On Thu, 26 Nov 2020, Jakub Jelinek wrote:
> Hi!
>
> The testcase in the PR
> constexpr bool a = __builtin_nan ("") > 0.0;
> constexpr bool b = __builtin_nans ("") > 0.0;
> constexpr bool c = __builtin_nan ("") < 0.0;
> constexpr bool d = __builtin_nans ("") < 0.0;
> constexpr bool e = __builtin_nan ("") >= 0.0;
> constexpr bool f = __builtin_nans ("") >= 0.0;
> constexpr bool g = __builtin_nan ("") <= 0.0;
> constexpr bool h = __builtin_nans ("") <= 0.0;
> has inconsistent behavior, we fold c and d initializers to 0 and don't fold
> any other comparisons to zero.
> Not including the testcase in the testsuite because I really don't know
> if it should be accepted or rejected (it is all accepted with
> -fno-trapping-math).
> The reason we optimize those < 0.0 comparisons to 0 is:
> /* Convert ABS_EXPR<x> < 0 to false. */
> strict_overflow_p = false;
> if (code == LT_EXPR
> && (integer_zerop (arg1) || real_zerop (arg1))
> && tree_expr_nonnegative_warnv_p (arg0, &strict_overflow_p))
> and we return true for __builtin_nan ("") (but not for -__builtin_nan ("").
>
> Now, it just feels wrong to me to say that NaN with sign bit clear is
> non-negative, but other than the above inconsistency I haven't been able to
> construct a miscompiled testcase, only questionable thing is that
> we fold away also comparisons of sNaN < 0.0 - but then while for sNaN >= 0.0
> we don't fold that away during gimple optimizations, we fold it during RTL
> optimizations at least on x86_64.
>
> So, I really don't know if we want this or not, posting it for discussions.
Is copysign (x, NaN) supposed to be well-defined? We'd stop folding
this then, no? I think the ABS_EXPR<x> < 0 to false folding is
simply incomplete and should first check whether the operands are
ordered? That said, NaN is nonnegative but NaN < 0 isn't false(?)
So I don't think the patch is good.
Richard.
> 2020-11-26 Jakub Jelinek <ja...@redhat.com>
>
> PR c++/97965
> * fold-const.c (tree_single_nonnegative_warnv_p): Don't return true
> for NaNs with sign bit clear.
>
> --- gcc/fold-const.c.jj 2020-11-24 09:02:25.330419895 +0100
> +++ gcc/fold-const.c 2020-11-25 15:37:14.426229476 +0100
> @@ -14186,7 +14186,9 @@ tree_single_nonnegative_warnv_p (tree t,
> return tree_int_cst_sgn (t) >= 0;
>
> case REAL_CST:
> - return ! REAL_VALUE_NEGATIVE (TREE_REAL_CST (t));
> + return (! REAL_VALUE_NEGATIVE (TREE_REAL_CST (t))
> + /* Don't consider NaNs non-negative. */
> + && ! REAL_VALUE_ISNAN (TREE_REAL_CST (t)));
>
> case FIXED_CST:
> return ! FIXED_VALUE_NEGATIVE (TREE_FIXED_CST (t));
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imend
