> 2020-12-10 Jakub Jelinek <ja...@redhat.com>
>
> PR rtl-optimization/98212
> * dojump.c (do_compare_rtx_and_jump): Change computation of
> first_prob for and_them. Add comment explaining and_them case.
>
> * gcc.dg/predict-8.c: Adjust expected probability.
>
> --- gcc/dojump.c.jj 2020-12-10 12:30:50.677948803 +0100
> +++ gcc/dojump.c 2020-12-10 13:17:30.568082332 +0100
> @@ -1138,19 +1138,38 @@ do_compare_rtx_and_jump (rtx op0, rtx op
> cprob = cprob.apply_scale (99, 100);
> else
> cprob = profile_probability::even ();
> - /* We want to split:
> + /* For and_them we want to split:
> if (x) goto t; // prob;
> + goto f;
> into
> - if (a) goto t; // first_prob;
> - if (b) goto t; // prob;
> + if (a) ; else goto f; // first_prob for ;
> + // 1 - first_prob for goto
f;
> + if (b) goto t; // adjusted prob;
> + goto f;
> such that the overall probability of jumping to t
> - remains the same and first_prob is prob * cprob. */
> + remains the same. The and_them case should be
> + probability-wise equivalent to the !and_them case with
> + f and t swapped and also the conditions inverted, i.e.
> + if (!a) goto f;
> + if (!b) goto f;
> + goto t;
> + where the overall probability of jumping to f is
> + 1 - prob (thus the first prob.invert () below).
> + cprob.invert () is because the a condition is inverted,
> + so if it was originally ORDERED, !a is UNORDERED and
> + thus should be relative 1% rather than 99%.
> + The invert () on assignment to first_prob is because
> + first_prob represents the probability of fallthru,
> + rather than goto f. And the last prob.invert () is
> + because the adjusted prob represents the probability of
> + jumping to t rather than to f. */
> if (and_them)
> {
> rtx_code_label *dest_label;
> - prob = prob.invert ();
> - profile_probability first_prob = prob.split
(cprob).invert ();
> - prob = prob.invert ();
> + prob = prob.invert ();
> + profile_probability first_prob
> + = prob.split (cprob.invert ()).invert ();
> + prob = prob.invert ();
> /* If we only jump if true, just bypass the second
jump. */
> if (! if_false_label)
> {
OK, thanks, but aren't there missing TABs in the new version? Only one line
is changed in the end AFAICS.
--
Eric Botcazou
