This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix combiner i3/i2 pattern splitting (PR rtl-optimization/45695)
On Sat, Sep 18, 2010 at 05:17:58PM +0200, Eric Botcazou wrote:
> > In the PR rtl-optimization/44858 fix I've been assuming that the comment
> > claiming the order doesn't really matter and HAVE_cc0 implementation are
> > right, unfortunately it seems it was wrong. We tested unconditionally
> > && ! use_crosses_set_p (SET_SRC (XVECEXP (newpat, 0, 1)),
> > DF_INSN_LUID (i2))
> > but for HAVE_cc0 could have chosen to put XVECEXP (newpat, 0, 0)
> > into i2 instead of i3 (then the use_crosses_set_p test is irrelevant
> > to it and nothing tests that the pattern we want to put into i2
> > doesn't depend on something set by insns in between i2 and i3.
> > And my patch of course made things even worse, as it tries to swap
> > the patterns if added clobbers disallow the initially chosen order, even
> > on !HAVE_cc0 targets.
>
> How about erring on the side of caution and trimming down the fix for 44858
> to make it identical to that of 20322? Would that pessimize too much?
HAVE_cc0 targets would still be buggy. If you want I can gather statistics
on how many cases would match with the swapped patterns compared to without
that.
Jakub