On Tue, Oct 4, 2011 at 9:17 AM, Maxim Kuvyrkov <ma...@codesourcery.com> wrote:
> Richard,
>
> The following patch fixes a CFG consistency problem.
>
> When early IPA optimizations (e.g., early SRA) create a version of a function
> that no longer throws, versioning machinery deletes EH landings pads in
> _callers_ of the function [*]. However, the EH cfg edges are not updated in
> the callers, which eventually leads to an ICE in verify_eh_edges.
It needs to update EH edges in the callers via the common idiom
if (maybe_clean_or_replace_eh_stmt (stmt, stmt)
&& gimple_purge_dead_eh_edges (gimple_bb (stmt)))
... arrange for cfg-cleanup to be run ...
where the cfg-cleanup part probably explains why it doesn't. So one option
is to not delete the landing pads. Where does it do that? I suppose it
merely fails to properly transfer the EH info from the old to the new call
(thus, fails to forcefully do what maybe_duplicate_eh_stmt does).
Richard.
> I've tried several ways to fix this, including trying to make early
> optimizations not delete callers' landing pads. You are the original author
> of the gimple EH code, and I wonder if you know the rationale behind removing
> EH landing pads from callers or a function being optimized.
>
> The failure triggers on several libstdc++ testcases, but requires an
> unrelated patch of mine (soon to be posted).
>
> In any case, attached is what appears to be the best way to fix this up.
>
> Bootstrapped and regrested on x86_64-linux-gnu. OK for trunk?
>
> Thank you,
>
> [*] this is done in cgraphunit.c: update_call_expr() and
> ipa_modify_call_arguments() via maybe_clean[_and_replace]_eh_stmt().
>
> --
> Maxim Kuvyrkov
> CodeSourcery / Mentor Graphics
>
>
>
