Re: decrement and branch optimization broken?

[Roman Zippel <zippel at linux-m68k dot org>]

Hi,

Richard Henderson wrote:

> > I've found the problem, recog_for_combine removes the old notes prior to
> > calling recog, but this pattern looks for REG_NONNEG.
> 
> You should really update to the new doloop patterns then.
> We don't need those notes for that pass, and the reason the
> notes were removed here is that they were no longer correct.

I've looked into this and while it's not difficult to add this pattern,
it's not really useful. :(
doloop_valid_p is very restrictive, so this pattern is only generated
for a minority of possible cases and I only get a loop register with
SImode, instead of a possibly better HImode.
If the REG_NONNEG note is really not usable anymore, the documentation
about loop patterns needs a serious update.

bye, Roman