backend question in alpha nobwx

backend question in alpha nobwx

[Spundun Bhatt]

Hi gcc
6 months back I was pointed to alpha backend with nobwx for an example
of architecture with no HI or QI memory accesses. I went through the mov
functions and helper templates.... 

So my question is how is gcc making sure (in alpha backend) that during
reload, HI and QI moves from registers to memory dont take place(They
are not handled by the alpha_expand_mov_nobwx)? To me it looks like,
unless you take special precautions to avoid that.. you will probably
run in to situations where this would happen. 

I tried to merge these alpha hacks  into rs6000(thats what I wanted to
do)... its working for many cases... but I get ICE in gen_reg_rtx for
many other cases. Apparantly its calling "gen_reg_rtx" during reload
(used ddd). 
In alpha_expand_mov_nobwx around line 3390.... its calling gen_reg_rtx.
Which means that its trying to move HI reg contents to memory during
reload and theres no clean way of handling it. This is happening when I
use that code in the rs6000 backend.. but not at its original alpha
backend. 
I have spent quite a lot of time after this during past couple of
months.... I would really appreciate if some one could provide some clue
to this.
Thanx a lot
Spundun
ps: Good luck with the gcc 3.3 release.

Re: backend question in alpha nobwx

[Richard Henderson]

On Wed, Mar 12, 2003 at 10:00:12PM -0800, Spundun Bhatt wrote:
> So my question is how is gcc making sure (in alpha backend) that during
> reload, HI and QI moves from registers to memory dont take place

reload_{in,out}{h,q}i.


r~

Re: backend question in alpha nobwx

[Spundun Bhatt]

Thanx a lot for your reply. The thing that was wrong was that I had to
modify SECONDARY_RELOAD_OUTPUT/INPUT_CLASS macros.
Desperate for a solution... I searched on google and found a clue to
this from a Changelog entry that you made in 1997 ! 
Thanx again.
Spundun.
On Thu, 2003-03-13 at 14:35, Richard Henderson wrote:
> On Wed, Mar 12, 2003 at 10:00:12PM -0800, Spundun Bhatt wrote:
> > So my question is how is gcc making sure (in alpha backend) that during
> > reload, HI and QI moves from registers to memory dont take place
> 
> reload_{in,out}{h,q}i.
> 
> 
> r~