wrong-code if stdcall is applied via template parameters

[Kevin Puetz]

g++ 9.3.0-17ubuntu1~20.04 miscompiles the following example for x86 (bar_stdcall should pop 4 bytes fewer than an bar_cdecl, as stdcall is callee-cleanup for the argument). To see it in this minimized example one must use `-O0` as it will otherwise be inlined and the calling-convention will no longer be meaningful. Larger examples it can hit with optimizations on.

>>>>>>>>>>>>>>>

// g++-9 -c -O0 simple.cpp -m32
//
// objdump -S simple.o

template<bool bIsStdcall, typename T> struct func_ptr_t;

template<typename T> struct func_ptr_t<false,T> {
 using type = void(*)(T);
};

template<typename T> struct func_ptr_t<true,T> {
#if 1
 using type = void(__attribute__((__stdcall__))*)(T);
#else
 using type = void(__attribute__((__stdcall__))*)(int); // this works, using T is important somehow
#endif
};

#if 1
 using foo_stdcall_ptr = func_ptr_t<true,int>::type;
 using foo_cdecl_ptr = func_ptr_t<false,int>::type;
#else
 using foo_stdcall_ptr = void(__attribute__((__stdcall__))*)(int,int);
 using foo_cdecl_ptr = void(*)(int,int);
#endif

foo_stdcall_ptr foo_stdcall;
foo_cdecl_ptr foo_cdecl;

void bar_stdcall() {
 foo_stdcall(1);
}

void bar_cdecl() {
 foo_cdecl(1);
}

>>>>>>>>>>>>>>>

--- bar_cdecl.S 2021-03-09 00:54:58.404022904 +0000
+++ bar_stdcall.S 2021-03-09 00:52:07.900015002 +0000
@@ -1,15 +1,15 @@
-<bar_cdecl()>:
+<bar_stdcall()>:
    0: f3 0f 1e fb endbr32
    4: 55 push %ebp
    5: 89 e5 mov %esp,%ebp
    7: 83 ec 08 sub $0x8,%esp
- a: e8 fc ff ff ff call 32 <bar_cdecl()+0xb>
+ a: e8 fc ff ff ff call b <bar_stdcall()+0xb>
    f: 05 01 00 00 00 add $0x1,%eax
   14: 8b 80 00 00 00 00 mov 0x0(%eax),%eax
   1a: 83 ec 0c sub $0xc,%esp
   1d: 6a 01 push $0x1
   1f: ff d0 call *%eax
- 21: 83 c4 10 add $0x10,%esp
+ 21: 83 c4 0c add $0xc,%esp
   24: 90 nop
   25: c9 leave
   26: c3 ret
```

Upstream gcc-9.3.0 does not have this bug (See https://godbolt.org/z/vK3TdM318), but 9.4.0, 10.0, and 10.1 do). The regression seems to have happened upstream in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90750 - https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=aa988998be8f85334665a6b049d5d9139408c250 - which ubuntu has backported into git-updates.patch (it's in in gcc-9_9.3.0-17ubuntu1~20.04.debian.tar.xz, I have not pinned down when ubuntu picked this up). The regression was fixed in 10.2 by https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95222 - https://gcc.gnu.org/git/?p=gcc.git;a=commit;efdbd4fbea08005091e490ec3f9972aa9c946374 and https://gcc.gnu.org/git/?p=gcc.git;a=commit;0c473d8f32510fcc96d584ee5099b856cfd3d8d6).

I don't think it was realized (when the was found/fixed upstream) that it could also cause wrong-code, so I'm just posting this to make you aware that ubuntu has backported a regression and not the fix. I assume you'll want to align with upstream on whether gcc-9 branch ends up getting this fixed or not.

[Matthias Klose (doko)]

9.4.0 and 10.5.0 have been backported to 20.04 LTS