Re: Branch created: gcc-3_3-e500-branch

[Jan Hubicka <jh at suse dot cz>]

> 
> On Wednesday, April 16, 2003, at 11:53  AM, Kumar Gala wrote:
> 
> >What is left to submit / get working in the mainline w/respect to the 
> >ABI?
> >
> 
> varargs are broken because of GCC infrastructure problems for subregs 
> of simd types on hard registers.  I suggested an approach, but Jan 
> Hubicka had another idea which he hasn't implemented.  I may kludge it 
> on the e500 branch, since there doesn't seem to be an agreement on how 
> to fix GCC.

We should proceed somewhat on the SIMD stuff.  I probably won't be able
to work on it in next 3 weeks, so it is bit long since I am burned alive
in my TODO list (bugfixes, merging of patches for 3.3 and so on) and I
am leaving for study in Barcelona.  In case for Motorola bits
implementing the named patterns to extract/set the specific field of
vector (not the RTL construct as I believe we don't need it because we
do have VEC_SELECT/VEC_MERGE/VEC_DUPLICATE), please go forward for it.
I beleive there will be needed anyway even when we will need some more
complex machinery to fit all the weird side corners of SSE instruction
set.

Sorry if I delayed you by being overcomplicating the issue.

Honza
> 
> long doubles are broken on purpose on mainline.  The ABI specifies 128 
> bit long doubles, but I have left them as 64 bits since there are no 
> emulation bits for powerpc-eabi*.
> 
> vector types going on varargs, in the named section of the arguments, 
> are supposed to be split.  They currently are not.
> 
> Everything above I'm "working on".
> 
> Finally, there are a mess of multilibs to be built by default.  I have 
> a patch, but I'm still testing the multilibs before I commit to 
> mainline.