Hi,
> gcc/ChangeLog:
>
> PR tree-optimization/117790
> * cfgloopmanip.cc (can_flow_scale_loop_freqs_p): New.
> (flow_scale_loop_freqs): New.
> (scale_loop_freqs_with_exit_counts): New.
> (scale_loop_freqs_hold_exit_counts): New.
> (scale_loop_profile): Refactor to use the newly-added
> scale_loop_profile_1, and use scale_loop_freqs_hold_exit_counts to
> correctly handle reducing the expected niters for loops with multiple
> exits.
> (scale_loop_freqs_with_new_exit_count): New.
> (scale_loop_profile_1): New.
> (scale_loop_profile_hold_exit_counts): New.
> * cfgloopmanip.h (scale_loop_profile_hold_exit_counts): New.
> (scale_loop_freqs_with_new_exit_count): New.
+template<typename ExitCountFn>
+static bool
+can_flow_scale_loop_freqs_p (class loop *loop,
+ ExitCountFn get_exit_count)
+{
+ basic_block bb = loop->header;
+
+ const profile_count count_in = loop_count_in (loop);
+ profile_count exit_count = profile_count::zero ();
+
+ while (bb != loop->latch)
+ {
+ /* Punt if any of the BB counts are uninitialized. */
+ if (!bb->count.initialized_p ())
+ return false;
+
+ bool found_exit = false;
+ edge internal_edge = nullptr;
+ for (auto e : bb->succs)
+ if (flow_bb_inside_loop_p (loop, e->dest))
+ {
+ if (internal_edge)
+ return false;
+ internal_edge = e;
This assumes that there are at most 2 edges out which is not always the
case (i.e. for EH and switch). I suppose vectorizer never calls it
there but probably you want to test that there are precisely two edges
in can_flow_scale_loop_freqs and if not drop message to dump file, so in
case we encounter such loops we notice.
+ }
+ else
+ {
+ if (found_exit)
+ return false;
+ found_exit = true;
+ exit_count += get_exit_count (e);
+ }
+
+ bb = internal_edge->dest;
+ }
+
+ /* Punt if any exit edge had an uninitialized count. */
+ if (!exit_count.initialized_p ())
+ return false;
You already early rturn once you hit bb with uninitialized count, so
perhaps you can move this check just after call of get_exit_count?
+ const profile_count new_exit_count = get_exit_count (exit_edge);
+ profile_probability new_exit_prob;
+ if (new_block_count.nonzero_p ())
+ new_exit_prob = new_exit_count.probability_in (new_block_count);
If new_exit_count > new_block_count probability_in will return 1. I
guess there is not much to do, but pehraps logging inconsistency into
dump is not a bad idea here.
+ else
+ {
+ /* NEW_BLOCK_COUNT is zero, so the only way we can make the profile
+ consistent is if NEW_EXIT_COUNT is zero too. */
+ if (dump_file && new_exit_count.nonzero_p ())
+ fprintf (dump_file,
+ ";; flow_scale_loop_freqs wants non-zero exit count "
+ "but bb count is zero/uninit: profile is inconsistent\n");
+
+ /* Arbitrarily set the exit probability to 0. */
+ new_exit_prob = profile_probability::never ();
never is kind of strong hint to optimize the other patch (it has
RELIABLE reliability). Since we have no info I would just keep the
probability which was there before.
+ }
Patch is OK with these changes.
Honza
