[Bug lto/81847] ICE with LTO enabled

["marxin at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81847

--- Comment #6 from Martin Liška <marxin at gcc dot gnu.org> ---
Created attachment 42004
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42004&action=edit
A bit smaller test-case

Smaller test-cast which needs to add --param lto-min-partition=1.

Problem is following, we create an isra clone and we'll put it into a comdat
group:

_Z11format_packIJiiN1a1b1fIcccEEEEvDpT_.isra.37/2088
(_Z11format_packIJiiN1a1b1fIcccEEEEvDpT_.isra.37) @0x2b2b65795b80
  Type: function definition analyzed
  Visibility: prevailing_def_ironly
comdat_group:_ZN3PigC5IiJiN1a1b1fIcccEEEEET_DpT0_ artificial
  Same comdat group as: _ZN3PigC2IiJiN1a1b1fIcccEEEEET_DpT0_/371
  References: __gxx_personality_v0/1347 (addr)
  Referring: 
  Read from file: gccbug.o
  Availability: local
  First run: 0
  Function flags: local
  Called by: _ZN3PigC2IiJiN1a1b1fIcccEEEEET_DpT0_/371 (1.00 per call) 
  Calls: _ZN1a1b1fIcccE2crEl/1473 _ZN1a5tupleIJiiNS_1b1fIcccEEEED1Ev/607
_ZN1a1b1fIcccE2crEl/1473 (0.37 per call)
_ZN1a5tupleIJiiNS_1b1fIcccEEEED2Ev/2938 (inlined) (1.00 per call)
_Z12format_tupleIN1a5tupleIJiiNS0_1b1fIcccEEEEELi3EEvT_NS0_1lIiXT0_EEE.isra.10/1803
(inlined) (1.00 per call)
_ZN1a5tupleIJiiNS_1b1fIcccEEEEC2IJiiS3_ELj1EEEDpT_/798 (inlined) (1.00 per
call) _ZN1a1b1fIcccE2coC1EPcNS_9allocatorIcEE/1471 (1.00 per call) (can throw
external) _ZN1a1b1fIcccE2csEv/1470 (1.00 per call) (can throw external) 

Then we start partitioning and following symbols are put to a partition:

add_symbol_to_partition: __base_ctor /371
Step 1: added _Z11format_packIJiiN1a1b1fIcccEEEEvDpT_.isra.37/2088, size 931,
cost 3013/1012 best 3013/1012, step 1

the partition is finished and in a next partition:
Symbol node __base_ctor /371 now used in multiple partitions

and then we attempt to add again the isra clone and we fail. That triggers the
assert.
We have:

   113  static bool
   114  add_symbol_to_partition_1 (ltrans_partition part, symtab_node *node)
   115  {
...
   124    /* non-duplicated aliases or tunks of a duplicated symbol needs to be
output
   125       just once.
   126  
   127       Be lax about comdats; they may or may not be duplicated and we may
   128       end up in need to duplicate keyed comdat because it has unkeyed
alias.  */
   129    if (c == SYMBOL_PARTITION && !DECL_COMDAT (node->decl)
   130        && symbol_partitioned_p (node))
   131      return false;

Shouldn't be this place somehow adjusted in order to handle symbols in comdat
group that are actually
not DECL_COMDATs? Honza?