Re: [PATCH 3/6] Allow jumps in epilogues

[Richard Henderson <rth at redhat dot com>]

On 03/25/2011 10:34 AM, Bernd Schmidt wrote:
> I don't know much about the unwinding code. I'm currently thinking about
> writing out a cfi_remember_state at the start of the function, restoring
> that clean state when necessary at the start of a new block and emitting
> the necessary directives to reach the correct state. What directives
> should I expect to be required? Can I get by just with cfi_offset and
> cfi_def_cfa_offset, or will something else be necessary?

Yes, several things: register, expression, gnu_args_size, perhaps a few more.

I think the ideal thing would be a pass while the cfg is still extant that
captures the unwind info into notes; these can be recorded at basic block
boundaries, so that they persist until the end of compilation.

So long as late late compilation passes continue to not move frame-related
insns across basic block boundaries, we should be fine.

Irritatingly, the exact place to locate this pass is difficult to pin down.
Immediately before md_reorg is the last place we have the cfg.  But we do
strange things in, e.g. ia64 where we rebuild the cfg and run sched_ebb
during md_reorg.

Of course, ia64 is a bad example because its unwind info is target-specific,
and quite a lot of the possible benefit of shrink wrapping is lost via the
register windowing.

I'm willing to work with you on the problem of cfg-aware unwind info.  We
have needed this for a really long time; there are existing bugs related 
to exception handling and !ACCUMULATE_OUTGOING_ARGS that would be fixed by
this.


r~