https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117095
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I'm afraid we are back to the REG_UNUSED mess :(
The equivalence that pseudo 165 is 1 is recorded in cse.cc (record_jump_cond)
called from record_jump_equiv.
This is guarded by
/* If this is a conditional jump insn, record any known
equivalences due to the condition being tested. */
insn = BB_END (bb);
if (path_entry < path_size - 1
&& EDGE_COUNT (bb->succs) == 2
&& JUMP_P (insn)
&& single_set (insn)
&& any_condjump_p (insn))
{
basic_block next_bb = ebb_data->path[path_entry + 1].bb;
bool taken = (next_bb == BRANCH_EDGE (bb)->dest);
record_jump_equiv (insn, taken);
}
That the insn has a single set (the pc_set) is a reasonable precondition of the
optimization, it doesn't really walk further sets and try to figure out how
they affect the result.
But the insn is
(jump_insn 217 78 216 10 (parallel [
(set (pc)
(if_then_else (ne (reg:SI 165)
(const_int 1 [0x1]))
(label_ref 216)
(pc)))
(set (reg:SI 165)
(plus:SI (reg:SI 165)
(const_int -1 [0xffffffffffffffff])))
(clobber (scratch:SI))
(clobber (reg:CC 33 %cc))
]) "pr117095.c":14:17 discrim 1 2192 {doloop_si64}
(expr_list:REG_UNUSED (reg:SI 165)
(expr_list:REG_UNUSED (reg:CC 33 %cc)
(int_list:REG_BR_PROB 955630228 (nil))))
and because of the REG_UNUSED note (note, it actually even isn't stale, cse.cc
has
df_note_add_problem (); right before df_analyze ()) single_set returns
non-NULL.
But we actually want to extend the lifetime of the pseudo if the recorded
equivalence is ever used, so we need to disregard REG_UNUSED notes there.
I think this got broken with r0-77890, before that record_jump_equiv was done
in cse_insn and guarded on n_sets == 1.
