SSE intrinsics + auto-alignment (3.2.2)

[Jan Hubicka]

> Heya dudes,
> 
> 
> Thus, Is it planned to correct this issue for GCC3.3 release and above? In the 
> negative, why 
> and what does it involve? :-p

I've implemented aligning of structures and it is for sure in 3.3 and
mainilne branches.  I am surprised it is not in 3.2, I will check.
(hope it works in all cases, I would be interested in seeing any
testcases that fails)

There also has been number of misscompilations of the intrincisc even in
last official release.  For 3.3, I can now build quite large SSE
testsuite (involving every builtin with several possible inputs) and get
no failures.
There are problems remaining tought.

Most importantly function arguments are not aligned at all, so SSE
argumetns must come first, otherwise program crash.  Additionally
there is no dynamic alignment of stack frames, but GCC can maintain the
aligned stack.  This does not work for function main, so you must avoid
any use of SSE in main().
I have patches for the first.  The problem is that we didn't came to
conclusion on the ABI.  Apparently new ICC is passing SSE arugments in
registers, while old one in memory.
Redhat is having patches for the second.  Not sure about the status of
them.

Honza
> 
> Thank u!
> 
> tchOo!
> 
> Ozzy.
> 
> ps: (is there any forum btw just like gl forums @opengl.org  for gcc purposes?)