On Mar 16, 2011, at 10:12 AM, Richard Henderson wrote:
> On 03/16/2011 04:48 AM, Nick Clifton wrote:
>> (peephole): Add peephole to combine zero- and sign- extending
>> loads with arithmetic instructions.
>
> Do you really need peepholes for this? I would have thought that
> merely adding the instruction patterns for this would have been
> enough to encourage combine to do its job merging these patterns...
I've caught 4.5.x not handling really, really simple things like:
unsigned char ub;
unsigned short us;
unsigned int ui;
unsigned long ul;
signed char sb;
signed short ss;
signed int si;
signed long sl;
main() {
us = ub;
ui = us;
}
because there were two zero_extends, instead of just one, my patterns only
covered 1. I've love a more comprehensive solution, where I can explain my
loads can do either sign or zero extend, and simply not have to worry about all
the knock on patterns. Another nit, I had a QI->DI extending load pattern, but
when asked to do a QI-SI load, it didn't want to widen to a QI->DI load and
then just subreg the SI part. :-( The scary thing, it isn't like zero
extension is a brand new instruction that had been invented last week.
