radlink: perf + feature series — ICF/ICFSTATIC, GCTYPES, header-units, link-time + memory (rebased on dev, taken commits dropped)

build.bat

@@ -46,6 +46,7 @@ if "%spall%"=="1"                   set auto_compile_flags=%auto_compile_flags%
 if "%asan%"=="1"                    set auto_compile_flags=%auto_compile_flags% -fsanitize=address && echo [asan enabled]
 if "%ubsan%"=="1"                   set auto_compile_flags=%auto_compile_flags% -fsanitize=undefined && echo [ubsan enabled]
 if "%opengl%"=="1"                  set auto_compile_flags=%auto_compile_flags% -DR_BACKEND=R_BACKEND_OPENGL && echo [opengl render backend]
+if "%worklist_selfcheck%"=="1"      set auto_compile_flags=%auto_compile_flags% -DICF_WORKLIST_SELFCHECK=1 && echo [icf worklist selfcheck enabled]
 if "%dwarf%"=="1" if "%clang%"=="1" set auto_compile_flags=%auto_compile_flags% -gdwarf && echo [dwarf debug info]
 if "%dwarf%"==""  if "%clang%"=="1" set auto_compile_flags=%auto_compile_flags% -gcodeview
 if "%pgo%"=="1" (
@@ -142,7 +143,14 @@ pushd build
 if "%raddbg%"=="1"                     set didbuild=1 && %compile% ..\src\raddbg\raddbg_main.c                               %compile_link% %link_icon% %out%raddbg.exe || exit /b 1
 if "%raddbg_non_graphical%"=="1"       set didbuild=1 && %compile% -DWM_STUB=1 -DR_BACKEND=R_BACKEND_STUB ..\src\raddbg\raddbg_main.c %compile_link% %link_icon% %out%raddbg_non_graphical.exe || exit /b 1
 if "%com_shim%"=="1"                   set didbuild=1 && %compile% ..\src\com_shim\com_shim_main.c                           %compile_link% %out%com_shim.exe || exit /b 1
-if "%radlink%"=="1"                    set didbuild=1 && %compile% ..\src\linker\lnk.c                                       %compile_link% %linker% /NOIMPLIB %linker% /NATVIS:"%~dp0\src\linker\linker.natvis" %out%radlink.exe || exit /b 1
+:: NOTE: -DBLAKE3_ATOMICS=1 makes BLAKE3 use C11 _Atomic (plain atomic load) for
+:: get_cpu_features instead of MSVC's _InterlockedOr `lock or` barrier on every
+:: compress dispatch (was a ~5.5s main-thread hot spot). MSVC C11 atomics require
+:: /std:c11 /experimental:c11atomics. Kept external so the vendored blake3 source
+:: stays pristine.
+set radlink_msvc_flags=
+if "%msvc%"=="1" set radlink_msvc_flags=/std:c11 /experimental:c11atomics -DBLAKE3_ATOMICS=1
+if "%radlink%"=="1"                    set didbuild=1 && %compile% %radlink_msvc_flags% ..\src\linker\lnk.c                    %compile_link% %linker% /NOIMPLIB %linker% /NATVIS:"%~dp0\src\linker\linker.natvis" %out%radlink.exe || exit /b 1
 if "%radbin%"=="1"                     set didbuild=1 && %compile% ..\src\radbin\radbin_main.c                               %compile_link% %out%radbin.exe || exit /b 1
 if "%raddump%"=="1"                    set didbuild=1 && %compile% ..\src\raddump\raddump_main.c                             %compile_link% %out%raddump.exe || exit /b 1
 if "%ryan_scratch%"=="1"               set didbuild=1 && %compile% ..\src\scratch\ryan_scratch.c                             %compile_link% %out%ryan_scratch.exe || exit /b 1

src/base/base_arena.c

@@ -346,6 +346,54 @@ arena_pop_to(Arena *arena, U64 pos)
 
 }
 
+//- rjf: arena decommit of unused (rewound/free) pages
+
+internal void
+arena_decommit_unused(Arena *arena)
+{
+  // NOTE(perf): decommit committed-but-unused pages so they stop counting against
+  // working set, while keeping the reservation. Only touches pages strictly above
+  // the live `pos` high-water of each block in the active chain, and the unused
+  // bodies of free-list blocks. Live data (<= pos) is never touched. The push path
+  // re-commits on demand (arena_push grows `cmt`), so reuse is transparent.
+  if(arena->flags & ArenaFlag_LargePages)
+  {
+    // large pages cannot be partially decommitted safely; skip.
+    return;
+  }
+  U64 page_size = get_system_info()->page_size;
+
+  // rjf: active chain -- decommit committed region above each block's live pos
+  for(Arena *n = arena->current; n != 0; n = n->prev)
+  {
+    U64 pos_aligned = AlignPow2(n->pos, page_size);
+    if(pos_aligned < n->cmt)
+    {
+      U8 *decommit_ptr = (U8 *)n + pos_aligned;
+      U64 decommit_size = n->cmt - pos_aligned;
+      AsanPoisonMemoryRegion(decommit_ptr, decommit_size);
+      decommit_memory(decommit_ptr, decommit_size);
+      n->cmt = pos_aligned;
+    }
+  }
+
+#if ARENA_FREE_LIST
+  // rjf: free chain -- decommit everything above the first (header) page
+  for(Arena *n = arena->free_last; n != 0; n = n->prev)
+  {
+    U64 keep = AlignPow2(ARENA_HEADER_SIZE, page_size);
+    if(keep < n->cmt)
+    {
+      U8 *decommit_ptr = (U8 *)n + keep;
+      U64 decommit_size = n->cmt - keep;
+      AsanPoisonMemoryRegion(decommit_ptr, decommit_size);
+      decommit_memory(decommit_ptr, decommit_size);
+      n->cmt = keep;
+    }
+  }
+#endif
+}
+
 //- rjf: arena push/pop helpers
 
 internal void

src/base/base_arena.h

@@ -76,6 +76,9 @@ internal void *arena_push(Arena *arena, U64 size, U64 align, B32 zero);
 internal U64   arena_pos(Arena *arena);
 internal void  arena_pop_to(Arena *arena, U64 pos);
 
+//- rjf: arena decommit of unused (rewound/free) pages
+internal void arena_decommit_unused(Arena *arena);
+
 //- rjf: arena push/pop helpers
 internal void arena_clear(Arena *arena);
 internal void arena_pop(Arena *arena, U64 amt);

src/base/base_thread_context.c

@@ -82,6 +82,21 @@ tctx_get_scratch(Arena **conflicts, U64 count)
   return result;
 }
 
+//- rjf: scratch decommit (release committed-but-unused scratch pages back to OS)
+
+internal void
+tctx_scratch_decommit(void)
+{
+  TCTX *tctx = tctx_selected();
+  for(U64 i = 0; i < ArrayCount(tctx->arenas); i += 1)
+  {
+    if(tctx->arenas[i] != 0)
+    {
+      arena_decommit_unused(tctx->arenas[i]);
+    }
+  }
+}
+
 //- rjf: lane metadata
 
 internal LaneCtx

src/base/base_thread_context.h

@@ -90,6 +90,7 @@ internal TCTX *tctx_selected(void);
 
 //- rjf: scratch arenas
 internal Arena *tctx_get_scratch(Arena **conflicts, U64 count);
+internal void tctx_scratch_decommit(void);
 #define scratch_begin(conflicts, count) temp_begin(tctx_get_scratch((conflicts), (count)))
 #define scratch_end(scratch) temp_end(scratch)
 

src/coff/coff_parse.c

@@ -166,11 +166,16 @@ coff_section_header_array_from_name(Arena *arena, String8 string_table, COFF_Sec
 }
 
 
+// NOTE: name-skipping variants. coff_read_symbol_name does a cstr scan over the
+// memory-mapped string table, which is the dominant cost when parsing symbols in
+// bulk; callers that only need the scalar fields (value/section/storage_class/aux,
+// e.g. symbol-value interpretation) should use these to avoid that scan. The full
+// coff_parse_symbol{16,32} below are these plus the name read, so the scalar-field
+// logic lives in exactly one place.
 internal COFF_ParsedSymbol
-coff_parse_symbol32(String8 string_table, COFF_Symbol32 *sym32)
+coff_parse_symbol32_no_name(COFF_Symbol32 *sym32)
 {
   COFF_ParsedSymbol result = {0};
-  result.name              = coff_read_symbol_name(string_table, &sym32->name);
   result.value             = sym32->value;
   result.section_number    = sym32->section_number;
   result.type              = sym32->type;
@@ -181,10 +186,9 @@ coff_parse_symbol32(String8 string_table, COFF_Symbol32 *sym32)
 }
 
 internal COFF_ParsedSymbol
-coff_parse_symbol16(String8 string_table, COFF_Symbol16 *sym16)
+coff_parse_symbol16_no_name(COFF_Symbol16 *sym16)
 {
   COFF_ParsedSymbol result = {0};
-  result.name              = coff_read_symbol_name(string_table, &sym16->name);
   result.value             = sym16->value;
   if (sym16->section_number == COFF_Symbol_DebugSection16) {
     result.section_number = COFF_Symbol_DebugSection32;
@@ -200,6 +204,22 @@ coff_parse_symbol16(String8 string_table, COFF_Symbol16 *sym16)
   return result;
 }
 
+internal COFF_ParsedSymbol
+coff_parse_symbol32(String8 string_table, COFF_Symbol32 *sym32)
+{
+  COFF_ParsedSymbol result = coff_parse_symbol32_no_name(sym32);
+  result.name              = coff_read_symbol_name(string_table, &sym32->name);
+  return result;
+}
+
+internal COFF_ParsedSymbol
+coff_parse_symbol16(String8 string_table, COFF_Symbol16 *sym16)
+{
+  COFF_ParsedSymbol result = coff_parse_symbol16_no_name(sym16);
+  result.name              = coff_read_symbol_name(string_table, &sym16->name);
+  return result;
+}
+
 internal COFF_ParsedSymbol
 coff_parse_symbol(COFF_FileHeaderInfo header, String8 string_table, String8 symbol_table, U32 symbol_idx)
 {

src/coff/coff_parse.h

@@ -261,6 +261,8 @@ internal String8                 coff_name_from_section_header      (String8 str
 
 internal COFF_ParsedSymbol coff_parse_symbol32(String8 string_table, COFF_Symbol32 *sym32);
 internal COFF_ParsedSymbol coff_parse_symbol16(String8 string_table, COFF_Symbol16 *sym16);
+internal COFF_ParsedSymbol coff_parse_symbol32_no_name(COFF_Symbol32 *sym32);
+internal COFF_ParsedSymbol coff_parse_symbol16_no_name(COFF_Symbol16 *sym16);
 internal COFF_ParsedSymbol coff_parse_symbol  (COFF_FileHeaderInfo header, String8 string_table, String8 symbol_table, U32 symbol_idx);
 
 internal COFF_Symbol32Array coff_symbol_array_from_data_16(Arena *arena, String8 data, U64 symbol_array_off, U64 symbol_count);

src/linker/codeview_ext/ifc.c

@@ -0,0 +1,107 @@
+// Copyright (c) Epic Games Tools
+// Licensed under the MIT license (https://opensource.org/license/mit/)
+
+read_only global U8 g_ifc_signature[4] = { 0x54, 0x51, 0x45, 0x1A };
+read_only global U8 g_uba_signature[4] = { 0x55, 0x42, 0x41, 0x01 }; // "UBA\x01"
+
+internal IFC_File
+ifc_file_read(Arena *arena, String8 path, String8 *error_out)
+{
+  IFC_File ifc = {0};
+  ifc.path = push_str8_copy(arena, path);
+
+  String8 data = lnk_read_data_from_file_path(arena, 0, path);
+  ifc.data = data;
+  if (data.size < 4) {
+    *error_out = push_str8f(arena, "IFC '%S' is too small (%llu bytes)", path, data.size);
+    return ifc;
+  }
+
+  // detect UBA-compressed input (out of scope)
+  if (MemoryMatch(data.str, g_uba_signature, sizeof(g_uba_signature))) {
+    *error_out = push_str8f(arena, "IFC '%S' is UBA-compressed (magic 'UBA\\x01'); materialize a raw .ifc (UBA decompress unsupported)", path);
+    return ifc;
+  }
+
+  // validate signature
+  if ( ! MemoryMatch(data.str, g_ifc_signature, sizeof(g_ifc_signature))) {
+    *error_out = push_str8f(arena, "IFC '%S' has bad signature (expected 54 51 45 1A)", path);
+    return ifc;
+  }
+
+  // --- parse header ---
+  U64 off = 4;
+  if (off + 32 > data.size) { goto truncated; }
+  MemoryCopy(ifc.content_hash, data.str + off, 32);
+  off += 32;
+
+  if (off + 4 > data.size) { goto truncated; }
+  U8 major = data.str[off+0];
+  U8 minor = data.str[off+1];
+  U8 abi   = data.str[off+2]; (void)abi;
+  U8 arch  = data.str[off+3];
+  off += 4;
+
+  // assert version 0.44 + x64; error otherwise (encoding proven only for these)
+  if ( ! (major == 0 && minor == 44)) {
+    *error_out = push_str8f(arena, "IFC '%S' unsupported version %u.%u (expected 0.44)", path, major, minor);
+    return ifc;
+  }
+  if (arch != 2) {
+    *error_out = push_str8f(arena, "IFC '%S' unsupported architecture %u (expected 2 == x64)", path, arch);
+    return ifc;
+  }
+
+  U32 cplusplus;           off += str8_deserial_read_struct(data, off, &cplusplus); (void)cplusplus;
+  U32 string_table_bytes;  off += str8_deserial_read_struct(data, off, &string_table_bytes);
+  U32 string_table_size;   off += str8_deserial_read_struct(data, off, &string_table_size);
+  U32 unit;                off += str8_deserial_read_struct(data, off, &unit); (void)unit;
+  U32 src_path;            off += str8_deserial_read_struct(data, off, &src_path); (void)src_path;
+  U32 global_scope;        off += str8_deserial_read_struct(data, off, &global_scope); (void)global_scope;
+  U32 toc;                 off += str8_deserial_read_struct(data, off, &toc);
+  U32 partition_count;     off += str8_deserial_read_struct(data, off, &partition_count);
+  if (off > data.size) { goto truncated; }
+
+  // string table
+  if ((U64)string_table_bytes + string_table_size > data.size) {
+    *error_out = push_str8f(arena, "IFC '%S' string table out of bounds", path);
+    return ifc;
+  }
+  String8 string_table = str8(data.str + string_table_bytes, string_table_size);
+
+  // --- partition summary table ---
+  String8 needle = str8_lit(".msvc.trait.debug-records");
+  U64     po     = toc;
+  for (U32 i = 0; i < partition_count; ++i, po += 16) {
+    if (po + 16 > data.size) { goto truncated; }
+    U32 name_off, p_off, count, entity_size;
+    str8_deserial_read_struct(data, po + 0,  &name_off);
+    str8_deserial_read_struct(data, po + 4,  &p_off);
+    str8_deserial_read_struct(data, po + 8,  &count);
+    str8_deserial_read_struct(data, po + 12, &entity_size);
+
+    if (name_off >= string_table.size) { continue; }
+    String8 name = str8_cstring((char *)string_table.str + name_off);
+    if (str8_match(name, needle, 0)) {
+      // entity_size == 1 -> count is a byte length
+      if ((U64)p_off + count > data.size) {
+        *error_out = push_str8f(arena, "IFC '%S' debug-records partition out of bounds", path);
+        return ifc;
+      }
+      ifc.debug_records = str8(data.str + p_off, count);
+      break;
+    }
+  }
+
+  if (ifc.debug_records.size == 0) {
+    *error_out = push_str8f(arena, "IFC '%S' has no '.msvc.trait.debug-records' partition", path);
+    return ifc;
+  }
+
+  ifc.is_valid = 1;
+  return ifc;
+
+truncated:
+  *error_out = push_str8f(arena, "IFC '%S' is truncated", path);
+  return ifc;
+}

src/linker/codeview_ext/ifc.h

@@ -0,0 +1,43 @@
+// Copyright (c) Epic Games Tools
+// Licensed under the MIT license (https://opensource.org/license/mit/)
+
+#pragma once
+
+////////////////////////////////
+// MSVC IFC (header-unit module interface) reader
+//
+// radlink consumes the `.msvc.trait.debug-records` partition embedded in an
+// MSVC `.ifc` (C++20 module / header-unit interface) file. That partition is a
+// raw CodeView type-leaf stream (entity_size == 1, NO u32 signature, first leaf
+// at offset 0, TI base 0x1000). A consuming `.obj` references this stream via
+// LF_IFC_RECORD (0x1522) leaves -- see lnk_debug_info.c.
+//
+// File layout (microsoft/ifc-spec, FileHeader):
+//   u8[4]  signature = { 0x54,0x51,0x45,0x1A }   ("TQE\x1a")
+//   u8[32] content_hash (sha256)  -- record.GUID(16)++record.hash(16) == first 32 bytes here
+//   u8     major, minor           -- assert 0.44
+//   u8     abi
+//   u8     arch                   -- 2 == x64
+//   u32    cplusplus
+//   u32    string_table_bytes (off), u32 string_table_size
+//   u32    unit
+//   u32    src_path (textoffset)
+//   u32    global_scope
+//   u32    toc                    -- offset to partition summary table
+//   u32    partition_count
+//   u8     internal_partition
+// Partition summary entry (16 bytes): { u32 name(textoffset); u32 offset; u32 count; u32 entity_size }
+
+typedef struct IFC_File
+{
+  String8 data;          // whole .ifc bytes (owning view into arena)
+  String8 path;          // .ifc path (copied)
+  U8      content_hash[32];
+  String8 debug_records; // .msvc.trait.debug-records blob {ptr,size}; size 0 if absent
+  B32     is_valid;
+} IFC_File;
+
+// Reads `path`, validates magic/version/arch, locates `.msvc.trait.debug-records`.
+// On error fills *error_out and returns is_valid=0. Detects UBA-compressed inputs
+// (magic "UBA\x01") and reports them (decompression is out of scope).
+internal IFC_File ifc_file_read(Arena *arena, String8 path, String8 *error_out);