pre_inc/pre_dec/PUSH_ROUNDING inconsistency

[Jan Hubicka]

> On Mon, Jul 17, 2000 at 11:10:40PM +0200, Jan Hubicka wrote:
> > Yes, but whats about the floating point stores. I remember Jeff claiming
> > that (subreg:TF (reg:XF)) is incorrect...
> > is (subreg:TI (reg:XF)) better?
> 
> Yes, the TFmode subreg is incorrect.  As for the other... well,
> I guess it's not too bad.  I'm not sure what to put in its place.
OK. I will go with the integer subregs (it will require some care
in machine description, since integer pushdi instruction will be
receiving floats too... looks quite ugly.
It can also force reload to move value from FP reigsters, not being able
to hold TImode to integer before actually pushing. :(
Also it will probably inhibit the memory operand for such push instructions.
While most architectures are probably not crazy enought to define
"zero extend and push from memory" instruction, one I am working at
can handle such stuff by doing actually two instruction, one for data
and one for padding, saving an register.

Honza
> 
> 
> 
> r~