Re: [PATCH] Do not set BB_SUPERBLOCK in loop-unroll.c:split_edge_and_insert

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

> This patch is a follow-up to a call-for-help I sent out last week:
> http://gcc.gnu.org/ml/gcc/2007-02/msg00379.html
> 
> Further testing still doesn't show any breakage.  Also, it turns out that
> loop-unswitch.c also uses compare_and_jump_seq(), without considering the
> possibility that compare_and_jump_seq may produce insn chains with more
> than one jump.  Apparently the BB_SUPERBLOCK trick in loop-unroll.c is
> there to handle that possibility, but if loop-unswitch doesn't have to do
> this, then I'm guessing that loop-unroll also doesn't need it.
> 
> So I'd like to just remove it, and replace it with a sanity check.
> 
> The worst thing that could happen with the attached patch is more test
> suite failures iff something is broken by this change.  I will revert the
> patch and figure out a fix in case something does break.
> 
> Bootstrapped&tested on {i686,x86_64,ia64}-unknown-linux-gnu, plus the
> testing mentioned in the replies to my cfg (including bootstrap on alpha
> with loop unrolling enabled).
> OK for the trunk?

I do not dare to approve this; introducing code that relies on
"blackbox" understanding of what some functions do seems a bit dangerous
to me.

Although it may be quite difficult, I think that the only clean way is
to find out why force_operand/compare_and_jump_seq do not emit control
flow insns when they are used as in loop-unroll.c, and base a solution
on this (which may turn out to be just extending the patch below by
documenting in which cases these functions are guaranteed not to
introduce control flow insns).

Zdenek