On Fri, 10 Mar 2023, Jakub Jelinek wrote:
> Hi!
>
> This patch, incremental to the just posted one, improves the reverse
> operation ranges significantly by widening just by 0.5ulp in each
> direction rather than 1ulp. Again, REAL_VALUE_TYPE has both wider
> exponent range and wider mantissa precision (160 bits) than any
> supported type, this patch uses the latter property.
>
> The patch doesn't do it if -frounding-math, because then the rounding
> can be +-1ulp in each direction depending on the rounding mode which
> we don't know, or for IBM double double because that type is just weird
> and we can't trust in sane properties.
>
> I've performed testing of these 2 patches on 300000 random tests as with
> yesterday's patch, exact numbers are in the PR, but I see very significant
> improvement in the precision of the ranges while keeping it conservatively
> correct.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
The flow is a little bit obfuscated, but OK.
Thanks,
Richard.
> 2023-03-10 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/109008
> * range-op-float.cc (float_widen_lhs_range): If not
> -frounding-math and not IBM double double format, extend lhs
> range just by 0.5ulp rather than 1ulp in each direction.
>
> --- gcc/range-op-float.cc.jj 2023-03-09 12:13:57.189790814 +0100
> +++ gcc/range-op-float.cc 2023-03-09 13:12:05.248873234 +0100
> @@ -2205,8 +2205,8 @@ zero_to_inf_range (REAL_VALUE_TYPE &lb,
> [1., 1.] = op1 + [1., 1.]. op1's range is not [0., 0.], but
> [-0x1.0p-54, 0x1.0p-53] (when not -frounding-math), any value for
> which adding 1. to it results in 1. after rounding to nearest.
> - So, for op1_range/op2_range extend the lhs range by 1ulp in each
> - direction. See PR109008 for more details. */
> + So, for op1_range/op2_range extend the lhs range by 1ulp (or 0.5ulp)
> + in each direction. See PR109008 for more details. */
>
> static frange
> float_widen_lhs_range (tree type, const frange &lhs)
> @@ -2230,6 +2230,14 @@ float_widen_lhs_range (tree type, const
> lb = dconstm1;
> SET_REAL_EXP (&lb, FLOAT_MODE_FORMAT (TYPE_MODE (type))->emax + 1);
> }
> + if (!flag_rounding_math && !MODE_COMPOSITE_P (TYPE_MODE (type)))
> + {
> + /* If not -frounding-math nor IBM double double, actually widen
> + just by 0.5ulp rather than 1ulp. */
> + REAL_VALUE_TYPE tem;
> + real_arithmetic (&tem, PLUS_EXPR, &lhs.lower_bound (), &lb);
> + real_arithmetic (&lb, RDIV_EXPR, &tem, &dconst2);
> + }
> }
> if (real_isfinite (&ub))
> {
> @@ -2240,6 +2248,14 @@ float_widen_lhs_range (tree type, const
> ub = dconst1;
> SET_REAL_EXP (&ub, FLOAT_MODE_FORMAT (TYPE_MODE (type))->emax + 1);
> }
> + if (!flag_rounding_math && !MODE_COMPOSITE_P (TYPE_MODE (type)))
> + {
> + /* If not -frounding-math nor IBM double double, actually widen
> + just by 0.5ulp rather than 1ulp. */
> + REAL_VALUE_TYPE tem;
> + real_arithmetic (&tem, PLUS_EXPR, &lhs.upper_bound (), &ub);
> + real_arithmetic (&ub, RDIV_EXPR, &tem, &dconst2);
> + }
> }
> /* Temporarily disable -ffinite-math-only, so that frange::set doesn't
> reduce the range back to real_min_representable (type) as lower bound
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
