diff options
| author | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2018-05-14 10:14:49 +0100 |
|---|---|---|
| committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2018-05-16 13:52:13 +0100 |
| commit | 6c4401ca9ec7bd11efc8ef33e1fb0e3fd1fb6568 (patch) | |
| tree | bec88c2dc9c1774267ef980153d8d1f4d740d96e /math/powf.c | |
| parent | c1adc91f8af052f293a01327be335b66870e24b4 (diff) | |
| download | arm-optimized-routines-6c4401ca9ec7bd11efc8ef33e1fb0e3fd1fb6568.tar.gz | |
Fix spurious divbyzero exception with clang
A division is hoisted from behind a check (to use fcsel on the result) by
several versions of clang, and it can signal spurious divbyzero exception.
Diffstat (limited to 'math/powf.c')
| -rw-r--r-- | math/powf.c | 8 |
1 files changed, 8 insertions, 0 deletions
diff --git a/math/powf.c b/math/powf.c index 42d3e67..69a57db 100644 --- a/math/powf.c +++ b/math/powf.c @@ -182,6 +182,14 @@ powf (float x, float y) if (2 * ix == 0 && iy & 0x80000000) return __math_divzerof (sign_bias); #endif +#ifdef CLANG_EXCEPTIONS + /* Some versions of clang hoist the 1/x2 and thus division + by zero exception can be signaled spuriously, this is + not a real fix (that would require clang to model fenv + correctly), but seems to work in practice. */ + if (2 * ix == 0 && !(iy & 0x80000000)) + return x2; +#endif return iy & 0x80000000 ? 1 / x2 : x2; } /* x and y are non-zero finite. */ |