diff options
Diffstat (limited to 'src/math/__fpclassifyl.c')
| -rw-r--r-- | src/math/__fpclassifyl.c | 12 |
1 files changed, 10 insertions, 2 deletions
diff --git a/src/math/__fpclassifyl.c b/src/math/__fpclassifyl.c index 481c0b94..e41781b6 100644 --- a/src/math/__fpclassifyl.c +++ b/src/math/__fpclassifyl.c @@ -13,10 +13,18 @@ int __fpclassifyl(long double x) int msb = u.i.m>>63; if (!e && !msb) return u.i.m ? FP_SUBNORMAL : FP_ZERO; + if (e == 0x7fff) { + /* The x86 variant of 80-bit extended precision only admits + * one representation of each infinity, with the mantissa msb + * necessarily set. The version with it clear is invalid/nan. + * The m68k variant, however, allows either, and tooling uses + * the version with it clear. */ + if (__BYTE_ORDER == __LITTLE_ENDIAN && !msb) + return FP_NAN; + return u.i.m << 1 ? FP_NAN : FP_INFINITE; + } if (!msb) return FP_NAN; - if (e == 0x7fff) - return u.i.m << 1 ? FP_NAN : FP_INFINITE; return FP_NORMAL; } #elif LDBL_MANT_DIG == 113 && LDBL_MAX_EXP == 16384 |