Re: PR tree-optimzation/31081

["Richard Guenther" <richard dot guenther at gmail dot com>]

On Dec 30, 2007 2:45 PM, Jan Hubicka <jh@suse.cz> wrote:
> Hi,
> this testcase in ICE in out-of-ssa pass because we introduce abnormal
> PHI with overlapping liferange.  It however just points to the fact that
> inlining into loops increase lifetime of uninitialized variables of
> function being inlinined.
>
> This patch handle it by initializing to 0 all uninitialized variables
> at the beggining of inlined function body.  It is possible to do better
> - we can detect strongly connected regions of CFG and do initialization
> only in those and we can also do initialization to 0 lazilly, but this
> is bit expensive to do (one would need to maintain SCC and dominator
> information during inlining that would need some work) and I am not
> convinced it is worth the extra compilation time - I didn't measured any
> noticeable differences with this patch.
>
> Note that I am not quite convinced that
>
> +                 = build_gimple_modify_stmt (new,
> +                                             fold_convert (TREE_TYPE (new),
> +                                                           integer_zero_node));
>
> is safe way to initialize all gimple regs to 0, but I don't know of
> better alternative.

This should work (it's used in other places as well).  I suggested to
create a build_zero_cst () function for this at some point (to get
rid of the special coding in fold_convert), like I introduced
build_one_cst () sometime ago (because that isn't special-cased
in fold_convert ;)).

> Bootstrapped/regtsted i686-linux, OK?

Yes, this look reasonable.  Please watch for fallout.

Thanks,
Richard.