>
> gcc/ChangeLog:
>
> 2018-04-24 Martin Liska <mli...@suse.cz>
>
> * ipa-pure-const.c (struct funct_state_d): Do it class instead
> of struct.
> (class funct_state_summary_t): New function_summary class.
> (has_function_state): Remove.
> (get_function_state): Likewise.
> (set_function_state): Likewise.
> (add_new_function): Likewise.
> (funct_state_summary_t::insert): New function.
> (duplicate_node_data): Remove.
> (remove_node_data): Remove.
> (funct_state_summary_t::duplicate): New function.
> (register_hooks): Create new funct_state_summaries.
> (pure_const_generate_summary): Use it.
> (pure_const_write_summary): Likewise.
> (pure_const_read_summary): Likewise.
> (propagate_pure_const): Likewise.
> (propagate_nothrow): Likewise.
> (dump_malloc_lattice): Likewise.
> (propagate_malloc): Likewise.
> (execute): Do not register hooks, just remove summary
> instead.
> (pass_ipa_pure_const::pass_ipa_pure_const): Simplify
> constructor.
OK with changes below.
In general, it would be cool to reorg this pass into simple SCC propagation
template
because it does same things over and over again (sometimes on slightly
different graph
because it has feature to skip uninteresting edges e.g. for nothrow
propagation).
> @@ -1485,7 +1439,7 @@ propagate_pure_const (void)
> int i;
> struct ipa_ref *ref = NULL;
>
> - funct_state w_l = get_function_state (w);
> + funct_state w_l = funct_state_summaries->get_create (w);
> if (dump_file && (dump_flags & TDF_DETAILS))
> fprintf (dump_file, " Visiting %s state:%s looping %i\n",
> w->dump_name (),
> @@ -1527,7 +1481,7 @@ propagate_pure_const (void)
> }
> if (avail > AVAIL_INTERPOSABLE)
> {
> - funct_state y_l = get_function_state (y);
> + funct_state y_l = funct_state_summaries->get_create (y);
> if (dump_file && (dump_flags & TDF_DETAILS))
> {
> fprintf (dump_file,
The functions are organized in a way that it goes cycle by cycle. First loop
initializes
everything in the cycle and in this case you want get_create (where w_l is set)
y_l looks for calls and because it works in SCC order all calls are either in
same cycle
or processed. I do not see any guard checking that y_l is initialized and
because you
now initialize it to pesimistic state I think it will turn any SCC component
into non-pure
now (while it originaly worked by initializing to 0 which is optimistic state
CONST)
So i would add guard that y is in different SCC component and make code to crash
if get() returns NULL.
Similarly for all other propagators.
> @@ -1642,7 +1596,7 @@ propagate_pure_const (void)
> while (w && !can_free)
> {
> struct cgraph_edge *e;
> - funct_state w_l = get_function_state (w);
> + funct_state w_l = funct_state_summaries->get_create (w);
>
> if (w_l->can_free
> || w->get_availability () == AVAIL_INTERPOSABLE
> @@ -1657,7 +1611,7 @@ propagate_pure_const (void)
> e->caller);
>
> if (avail > AVAIL_INTERPOSABLE)
> - can_free = get_function_state (y)->can_free;
> + can_free = funct_state_summaries->get_create (y)->can_free;
> else
> can_free = true;
> }
Here everything should be computed already.
OK with those changes provides it does not affect generated code.
Honza
