Skip to contentSkip to conversation

softfloat: add float128_is_{normal, denormal}

PATCH Superseded David Hildenbrand 2 participants 3 messages Feb 5, 2019 – Feb 6, 2019

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 softfloat: add float128_is_{normal, denormal}

Needed on s390x, to test for the data class of a number. So it will gain soon a user.

---
 include/fpu/softfloat.h | 10 ++++++++++
 1 file changed, 10 insertions(+)
include/fpu/softfloat.h +10 -0
index 38a5e99cf3..10c31557b4 100644
@@ -940,6 +940,16 @@ static inline int float128_is_zero_or_denormal(float128 a)
     return (a.high & 0x7fff000000000000LL) == 0;
 }
 
+static inline bool float128_is_normal(float128 a)
+{
+    return ((a.high + (1ULL << 47)) & -1ULL >> 1) >= 1ULL << 48;
+}
+
+static inline bool float128_is_denormal(float128 a)
+{
+    return float128_is_zero_or_denormal(a) && !float128_is_zero(a);
+}
+
 static inline int float128_is_any_nan(float128 a)
 {
     return ((a.high >> 48) & 0x7fff) == 0x7fff &&
Signature · 1 line
2.17.2
#1 lore raw

include/fpu/softfloat.h static inline int float128_is_zero_or_denormal(float128 a) · line 945

Quoted 3 lines · David Hildenbrand
include/fpu/softfloat.h +3 -0
+static inline bool float128_is_normal(float128 a)
+{
+    return ((a.high + (1ULL << 47)) & -1ULL >> 1) >= 1ULL << 48;

I believe this is off by one: 1 << 48 and >= 1 << 49.

The exponent is at [62:48]. The trick is adding 1, letting Inf+NaN overflow into the sign, masking out the sign, and checking that the result >= 2 to eliminate Inf+NaN (0) and Zero+Denormal (1).

I think this is clearer as

(((a.high >> 48) + 1) & 0x7fff) >= 2.

It might be worth applying this to the other formats for clarity...

r~

#2 lore raw
Quoted 10 lines · Richard Henderson

include/fpu/softfloat.h static inline int float128_is_zero_or_denormal(float128 a) · line 945

Quoted 3 lines · David Hildenbrand
include/fpu/softfloat.h +3 -0
+static inline bool float128_is_normal(float128 a)
+{
+    return ((a.high + (1ULL << 47)) & -1ULL >> 1) >= 1ULL << 48;

I believe this is off by one: 1 << 48 and >= 1 << 49.

The exponent is at [62:48]. The trick is adding 1, letting Inf+NaN overflow into the sign, masking out the sign, and checking that the result >= 2 to eliminate Inf+NaN (0) and Zero+Denormal (1).

Right, the exponent must not be 0 or all 1's. so after adding 1, it must be >=2.

Yesterday it all made sense on my sheet of paper :) Let's verify against float64

float64 has an exponent of 11:

return ((float64_val(a) + (1ULL << 52)) & -1ULL >> 1) >= 1ULL << 53;

float128 has an exponent of 15, the difference is 4. So 52 -> 48, 53 -> 49.

So you're right.

Quoted 6 lines

I think this is clearer as

(((a.high >> 48) + 1) & 0x7fff) >= 2.>
It might be worth applying this to the other formats for clarity...

Yes, this makes it much clearer, will send a patch for the others as well, thanks!

Quoted 3 lines

r~

Signature · 3 lines
Thanks,

David / dhildenb