------- Comment #3 from rguenther at suse dot de 2008-06-29 15:31 -------
Subject: Re: [4.4 Regression] ICE in
process_constraint, at tree-ssa-structalias.c:2573
On Sun, 29 Jun 2008, dberlin at gcc dot gnu dot org wrote:
> ------- Comment #2 from dberlin at gcc dot gnu dot org 2008-06-29 15:23
> -------
> You can't remove the assert, the assert is there because things will be broken
> if it hasn't been handed off to do_structure_copy.
Well... not. 1) we don't have subfields here, 2) the constraint is
supposed to do what it says - what the lhs points to should point to
what the rhs points to.
The remaining bug, in the case we have subfields, is that
get_constraint_for for struct-valued references gets us just a single
constraint for the first field it finds.
The bug is from handle_const_call:
/* May return arguments. */
FOR_EACH_CALL_EXPR_ARG (arg, iter, call)
if (could_have_pointers (arg))
{
VEC(ce_s, heap) *argc = NULL;
struct constraint_expr *argp;
int i;
get_constraint_for (arg, &argc);
for (i = 0; VEC_iterate (ce_s, argc, i, argp); i++)
for (j = 0; VEC_iterate (ce_s, lhsc, j, lhsp); j++)
process_constraint (new_constraint (*lhsp, *argp));
VEC_free (ce_s, heap, argc);
}
for subfields I expected that (for example for two) we end up with
the following constraints generated in this loop
x->a = y->a
x->b = y->a
x->a = y->b
x->b = y->b
which is _not_ a structure copy! But of course in the end it only generates
x->a = y->a
oops.
IMHO get_constraint_for for a struct-valued reference needs to fill the
array with all subvars the reference touches. The do_structure_copy
handling is a special case that handles "real" assignments.
In the above light the assert is bogus, because it expects all
(struct) assignments coming from real instructions.
But maybe I'm missing sth? ;)
Richard.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36666
