Re: [rfc] exception handling in gimple

[Richard Henderson <rth at redhat dot com>]

On 09/09/2009 07:34 AM, Michael Matz wrote:
> While reading the except.c hunks I noticed this:
>
> +      /* Handle three cases here:
> +        (1) no reachable handler and no edge,
> +        (2) reachable handler and an existing edge to post-landing-pad,
> +        (3) reachable handler and a missing edge.
> +        ??? Not sure why case 3 slips through to this point, it sure
> +        seems like we shouldn't be introducing new statements that can
> +        throw during the gimple->rtl lowering process.  The alternative
> +        is that we're losing edges somewhere, which is equally bad.  */
>
> Have you meanwhile found out why case 3 can happen there?

No, though I have a theory.  I'll experiment with that today.

> Also, in the sjlj RTL expanders it's still building a cmp-jump sequence
> in the (one) landing pad.  Shouldn't this eventually also be done on the
> tree-side already (to be able to use a switch)?

I'm not sure I can generate the same rtl from trees, because of where I
go off and place all the bits from the parts of 
expand_builtin_setjmp_receiver.

I do think I can rearrange things such that I can call back into 
expand_case though.  We'll see.


> +/* YOUAREHERE: Replace with expand_resx.  */

Oops.  =)


r~