[Bug 50171] New: 32-bit PE entry point no longer called through BaseThreadInitThunk() assembly wrapper when Wine is built with LLVM MinGW

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=50171

Bug ID: 50171
           Summary: 32-bit PE entry point no longer called through
                    BaseThreadInitThunk() assembly wrapper when Wine is
                    built with LLVM MinGW
           Product: Wine
           Version: 5.22
          Hardware: x86-64
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P2
         Component: kernel32
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

Hello folks,

while investigating bug 28332 I found a regression introduced by
https://source.winehq.org/git/wine.git/commitdiff/0c631ebb2354334eaf309bc076...
("kernel32: Build with msvcrt.").

With 32-bit kernel32 cross-compiled to PE format, the stack on app entry point
is now aligned to 16-byte boundary. This makes the game from bug 28332
"magically" work while the 32-bit non-PE kernel32 causes a crash (expected due
to Wine quartz bug).

https://bugs.winehq.org/show_bug.cgi?id=28332#c17

The stack alignment from entry point, including selected functions from the
call chain up to alloca() caller.

--- snip ---
EIP      | PE ESP        | non-PE ESP    | comment
0x42D5E3 | 0x31FF60 +00  | 0x31FF44 +00  | entry point main thread
0x42A720 | 0x31FEA0 +C0  | 0x31FE84 +C0  |
0x4200C5 | 0x31FE54 +4C  | 0x31FE38 +4C  |
0x40D8B7 | 0x31FE4C +08  | 0x31FE30 +08  |
0x40D570 | 0x31FE48 +04  | 0x31FE2C +04  |
0x40D6FF | 0x31FB1C +32C | 0x31FB00 +32C | before alloca(0)
0x40D704 | 0x31FB10 +0C  | 0x31FB00 +00  | after alloca(0)
--- snip ---

It seems in 32-bit PE kernel32.dll, the thread entry point is called through
the fastcall "C" version of BaseThreadInitThunk() instead of the fastcall
assembly wrapper.

32-bit non-PE kernel32.dll.so works correctly, the thread entry point is called
through __fastcall_BaseThreadInitThunk() assembly wrapper.

https://source.winehq.org/git/wine.git/blob/bedfb9cae224a369efa4588332a5518d...

--- snip ---
  36 #ifdef __i386__
  37 __ASM_STDCALL_FUNC( __fastcall_BaseThreadInitThunk, 12,
  38                     "pushl %ebp\n\t"
  39                     __ASM_CFI(".cfi_adjust_cfa_offset 4\n\t")
  40                     __ASM_CFI(".cfi_rel_offset %ebp,0\n\t")
  41                     "movl %esp,%ebp\n\t"
  42                     __ASM_CFI(".cfi_def_cfa_register %ebp\n\t")
  43                     "pushl %ebx\n\t"
  44                     __ASM_CFI(".cfi_rel_offset %ebx,-4\n\t")
  45                     "movl 8(%ebp),%ebx\n\t"
  46                     /* deliberately mis-align the stack by 8, Doom 3 needs
this */
  47                     "pushl 4(%ebp)\n\t"  /* Driller expects readable
address at this offset */
  48                     "pushl 4(%ebp)\n\t"
  49                     "pushl %ebx\n\t"
  50                     "call *%edx\n\t"
  51                     "movl %eax,(%esp)\n\t"
  52                     "call " __ASM_STDCALL( "RtlExitUserThread", 4 ))
  53 #endif
  54 
  55 /***********************************************************************
  56  *           BaseThreadInitThunk (KERNEL32.@)
  57  */
  58 void __fastcall BaseThreadInitThunk( DWORD unknown, LPTHREAD_START_ROUTINE
entry, void *arg )
  59 {
  60     RtlExitUserThread( entry( arg ) );
  61 }
--- snip ---

https://source.winehq.org/git/wine.git/blob/bedfb9cae224a369efa4588332a5518d...

--- snip ---
 185 @ stdcall -fastcall BaseThreadInitThunk(long ptr ptr)
--- snip ---

cross-toolchain used: LLVM MinGW 20201020 with LLVM stable 11.0.0

https://github.com/mstorsjo/llvm-mingw

https://github.com/mstorsjo/llvm-mingw/releases

https://github.com/mstorsjo/llvm-mingw/releases/download/20201020/llvm-mingw...

$ wine --version
wine-5.22

Regards