Hi,
On Mon, Feb 11 2019, Jan Hubicka wrote:
>> Hi.
>>
>> IPA pure const should always construct ipa_reduced_postorder with
>> possibility to cross AVAIL_INTERPOSABLE boundary. The pass itself
>> can then properly stop propagation on these symbols.
>>
>> The patch is pre-approved by Honza.
>> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
>>
>> Thanks,
>> Martin
>>
>> gcc/ChangeLog:
>>
>> 2019-02-08 Martin Liska <mli...@suse.cz>
>>
>> PR ipa/89009
>> * ipa-cp.c (build_toporder_info): Remove usage of a param.
>> * ipa-inline.c (inline_small_functions): Likewise.
>> * ipa-pure-const.c (propagate_pure_const): Likewise.
>> (propagate_nothrow): Likewise.
>> * ipa-reference.c (propagate): Likewise.
>> * ipa-utils.c (struct searchc_env): Remove unused field.
>> (searchc): Always search across AVAIL_INTERPOSABLE.
>> (ipa_reduced_postorder): Always allow AVAIL_INTERPOSABLE as
>> the only called IPA pure const can properly not propagate
>> across interposable boundary.
>> * ipa-utils.h (ipa_reduced_postorder): Remove param.
>> @@ -105,7 +104,7 @@ searchc (struct searchc_env* env, struct cgraph_node *v,
>>
>> if (w->aux
>> && (avail > AVAIL_INTERPOSABLE
>> - || (env->allow_overwritable && avail == AVAIL_INTERPOSABLE)))
>> + || avail == AVAIL_INTERPOSABLE))
This looks a tad too mechanical, surely
(avail > AVAIL_INTERPOSABLE || avail == AVAIL_INTERPOSABLE)
can be simplified into just one comparison :-)
>> {
>> w_info = (struct ipa_dfs_info *) w->aux;
>> if (w_info->new_node)
>
> This optimizes >= AVAIL_INTERPOSABLE then, but for ipa-cp, ipa-inline,
> ipa-pure-const we do not want interposable calls to close SCC regions
> (because we can not propagate here). ipa-reference is more subtle -
> interposable calls close SCC regions IFF they are declared leaf.
>
> Can you, please, update ignore_edge_p of indiviual optimization passes
> this way? (I.e. for all but ipa-reference add aval <= AVAIL_INTERPOSABLE
> check for the availability of the target function and for ipa-reference
> test also leaf attribute?).
At least IPA-CP does not have any ignore_edge_p... does it really need
any? Such node can incur an extra iteration of the propagator over an
entire SCC, that is true, but is it worth adding a callback? Are SCCs
with an interposable node common?
>
> In all those cases we also want to check that given optimization pass is
> enabled for both caller and callee. In other cases we do not want to
> consider them part of SCCs as well. It may make sense to do this as one
> change or I can do that incrementally.
Likewise.
Martin
