On Wed, Dec 16, 2020 at 11:17:29AM -0600, Segher Boessenkool wrote:
> It was just a bootstrap+regression check, so no new testcase was needed.
> I don't remember what target, but powerpc (32+64, BE) probably.
I'll bootstrap/regtest the patch on powerpc64 and powerpc64le then.
>
> > The testcase in the patch (which I've excended today compared to what
> > the patch had yesterday) has the cases of a single as well as multiple
> > edges to the ideal prologue bb in the cold partition, yet I can't reproduce
> > the ICE.
> > As for what changed, I remember fixing PR87475 which is related to
> > redirection of crossing jumps.
>
> But that is for cfglayout mode? Shrink-wrap has to do everything in
> cfgrtl mode unfortunately, because of the partitioning stuff. Sounds
> like a good cancidate otherwise.
The patch affects both modes. patch_jump_insn would in some cases ICE
and now it doesn't, instead returns false and rtl_redirect_edge_and_branch_force
then can deal with it another way.
Jakub
