Re: [PATCH 26/72] softfloat: Convert float128_silence_nan to parts

[Richard Henderson]

On 5/13/21 3:34 AM, Alex Bennée wrote:
> Richard Henderson <richard.henderson@linaro.org> writes:
>
> > This is the minimal change that also introduces float128_params,
> > float128_unpack_raw, and float128_pack_raw without running into
> > unused symbol Werrors.
> >
> > Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> > ---
> >   fpu/softfloat.c                | 96 +++++++++++++++++++++++++++++-----
> >   fpu/softfloat-specialize.c.inc | 25 +++------
> >   2 files changed, 89 insertions(+), 32 deletions(-)
> >
> > diff --git a/fpu/softfloat.c b/fpu/softfloat.c
> > index 2d6f61ee7a..073b80d502 100644
> > --- a/fpu/softfloat.c
> > +++ b/fpu/softfloat.c
> > @@ -500,14 +500,12 @@ static inline __attribute__((unused)) bool 
> > is_qnan(FloatClass c)
> >   }
> >   
> >   /*
> > - * Structure holding all of the decomposed parts of a float. The
> > - * exponent is unbiased and the fraction is normalized. All
> > - * calculations are done with a 64 bit fraction and then rounded as
> > - * appropriate for the final format.
> > + * Structure holding all of the decomposed parts of a float.
> > + * The exponent is unbiased and the fraction is normalized.
> >    *
> > - * Thanks to the packed FloatClass a decent compiler should be able to
> > - * fit the whole structure into registers and avoid using the stack
> > - * for parameter passing.
> > + * The fraction words are stored in big-endian word ordering,
> > + * so that truncation from a larger format to a smaller format
> > + * can be done simply by ignoring subsequent elements.
> >    */
> >   
> >   typedef struct {
> > @@ -526,6 +524,15 @@ typedef struct {
> >       };
> >   } FloatParts64;
> >   
> > +typedef struct {
> > +    FloatClass cls;
> > +    bool sign;
> > +    int32_t exp;
> > +    uint64_t frac_hi;
> > +    uint64_t frac_lo;
> > +} FloatParts128;
> > +
> > +/* These apply to the most significant word of each FloatPartsN. */
> >   #define DECOMPOSED_BINARY_POINT    63
> >   #define DECOMPOSED_IMPLICIT_BIT    (1ull << DECOMPOSED_BINARY_POINT)
> >   
> > @@ -561,11 +568,11 @@ typedef struct {
> >       .exp_bias       = ((1 << E) - 1) >> 1,                           \
> >       .exp_max        = (1 << E) - 1,                                  \
> >       .frac_size      = F,                                             \
> > -    .frac_shift     = DECOMPOSED_BINARY_POINT - F,                   \
> > -    .frac_lsb       = 1ull << (DECOMPOSED_BINARY_POINT - F),         \
> > -    .frac_lsbm1     = 1ull << ((DECOMPOSED_BINARY_POINT - F) - 1),   \
> > -    .round_mask     = (1ull << (DECOMPOSED_BINARY_POINT - F)) - 1,   \
> > -    .roundeven_mask = (2ull << (DECOMPOSED_BINARY_POINT - F)) - 1
> > +    .frac_shift     = (-F - 1) & 63,                                 \
> > +    .frac_lsb       = 1ull << ((-F - 1) & 63),                       \
> > +    .frac_lsbm1     = 1ull << ((-F - 2) & 63),                       \
> > +    .round_mask     = (1ull << ((-F - 1) & 63)) - 1,                 \
> > +    .roundeven_mask = (2ull << ((-F - 1) & 63)) - 1
>
> I have to admit I find the switch to (-F - 1) & 63 a little black
> magical. Isn't the shift always going to end up a factor of the number
> of exponent bits we need to move past and the natural size of the
> original float?

Yep.  But now we're looking to compute the number relative to .frac_lo, rather 
than the entire logical fraction.


r~

> Anyway my personal brain twisting aside it obviously works and
> everything else looks fine so:
>
> Reviewed-by: Alex Bennée <alex.bennee@linaro.org>