https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124260
--- Comment #8 from Iain Sandoe <iains at gcc dot gnu.org> ---
(In reply to Iain Sandoe from comment #7)
> (In reply to Martin Jambor from comment #6)
> > (In reply to Iain Sandoe from comment #5)
> > > (In reply to Richard Biener from comment #4)
> > > > So varasm.cc:output_constructor_regular_field "supports" a NULL
> > > > FIELD_DECL in
> > > > CTORs by "appending" and advancing according to the types size. That
> > > > suggests
> > > > that ipa_analyze_var_static_initializer would need to support that as
> > > > well,
> > > > or that it had to get a "failure" mode. OTOH it does not seem to be
> > > > conservative given it's TODO, so a do-nothing is correct here? In that
> > > > case adding || !index to the continue case would be correct?
> > >
> > > I have not looked at the IPA stuff .. it's outside what I'm comfortable
> > > with.
> > >
> >
> > I assume Richi meant to do the following. I cannot test it helps now since
> > I no longer have an easy access to a 32bit host. If an address to a
> > function is skipped, it may result in a "bad" decision to speculatively
> > "devirtualize" or not. But since it is speculative, things will remain
> > correct.
> >
> > diff --git a/gcc/ipa-prop.cc b/gcc/ipa-prop.cc
> > index 385912c3790..e5b401b9d3f 100644
> > --- a/gcc/ipa-prop.cc
> > +++ b/gcc/ipa-prop.cc
> > @@ -3637,7 +3637,8 @@ ipa_analyze_var_static_initializer (varpool_node
> > *node)
> > val)
> > {
> > if (TREE_CODE (val) != ADDR_EXPR
> > - || TREE_CODE (TREE_OPERAND (val, 0)) != FUNCTION_DECL)
> > + || TREE_CODE (TREE_OPERAND (val, 0)) != FUNCTION_DECL
> > + || !index)
> > continue;
> > HOST_WIDE_INT elt_offset = int_bit_position (index);
> > if ((elt_offset % BITS_PER_UNIT) != 0)
>
> I am testing this (separately from the FE fixes).
It fixes the regression on that target (as does the FE fixes) - obv. this is a
much smaller change.
