Hi,
On Tue, Jul 13 2021, Alexandre Oliva wrote:
> The logic in fill_vector_of_new_param_types may skip some parameters
> when pushing into m_new_types, but common_initialization doesn't take
> that into account, and may end up attempting to access the vector past
> its end when IPA_PARAM_OP_(NEW|SPLIT) operands appear after skipped
> _COPY ones.
>
> This patch adjusts the consumer logic to match the indexing in the
> producer. It came up in libstdc++-v3's testsuite, in
> std/ranges/adaptors/filter.cc, but only with wrappers introduced by a pass
> I'm working on. The _NEW parameters were reference-typed replacements
> for some by-value ones in my function-wrapping logic, and other IPA
> transformations cause plenty of unused/irrelevant arguments to be
> dropped for certain calls.
>
> Regstrapped on x86_64-linux-gnu. Ok to install?
I agree that there is a mismatch but I do not like proposed change.
Indeed fill_vector_of_new_param_types can skip some parameters when they
appear to be copies of something that is not there - but that bail-out
is not supposed to ever happen when fill_vector_of_new_param_types is
called from ipa_param_body_adjustments.
It is there for cases where one compilation unit has a completely bogus
idea of a type defined in another compilation unit and so creates a bad
type for gimple_call_fntype, but then LTO decides to do something with
the parameters and modifying the gimple_call_fntype becomes a
garbage-in-garbage-out operation (but avoids ICE).
That should never happen when you actually have a body and so presumably
the type and parameters match. So I would first check how come that you
request IPA_PARAM_OP_COPY of something that does not seem to have a
corresponding type but there is a DECL, otherwise the code would have
ICEd when attempting to "carry over" that.
If you believe that what you're doing is correct (but check twice), I
would prefer the following change (only mildly tested) that is a
clean-up and should help you with your problem too.
Thanks,
Martin
diff --git a/gcc/ipa-param-manipulation.c b/gcc/ipa-param-manipulation.c
index 26b02d7aa95..c5285b7cdf3 100644
--- a/gcc/ipa-param-manipulation.c
+++ b/gcc/ipa-param-manipulation.c
@@ -1067,14 +1067,15 @@ ipa_param_body_adjustments::common_initialization (tree
old_fndecl,
auto_vec<tree,16> otypes;
if (TYPE_ARG_TYPES (TREE_TYPE (old_fndecl)) != NULL_TREE)
push_function_arg_types (&otypes, TREE_TYPE (old_fndecl));
- else
+ if (m_oparms.length () != otypes.length ())
{
- auto_vec<tree,16> oparms;
- push_function_arg_decls (&oparms, old_fndecl);
- unsigned ocount = oparms.length ();
+ /* Parameter type information is not available or there is a mismatch
+ between it and the real parameters (probably K&R code or weird LTO
+ mismatched decls). */
+ unsigned ocount = m_oparms.length ();
otypes.reserve_exact (ocount);
for (unsigned i = 0; i < ocount; i++)
- otypes.quick_push (TREE_TYPE (oparms[i]));
+ otypes.quick_push (TREE_TYPE (m_oparms[i]));
}
fill_vector_of_new_param_types (&m_new_types, &otypes, m_adj_params, true);
