Re: ix86 qi/hi to si mode promoting patch.

[Jan Hubicka <hubicka at atrey dot karlin dot mff dot cuni dot cz>]

> On Mon, Feb 28, 2000 at 03:39:43PM +0100, Jan Hubicka wrote:
> > I believe that most speedups are now due to avoided prefixes in spilling
> > himodes to stack. It would be nice to reach this w/o extra movzx overhead.
> > (IE teach reload and stuff to use SImode stack slots).
> 
> Yes, this would be very useful elsewhere too.  Especially Alphas
> without byte load/store instructions.
I guess Sparc is another one.
I will take a look how to implement this. It would be nice to make
XFmode spilling of fp registers as the side effect too.
Do you have some suggestions where to start?
I was taking a look at this once, but failed to understand how does
reload asks for the stack slots.

Another scheme I was thinking about is to modify rtl generation pass,
to use SImode registers, but to not preserve invariant, that it's
value is extended HImode value and just keep upper half undefined
and emit extension before instruction that needs it, not after the instruciton,
as it does now.
The rationale behind is that most common instruction s (moves, additions,
bit operations) don't nbeed it and machine description would provide limited
set of HImode/QImode/Simode for 64 bit patterns for thinks such as comparisons
(sparc is such case).
This may result in fewer extensions out of rtl generation pass, but on the other
hand it will most probably confuse thinks such as loop optimization and
value range propagation in near future.
So it is probably missleading, but interesting.
Honza