Re: SSE fix 12 - argument alignment PR pending/8212

[Jan Hubicka <jh at suse dot cz>]

> On Mon, Oct 21, 2002 at 11:41:11PM +0200, Jan Hubicka wrote:
> > How do we handle MEMs pointing to stack with PREFERRED_STACK_BOUNDARY?
> 
> We should be marking as STACK_BOUNDARY aligned.
> 
> > Or is there mechanism to say when PREFERRED_STACK_BOUDNARY is seen as
> > requirement and when as recommendation?
> 
> No.  That would be the alternate entry point stuff that
> never got in.

However only way to deal with SSE with sane performance is to assume
that entry point is aligned.  Otherwise we need to avoid memory operands
in all operations except for loads that would pretty much invalidate
the use of SSE builtins to get programs faster.

ICC and MSVC for some time didn't have alternative entry points eighter
and used same strategy as GCC did I believe. Our default setting of
preferred-stack-boundary and the fact that whole distros are built with
it pretty much ensures that this scheme works in practice now.

Honza
> 
> 
> r~