[PATCH v2 04/69] fpu: Implement float_flag_input_denormal_used

[Peter Maydell <peter.maydell@linaro.org>]

For the x86 and the Arm FEAT_AFP semantics, we need to be able to
tell the target code that the FPU operation has used an input
denormal.  Implement this; when it happens we set the new
float_flag_denormal_input_used.

Note that we only set this when an input denormal is actually used by
the operation: if the operation results in Invalid Operation or
Divide By Zero or the result is a NaN because some other input was a
NaN then we never needed to look at the input denormal and do not set
denormal_input_used.

We mostly do not need to adjust the hardfloat codepaths to deal with
this flag, because almost all hardfloat operations are already gated
on the input not being a denormal, and will fall back to softfloat
for a denormal input.  The only exception is the comparison
operations, where we need to add the check for input denormals, which
must now fall back to softfloat where they did not before.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
v2: drop the now-unnecessary float*_input_flush2() calls in
the float*_hs_compare() functions
---
 include/fpu/softfloat-types.h |  7 ++++
 fpu/softfloat.c               | 38 +++++++++++++++++---
 fpu/softfloat-parts.c.inc     | 68 ++++++++++++++++++++++++++++++++++-
 3 files changed, 107 insertions(+), 6 deletions(-)

diff --git a/include/fpu/softfloat-types.h b/include/fpu/softfloat-types.h
index c177923e319..b9b4e8e55fc 100644
--- a/include/fpu/softfloat-types.h
+++ b/include/fpu/softfloat-types.h
@@ -165,6 +165,13 @@ enum {
     float_flag_invalid_sqrt    = 0x0800,  /* sqrt(-x) */
     float_flag_invalid_cvti    = 0x1000,  /* non-nan to integer */
     float_flag_invalid_snan    = 0x2000,  /* any operand was snan */
+    /*
+     * An input was denormal and we used it (without flushing it to zero).
+     * Not set if we do not actually use the denormal input (e.g.
+     * because some other input was a NaN, or because the operation
+     * wasn't actually carried out (divide-by-zero; invalid))
+     */
+    float_flag_input_denormal_used = 0x4000,
 };
 
 /*
diff --git a/fpu/softfloat.c b/fpu/softfloat.c
index 03a604c38ec..f4fed9bfda9 100644
--- a/fpu/softfloat.c
+++ b/fpu/softfloat.c
@@ -2718,8 +2718,10 @@ static void parts_float_to_ahp(FloatParts64 *a, float_status *s)
                                   float16_params_ahp.frac_size + 1);
         break;
 
-    case float_class_normal:
     case float_class_denormal:
+        float_raise(float_flag_input_denormal_used, s);
+        break;
+    case float_class_normal:
     case float_class_zero:
         break;
 
@@ -2733,6 +2735,9 @@ static void parts64_float_to_float(FloatParts64 *a, float_status *s)
     if (is_nan(a->cls)) {
         parts_return_nan(a, s);
     }
+    if (a->cls == float_class_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
 }
 
 static void parts128_float_to_float(FloatParts128 *a, float_status *s)
@@ -2740,6 +2745,9 @@ static void parts128_float_to_float(FloatParts128 *a, float_status *s)
     if (is_nan(a->cls)) {
         parts_return_nan(a, s);
     }
+    if (a->cls == float_class_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
 }
 
 #define parts_float_to_float(P, S) \
@@ -2752,12 +2760,21 @@ static void parts_float_to_float_narrow(FloatParts64 *a, FloatParts128 *b,
     a->sign = b->sign;
     a->exp = b->exp;
 
-    if (is_anynorm(a->cls)) {
+    switch (a->cls) {
+    case float_class_denormal:
+        float_raise(float_flag_input_denormal_used, s);
+        /* fall through */
+    case float_class_normal:
         frac_truncjam(a, b);
-    } else if (is_nan(a->cls)) {
+        break;
+    case float_class_snan:
+    case float_class_qnan:
         /* Discard the low bits of the NaN. */
         a->frac = b->frac_hi;
         parts_return_nan(a, s);
+        break;
+    default:
+        break;
     }
 }
 
@@ -2772,6 +2789,9 @@ static void parts_float_to_float_widen(FloatParts128 *a, FloatParts64 *b,
     if (is_nan(a->cls)) {
         parts_return_nan(a, s);
     }
+    if (a->cls == float_class_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
 }
 
 float32 float16_to_float32(float16 a, bool ieee, float_status *s)
@@ -4411,7 +4431,11 @@ float32_hs_compare(float32 xa, float32 xb, float_status *s, bool is_quiet)
         goto soft;
     }
 
-    float32_input_flush2(&ua.s, &ub.s, s);
+    if (unlikely(float32_is_denormal(ua.s) || float32_is_denormal(ub.s))) {
+        /* We may need to set the input_denormal_used flag */
+        goto soft;
+    }
+
     if (isgreaterequal(ua.h, ub.h)) {
         if (isgreater(ua.h, ub.h)) {
             return float_relation_greater;
@@ -4461,7 +4485,11 @@ float64_hs_compare(float64 xa, float64 xb, float_status *s, bool is_quiet)
         goto soft;
     }
 
-    float64_input_flush2(&ua.s, &ub.s, s);
+    if (unlikely(float64_is_denormal(ua.s) || float64_is_denormal(ub.s))) {
+        /* We may need to set the input_denormal_used flag */
+        goto soft;
+    }
+
     if (isgreaterequal(ua.h, ub.h)) {
         if (isgreater(ua.h, ub.h)) {
             return float_relation_greater;
diff --git a/fpu/softfloat-parts.c.inc b/fpu/softfloat-parts.c.inc
index 8621cb87185..0122b35008a 100644
--- a/fpu/softfloat-parts.c.inc
+++ b/fpu/softfloat-parts.c.inc
@@ -433,6 +433,15 @@ static FloatPartsN *partsN(addsub)(FloatPartsN *a, FloatPartsN *b,
     bool b_sign = b->sign ^ subtract;
     int ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
 
+    /*
+     * For addition and subtraction, we will consume an
+     * input denormal unless the other input is a NaN.
+     */
+    if ((ab_mask & (float_cmask_denormal | float_cmask_anynan)) ==
+        float_cmask_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
+
     if (a->sign != b_sign) {
         /* Subtraction */
         if (likely(cmask_is_only_normals(ab_mask))) {
@@ -516,6 +525,10 @@ static FloatPartsN *partsN(mul)(FloatPartsN *a, FloatPartsN *b,
     if (likely(cmask_is_only_normals(ab_mask))) {
         FloatPartsW tmp;
 
+        if (ab_mask & float_cmask_denormal) {
+            float_raise(float_flag_input_denormal_used, s);
+        }
+
         frac_mulw(&tmp, a, b);
         frac_truncjam(a, &tmp);
 
@@ -541,6 +554,10 @@ static FloatPartsN *partsN(mul)(FloatPartsN *a, FloatPartsN *b,
     }
 
     /* Multiply by 0 or Inf */
+    if (ab_mask & float_cmask_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
+
     if (ab_mask & float_cmask_inf) {
         a->cls = float_class_inf;
         a->sign = sign;
@@ -664,6 +681,16 @@ static FloatPartsN *partsN(muladd_scalbn)(FloatPartsN *a, FloatPartsN *b,
     if (flags & float_muladd_negate_result) {
         a->sign ^= 1;
     }
+
+    /*
+     * All result types except for "return the default NaN
+     * because this is an Invalid Operation" go through here;
+     * this matches the set of cases where we consumed a
+     * denormal input.
+     */
+    if (abc_mask & float_cmask_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
     return a;
 
  return_sub_zero:
@@ -693,6 +720,9 @@ static FloatPartsN *partsN(div)(FloatPartsN *a, FloatPartsN *b,
     bool sign = a->sign ^ b->sign;
 
     if (likely(cmask_is_only_normals(ab_mask))) {
+        if (ab_mask & float_cmask_denormal) {
+            float_raise(float_flag_input_denormal_used, s);
+        }
         a->sign = sign;
         a->exp -= b->exp + frac_div(a, b);
         return a;
@@ -713,6 +743,10 @@ static FloatPartsN *partsN(div)(FloatPartsN *a, FloatPartsN *b,
         return parts_pick_nan(a, b, s);
     }
 
+    if ((ab_mask & float_cmask_denormal) && b->cls != float_class_zero) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
+
     a->sign = sign;
 
     /* Inf / X */
@@ -751,6 +785,9 @@ static FloatPartsN *partsN(modrem)(FloatPartsN *a, FloatPartsN *b,
     int ab_mask = float_cmask(a->cls) | float_cmask(b->cls);
 
     if (likely(cmask_is_only_normals(ab_mask))) {
+        if (ab_mask & float_cmask_denormal) {
+            float_raise(float_flag_input_denormal_used, s);
+        }
         frac_modrem(a, b, mod_quot);
         return a;
     }
@@ -771,6 +808,10 @@ static FloatPartsN *partsN(modrem)(FloatPartsN *a, FloatPartsN *b,
         return a;
     }
 
+    if (ab_mask & float_cmask_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
+
     /* N % Inf; 0 % N */
     g_assert(b->cls == float_class_inf || a->cls == float_class_zero);
     return a;
@@ -801,6 +842,10 @@ static void partsN(sqrt)(FloatPartsN *a, float_status *status,
     if (unlikely(a->cls != float_class_normal)) {
         switch (a->cls) {
         case float_class_denormal:
+            if (!a->sign) {
+                /* -ve denormal will be InvalidOperation */
+                float_raise(float_flag_input_denormal_used, status);
+            }
             break;
         case float_class_snan:
         case float_class_qnan:
@@ -1431,6 +1476,9 @@ static FloatPartsN *partsN(minmax)(FloatPartsN *a, FloatPartsN *b,
         if ((flags & (minmax_isnum | minmax_isnumber))
             && !(ab_mask & float_cmask_snan)
             && (ab_mask & ~float_cmask_qnan)) {
+            if (ab_mask & float_cmask_denormal) {
+                float_raise(float_flag_input_denormal_used, s);
+            }
             return is_nan(a->cls) ? b : a;
         }
 
@@ -1455,6 +1503,10 @@ static FloatPartsN *partsN(minmax)(FloatPartsN *a, FloatPartsN *b,
         return parts_pick_nan(a, b, s);
     }
 
+    if (ab_mask & float_cmask_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
+
     a_exp = a->exp;
     b_exp = b->exp;
 
@@ -1524,6 +1576,10 @@ static FloatRelation partsN(compare)(FloatPartsN *a, FloatPartsN *b,
     if (likely(cmask_is_only_normals(ab_mask))) {
         FloatRelation cmp;
 
+        if (ab_mask & float_cmask_denormal) {
+            float_raise(float_flag_input_denormal_used, s);
+        }
+
         if (a->sign != b->sign) {
             goto a_sign;
         }
@@ -1549,6 +1605,10 @@ static FloatRelation partsN(compare)(FloatPartsN *a, FloatPartsN *b,
         return float_relation_unordered;
     }
 
+    if (ab_mask & float_cmask_denormal) {
+        float_raise(float_flag_input_denormal_used, s);
+    }
+
     if (ab_mask & float_cmask_zero) {
         if (ab_mask == float_cmask_zero) {
             return float_relation_equal;
@@ -1588,8 +1648,10 @@ static void partsN(scalbn)(FloatPartsN *a, int n, float_status *s)
     case float_class_zero:
     case float_class_inf:
         break;
-    case float_class_normal:
     case float_class_denormal:
+        float_raise(float_flag_input_denormal_used, s);
+        /* fall through */
+    case float_class_normal:
         a->exp += MIN(MAX(n, -0x10000), 0x10000);
         break;
     default:
@@ -1609,6 +1671,10 @@ static void partsN(log2)(FloatPartsN *a, float_status *s, const FloatFmt *fmt)
     if (unlikely(a->cls != float_class_normal)) {
         switch (a->cls) {
         case float_class_denormal:
+            if (!a->sign) {
+                /* -ve denormal will be InvalidOperation */
+                float_raise(float_flag_input_denormal_used, s);
+            }
             break;
         case float_class_snan:
         case float_class_qnan:
-- 
2.34.1