated_offset1 = offset; + break; + case 3: + offset = repeated_offset3; + repeated_offset3 = repeated_offset2; + repeated_offset2 = repeated_offset1; + repeated_offset1 = offset; + break; + case 4: + offset = repeated_offset1 - 1; + repeated_offset3 = repeated_offset2; + repeated_offset2 = repeated_offset1; + repeated_offset1 = offset; + break; + } + } + + ++seq; + if (seq < seq_count) + { + uint32_t v; + + /* Update the three states. */ + + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; + + need = literal_bits; + bits -= need; + v = (val >> bits) & (((uint32_t)1 << need) - 1); - if (unlikely ((unsigned int) (poutend - pout) < len)) - { - elf_uncompress_failed (); - return 0; - } + literal_state = literal_base + v; - memset (pout, pout[-1], len); - pout += len; - } - else if (unlikely (dist > 29)) - { - elf_uncompress_failed (); + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) return 0; - } - else - { - if (dist < 4) - dist = dist + 1; - else - { - unsigned int extra; - if (!elf_zlib_fetch (&pin, pinend, &val, &bits)) - return 0; + need = match_bits; + bits -= need; + v = (val >> bits) & (((uint32_t)1 << need) - 1); - /* This is an expression for the table of - distance codes in RFC 1951 3.2.5. */ - dist -= 4; - extra = (dist >> 1) + 1; - dist = (dist & 1) << extra; - dist += 5; - dist += ((1U << (extra - 1)) - 1) << 2; - dist += val & ((1U << extra) - 1); - val >>= extra; - bits -= extra; - } + match_state = match_base + v; - /* Go back dist bytes, and copy len bytes from - there. */ + if (!elf_fetch_bits_backward (&pback, pin, &val, &bits)) + return 0; - if (unlikely ((unsigned int) (pout - porigout) < dist)) - { - elf_uncompress_failed (); - return 0; - } + need = offset_bits; + bits -= need; + v = (val >> bits) & (((uint32_t)1 << need) - 1); - if (unlikely ((unsigned int) (poutend - pout) < len)) - { - elf_uncompress_failed (); - return 0; - } + offset_state = offset_base + v; + } - if (dist >= len) - { - memcpy (pout, pout - dist, len); - pout += len; - } - else - { - while (len > 0) - { - unsigned int copy; + /* The next sequence is now in LITERAL, OFFSET, MATCH. */ - copy = len < dist ? len : dist; - memcpy (pout, pout - dist, copy); - len -= copy; - pout += copy; - } - } - } - } - } - } - } + /* Copy LITERAL bytes from the literals. */ - /* We should have filled the output buffer. */ - if (unlikely (pout != poutend)) - { - elf_uncompress_failed (); - return 0; - } + if (unlikely ((size_t)(poutend - pout) < literal)) + { + elf_uncompress_failed (); + return 0; + } - return 1; -} + if (unlikely (literal_count < literal)) + { + elf_uncompress_failed (); + return 0; + } + + literal_count -= literal; + + /* Often LITERAL is small, so handle small cases quickly. */ + switch (literal) + { + case 8: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 7: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 6: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 5: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 4: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 3: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 2: + *pout++ = *plit++; + /* FALLTHROUGH */ + case 1: + *pout++ = *plit++; + break; + + case 0: + break; + + default: + if (unlikely ((size_t)(plit - pout) < literal)) + { + uint32_t move; + + move = plit - pout; + while (literal > move) + { + memcpy (pout, plit, move); + pout += move; + plit += move; + literal -= move; + } + } + + memcpy (pout, plit, literal); + pout += literal; + plit += literal; + } + + if (match > 0) + { + /* Copy MATCH bytes from the decoded output at OFFSET. */ + + if (unlikely ((size_t)(poutend - pout) < match)) + { + elf_uncompress_failed (); + return 0; + } -/* Verify the zlib checksum. The checksum is in the 4 bytes at - CHECKBYTES, and the uncompressed data is at UNCOMPRESSED / - UNCOMPRESSED_SIZE. Returns 1 on success, 0 on failure. */ + if (unlikely ((size_t)(pout - poutstart) < offset)) + { + elf_uncompress_failed (); + return 0; + } + + if (offset >= match) + { + memcpy (pout, pout - offset, match); + pout += match; + } + else + { + while (match > 0) + { + uint32_t copy; + + copy = match < offset ? match : offset; + memcpy (pout, pout - offset, copy); + match -= copy; + pout += copy; + } + } + } + + if (unlikely (seq >= seq_count)) + { + /* Copy remaining literals. */ + if (literal_count > 0 && plit != pout) + { + if (unlikely ((size_t)(poutend - pout) + < literal_count)) + { + elf_uncompress_failed (); + return 0; + } -static int -elf_zlib_verify_checksum (const unsigned char *checkbytes, - const unsigned char *uncompressed, - size_t uncompressed_size) -{ - unsigned int i; - unsigned int cksum; - const unsigned char *p; - uint32_t s1; - uint32_t s2; - size_t hsz; + if ((size_t)(plit - pout) < literal_count) + { + uint32_t move; - cksum = 0; - for (i = 0; i < 4; i++) - cksum = (cksum << 8) | checkbytes[i]; + move = plit - pout; + while (literal_count > move) + { + memcpy (pout, plit, move); + pout += move; + plit += move; + literal_count -= move; + } + } - s1 = 1; - s2 = 0; + memcpy (pout, plit, literal_count); + } - /* Minimize modulo operations. */ + pout += literal_count; - p = uncompressed; - hsz = uncompressed_size; - while (hsz >= 5552) - { - for (i = 0; i < 5552; i += 16) - { - /* Manually unroll loop 16 times. */ - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; + break; + } + } + + pin = pblockend; + } + break; + + case 3: + default: + elf_uncompress_failed (); + return 0; } - hsz -= 5552; - s1 %= 65521; - s2 %= 65521; } - while (hsz >= 16) + if (has_checksum) { - /* Manually unroll loop 16 times. */ - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; - s1 = s1 + *p++; - s2 = s2 + s1; + if (unlikely (pin + 4 > pinend)) + { + elf_uncompress_failed (); + return 0; + } - hsz -= 16; - } + /* We don't currently verify the checksum. Currently running GNU ld with + --compress-debug-sections=zstd does not seem to generate a + checksum. */ - for (i = 0; i < hsz; ++i) - { - s1 = s1 + *p++; - s2 = s2 + s1; + pin += 4; } - s1 %= 65521; - s2 %= 65521; - - if (unlikely ((s2 << 16) + s1 != cksum)) + if (pin != pinend) { elf_uncompress_failed (); return 0; @@ -2544,20 +5022,8 @@ elf_zlib_verify_checksum (const unsigned char *checkbytes, return 1; } -/* Inflate a zlib stream from PIN/SIN to POUT/SOUT, and verify the - checksum. Return 1 on success, 0 on error. */ - -static int -elf_zlib_inflate_and_verify (const unsigned char *pin, size_t sin, - uint16_t *zdebug_table, unsigned char *pout, - size_t sout) -{ - if (!elf_zlib_inflate (pin, sin, zdebug_table, pout, sout)) - return 0; - if (!elf_zlib_verify_checksum (pin + sin - 4, pout, sout)) - return 0; - return 1; -} +#define ZDEBUG_TABLE_SIZE \ + (ZLIB_TABLE_SIZE > ZSTD_TABLE_SIZE ? ZLIB_TABLE_SIZE : ZSTD_TABLE_SIZE) /* Uncompress the old compressed debug format, the one emitted by --compress-debug-sections=zlib-gnu. The compressed data is in @@ -2628,6 +5094,8 @@ elf_uncompress_chdr (struct backtrace_state *state, unsigned char **uncompressed, size_t *uncompressed_size) { const b_elf_chdr *chdr; + char *alc; + size_t alc_len; unsigned char *po; *uncompressed = NULL; @@ -2639,31 +5107,50 @@ elf_uncompress_chdr (struct backtrace_state *state, chdr = (const b_elf_chdr *) compressed; - if (chdr->ch_type != ELFCOMPRESS_ZLIB) - { - /* Unsupported compression algorithm. */ - return 1; - } - + alc = NULL; + alc_len = 0; if (*uncompressed != NULL && *uncompressed_size >= chdr->ch_size) po = *uncompressed; else { - po = (unsigned char *) backtrace_alloc (state, chdr->ch_size, - error_callback, data); - if (po == NULL) + alc_len = chdr->ch_size; + alc = (char*)backtrace_alloc (state, alc_len, error_callback, data); + if (alc == NULL) return 0; + po = (unsigned char *) alc; } - if (!elf_zlib_inflate_and_verify (compressed + sizeof (b_elf_chdr), - compressed_size - sizeof (b_elf_chdr), - zdebug_table, po, chdr->ch_size)) - return 1; + switch (chdr->ch_type) + { + case ELFCOMPRESS_ZLIB: + if (!elf_zlib_inflate_and_verify (compressed + sizeof (b_elf_chdr), + compressed_size - sizeof (b_elf_chdr), + zdebug_table, po, chdr->ch_size)) + goto skip; + break; + + case ELFCOMPRESS_ZSTD: + if (!elf_zstd_decompress (compressed + sizeof (b_elf_chdr), + compressed_size - sizeof (b_elf_chdr), + (unsigned char *)zdebug_table, po, + chdr->ch_size)) + goto skip; + break; + + default: + /* Unsupported compression algorithm. */ + goto skip; + } *uncompressed = po; *uncompressed_size = chdr->ch_size; return 1; + + skip: + if (alc != NULL && alc_len > 0) + backtrace_free (state, alc, alc_len, error_callback, data); + return 1; } /* This function is a hook for testing the zlib support. It is only @@ -2692,6 +5179,31 @@ backtrace_uncompress_zdebug (struct backtrace_state *state, return ret; } +/* This function is a hook for testing the zstd support. It is only used by + tests. */ + +int +backtrace_uncompress_zstd (struct backtrace_state *state, + const unsigned char *compressed, + size_t compressed_size, + backtrace_error_callback error_callback, + void *data, unsigned char *uncompressed, + size_t uncompressed_size) +{ + unsigned char *zdebug_table; + int ret; + + zdebug_table = ((unsigned char *) backtrace_alloc (state, ZDEBUG_TABLE_SIZE, + error_callback, data)); + if (zdebug_table == NULL) + return 0; + ret = elf_zstd_decompress (compressed, compressed_size, + zdebug_table, uncompressed, uncompressed_size); + backtrace_free (state, zdebug_table, ZDEBUG_TABLE_SIZE, + error_callback, data); + return ret; +} + /* Number of LZMA states. */ #define LZMA_STATES (12) @@ -4688,7 +7200,7 @@ elf_add (struct backtrace_state *state, const char *filename, int descriptor, if (zdebug_table == NULL) { zdebug_table = ((uint16_t *) - backtrace_alloc (state, ZDEBUG_TABLE_SIZE, + backtrace_alloc (state, ZLIB_TABLE_SIZE, error_callback, data)); if (zdebug_table == NULL) goto fail; @@ -4714,8 +7226,15 @@ elf_add (struct backtrace_state *state, const char *filename, int descriptor, } } + if (zdebug_table != NULL) + { + backtrace_free (state, zdebug_table, ZLIB_TABLE_SIZE, + error_callback, data); + zdebug_table = NULL; + } + /* Uncompress the official ELF format - (--compress-debug-sections=zlib-gabi). */ + (--compress-debug-sections=zlib-gabi, --compress-debug-sections=zstd). */ for (i = 0; i < (int) DEBUG_MAX; ++i) { unsigned char *uncompressed_data; diff --git a/src/external/tracy/libbacktrace/internal.hpp b/src/external/tracy/libbacktrace/internal.hpp index 96c097e02..f871844b6 100644 --- a/src/external/tracy/libbacktrace/internal.hpp +++ b/src/external/tracy/libbacktrace/internal.hpp @@ -371,6 +371,15 @@ extern int backtrace_uncompress_zdebug (struct backtrace_state *, unsigned char **uncompressed, size_t *uncompressed_size); +/* A test-only hook for elf_zstd_decompress. */ + +extern int backtrace_uncompress_zstd (struct backtrace_state *, + const unsigned char *compressed, + size_t compressed_size, + backtrace_error_callback, void *data, + unsigned char *uncompressed, + size_t uncompressed_size); + /* A test-only hook for elf_uncompress_lzma. */ extern int backtrace_uncompress_lzma (struct backtrace_state *, diff --git a/src/external/tracy/tracy/Tracy.hpp b/src/external/tracy/tracy/Tracy.hpp index cd7e085d4..d42f4bf3b 100644 --- a/src/external/tracy/tracy/Tracy.hpp +++ b/src/external/tracy/tracy/Tracy.hpp @@ -4,6 +4,18 @@ #include "../common/TracyColor.hpp" #include "../common/TracySystem.hpp" +#ifndef TracyFunction +# define TracyFunction __FUNCTION__ +#endif + +#ifndef TracyFile +# define TracyFile __FILE__ +#endif + +#ifndef TracyLine +# define TracyLine __LINE__ +#endif + #ifndef TRACY_ENABLE #define ZoneNamed(x,y) @@ -37,14 +49,14 @@ #define FrameImage(x,y,z,w,a) -#define TracyLockable( type, varname ) type varname; -#define TracyLockableN( type, varname, desc ) type varname; -#define TracySharedLockable( type, varname ) type varname; -#define TracySharedLockableN( type, varname, desc ) type varname; +#define TracyLockable( type, varname ) type varname +#define TracyLockableN( type, varname, desc ) type varname +#define TracySharedLockable( type, varname ) type varname +#define TracySharedLockableN( type, varname, desc ) type varname #define LockableBase( type ) type #define SharedLockableBase( type ) type -#define LockMark(x) (void)x; -#define LockableName(x,y,z); +#define LockMark(x) (void)x +#define LockableName(x,y,z) #define TracyPlot(x,y) #define TracyPlotConfig(x,y,z,w,a) @@ -110,21 +122,21 @@ #include "../client/TracyScoped.hpp" #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK -# define ZoneNamed( varname, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ) -# define ZoneNamedN( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ) -# define ZoneNamedC( varname, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ) -# define ZoneNamedNC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ) +# define ZoneNamed( varname, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ) +# define ZoneNamedN( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ) +# define ZoneNamedC( varname, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ) +# define ZoneNamedNC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ) -# define ZoneTransient( varname, active ) tracy::ScopedZone varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), nullptr, 0, TRACY_CALLSTACK, active ) -# define ZoneTransientN( varname, name, active ) tracy::ScopedZone varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), TRACY_CALLSTACK, active ) +# define ZoneTransient( varname, active ) tracy::ScopedZone varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), nullptr, 0, TRACY_CALLSTACK, active ) +# define ZoneTransientN( varname, name, active ) tracy::ScopedZone varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), TRACY_CALLSTACK, active ) #else -# define ZoneNamed( varname, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), active ) -# define ZoneNamedN( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), active ) -# define ZoneNamedC( varname, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), active ) -# define ZoneNamedNC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), active ) +# define ZoneNamed( varname, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), active ) +# define ZoneNamedN( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), active ) +# define ZoneNamedC( varname, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), active ) +# define ZoneNamedNC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), active ) -# define ZoneTransient( varname, active ) tracy::ScopedZone varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), nullptr, 0, active ) -# define ZoneTransientN( varname, name, active ) tracy::ScopedZone varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), active ) +# define ZoneTransient( varname, active ) tracy::ScopedZone varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), nullptr, 0, active ) +# define ZoneTransientN( varname, name, active ) tracy::ScopedZone varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), active ) #endif #define ZoneScoped ZoneNamed( ___tracy_scoped_zone, true ) @@ -150,13 +162,13 @@ #define FrameImage( image, width, height, offset, flip ) tracy::Profiler::SendFrameImage( image, width, height, offset, flip ) -#define TracyLockable( type, varname ) tracy::Lockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, #type " " #varname, __FILE__, __LINE__, 0 }; return &srcloc; }() } -#define TracyLockableN( type, varname, desc ) tracy::Lockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, desc, __FILE__, __LINE__, 0 }; return &srcloc; }() } -#define TracySharedLockable( type, varname ) tracy::SharedLockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, #type " " #varname, __FILE__, __LINE__, 0 }; return &srcloc; }() } -#define TracySharedLockableN( type, varname, desc ) tracy::SharedLockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, desc, __FILE__, __LINE__, 0 }; return &srcloc; }() } +#define TracyLockable( type, varname ) tracy::Lockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, #type " " #varname, TracyFile, TracyLine, 0 }; return &srcloc; }() } +#define TracyLockableN( type, varname, desc ) tracy::Lockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, desc, TracyFile, TracyLine, 0 }; return &srcloc; }() } +#define TracySharedLockable( type, varname ) tracy::SharedLockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, #type " " #varname, TracyFile, TracyLine, 0 }; return &srcloc; }() } +#define TracySharedLockableN( type, varname, desc ) tracy::SharedLockable varname { [] () -> const tracy::SourceLocationData* { static constexpr tracy::SourceLocationData srcloc { nullptr, desc, TracyFile, TracyLine, 0 }; return &srcloc; }() } #define LockableBase( type ) tracy::Lockable #define SharedLockableBase( type ) tracy::SharedLockable -#define LockMark( varname ) static constexpr tracy::SourceLocationData __tracy_lock_location_##varname { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; varname.Mark( &__tracy_lock_location_##varname ) +#define LockMark( varname ) static constexpr tracy::SourceLocationData __tracy_lock_location_##varname { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; varname.Mark( &__tracy_lock_location_##varname ) #define LockableName( varname, txt, size ) varname.CustomName( txt, size ) #define TracyPlot( name, val ) tracy::Profiler::PlotData( name, val ) @@ -197,13 +209,13 @@ #endif #ifdef TRACY_HAS_CALLSTACK -# define ZoneNamedS( varname, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), depth, active ) -# define ZoneNamedNS( varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), depth, active ) -# define ZoneNamedCS( varname, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { nullptr, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), depth, active ) -# define ZoneNamedNCS( varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,__LINE__), depth, active ) +# define ZoneNamedS( varname, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), depth, active ) +# define ZoneNamedNS( varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), depth, active ) +# define ZoneNamedCS( varname, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { nullptr, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), depth, active ) +# define ZoneNamedNCS( varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::ScopedZone varname( &TracyConcat(__tracy_source_location,TracyLine), depth, active ) -# define ZoneTransientS( varname, depth, active ) tracy::ScopedZone varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), nullptr, 0, depth, active ) -# define ZoneTransientNS( varname, name, depth, active ) tracy::ScopedZone varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), depth, active ) +# define ZoneTransientS( varname, depth, active ) tracy::ScopedZone varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), nullptr, 0, depth, active ) +# define ZoneTransientNS( varname, name, depth, active ) tracy::ScopedZone varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), depth, active ) # define ZoneScopedS( depth ) ZoneNamedS( ___tracy_scoped_zone, depth, true ) # define ZoneScopedNS( name, depth ) ZoneNamedNS( ___tracy_scoped_zone, name, depth, true ) diff --git a/src/external/tracy/tracy/TracyC.h b/src/external/tracy/tracy/TracyC.h index 6268f9e9c..bedf5e162 100644 --- a/src/external/tracy/tracy/TracyC.h +++ b/src/external/tracy/tracy/TracyC.h @@ -15,6 +15,17 @@ TRACY_API void ___tracy_set_thread_name( const char* name ); #define TracyCSetThreadName( name ) ___tracy_set_thread_name( name ); +#ifndef TracyFunction +# define TracyFunction __FUNCTION__ +#endif + +#ifndef TracyFile +# define TracyFile __FILE__ +#endif + +#ifndef TracyLine +# define TracyLine __LINE__ +#endif #ifndef TRACY_ENABLE @@ -47,6 +58,8 @@ typedef const void* TracyCZoneCtx; #define TracyCFrameImage(x,y,z,w,a) #define TracyCPlot(x,y) +#define TracyCPlotF(x,y) +#define TracyCPlotI(x,y) #define TracyCMessage(x,y) #define TracyCMessageL(x) #define TracyCMessageC(x,y,z) @@ -195,15 +208,15 @@ TRACY_API void ___tracy_emit_gpu_calibration_serial( const struct ___tracy_gpu_c TRACY_API int ___tracy_connected(void); #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK -# define TracyCZone( ctx, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { NULL, __func__, __FILE__, (uint32_t)__LINE__, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ); -# define TracyCZoneN( ctx, name, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { name, __func__, __FILE__, (uint32_t)__LINE__, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ); -# define TracyCZoneC( ctx, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { NULL, __func__, __FILE__, (uint32_t)__LINE__, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ); -# define TracyCZoneNC( ctx, name, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { name, __func__, __FILE__, (uint32_t)__LINE__, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), TRACY_CALLSTACK, active ); +# define TracyCZone( ctx, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__, TracyFile, (uint32_t)TracyLine, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ); +# define TracyCZoneN( ctx, name, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { name, __func__, TracyFile, (uint32_t)TracyLine, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ); +# define TracyCZoneC( ctx, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__, TracyFile, (uint32_t)TracyLine, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ); +# define TracyCZoneNC( ctx, name, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { name, __func__, TracyFile, (uint32_t)TracyLine, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), TRACY_CALLSTACK, active ); #else -# define TracyCZone( ctx, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { NULL, __func__, __FILE__, (uint32_t)__LINE__, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,__LINE__), active ); -# define TracyCZoneN( ctx, name, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { name, __func__, __FILE__, (uint32_t)__LINE__, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,__LINE__), active ); -# define TracyCZoneC( ctx, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { NULL, __func__, __FILE__, (uint32_t)__LINE__, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,__LINE__), active ); -# define TracyCZoneNC( ctx, name, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { name, __func__, __FILE__, (uint32_t)__LINE__, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,__LINE__), active ); +# define TracyCZone( ctx, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__, TracyFile, (uint32_t)TracyLine, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,TracyLine), active ); +# define TracyCZoneN( ctx, name, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { name, __func__, TracyFile, (uint32_t)TracyLine, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,TracyLine), active ); +# define TracyCZoneC( ctx, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__, TracyFile, (uint32_t)TracyLine, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,TracyLine), active ); +# define TracyCZoneNC( ctx, name, color, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { name, __func__, TracyFile, (uint32_t)TracyLine, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin( &TracyConcat(__tracy_source_location,TracyLine), active ); #endif #define TracyCZoneEnd( ctx ) ___tracy_emit_zone_end( ctx ); @@ -274,17 +287,21 @@ TRACY_API void ___tracy_emit_frame_image( const void* image, uint16_t w, uint16_ TRACY_API void ___tracy_emit_plot( const char* name, double val ); +TRACY_API void ___tracy_emit_plot_float( const char* name, float val ); +TRACY_API void ___tracy_emit_plot_int( const char* name, int64_t val ); TRACY_API void ___tracy_emit_message_appinfo( const char* txt, size_t size ); #define TracyCPlot( name, val ) ___tracy_emit_plot( name, val ); +#define TracyCPlotF( name, val ) ___tracy_emit_plot_float( name, val ); +#define TracyCPlotI( name, val ) ___tracy_emit_plot_int( name, val ); #define TracyCAppInfo( txt, size ) ___tracy_emit_message_appinfo( txt, size ); #ifdef TRACY_HAS_CALLSTACK -# define TracyCZoneS( ctx, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { NULL, __func__, __FILE__, (uint32_t)__LINE__, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), depth, active ); -# define TracyCZoneNS( ctx, name, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { name, __func__, __FILE__, (uint32_t)__LINE__, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), depth, active ); -# define TracyCZoneCS( ctx, color, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { NULL, __func__, __FILE__, (uint32_t)__LINE__, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), depth, active ); -# define TracyCZoneNCS( ctx, name, color, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,__LINE__) = { name, __func__, __FILE__, (uint32_t)__LINE__, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,__LINE__), depth, active ); +# define TracyCZoneS( ctx, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__, TracyFile, (uint32_t)TracyLine, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), depth, active ); +# define TracyCZoneNS( ctx, name, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { name, __func__, TracyFile, (uint32_t)TracyLine, 0 }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), depth, active ); +# define TracyCZoneCS( ctx, color, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { NULL, __func__, TracyFile, (uint32_t)TracyLine, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), depth, active ); +# define TracyCZoneNCS( ctx, name, color, depth, active ) static const struct ___tracy_source_location_data TracyConcat(__tracy_source_location,TracyLine) = { name, __func__, TracyFile, (uint32_t)TracyLine, color }; TracyCZoneCtx ctx = ___tracy_emit_zone_begin_callstack( &TracyConcat(__tracy_source_location,TracyLine), depth, active ); # define TracyCAllocS( ptr, size, depth ) ___tracy_emit_memory_alloc_callstack( ptr, size, depth, 0 ) # define TracyCFreeS( ptr, depth ) ___tracy_emit_memory_free_callstack( ptr, depth, 0 ) @@ -324,10 +341,10 @@ TRACY_API void ___tracy_emit_message_appinfo( const char* txt, size_t size ); #define TracyCIsConnected ___tracy_connected() +#ifdef TRACY_FIBERS TRACY_API void ___tracy_fiber_enter( const char* fiber ); TRACY_API void ___tracy_fiber_leave( void ); -#ifdef TRACY_FIBERS # define TracyCFiberEnter( fiber ) ___tracy_fiber_enter( fiber ); # define TracyCFiberLeave ___tracy_fiber_leave(); #endif diff --git a/src/external/tracy/tracy/TracyD3D11.hpp b/src/external/tracy/tracy/TracyD3D11.hpp index 75d93bf2a..9f358c4a5 100644 --- a/src/external/tracy/tracy/TracyD3D11.hpp +++ b/src/external/tracy/tracy/TracyD3D11.hpp @@ -42,6 +42,8 @@ using TracyD3D11Ctx = void*; #include "../common/TracyAlign.hpp" #include "../common/TracyAlloc.hpp" +#include + namespace tracy { @@ -410,23 +412,23 @@ using TracyD3D11Ctx = tracy::D3D11Ctx*; #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK # define TracyD3D11Zone( ctx, name ) TracyD3D11NamedZoneS( ctx, ___tracy_gpu_zone, name, TRACY_CALLSTACK, true ) # define TracyD3D11ZoneC( ctx, name, color ) TracyD3D11NamedZoneCS( ctx, ___tracy_gpu_zone, name, color, TRACY_CALLSTACK, true ) -# define TracyD3D11NamedZone( ctx, varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), TRACY_CALLSTACK, active ); -# define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), TRACY_CALLSTACK, active ); +# define TracyD3D11NamedZone( ctx, varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); +# define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); # define TracyD3D11ZoneTransient(ctx, varname, name, active) TracyD3D11ZoneTransientS(ctx, varname, cmdList, name, TRACY_CALLSTACK, active) #else # define TracyD3D11Zone( ctx, name ) TracyD3D11NamedZone( ctx, ___tracy_gpu_zone, name, true ) # define TracyD3D11ZoneC( ctx, name, color ) TracyD3D11NamedZoneC( ctx, ___tracy_gpu_zone, name, color, true ) -# define TracyD3D11NamedZone( ctx, varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), active ); -# define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), active ); -# define TracyD3D11ZoneTransient(ctx, varname, name, active) tracy::D3D11ZoneScope varname{ ctx, __LINE__, __FILE__, strlen(__FILE__), __FUNCTION__, strlen(__FUNCTION__), name, strlen(name), active }; +# define TracyD3D11NamedZone( ctx, varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), active ); +# define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), active ); +# define TracyD3D11ZoneTransient(ctx, varname, name, active) tracy::D3D11ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), active }; #endif #ifdef TRACY_HAS_CALLSTACK # define TracyD3D11ZoneS( ctx, name, depth ) TracyD3D11NamedZoneS( ctx, ___tracy_gpu_zone, name, depth, true ) # define TracyD3D11ZoneCS( ctx, name, color, depth ) TracyD3D11NamedZoneCS( ctx, ___tracy_gpu_zone, name, color, depth, true ) -# define TracyD3D11NamedZoneS( ctx, varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), depth, active ); -# define TracyD3D11NamedZoneCS( ctx, varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), depth, active ); -# define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active) tracy::D3D11ZoneScope varname{ ctx, __LINE__, __FILE__, strlen(__FILE__), __FUNCTION__, strlen(__FUNCTION__), name, strlen(name), depth, active }; +# define TracyD3D11NamedZoneS( ctx, varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active ); +# define TracyD3D11NamedZoneCS( ctx, varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active ); +# define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active) tracy::D3D11ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), depth, active }; #else # define TracyD3D11ZoneS( ctx, name, depth, active ) TracyD3D11Zone( ctx, name ) # define TracyD3D11ZoneCS( ctx, name, color, depth, active ) TracyD3D11ZoneC( name, color ) diff --git a/src/external/tracy/tracy/TracyD3D12.hpp b/src/external/tracy/tracy/TracyD3D12.hpp index 66f20bc0c..d7944cb8e 100644 --- a/src/external/tracy/tracy/TracyD3D12.hpp +++ b/src/external/tracy/tracy/TracyD3D12.hpp @@ -474,23 +474,23 @@ using TracyD3D12Ctx = tracy::D3D12QueueCtx*; #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK # define TracyD3D12Zone(ctx, cmdList, name) TracyD3D12NamedZoneS(ctx, ___tracy_gpu_zone, cmdList, name, TRACY_CALLSTACK, true) # define TracyD3D12ZoneC(ctx, cmdList, name, color) TracyD3D12NamedZoneCS(ctx, ___tracy_gpu_zone, cmdList, name, color, TRACY_CALLSTACK, true) -# define TracyD3D12NamedZone(ctx, varname, cmdList, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, __LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, __LINE__), TRACY_CALLSTACK, active }; -# define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, __LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, __LINE__), TRACY_CALLSTACK, active }; +# define TracyD3D12NamedZone(ctx, varname, cmdList, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, TracyLine), TRACY_CALLSTACK, active }; +# define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, TracyLine), TRACY_CALLSTACK, active }; # define TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active) TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, TRACY_CALLSTACK, active) #else # define TracyD3D12Zone(ctx, cmdList, name) TracyD3D12NamedZone(ctx, ___tracy_gpu_zone, cmdList, name, true) # define TracyD3D12ZoneC(ctx, cmdList, name, color) TracyD3D12NamedZoneC(ctx, ___tracy_gpu_zone, cmdList, name, color, true) -# define TracyD3D12NamedZone(ctx, varname, cmdList, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, __LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, __LINE__), active }; -# define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, __LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, __LINE__), active }; -# define TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active) tracy::D3D12ZoneScope varname{ ctx, __LINE__, __FILE__, strlen(__FILE__), __FUNCTION__, strlen(__FUNCTION__), name, strlen(name), cmdList, active }; +# define TracyD3D12NamedZone(ctx, varname, cmdList, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, TracyLine), active }; +# define TracyD3D12NamedZoneC(ctx, varname, cmdList, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, TracyLine), active }; +# define TracyD3D12ZoneTransient(ctx, varname, cmdList, name, active) tracy::D3D12ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), cmdList, active }; #endif #ifdef TRACY_HAS_CALLSTACK # define TracyD3D12ZoneS(ctx, cmdList, name, depth) TracyD3D12NamedZoneS(ctx, ___tracy_gpu_zone, cmdList, name, depth, true) # define TracyD3D12ZoneCS(ctx, cmdList, name, color, depth) TracyD3D12NamedZoneCS(ctx, ___tracy_gpu_zone, cmdList, name, color, depth, true) -# define TracyD3D12NamedZoneS(ctx, varname, cmdList, name, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, __LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, __LINE__), depth, active }; -# define TracyD3D12NamedZoneCS(ctx, varname, cmdList, name, color, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, __LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, __LINE__), depth, active }; -# define TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, depth, active) tracy::D3D12ZoneScope varname{ ctx, __LINE__, __FILE__, strlen(__FILE__), __FUNCTION__, strlen(__FUNCTION__), name, strlen(name), cmdList, depth, active }; +# define TracyD3D12NamedZoneS(ctx, varname, cmdList, name, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, TracyLine), depth, active }; +# define TracyD3D12NamedZoneCS(ctx, varname, cmdList, name, color, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location, TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D12ZoneScope varname{ ctx, cmdList, &TracyConcat(__tracy_gpu_source_location, TracyLine), depth, active }; +# define TracyD3D12ZoneTransientS(ctx, varname, cmdList, name, depth, active) tracy::D3D12ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), cmdList, depth, active }; #else # define TracyD3D12ZoneS(ctx, cmdList, name, depth) TracyD3D12Zone(ctx, cmdList, name) # define TracyD3D12ZoneCS(ctx, cmdList, name, color, depth) TracyD3D12Zone(ctx, cmdList, name, color) diff --git a/src/external/tracy/tracy/TracyOpenCL.hpp b/src/external/tracy/tracy/TracyOpenCL.hpp index c13471a8b..34466ccc9 100644 --- a/src/external/tracy/tracy/TracyOpenCL.hpp +++ b/src/external/tracy/tracy/TracyOpenCL.hpp @@ -47,12 +47,12 @@ using TracyCLCtx = void*; #define TRACY_CL_TO_STRING_INDIRECT(T) #T #define TRACY_CL_TO_STRING(T) TRACY_CL_TO_STRING_INDIRECT(T) #define TRACY_CL_ASSERT(p) if(!(p)) { \ - TracyMessageL( "TRACY_CL_ASSERT failed on " __FILE__ ":" TRACY_CL_TO_STRING(__LINE__) ); \ + TracyMessageL( "TRACY_CL_ASSERT failed on " TracyFile ":" TRACY_CL_TO_STRING(TracyLine) ); \ assert(false && "TRACY_CL_ASSERT failed"); \ } #define TRACY_CL_CHECK_ERROR(err) if(err != CL_SUCCESS) { \ std::ostringstream oss; \ - oss << "TRACY_CL_CHECK_ERROR failed on " << __FILE__ << ":" << __LINE__ \ + oss << "TRACY_CL_CHECK_ERROR failed on " << TracyFile << ":" << TracyLine \ << ": error code " << err; \ auto msg = oss.str(); \ TracyMessage(msg.data(), msg.size()); \ @@ -377,25 +377,25 @@ using TracyCLCtx = tracy::OpenCLCtx*; #define TracyCLDestroy(ctx) tracy::DestroyCLContext(ctx); #define TracyCLContextName(context, name, size) ctx->Name(name, size); #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK -# define TracyCLNamedZone(ctx, varname, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), TRACY_CALLSTACK, active ); -# define TracyCLNamedZoneC(ctx, varname, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), TRACY_CALLSTACK, active ); +# define TracyCLNamedZone(ctx, varname, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); +# define TracyCLNamedZoneC(ctx, varname, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); # define TracyCLZone(ctx, name) TracyCLNamedZoneS(ctx, __tracy_gpu_zone, name, TRACY_CALLSTACK, true) # define TracyCLZoneC(ctx, name, color) TracyCLNamedZoneCS(ctx, __tracy_gpu_zone, name, color, TRACY_CALLSTACK, true) -# define TracyCLZoneTransient( ctx, varname, name, active ) tracy::OpenCLCtxScope varname( ctx, __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), TRACY_CALLSTACK, active ); +# define TracyCLZoneTransient( ctx, varname, name, active ) tracy::OpenCLCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), TRACY_CALLSTACK, active ); #else -# define TracyCLNamedZone(ctx, varname, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__){ name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), active); -# define TracyCLNamedZoneC(ctx, varname, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__){ name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), active); +# define TracyCLNamedZone(ctx, varname, name, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine){ name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), active); +# define TracyCLNamedZoneC(ctx, varname, name, color, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine){ name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), active); # define TracyCLZone(ctx, name) TracyCLNamedZone(ctx, __tracy_gpu_zone, name, true) # define TracyCLZoneC(ctx, name, color) TracyCLNamedZoneC(ctx, __tracy_gpu_zone, name, color, true ) -# define TracyCLZoneTransient( ctx, varname, name, active ) tracy::OpenCLCtxScope varname( ctx, __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), active ); +# define TracyCLZoneTransient( ctx, varname, name, active ) tracy::OpenCLCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), active ); #endif #ifdef TRACY_HAS_CALLSTACK -# define TracyCLNamedZoneS(ctx, varname, name, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__){ name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), depth, active); -# define TracyCLNamedZoneCS(ctx, varname, name, color, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__){ name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), depth, active); +# define TracyCLNamedZoneS(ctx, varname, name, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine){ name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active); +# define TracyCLNamedZoneCS(ctx, varname, name, color, depth, active) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine){ name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::OpenCLCtxScope varname(ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active); # define TracyCLZoneS(ctx, name, depth) TracyCLNamedZoneS(ctx, __tracy_gpu_zone, name, depth, true) # define TracyCLZoneCS(ctx, name, color, depth) TracyCLNamedZoneCS(ctx, __tracy_gpu_zone, name, color, depth, true) -# define TracyCLZoneTransientS( ctx, varname, name, depth, active ) tracy::OpenCLCtxScope varname( ctx, __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), depth, active ); +# define TracyCLZoneTransientS( ctx, varname, name, depth, active ) tracy::OpenCLCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), depth, active ); #else # define TracyCLNamedZoneS(ctx, varname, name, depth, active) TracyCLNamedZone(ctx, varname, name, active) # define TracyCLNamedZoneCS(ctx, varname, name, color, depth, active) TracyCLNamedZoneC(ctx, varname, name, color, active) diff --git a/src/external/tracy/tracy/TracyOpenGL.hpp b/src/external/tracy/tracy/TracyOpenGL.hpp index 68b4cf44b..3bdadccee 100644 --- a/src/external/tracy/tracy/TracyOpenGL.hpp +++ b/src/external/tracy/tracy/TracyOpenGL.hpp @@ -52,26 +52,26 @@ public: #define TracyGpuContext tracy::GetGpuCtx().ptr = (tracy::GpuCtx*)tracy::tracy_malloc( sizeof( tracy::GpuCtx ) ); new(tracy::GetGpuCtx().ptr) tracy::GpuCtx; #define TracyGpuContextName( name, size ) tracy::GetGpuCtx().ptr->Name( name, size ); #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK -# define TracyGpuNamedZone( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,__LINE__), TRACY_CALLSTACK, active ); -# define TracyGpuNamedZoneC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,__LINE__), TRACY_CALLSTACK, active ); +# define TracyGpuNamedZone( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); +# define TracyGpuNamedZoneC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); # define TracyGpuZone( name ) TracyGpuNamedZoneS( ___tracy_gpu_zone, name, TRACY_CALLSTACK, true ) # define TracyGpuZoneC( name, color ) TracyGpuNamedZoneCS( ___tracy_gpu_zone, name, color, TRACY_CALLSTACK, true ) -# define TracyGpuZoneTransient( varname, name, active ) tracy::GpuCtxScope varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), TRACY_CALLSTACK, active ); +# define TracyGpuZoneTransient( varname, name, active ) tracy::GpuCtxScope varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), TRACY_CALLSTACK, active ); #else -# define TracyGpuNamedZone( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,__LINE__), active ); -# define TracyGpuNamedZoneC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,__LINE__), active ); +# define TracyGpuNamedZone( varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,TracyLine), active ); +# define TracyGpuNamedZoneC( varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,TracyLine), active ); # define TracyGpuZone( name ) TracyGpuNamedZone( ___tracy_gpu_zone, name, true ) # define TracyGpuZoneC( name, color ) TracyGpuNamedZoneC( ___tracy_gpu_zone, name, color, true ) -# define TracyGpuZoneTransient( varname, name, active ) tracy::GpuCtxScope varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), active ); +# define TracyGpuZoneTransient( varname, name, active ) tracy::GpuCtxScope varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), active ); #endif #define TracyGpuCollect tracy::GetGpuCtx().ptr->Collect(); #ifdef TRACY_HAS_CALLSTACK -# define TracyGpuNamedZoneS( varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,__LINE__), depth, active ); -# define TracyGpuNamedZoneCS( varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,__LINE__), depth, active ); +# define TracyGpuNamedZoneS( varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active ); +# define TracyGpuNamedZoneCS( varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::GpuCtxScope varname( &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active ); # define TracyGpuZoneS( name, depth ) TracyGpuNamedZoneS( ___tracy_gpu_zone, name, depth, true ) # define TracyGpuZoneCS( name, color, depth ) TracyGpuNamedZoneCS( ___tracy_gpu_zone, name, color, depth, true ) -# define TracyGpuZoneTransientS( varname, name, depth, active ) tracy::GpuCtxScope varname( __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), depth, active ); +# define TracyGpuZoneTransientS( varname, name, depth, active ) tracy::GpuCtxScope varname( TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), depth, active ); #else # define TracyGpuNamedZoneS( varname, name, depth, active ) TracyGpuNamedZone( varname, name, active ) # define TracyGpuNamedZoneCS( varname, name, color, depth, active ) TracyGpuNamedZoneC( varname, name, color, active ) diff --git a/src/external/tracy/tracy/TracyVulkan.hpp b/src/external/tracy/tracy/TracyVulkan.hpp index 351d4f19e..3f4f6a31c 100644 --- a/src/external/tracy/tracy/TracyVulkan.hpp +++ b/src/external/tracy/tracy/TracyVulkan.hpp @@ -479,26 +479,26 @@ using TracyVkCtx = tracy::VkCtx*; #define TracyVkDestroy( ctx ) tracy::DestroyVkContext( ctx ); #define TracyVkContextName( ctx, name, size ) ctx->Name( name, size ); #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK -# define TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), cmdbuf, TRACY_CALLSTACK, active ); -# define TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), cmdbuf, TRACY_CALLSTACK, active ); +# define TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, TRACY_CALLSTACK, active ); +# define TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, TRACY_CALLSTACK, active ); # define TracyVkZone( ctx, cmdbuf, name ) TracyVkNamedZoneS( ctx, ___tracy_gpu_zone, cmdbuf, name, TRACY_CALLSTACK, true ) # define TracyVkZoneC( ctx, cmdbuf, name, color ) TracyVkNamedZoneCS( ctx, ___tracy_gpu_zone, cmdbuf, name, color, TRACY_CALLSTACK, true ) # define TracyVkZoneTransient( ctx, varname, cmdbuf, name, active ) TracyVkZoneTransientS( ctx, varname, cmdbuf, name, TRACY_CALLSTACK, active ) #else -# define TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), cmdbuf, active ); -# define TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), cmdbuf, active ); +# define TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, active ); +# define TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, active ); # define TracyVkZone( ctx, cmdbuf, name ) TracyVkNamedZone( ctx, ___tracy_gpu_zone, cmdbuf, name, true ) # define TracyVkZoneC( ctx, cmdbuf, name, color ) TracyVkNamedZoneC( ctx, ___tracy_gpu_zone, cmdbuf, name, color, true ) -# define TracyVkZoneTransient( ctx, varname, cmdbuf, name, active ) tracy::VkCtxScope varname( ctx, __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), cmdbuf, active ); +# define TracyVkZoneTransient( ctx, varname, cmdbuf, name, active ) tracy::VkCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), cmdbuf, active ); #endif #define TracyVkCollect( ctx, cmdbuf ) ctx->Collect( cmdbuf ); #ifdef TRACY_HAS_CALLSTACK -# define TracyVkNamedZoneS( ctx, varname, cmdbuf, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), cmdbuf, depth, active ); -# define TracyVkNamedZoneCS( ctx, varname, cmdbuf, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,__LINE__) { name, __FUNCTION__, __FILE__, (uint32_t)__LINE__, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,__LINE__), cmdbuf, depth, active ); +# define TracyVkNamedZoneS( ctx, varname, cmdbuf, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, depth, active ); +# define TracyVkNamedZoneCS( ctx, varname, cmdbuf, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::VkCtxScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), cmdbuf, depth, active ); # define TracyVkZoneS( ctx, cmdbuf, name, depth ) TracyVkNamedZoneS( ctx, ___tracy_gpu_zone, cmdbuf, name, depth, true ) # define TracyVkZoneCS( ctx, cmdbuf, name, color, depth ) TracyVkNamedZoneCS( ctx, ___tracy_gpu_zone, cmdbuf, name, color, depth, true ) -# define TracyVkZoneTransientS( ctx, varname, cmdbuf, name, depth, active ) tracy::VkCtxScope varname( ctx, __LINE__, __FILE__, strlen( __FILE__ ), __FUNCTION__, strlen( __FUNCTION__ ), name, strlen( name ), cmdbuf, depth, active ); +# define TracyVkZoneTransientS( ctx, varname, cmdbuf, name, depth, active ) tracy::VkCtxScope varname( ctx, TracyLine, TracyFile, strlen( TracyFile ), TracyFunction, strlen( TracyFunction ), name, strlen( name ), cmdbuf, depth, active ); #else # define TracyVkNamedZoneS( ctx, varname, cmdbuf, name, depth, active ) TracyVkNamedZone( ctx, varname, cmdbuf, name, active ) # define TracyVkNamedZoneCS( ctx, varname, cmdbuf, name, color, depth, active ) TracyVkNamedZoneC( ctx, varname, cmdbuf, name, color, active )