On Wed, Jan 17, 2024 at 01:45:18PM +0100, Jan Hubicka wrote:
> Hi,
> expr_expected_value is doing some guesswork when it is merging two or more
> independent value predictions either in PHI node or in binary operation.
> Since we do not know how the predictions interact with each other, we can
> not really merge the values precisely.
>
> The previous logic merged the prediciton and picked the later predictor
> (since predict.def is sorted by reliability). This however leads to troubles
> with __builtin_expect_with_probability since it is special cased as a
> predictor
> with custom probabilities. If this predictor is downgraded to something else,
> we ICE since we have prediction given by predictor that is not expected
> to have customprobability.
>
> This patch fixies it by inventing new predictors
> PRED_COMBINED_VALUE_PREDICTIONS
> and PRED_COMBINED_VALUE_PREDICTIONS_PHI which also allows custom values but
> are considered less reliable then __builtin_expect_with_probability (they
> are combined by ds theory rather then by first match). This is less likely
> going to lead to very stupid decisions if combining does not work as expected.
>
> I also updated the code to be bit more careful about merging values and do not
> downgrade the precision when unnecesary (as tested by new testcases).
>
> Bootstrapped/regtested x86_64-linux, will commit it tomorrow if there are
> no complains.
>
> 2024-01-17 Jan Hubicka <j...@suse.cz>
> Jakub Jelinek <ja...@redhat.com>
2 spaces before < rather than 1.
>
> PR tree-optimization/110852
>
> gcc/ChangeLog:
>
> * predict.cc (expr_expected_value_1):
> (get_predictor_value):
> * predict.def (PRED_COMBINED_VALUE_PREDICTIONS):
> (PRED_COMBINED_VALUE_PREDICTIONS_PHI):
>
> gcc/testsuite/ChangeLog:
>
> * gcc.dg/predict-18.c:
Please fill in what has changed, both for predict-18.c and predict.{cc,def}
changes.
> @@ -2613,24 +2658,40 @@ expr_expected_value_1 (tree type, tree op0, enum
> tree_code code,
> if (!nop1)
> nop1 = op1;
> }
> + /* We already checked if folding one of arguments to constant is good
> + enough. Consequently failing to fold both means that we will not
> + succeed determinging the value. */
s/determinging/determining/
Otherwise LGTM.
Jakub
