[v3,1/2] x86 long double: Consider pseudo numbers as signaling

  Add support to treat pseudo-numbers specially and implement x86
version to consider all of them as signaling.
---
 sysdeps/generic/nan-pseudo-number.h      | 32 ++++++++++++++++++++++++
 sysdeps/ieee754/ldbl-96/s_issignalingl.c | 11 +++++---
 sysdeps/x86/fpu/nan-pseudo-number.h      | 30 ++++++++++++++++++++++
 3 files changed, 70 insertions(+), 3 deletions(-)
 create mode 100644 sysdeps/generic/nan-pseudo-number.h
 create mode 100644 sysdeps/x86/fpu/nan-pseudo-number.h
  

On 23/12/2020 23:45, Siddhesh Poyarekar wrote:
> Add support to treat pseudo-numbers specially and implement x86
> version to consider all of them as signaling.

Ok with the above changes.

Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>

> ---
>  sysdeps/generic/nan-pseudo-number.h      | 32 ++++++++++++++++++++++++
>  sysdeps/ieee754/ldbl-96/s_issignalingl.c | 11 +++++---
>  sysdeps/x86/fpu/nan-pseudo-number.h      | 30 ++++++++++++++++++++++
>  3 files changed, 70 insertions(+), 3 deletions(-)
>  create mode 100644 sysdeps/generic/nan-pseudo-number.h
>  create mode 100644 sysdeps/x86/fpu/nan-pseudo-number.h
> 
> diff --git a/sysdeps/generic/nan-pseudo-number.h b/sysdeps/generic/nan-pseudo-number.h
> new file mode 100644
> index 0000000000..1562e06e4b
> --- /dev/null
> +++ b/sysdeps/generic/nan-pseudo-number.h
> @@ -0,0 +1,32 @@
> +/* Pseudo-normal number handling.  Generic version.
> +   Copyright (C) 2020 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#ifndef NAN_PSEUDO_NUMBER_H
> +#define NAN_PSEUDO_NUMBER_H	1
> +
> +/* Set this macro and override the definition of IS_PSEUDO_SIGNALING if pseudo
> +   numbers need special handling.  */
> +#define HANDLE_PSEUDO_NUMBERS 0
> +static inline int
> +is_pseudo_signaling (uint32_t exi __attribute__ ((unused)),
> +		     uint32_t hxi __attribute__ ((unused)))
> +{
> +  return 0;
> +}
> +
> +#endif /* nan-pseudo-number.h */

Afaik, there is no need to __attribute__ ((unused)) on arguments for static
inline functions.

> diff --git a/sysdeps/ieee754/ldbl-96/s_issignalingl.c b/sysdeps/ieee754/ldbl-96/s_issignalingl.c
> index ec542ad468..37e05aa259 100644
> --- a/sysdeps/ieee754/ldbl-96/s_issignalingl.c
> +++ b/sysdeps/ieee754/ldbl-96/s_issignalingl.c
> @@ -19,12 +19,19 @@
>  #include <math.h>
>  #include <math_private.h>
>  #include <nan-high-order-bit.h>
> +#include <nan-pseudo-number.h>
>  
>  int
>  __issignalingl (long double x)
>  {
>    uint32_t exi, hxi, lxi;
>    GET_LDOUBLE_WORDS (exi, hxi, lxi, x);
> +
> +  /* By default we do not recognize a pseudo NaN as sNaN.  However on 80387 and
> +     later all pseudo numbers including pseudo NaNs result in a signal and are
> +     hence recognized as signaling.  */
> +  int ret = is_pseudo_signaling (exi, hxi);
> +
>  #if HIGH_ORDER_BIT_IS_SET_FOR_SNAN
>  # error not implemented
>  #else
> @@ -34,11 +41,9 @@ __issignalingl (long double x)
>    hxi ^= 0x40000000;
>    /* If lxi != 0, then set any suitable bit of the significand in hxi.  */
>    hxi |= (lxi | -lxi) >> 31;
> -  /* We do not recognize a pseudo NaN as sNaN; they're invalid on 80387 and
> -     later.  */
>    /* We have to compare for greater (instead of greater or equal), because x's
>       significand being all-zero designates infinity not NaN.  */
> -  return ((exi & 0x7fff) == 0x7fff) && (hxi > 0xc0000000);
> +  return ret || (((exi & 0x7fff) == 0x7fff) && (hxi > 0xc0000000));
>  #endif
>  }
>  libm_hidden_def (__issignalingl)

Ok.

> diff --git a/sysdeps/x86/fpu/nan-pseudo-number.h b/sysdeps/x86/fpu/nan-pseudo-number.h
> new file mode 100644
> index 0000000000..ea43b175a1
> --- /dev/null
> +++ b/sysdeps/x86/fpu/nan-pseudo-number.h
> @@ -0,0 +1,30 @@
> +/* Pseudo-normal number handling.  x86 version.
> +   Copyright (C) 2020 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#ifndef NAN_PSEUDO_NUMBER_H
> +#define NAN_PSEUDO_NUMBER_H	1
> +
> +#define HANDLE_PSEUDO_NUMBERS 1
> +/* Pseudo numbers on x86 are always signaling.  */
> +static inline int
> +is_pseudo_signaling (uint32_t exi, uint32_t hxi)
> +{
> +  return ((exi & 0x7fff) && ((hxi & 0x80000000) == 0));
> +}
> +
> +#endif /* nan-pseudo-number.h */
> 

Ok.
  

On Thu, 24 Dec 2020, Siddhesh Poyarekar via Libc-alpha wrote:

> diff --git a/sysdeps/generic/nan-pseudo-number.h b/sysdeps/generic/nan-pseudo-number.h

I don't think this belongs in sysdeps/generic/.  The "pseudo-number" 
concept is specific to the ldbl-96 format; a "generic" header is really 
something specific to ldbl-96 and so should go in sysdeps/ieee754/ldbl-96/ 
instead.  (x86 and m68k have different rules regarding what encodings are 
valid.  ia64 manuals seem to suggest that only a subset of the encodings 
that are invalid on x86 are also invalid on ia64 but I don't have access 
to hardware to test.)
  

On Dez 31 2020, Joseph Myers wrote:

> On Thu, 24 Dec 2020, Siddhesh Poyarekar via Libc-alpha wrote:
>
>> diff --git a/sysdeps/generic/nan-pseudo-number.h b/sysdeps/generic/nan-pseudo-number.h
>
> I don't think this belongs in sysdeps/generic/.  The "pseudo-number" 
> concept is specific to the ldbl-96 format; a "generic" header is really 
> something specific to ldbl-96 and so should go in sysdeps/ieee754/ldbl-96/ 
> instead.  (x86 and m68k have different rules regarding what encodings are 
> valid.

m68k doesn't have the concept of pseudo-nan and pseudo-inf, as it treats
the explicit integer bit in those formats as don't-care.

Andreas.