November 5th is Clang-Day

[Nathan Whitehorn]

On 11/04/12 08:29, Dimitry Andric wrote:
> On 2012-11-04 14:18, Konstantin Belousov wrote:
>> On Sun, Nov 04, 2012 at 02:42:13PM +0200, David Naylor wrote:
> ...
>>> I tried building (using gcc) wine with your patch and now (at least) 
>>> winecfg
>>> and regedit work with a clang built lib32.  I'll email Gerald (wine's
>>> maintainer) about including your patch in wine.
>>
>> The wine is the wrong place to fix. If system libraries suddenly started
>> requiring 16-byte stack alignment on i386, it is unacceptable breakage
>> of the ABI.
>
> So we really must use 4 byte stack alignment on i386 by default? I have
> attached a diff to llvm for this, but I would like to verify that it is
> really correct.  Apparently Darwin, Linux and Solaris all use 16 byte
> alignment.
>
> The Sys V ABI seems to say only: "The stack is word aligned. Although
> the architecture does not require any alignment of the stack, software
> convention and the operating system requires that the stack be aligned
> on a word boundary".

This is an ugly business. The stack is really 4 bytes aligned on Linux 
and Solaris too, but GCC decided to unilaterally change the ABI a few 
years ago (also on FreeBSD). You can find a chunk of the sordid story of 
this in a number of GCC bugs marked WONTFIX (e.g. 
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38496). To pacify the people 
who were yelling about it, they added some __attribute__(()) foo and 
compiler flags to change what it uses for specific programs and add 
magic to realign the stack at boundaries. So much water has passed under 
the bridge at this point as to reach flood stage, so I think the correct 
solution here is to add the same __attribute__(()) and flags to clang 
rather than changing the default alignment back to 4. Changing it only 
on FreeBSD is especially wrong.
-Nathan