Re: i386.h fix for 4.0 branch (pr16888 fallout)

[Jakub Jelinek <jakub at redhat dot com>]

On Wed, May 11, 2005 at 05:34:39PM +0200, Michael Matz wrote:
> In 4.0 now.  In retrospect I actually think that Alexandres change might 
> not have been such a good idea after all.  The problem is, that now you 
> can't compile anymore programs which do the checking for MMX capability at 
> runtime, i.e. something like:
>    if (cpu_supports_mmx()) {
>      asm do_stuff_with_mmx;
>    } else {
>      do_stuff_in_plain_c();
>    }
> 
> like is done in some media code.  With Alexandres changes you have to 
> specify -mmmx (or another option of that effect) now to even compile the 
> above.  The problem being that this activates TARGET_MMX (or _SSE{,2} for 
> that matter), and hence activates that instruction set also generally, 
> i.e. potentially on code paths not guarded by the above if written by the 
> programmer.  Hence you are now unable to compile such code which still 
> runs (slower) on non-mmx/sse capable processors.

Without -msse etc. you can't use x/y constraints, so the only thing
you talk about here are __asm clobbers, right?
For those you can use
#ifdef __SSE__
  "xmm0", "xmm6"
#endif
around SSE/SSE2 clobbers and
#ifdef __MMX__
  "mm0", "mm5"
#endif
around MMX clobbers.
Without those options (and therefore defines), GCC doesn't know anything
about those registers, so there is no point telling GCC about them.

	Jakub