[Bug tree-optimization/110852] [14 Regression] ICE: in get_predictor_value, at predict.cc:2695 with -O -fno-tree-fre and __builtin_expect() since r14-2219-geab57b825bcc35

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110852

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Jan Hubicka from comment #9)
> By removing the logic we lose ability to optimize things like
>   a = b * c;
> where b is predicted to value 0 and c has no useful prediction on it.

No, that is what my unposted WIP patch did, but predict-18.c test catched that.

> > @@ -2631,6 +2623,9 @@ expr_expected_value_1 (tree type, tree o
> > 
> >           if (predictor2 < *predictor)
> >             *predictor = predictor2;
> > +         if (*predictor != PRED_BUILTIN_EXPECT
> > +             && *predictor != PRED_BUILTIN_EXPECT_WITH_PROBABILITY)
> > +           *probability = -1;
> 
> This still can "upgrade" prediction to a predictor of lower enm value
> but higher probability that is not conservative thing to do.
> > 
> >           return res;
> >         }
> I ended up with the folloing patch that also takes care of various cases
> of phi merging and downgrading the predictor to new
> PRED_COMBINED_VALUE_PREDICTION which can, like PRED_BUILTIN_EXPECT hold
> custom probability but it is not trued as FIRST_MATCH.
> What do you think?

> +	      int p1 = get_predictor_value (*predictor, *probability);
> +	      int p2 = get_predictor_value (predictor2, probability2);
> +	      /* If both predictors agrees, it does not matter from which

s/agrees/agree/

> +         Consequently failing to fold both means that we will not suceed
> determinging

s/suceed/succeed/;s/determinging/determining/

Otherwise yes, but I think the code could be still simplified the way I had in
my patch (i.e. drop parts of the r14-2219 changes, and simply assume that
failed recursion for one operand is PRED_UNCONDITIONAL instead of returning
early, and not requiring the operands are INTEGER_CSTs, just that the result of
the binop folds to INTEGER_CST.