On Fri, Oct 05, 2012 at 02:49:07PM +0200, Jakub Jelinek wrote:
> I believe that gsi_insert_before in another function
> isn't going to work well.
> E.g. update_modified_stmt starts with
> if (!ssa_operands_active (cfun))
> return;
>
> Or is it ok to use gsi_insert_before_without_update and expect that
> when compilation resumes for the modified callee, it will
> update all modified stmts? There is not much that needs
> to be updated on the stmts, source bind has not setters nor uses
> of any SSA_NAMEs or virtual defs/sets, the normal bind has a DEBUG_EXPR_DECL
> on the RHS and so is not a use or def of anything as well.
To reply to my own post, without_update doesn't work,
/usr/src/gcc/gcc/testsuite/gcc.dg/guality/pr54519-1.c: In function 'fn2.part.0':
/usr/src/gcc/gcc/testsuite/gcc.dg/guality/pr54519-1.c:12:1: error: stmt
(0x7f2d69c3c9d8) marked modified after optimization pass:
# DEBUG D#3 s=> z
/usr/src/gcc/gcc/testsuite/gcc.dg/guality/pr54519-1.c:12:1: internal
# compiler error: verify_ssa failed
then kicks in.
def_temp = gimple_build_debug_source_bind (vexpr, parm,
NULL);
/* DEBUG D#N s=> parm stmt doesn't have any SSA defs or uses,
no need to update it. */
gimple_set_modified (def_temp, false);
gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT);
def_temp = gimple_build_debug_bind (var, vexpr, NULL);
/* DEBUG decl => D#N' stmt doesn't have any SSA defs or uses,
no need to update it. */
gimple_set_modified (def_temp, false);
gsi_insert_before (&cgsi, def_temp, GSI_SAME_STMT);
seems to work, but not sure if that isn't too hackish to avoid one pair
of push_cfun/pop_cfun.
> If push_cfun/pop_cfun is not needed, guess the two loops could be merged,
> the reason for two of them was just that I didn't want to push_cfun/pop_cfun
> for each optimized away parameter.
I was wrong apparently, the two loops are still needed, as the second loop
needs to loop through the parameters in reverse order.
Jakub
