radlink: link-time performance pass (~1.9x on a Fortnite link)

build.bat

@@ -141,7 +141,14 @@ popd
 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 "%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_strings.c

@@ -213,7 +213,9 @@ str8_cstring_capped(void *cstr, void *cap)
 {
   char *ptr = (char *)cstr;
   char *opl = (char *)cap;
-  for (;ptr < opl && *ptr != 0; ptr += 1);
+  // memchr is typically SIMD-accelerated; much faster than a byte-by-byte scan
+  char *nul = (char *)memchr(ptr, 0, (U64)(opl - ptr));
+  ptr = nul ? nul : opl;
   U64 size = (U64)(ptr - (char *)cstr);
   String8 result = str8((U8*)cstr, size);
   return result;

src/base/base_threads.h

@@ -129,6 +129,7 @@ internal Semaphore semaphore_open(String8 name);
 internal void      semaphore_close(Semaphore semaphore);
 internal B32       semaphore_take(Semaphore semaphore, U64 endt_us);
 internal void      semaphore_drop(Semaphore semaphore);
+internal void      semaphore_drop_n(Semaphore semaphore, U32 count); // release `count` permits in one syscall
 
 //- rjf: barriers
 internal Barrier barrier_alloc(U64 count);

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);