[jh@suse.cz: Re: target/9929: [3.3/3.4 regression] Can't find spill register]

[Jan Hubicka]

> > > (2) If we don't allow DImode here, we shouldn't allow SUBREG at all.
> > 
> > Shall we allow them only for 64bit mode then (where they are common 
> > for 32bit arithmetics?)
> 
> I guess.  I'd rather we would limit this to leal when we're
> sure we're producing an SImode result.

What we do currently is to split the leas into DImode leas post reload
and I vaguely remember that in some cases the subregs may remain.  Will
double check.

legitimate_address does not know when the operand is lea and we must use
"p" constraint as it is handled specially by reload and regstack.  Not
sure how this can be implemented.
Also I seem to remember that the acceptance of subregs was required to
avoid ICE on 64bit compilation but do not recall why.

Why reload insist on loading the whole 64bit register at all?

Honza
> 
> 
> r~