[PATCH] Fix ICEs with "omp declare simd" attribute on versioned fns or omp_fn* (PR middle-end/83977)

[Richard Biener]

On January 24, 2018 6:51:54 PM GMT+01:00, Richard Biener <rguenther@suse.de> wrote:
>On January 24, 2018 6:40:25 PM GMT+01:00, Jakub Jelinek
><jakub@redhat.com> wrote:
>>On Wed, Jan 24, 2018 at 06:36:02PM +0100, Martin Jambor wrote:
>>> > I think there's already a set of attributes that prevent cloning
>>and
>>> > or are adjusted by the IPA param machinery. The Martins or Honza
>>> > should know better.
>>> 
>>> I am not sure I understand the problem but if
>>> tree_versionable_function_p returns false, the local.versionable bit
>>is
>>> not set and no cloning for that function happens.
>>> 
>>> If it is sufficient that IPA-CP and other IPA passes do not change
>>the
>>> function type in any way (in practice that they don't remove
>>> parameters), it is sufficient to clear the
>local.can_change_signature
>>> cgraph flag in compute_fn_summary() in ipa-fnsummary.c.  That is how
>>we
>>> handle, or rather avoid handling, fnspec attributes.
>>
>>Well, "omp declare simd" is a part of the ABI just for the original
>>exported
>>functions, for everything else it is a pure optimization, but I'm not
>>sure
>>if we want to deoptimize e.g. callers of these functions outside of
>>loops
>>by disabling the signature changing cloning for those.  For calls from
>>within OpenMP simd regions or other loops where we try hard to
>>vectorize
>>them, it might make sense not to change those callers, for callers
>from
>>other loops, a question.
>
>Until we can distinguish the cases I think not changing the signature
>by default might be a good thing. 

Otoh cloning for ipa cp and then inlining is OK. Not sure how the mitigation mechanism works. 

Richard. 

>Richard. 
>
>>	Jakub