Alignment testing

[Hans Leidekker]

Hi,

kernel/tests/generated.c currently shows one failure
when the executable is compiled with MinGW. Investigation
shows that MinGW and GCC differ when it comes to alignment
of doubles (i.e. a 64 bit wide type).

I have constructed the test below that shows the difference
between MinGW and GCC. My question is, could someone compile
this with MSVC >= 6, run it, and report the outcome here?

Thanks,

 -Hans

[Dmitry Timoshkov]

"Hans Leidekker" <hans(a)it.vu.nl> wrote:
> kernel/tests/generated.c currently shows one failure
> when the executable is compiled with MinGW. Investigation
> shows that MinGW and GCC differ when it comes to alignment
> of doubles (i.e. a 64 bit wide type).
>
> I have constructed the test below that shows the difference
> between MinGW and GCC. My question is, could someone compile
> this with MSVC >= 6, run it, and report the outcome here?

I had to add #include <windows.h> as the very first line to make
it compile.

C:\>cl
Microsoft (R) 32-bit C/C++ Optimizing Compiler Version 12.00.8804 for 80x86
Copyright (C) Microsoft Corp 1984-1998. All rights reserved.

C:\>align.exe
Alignment of __int64: 8
Size of __int64: 8
Alignment of large_int: 8
Size of large_int: 8
Alignment of container1: 8
Size of container1: 16
Alignment of container2: 8
Size of container2: 16
Alignment of container3: 8
Size of container3: 8

-- 
Dmitry.

[Hans Leidekker]

Hi,

[Use a fixed width font to view this message]

                          MSVC   MinGW  GCC

Alignment of __int64:     8      8      8
Size of __int64:          8      8      8

Alignment of large_int:   8      8      4
Size of large_int:        8      8      8

Alignment of container1:  8      8      4
Size of container1:       16     16     12

Alignment of container2:  8      8      4
Size of container2:       16     16     12

Alignment of container3:  8      8      4
Size of container3:       8      8      8

Clearly GCC is the odd one out. When a double is embedded in
a structure GCC does not adjust the alignment of the structure
to the size of a double.

So, should we add -malign-double to the compiler flags?

 -Hans

[Alexandre Julliard]

Hans Leidekker <hans(a)it.vu.nl> writes:
> Clearly GCC is the odd one out. When a double is embedded in
> a structure GCC does not adjust the alignment of the structure
> to the size of a double.
>
> So, should we add -malign-double to the compiler flags?

No, that would break Unix compatibility. We have to explicitly pad the
structures that need it.

-- 
Alexandre Julliard
julliard(a)winehq.org

[Dmitry Timoshkov]

"Hans Leidekker" <hans(a)it.vu.nl> wrote:
> [Use a fixed width font to view this message]
>
> MSVC   MinGW  GCC
>
> Alignment of __int64:     8      8      8
> Size of __int64:          8      8      8
>
> Alignment of large_int:   8      8      4
> Size of large_int:        8      8      8
>
> Alignment of container1:  8      8      4
> Size of container1:       16     16     12
>
> Alignment of container2:  8      8      4
> Size of container2:       16     16     12
>
> Alignment of container3:  8      8      4
> Size of container3:       8      8      8
>
> Clearly GCC is the odd one out. When a double is embedded in
> a structure GCC does not adjust the alignment of the structure
> to the size of a double.
>
> So, should we add -malign-double to the compiler flags?

After another inspection of Wine's dlls/kernel/tests/generated.c
I think that MSVC case was actually never tested!

Here is an example of the code:

#if defined(_MSC_VER) && (_MSC_VER >= 1300) && defined(__cplusplus)
# define FIELD_ALIGNMENT(type, field) __alignof(((type*)0)->field)
#elif defined(__GNUC__)
# define FIELD_ALIGNMENT(type, field) __alignof__(((type*)0)->field)
#else
/* FIXME: Not sure if is possible to do without compiler extension */
#endif

I don't know what version of MS cl.exe has _MSC_VER >= 1300, but one
that comes with Visual Studio 6.0 has _MSC_VER set to 1200. Moreover,
since the test is a .c file __cplusplus is never defined, therefore
FIELD_ALIGNMENT is not defined by the above code and later in that
file we have:

#ifdef FIELD_ALIGNMENT
# define TEST_FIELD_ALIGNMENT(type, field, align) \
   ok(FIELD_ALIGNMENT(type, field) == align, \
       "FIELD_ALIGNMENT(" #type ", " #field ") == %d (expected " #align ")\n", \
           FIELD_ALIGNMENT(type, field))
#else
# define TEST_FIELD_ALIGNMENT(type, field, align) do { } while (0)
#endif

i.e. FIELD_ALIGNMENT gets an empty definition. Same for _TYPE_ALIGNMENT.

Patrik, did you actually got it working with MS cl.exe?

-- 
Dmitry.