Re: [PATCH 01/21] softfloat: make NO_SIGNALING_NANS runtime property

["Alex Bennée" <alex.bennee@linaro.org>]

Max Filippov <jcmvbkbc@gmail.com> writes:

> target/xtensa, the only user of NO_SIGNALING_NANS macro has FPU
> implementations with and without the corresponding property. With
> NO_SIGNALING_NANS being a macro they cannot be a part of the same QEMU
> executable.
> Replace macro with new property in float_status to allow cores with
> different FPU implementations coexist.
>
> Cc: Peter Maydell <peter.maydell@linaro.org>
> Cc: "Alex Bennée" <alex.bennee@linaro.org>
> Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
> ---
>  fpu/softfloat-specialize.inc.c  | 228 ++++++++++++++++----------------
>  include/fpu/softfloat-helpers.h |   5 +
>  include/fpu/softfloat-types.h   |   1 +
>  3 files changed, 117 insertions(+), 117 deletions(-)
>
> diff --git a/fpu/softfloat-specialize.inc.c b/fpu/softfloat-specialize.inc.c
> index 44f5b661f831..b26bc039b0b6 100644
> --- a/fpu/softfloat-specialize.inc.c
> +++ b/fpu/softfloat-specialize.inc.c
> @@ -79,13 +79,6 @@ this code that are retained.
>   * version 2 or later. See the COPYING file in the top-level directory.
>   */
>  
> -/* Define for architectures which deviate from IEEE in not supporting
> - * signaling NaNs (so all NaNs are treated as quiet).
> - */
> -#if defined(TARGET_XTENSA)
> -#define NO_SIGNALING_NANS 1
> -#endif
> -
>  /* Define how the architecture discriminates signaling NaNs.
>   * This done with the most significant bit of the fraction.
>   * In IEEE 754-1985 this was implementation defined, but in IEEE 754-2008
> @@ -111,12 +104,12 @@ static inline bool snan_bit_is_one(float_status *status)
>  
>  static bool parts_is_snan_frac(uint64_t frac, float_status *status)
>  {
> -#ifdef NO_SIGNALING_NANS
> -    return false;
> -#else
> -    bool msb = extract64(frac, DECOMPOSED_BINARY_POINT - 1, 1);
> -    return msb == snan_bit_is_one(status);
> -#endif
> +    if (status->no_signaling_nans) {
> +        return false;

I have no objection in principle but seeing as we go to the trouble of
allowing snan_bit_is_one() to constant fold away I think it would be
worth doing the same with a no_signalling_nans(status). We can then
avoid an admittedly well predicted test for the non XTENSA case.

> +    } else {
> +        bool msb = extract64(frac, DECOMPOSED_BINARY_POINT - 1, 1);
> +        return msb == snan_bit_is_one(status);
> +    }
>  }
>  
>  /*----------------------------------------------------------------------------
> @@ -170,9 +163,10 @@ static FloatParts parts_default_nan(float_status *status)
>  
>  static FloatParts parts_silence_nan(FloatParts a, float_status *status)
>  {
> -#ifdef NO_SIGNALING_NANS
> -    g_assert_not_reached();

We could then keep the assert:

  g_assert(!no_signaling_nan(status))

<snip>
>      return status->tininess_before_rounding;
> diff --git a/include/fpu/softfloat-types.h b/include/fpu/softfloat-types.h
> index 7680193ebc1c..10bd208e559f 100644
> --- a/include/fpu/softfloat-types.h
> +++ b/include/fpu/softfloat-types.h
> @@ -167,6 +167,7 @@ typedef struct float_status {
>      bool default_nan_mode;
>      /* not always used -- see snan_bit_is_one() in
>      softfloat-specialize.h */

and then expand this comment:

    /* the flags bellow are not used on all specializations and may
     * constant fold away (see snan_bit_is_one()/no_signalling_nans() in
     * softfloat-specialize.h)
     */

>      bool snan_bit_is_one;
> +    bool no_signaling_nans;
>  } float_status;
>  
>  #endif /* SOFTFLOAT_TYPES_H */


-- 
Alex Bennée