https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92357
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I'll note that GCC doesn't yet implement the OpenMP 5.0 automatic declare
target for functions with definitions called from declare target functions or
target regions. I've tried to do that this summer, but came up with too many
questions for which there are no answers yet, so it is e.g. unclear where the
compiler should do it, the implementation I had (will attach) was done quite
late, which means e.g. that if it should affect in the end whether some
variables will be declare target or not implicitly, it would be too late to
change depending on that the implicit mapping or privatization behavior on
target constructs, because by the time target is resolved declare variant on
some vars still wouldn't be done implicitly.
In the #c1 testcase, I see only two templates with their single or so method
each wrapped in declare target and nothing else, so I bet the testcase is
invalid (dunno if that is in the original RAJA).
Anyway, with Honza's help I came up with:
--- gcc/ipa-fnsummary.c 2019-12-03 21:45:21.538147319 +0100
+++ gcc/ipa-fnsummary.c 2019-12-13 16:24:14.007917974 +0100
@@ -4372,7 +4372,10 @@ ipa_fn_summary_write (void)
{
symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
cgraph_node *cnode = dyn_cast <cgraph_node *> (snode);
- if (cnode && cnode->definition && !cnode->alias)
+ if (cnode
+ && cnode->definition
+ && !cnode->alias
+ && lto_symtab_encoder_in_partition_p (encoder, cnode))
count++;
}
streamer_write_uhwi (ob, count);
@@ -4381,7 +4384,10 @@ ipa_fn_summary_write (void)
{
symtab_node *snode = lto_symtab_encoder_deref (encoder, i);
cgraph_node *cnode = dyn_cast <cgraph_node *> (snode);
- if (cnode && cnode->definition && !cnode->alias)
+ if (cnode
+ && cnode->definition
+ && !cnode->alias
+ && lto_symtab_encoder_in_partition_p (encoder, cnode))
{
class ipa_fn_summary *info = ipa_fn_summaries->get (cnode);
class ipa_size_summary *size_info = ipa_size_summaries->get (cnode);
patch with which this ICE is gone but there is another later on, so more work
is needed.
