Re: INSN_SCOPE changes break mips-elf

[Jan Hubicka <jh at suse dot cz>]

>    From: law@redhat.com
>    Date: Wed, 19 Jun 2002 14:42:53 -0600
>    
>    mips-elf has had regressions ever since the INSN_SCOPE changes were installed
>    into the tree.  While the number of regressions has gone down since the
>    introduction of the INSN_SCOPE changes one regression still remains.
>    
>    The remaining regression is gcc.dg/debug/debug-6 where we manage to lose
>    all debug information for "xyzzy".  I believe the fundamental problem is your
>    code in scope_to_insns_initialize does not properly handle SEQUENCEs as
>    generated by the delay slot filling pass.
> 
> For the record, this is failing on the sparc targets too.  I bet it is
> failing elsewhere...

I am having entrance exams in half an hour, I will investigate the bug
afterwards.  What happends in this cases is that some optimization kills
the code sequence.  For i386 it happent to be ifcvt I already fixed, for
Sparc it can be something else.

The testcase is not really good, as the variable is just assigned once,
so the block gets easilly elliminated.  I believe that in case combine
succeeds to merge the set with previous instruction, there is no way to
preserve the scope region as there is no way to describe currently that
the instruction does stuff from multiple statements in any existing
debug output.

Honza