[Bug lto/94320] New: [OpenMP][Offloading] lto1 ICE during IPA pass: inline – offline_size at gcc/ipa-inline-analysis.c:453

[Bug lto/94320] New: [OpenMP][Offloading] lto1 ICE during IPA pass: inline – offline_size at gcc/ipa-inline-analysis.c:453

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94320

            Bug ID: 94320
           Summary: [OpenMP][Offloading] lto1 ICE during IPA pass: inline
                    – offline_size at gcc/ipa-inline-analysis.c:453
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Keywords: ice-on-invalid-code, openmp
          Severity: normal
          Priority: P3
         Component: lto
          Assignee: unassigned at gcc dot gnu.org
          Reporter: burnus at gcc dot gnu.org
                CC: marxin at gcc dot gnu.org
  Target Milestone: ---

May or may not be related to PR 92029

Still reducing the test case. When compiling with "-fopenmp -O0" it works, but
with "-fopenmp -O1" the offloading-compiler (AMDGCN or – has here – nvptx) ICEs
with:


during IPA pass: inline
lto1: internal compiler error: Segmentation fault
0xbdb6ff crash_signal
       
/net/build5-trusty-cs/scratch/tburnus/openacc.x86_64-linux-gnu-openacc-mainline/src/gcc-mainline/gcc/toplev.c:328
0x98074f offline_size
       
/net/build5-trusty-cs/scratch/tburnus/openacc.x86_64-linux-gnu-openacc-mainline/src/gcc-mainline/gcc/ipa-inline-analysis.c:453
0x980b58 growth_positive_p(cgraph_node*, cgraph_edge*, int)
       
/net/build5-trusty-cs/scratch/tburnus/openacc.x86_64-linux-gnu-openacc-mainline/src/gcc-mainline/gcc/ipa-inline-analysis.c:549
0x138610f want_inline_function_to_all_callers_p
       
/net/build5-trusty-cs/scratch/tburnus/openacc.x86_64-linux-gnu-openacc-mainline/src/gcc-mainline/gcc/ipa-inline.c:1088
0x138c37f ipa_inline
       
/net/build5-trusty-cs/scratch/tburnus/openacc.x86_64-linux-gnu-openacc-mainline/src/gcc-mainline/gcc/ipa-inline.c:2752
0x138d099 execute
       
/net/build5-trusty-cs/scratch/tburnus/openacc.x86_64-linux-gnu-openacc-mainline/src/gcc-mainline/gcc/ipa-inline.c:3091


It fails in ipa-inline-analysis.c's:

offline_size (struct cgraph_node *node, ipa_size_summary *info)
{
  if (!DECL_EXTERNAL (node->decl))
    {
      if (node->will_be_removed_from_program_if_no_direct_calls_p ())
        return info->size;

as info == NULL and decl is " <function_decl 0x7ffff73df500 __ct_base "

"info" is:
    ipa_size_summary *s = ipa_size_summaries->get (node);

And ->get returns:
369         return exists (node) ? (*m_vector)[node->get_summary_id ()] : NULL;

[Bug lto/94320] [OpenMP][Offloading] lto1 ICE during IPA pass: inline – offline_size at gcc/ipa-inline-analysis.c:453

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94320

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Perhaps related to PR92357?
I'm stuck there, tried
2020-02-07  Jakub Jelinek  <jakub@redhat.com>

        PR ipa/92357
        * symtab.c (symtab_node::get_partitioning_class): Don't return
        SYMBOL_DUPLICATE for DECL_ONE_ONLY symbols if lto_stream_offload_p.

--- gcc/symtab.c.jj     2020-01-15 11:05:19.589136168 +0100
+++ gcc/symtab.c        2020-02-07 19:57:55.625108414 +0100
@@ -2003,6 +2003,7 @@ symtab_node::get_partitioning_class (voi
   if (DECL_ONE_ONLY (decl)
       && !force_output
       && !forced_by_abi
+      && !lto_stream_offload_p
       && !used_from_object_file_p ())
     return SYMBOL_DUPLICATE;

but it didn't work out.  I'm afraid we need to actually implement properly the
OpenMP 5.0 automatic omp declare target discovery and before IPA run perhaps
another iteration of that and make sure that if we offload something, we either
offload everything it refers to too (unless DECL_EXTERNAL), or diagnose it.

[Bug lto/94320] [OpenMP][Offloading] lto1 ICE during IPA pass: inline – offline_size at gcc/ipa-inline-analysis.c:453

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94320

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #1)
> tried
> +      && !lto_stream_offload_p

With that patch, I get 6 times "has been referenced in offloaded code but
hasn't been marked to be included in the offloaded code".

Thus, that's a useful patch for the code I looked at. All those those functions
are template functions. (With -O0, the nvptx run-time compilers complains only
about 3 of those functions.)

> I'm afraid we need to actually implement properly the OpenMP 5.0
> automatic omp declare target discovery

Yes, it looks like that. I think C++ code is mostly affected as one quickly
ends up using 'operator[]' and similar functions, which are not always inlined.

Additionally, template programming makes it harder for the user to mark those
functions as 'declare target'.