https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90273
--- Comment #18 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to rsand...@gcc.gnu.org from comment #17)
> Created attachment 46261 [details]
> Doing the removal in find_obviously_necessary_stmts
>
> Just for the record: I'd written this over the weekend for completely
> unrelated reasons (was looking at debug info, and wanted to replicate the
> DCE that is already done in RTL). It doesn't include the DEBUG_EXPR_DECL
> stuff, so obviously can't go in as-is, but it means that we never even add
> the redundant instructions to the worklist.
Yeah, that looks nearly equivalent (same issue with DEBUG_EXPR_DECLs though).
I think doing it in eliminate_unnecessary_stmts is slightly better for the
case where we come from
# DEBUG i => NULL;
i_3 = 1;
# DEBUG i => i_3;
and are about to DCE i_3 (I think that's not handled by my patch either
without further tweaks). No debug-stmt support in the GIMPLE FE right now
so not so easy to test atm. For
int main()
{
int i;
i = 1, i = 2;
return i+1;
}
with -O -g -fdisable-tree-ccp1 -fdisable-tree-fre1 -fdisable-tree-evrp we
go from
<bb 2> :
# DEBUG BEGIN_STMT
# DEBUG BEGIN_STMT
i_1 = 1;
# DEBUG i => i_1
i_2 = 2;
# DEBUG i => i_2
# DEBUG BEGIN_STMT
_3 = 3;
return 3;
after DSE1 to
<bb 2> :
# DEBUG BEGIN_STMT
# DEBUG BEGIN_STMT
# DEBUG i => 1
# DEBUG i => 2
# DEBUG BEGIN_STMT
return 3;
after CDDCE1 which could be improved to just # DEBUG i => 2 (note the
missing DEBUG BEGIN_STMT for compound stmts). The following would do
that trick (doing a bit more debug DCE on the testcase in this PR)
Index: gcc/tree-ssa-dce.c
===================================================================
--- gcc/tree-ssa-dce.c (revision 270645)
+++ gcc/tree-ssa-dce.c (working copy)
@@ -1282,11 +1283,15 @@ eliminate_unnecessary_stmts (void)
if (!is_gimple_debug (stmt))
something_changed = true;
remove_dead_stmt (&gsi, bb, to_remove_edges);
+ continue;
}
else if (is_gimple_call (stmt))
{
