https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85213
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |NEW
CC| |aoliva at gcc dot gnu.org
Assignee|jakub at gcc dot gnu.org |unassigned at gcc dot
gnu.org
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The problem is that twoval_comparison_p breaks appart SAVE_EXPR if their
argument has no side effects.
First, cp_build_binary_op calls save_expr on:
(__builtin_expect ((long int) (# DEBUG BEGIN STMT;
x != 0; ? 0 : 1), (long int) NON_LVALUE_EXPR <3>) == 0) * -1;
and that resulting SAVE_EXPR is used in 2+ places, where one of them is a
comparison against INTEGER_CST, which is optimized using twoval_comparison_p as
another SAVE_EXPR with
__builtin_expect ((long int) (# DEBUG BEGIN STMT;
x != 0; ? 0 : 1), (long int) NON_LVALUE_EXPR <3>) == 0;
body. As mostly_copy_tree_r does not unshare STATEMENT_LISTs, we hand over to
the gimplifier the above two SAVE_EXPRs and both contain the same
STATEMENT_LIST, so we try to gimplify it twice, which as we know from other PRs
doesn't really work at all.
Now, not looking through SAVE_EXPRs in twoval_comparison_p if it contains a
STATEMENT_LIST might result in -fcompare-debug failures, because with -g0 there
could be no STATEMENT_LIST, while with -g we have one.
Not sure what our options are, besides just killing the SAVE_EXPR handling in
twoval_comparison_p.
