https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92788
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, what happens is that we have a bb like:
<bb 13> [local count: 7102878]:
# __result_2 = PHI <_10(3)>
*this_4(D).D.3185._M_impl.D.3135._M_finish = __result_2;
with EH and normal successor edges.
dom_opt_dom_walker::optimize_stmt decides to remove the stmt (the only one in
the bb):
2160 if (cached_lhs && operand_equal_p (rhs, cached_lhs, 0))
2161 {
2162 basic_block bb = gimple_bb (stmt);
2163 unlink_stmt_vdef (stmt);
2164 if (gsi_remove (si, true))
2165 {
2166 bitmap_set_bit (need_eh_cleanup, bb->index);
2167 if (dump_file && (dump_flags & TDF_DETAILS))
2168 fprintf (dump_file, " Flagged to clear EH
edges.\n");
2169 }
2170 release_defs (stmt);
2171 *removed_p = true;
2172 return retval;
2173 }
but before we perform
780 /* Removal of statements may make some EH edges dead. Purge
781 such edges from the CFG as needed. */
782 if (!bitmap_empty_p (need_eh_cleanup))
...
to clean up those EH edges,
774 /* Thread jumps, creating duplicate blocks as needed. */
775 cfg_altered |= thread_through_all_blocks (may_peel_loop_headers_p);
is called, calls thread_block_1 on the bb 13 and ICE trying to redirect EH edge
on a copy of that BB, because for EH edges we obviously assume there is some
statement that can throw but there are no statements in the bb at all.
Not really sure what can be done here.
