Re: [PATCH 50/72] softfloat: Move minmax_flags to softfloat-parts.c.inc

[David Hildenbrand <david@redhat.com>]

On 08.05.21 03:47, Richard Henderson wrote:
> Rename to parts$N_minmax.  Combine 3 bool arguments to a bitmask,
> return a tri-state value to indicate nan vs unchanged operand.
> Introduce ftype_minmax functions as a common optimization point.
> Fold bfloat16 expansions into the same macro as the other types.
> 
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>   fpu/softfloat.c           | 216 ++++++++++++++++----------------------
>   fpu/softfloat-parts.c.inc |  69 ++++++++++++
>   2 files changed, 158 insertions(+), 127 deletions(-)
> 
> diff --git a/fpu/softfloat.c b/fpu/softfloat.c
> index 586ea5d67a..4c04e88a3a 100644
> --- a/fpu/softfloat.c
> +++ b/fpu/softfloat.c
> @@ -482,6 +482,15 @@ enum {
>       float_cmask_anynan  = float_cmask_qnan | float_cmask_snan,
>   };
>   
> +/* Flags for parts_minmax. */
> +enum {
> +    /* Set for minimum; clear for maximum. */
> +    minmax_ismin = 1,
> +    /* Set for the IEEE 754-2008 minNum() and maxNum() operations. */
> +    minmax_isnum = 2,
> +    /* Set for the IEEE 754-2008 minNumMag() and minNumMag() operations. */
> +    minmax_ismag = 4 | minmax_isnum
> +};
>   
>   /* Simple helpers for checking if, or what kind of, NaN we have */
>   static inline __attribute__((unused)) bool is_nan(FloatClass c)
> @@ -864,6 +873,14 @@ static void parts128_uint_to_float(FloatParts128 *p, uint64_t a,
>   #define parts_uint_to_float(P, I, Z, S) \
>       PARTS_GENERIC_64_128(uint_to_float, P)(P, I, Z, S)
>   
> +static int parts64_minmax(FloatParts64 *a, FloatParts64 *b,
> +                          float_status *s, int flags, const FloatFmt *fmt);
> +static int parts128_minmax(FloatParts128 *a, FloatParts128 *b,
> +                           float_status *s, int flags, const FloatFmt *fmt);
> +
> +#define parts_minmax(A, B, S, Z, F) \
> +    PARTS_GENERIC_64_128(minmax, A)(A, B, S, Z, F)
> +
>   /*
>    * Helper functions for softfloat-parts.c.inc, per-size operations.
>    */
> @@ -3257,145 +3274,90 @@ float128 uint64_to_float128(uint64_t a, float_status *status)
>       return float128_round_pack_canonical(&p, status);
>   }
>   
> -/* Float Min/Max */
> -/* min() and max() functions. These can't be implemented as
> - * 'compare and pick one input' because that would mishandle
> - * NaNs and +0 vs -0.
> - *
> - * minnum() and maxnum() functions. These are similar to the min()
> - * and max() functions but if one of the arguments is a QNaN and
> - * the other is numerical then the numerical argument is returned.
> - * SNaNs will get quietened before being returned.
> - * minnum() and maxnum correspond to the IEEE 754-2008 minNum()
> - * and maxNum() operations. min() and max() are the typical min/max
> - * semantics provided by many CPUs which predate that specification.
> - *
> - * minnummag() and maxnummag() functions correspond to minNumMag()
> - * and minNumMag() from the IEEE-754 2008.
> +/*
> + * Minimum and maximum
>    */
> -static FloatParts64 minmax_floats(FloatParts64 a, FloatParts64 b, bool ismin,
> -                                bool ieee, bool ismag, float_status *s)
> +
> +static float16 float16_minmax(float16 a, float16 b, float_status *s, int flags)
>   {
> -    if (unlikely(is_nan(a.cls) || is_nan(b.cls))) {
> -        if (ieee) {
> -            /* Takes two floating-point values `a' and `b', one of
> -             * which is a NaN, and returns the appropriate NaN
> -             * result. If either `a' or `b' is a signaling NaN,
> -             * the invalid exception is raised.
> -             */
> -            if (is_snan(a.cls) || is_snan(b.cls)) {
> -                return *parts_pick_nan(&a, &b, s);
> -            } else if (is_nan(a.cls) && !is_nan(b.cls)) {
> -                return b;
> -            } else if (is_nan(b.cls) && !is_nan(a.cls)) {
> -                return a;
> -            }
> -        }
> -        return *parts_pick_nan(&a, &b, s);
> -    } else {
> -        int a_exp, b_exp;
> +    FloatParts64 pa, pb;
> +    int which;
>   
> -        switch (a.cls) {
> -        case float_class_normal:
> -            a_exp = a.exp;
> -            break;
> -        case float_class_inf:
> -            a_exp = INT_MAX;
> -            break;
> -        case float_class_zero:
> -            a_exp = INT_MIN;
> -            break;
> -        default:
> -            g_assert_not_reached();
> -            break;
> -        }
> -        switch (b.cls) {
> -        case float_class_normal:
> -            b_exp = b.exp;
> -            break;
> -        case float_class_inf:
> -            b_exp = INT_MAX;
> -            break;
> -        case float_class_zero:
> -            b_exp = INT_MIN;
> -            break;
> -        default:
> -            g_assert_not_reached();
> -            break;
> -        }
> -
> -        if (ismag && (a_exp != b_exp || a.frac != b.frac)) {
> -            bool a_less = a_exp < b_exp;
> -            if (a_exp == b_exp) {
> -                a_less = a.frac < b.frac;
> -            }
> -            return a_less ^ ismin ? b : a;
> -        }
> -
> -        if (a.sign == b.sign) {
> -            bool a_less = a_exp < b_exp;
> -            if (a_exp == b_exp) {
> -                a_less = a.frac < b.frac;
> -            }
> -            return a.sign ^ a_less ^ ismin ? b : a;
> -        } else {
> -            return a.sign ^ ismin ? b : a;
> -        }
> +    float16_unpack_canonical(&pa, a, s);
> +    float16_unpack_canonical(&pb, b, s);
> +    which = parts_minmax(&pa, &pb, s, flags, &float16_params);
> +    if (unlikely(which < 0)) {
> +        /* Some sort of nan, need to repack default and silenced nans. */
> +        return float16_round_pack_canonical(&pa, s);
>       }
> +    return which ? b : a;
>   }
>   
> -#define MINMAX(sz, name, ismin, isiee, ismag)                           \
> -float ## sz float ## sz ## _ ## name(float ## sz a, float ## sz b,      \
> -                                     float_status *s)                   \
> -{                                                                       \
> -    FloatParts64 pa, pb, pr;                                            \
> -    float ## sz ## _unpack_canonical(&pa, a, s);                        \
> -    float ## sz ## _unpack_canonical(&pb, b, s);                        \
> -    pr = minmax_floats(pa, pb, ismin, isiee, ismag, s);                 \
> -    return float ## sz ## _round_pack_canonical(&pr, s);                \
> +static bfloat16 bfloat16_minmax(bfloat16 a, bfloat16 b,
> +                                float_status *s, int flags)
> +{
> +    FloatParts64 pa, pb;
> +    int which;
> +
> +    bfloat16_unpack_canonical(&pa, a, s);
> +    bfloat16_unpack_canonical(&pb, b, s);
> +    which = parts_minmax(&pa, &pb, s, flags, &float16_params);
> +    if (unlikely(which < 0)) {
> +        /* Some sort of nan, need to repack default and silenced nans. */
> +        return bfloat16_round_pack_canonical(&pa, s);
> +    }
> +    return which ? b : a;
>   }
>   
> -MINMAX(16, min, true, false, false)
> -MINMAX(16, minnum, true, true, false)
> -MINMAX(16, minnummag, true, true, true)
> -MINMAX(16, max, false, false, false)
> -MINMAX(16, maxnum, false, true, false)
> -MINMAX(16, maxnummag, false, true, true)
> +static float32 float32_minmax(float32 a, float32 b, float_status *s, int flags)
> +{
> +    FloatParts64 pa, pb;
> +    int which;
>   
> -MINMAX(32, min, true, false, false)
> -MINMAX(32, minnum, true, true, false)
> -MINMAX(32, minnummag, true, true, true)
> -MINMAX(32, max, false, false, false)
> -MINMAX(32, maxnum, false, true, false)
> -MINMAX(32, maxnummag, false, true, true)
> -
> -MINMAX(64, min, true, false, false)
> -MINMAX(64, minnum, true, true, false)
> -MINMAX(64, minnummag, true, true, true)
> -MINMAX(64, max, false, false, false)
> -MINMAX(64, maxnum, false, true, false)
> -MINMAX(64, maxnummag, false, true, true)
> -
> -#undef MINMAX
> -
> -#define BF16_MINMAX(name, ismin, isiee, ismag)                          \
> -bfloat16 bfloat16_ ## name(bfloat16 a, bfloat16 b, float_status *s)     \
> -{                                                                       \
> -    FloatParts64 pa, pb, pr;                                            \
> -    bfloat16_unpack_canonical(&pa, a, s);                               \
> -    bfloat16_unpack_canonical(&pb, b, s);                               \
> -    pr = minmax_floats(pa, pb, ismin, isiee, ismag, s);                 \
> -    return bfloat16_round_pack_canonical(&pr, s);                       \
> +    float32_unpack_canonical(&pa, a, s);
> +    float32_unpack_canonical(&pb, b, s);
> +    which = parts_minmax(&pa, &pb, s, flags, &float32_params);
> +    if (unlikely(which < 0)) {
> +        /* Some sort of nan, need to repack default and silenced nans. */
> +        return float32_round_pack_canonical(&pa, s);
> +    }
> +    return which ? b : a;
>   }
>   
> -BF16_MINMAX(min, true, false, false)
> -BF16_MINMAX(minnum, true, true, false)
> -BF16_MINMAX(minnummag, true, true, true)
> -BF16_MINMAX(max, false, false, false)
> -BF16_MINMAX(maxnum, false, true, false)
> -BF16_MINMAX(maxnummag, false, true, true)
> +static float64 float64_minmax(float64 a, float64 b, float_status *s, int flags)
> +{
> +    FloatParts64 pa, pb;
> +    int which;
>   
> -#undef BF16_MINMAX
> +    float64_unpack_canonical(&pa, a, s);
> +    float64_unpack_canonical(&pb, b, s);
> +    which = parts_minmax(&pa, &pb, s, flags, &float64_params);
> +    if (unlikely(which < 0)) {
> +        /* Some sort of nan, need to repack default and silenced nans. */
> +        return float64_round_pack_canonical(&pa, s);
> +    }
> +    return which ? b : a;
> +}
> +
> +#define MINMAX_1(type, name, flags) \
> +    type type##_##name(type a, type b, float_status *s) \
> +    { return type##_minmax(a, b, s, flags); }
> +
> +#define MINMAX_2(type) \
> +    MINMAX_1(type, max, 0)                                      \
> +    MINMAX_1(type, maxnum, minmax_isnum)                        \
> +    MINMAX_1(type, maxnummag, minmax_ismag)                     \
> +    MINMAX_1(type, min, minmax_ismin)                           \
> +    MINMAX_1(type, minnum, minmax_ismin | minmax_isnum)         \
> +    MINMAX_1(type, minnummag, minmax_ismin | minmax_ismag)
> +
> +MINMAX_2(float16)
> +MINMAX_2(bfloat16)
> +MINMAX_2(float32)
> +MINMAX_2(float64)
> +
> +#undef MINMAX_1
> +#undef MINMAX_2
>   
>   /* Floating point compare */
>   static FloatRelation compare_floats(FloatParts64 a, FloatParts64 b, bool is_quiet,
> diff --git a/fpu/softfloat-parts.c.inc b/fpu/softfloat-parts.c.inc
> index f3c4f8c8d2..4d91ef0d32 100644
> --- a/fpu/softfloat-parts.c.inc
> +++ b/fpu/softfloat-parts.c.inc
> @@ -936,3 +936,72 @@ static void partsN(uint_to_float)(FloatPartsN *p, uint64_t a,
>           p->frac_hi = a << shift;
>       }
>   }
> +
> +/*
> + * Float min/max.
> + *
> + * Return -1 to return the chosen nan in *a;
> + * return 0 to use the a input unchanged; 1 to use the b input unchanged.
> + */
> +static int partsN(minmax)(FloatPartsN *a, FloatPartsN *b,
> +                          float_status *s, int flags, const FloatFmt *fmt)
> +{
> +    int ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
> +    int a_exp, b_exp;
> +    bool a_less;
> +
> +    if (unlikely(ab_mask & float_cmask_anynan)) {
> +        /*
> +         * For minnum/maxnum, if one operand is a QNaN, and the other
> +         * operand is numerical, then return numerical argument.
> +         */
> +        if ((flags & minmax_isnum)
> +            && !(ab_mask & float_cmask_snan)
> +            && (ab_mask & ~float_cmask_qnan)) {
> +            return is_nan(a->cls);
> +        }
> +        *a = *parts_pick_nan(a, b, s);
> +        return -1;
> +    }
> +
> +    a_exp = a->exp;
> +    b_exp = b->exp;
> +
> +    if (unlikely(ab_mask != float_cmask_normal)) {
> +        switch (a->cls) {
> +        case float_class_normal:
> +            break;
> +        case float_class_inf:
> +            a_exp = INT_MAX;
> +            break;
> +        case float_class_zero:
> +            a_exp = INT_MIN;
> +            break;
> +        default:
> +            g_assert_not_reached();
> +            break;
> +        }
> +        switch (b->cls) {
> +        case float_class_normal:
> +            break;
> +        case float_class_inf:
> +            b_exp = INT_MAX;
> +            break;
> +        case float_class_zero:
> +            b_exp = INT_MIN;
> +            break;
> +        default:
> +            g_assert_not_reached();
> +            break;
> +        }
> +    }
> +
> +    if (a->sign != b->sign && !(flags & minmax_ismag)) {
> +        a_less = a->sign;
> +    } else if (a_exp != b_exp) {
> +        a_less = a_exp < b_exp;
> +    } else {
> +        a_less = frac_cmp(a, b) < 0;
> +    }
> +    return a_less ^ !!(flags & minmax_ismin);
> +}

This patch introduces two issues:

1. Comparing two negative numbers is broken. We have to
invert the a_less result.

2. The check "flags & minmax_ismag" is broken because
"minmax_ismag = 4 | minmax_isnum" and it, therefore,
also triggers for "flags = minmax_isnum"

The following seems to make it work in my tests:


diff --git a/fpu/softfloat-parts.c.inc b/fpu/softfloat-parts.c.inc
index 1bc8db4cf1..d5a73dac00 100644
--- a/fpu/softfloat-parts.c.inc
+++ b/fpu/softfloat-parts.c.inc
@@ -1273,6 +1273,9 @@ static int partsN(minmax)(FloatPartsN *a, FloatPartsN *b,
      } else {
          a_less = frac_cmp(a, b) < 0;
      }
+    if (a->sign && b->sign && !(flags & minmax_ismag)) {
+        a_less = !a_less;
+    }
      return a_less ^ !!(flags & minmax_ismin);
  }
  
diff --git a/fpu/softfloat.c b/fpu/softfloat.c
index bfe5a6b975..649e8c6592 100644
--- a/fpu/softfloat.c
+++ b/fpu/softfloat.c
@@ -435,7 +435,7 @@ enum {
      /* Set for the IEEE 754-2008 minNum() and maxNum() operations. */
      minmax_isnum = 2,
      /* Set for the IEEE 754-2008 minNumMag() and minNumMag() operations. */
-    minmax_ismag = 4 | minmax_isnum
+    minmax_ismag = 4,
  };
  
  /* Simple helpers for checking if, or what kind of, NaN we have */
@@ -3916,10 +3916,10 @@ static float128 float128_minmax(float128 a, float128 b, float_status *s,
  #define MINMAX_2(type) \
      MINMAX_1(type, max, 0)                                      \
      MINMAX_1(type, maxnum, minmax_isnum)                        \
-    MINMAX_1(type, maxnummag, minmax_ismag)                     \
+    MINMAX_1(type, maxnummag, minmax_isnum | minmax_ismag)      \
      MINMAX_1(type, min, minmax_ismin)                           \
      MINMAX_1(type, minnum, minmax_ismin | minmax_isnum)         \
-    MINMAX_1(type, minnummag, minmax_ismin | minmax_ismag)
+    MINMAX_1(type, minnummag, minmax_ismin | minmax_isnum | minmax_ismag)
  
  MINMAX_2(float16)
  MINMAX_2(bfloat16)


-- 
Thanks,

David / dhildenb