Message ID | 20160114214145.GA22984@intel.com |
---|---|
State | New, archived |
Headers | show |
On Thu, 14 Jan 2016, H.J. Lu wrote: > Since floating-point operation may trigger floating-point exceptions, > we call math_opt_barrier inside if to prevent code motion. > > Tested on x86-64. OK for trunk? Please send a patch updating all implementations for which this issue is applicable.
On Thu, Jan 14, 2016 at 2:13 PM, Joseph Myers <joseph@codesourcery.com> wrote: > On Thu, 14 Jan 2016, H.J. Lu wrote: > >> Since floating-point operation may trigger floating-point exceptions, >> we call math_opt_barrier inside if to prevent code motion. >> >> Tested on x86-64. OK for trunk? > > Please send a patch updating all implementations for which this issue is > applicable. There are dbl-64/e_sqrt.c: libc_feholdexcept_setround (&env, FE_TONEAREST); dbl-64/s_fma.c: libc_feholdexcept_setround (&env, FE_TONEAREST); dbl-64/s_fmaf.c: libc_feholdexcept_setround (&env, FE_TOWARDZERO); under sysdeps/ieee754. dbl-64/s_fma.c is the only one with this problem.
On Thu, 14 Jan 2016, H.J. Lu wrote: > On Thu, Jan 14, 2016 at 2:13 PM, Joseph Myers <joseph@codesourcery.com> wrote: > > On Thu, 14 Jan 2016, H.J. Lu wrote: > > > >> Since floating-point operation may trigger floating-point exceptions, > >> we call math_opt_barrier inside if to prevent code motion. > >> > >> Tested on x86-64. OK for trunk? > > > > Please send a patch updating all implementations for which this issue is > > applicable. > > There are > > dbl-64/e_sqrt.c: libc_feholdexcept_setround (&env, FE_TONEAREST); > dbl-64/s_fma.c: libc_feholdexcept_setround (&env, FE_TONEAREST); > dbl-64/s_fmaf.c: libc_feholdexcept_setround (&env, FE_TOWARDZERO); > > under sysdeps/ieee754. dbl-64/s_fma.c is the only one with this > problem. I disagree with that analysis. The question isn't whether there's a call to a particular internal interface. The question is whether there are conditionals on arithmetic, such that the arithmetic is exact in the conditional case but may be inexact if moved before the conditional. I think at least ldbl-96/s_fmal.c, ldbl-96/s_fma.c and ldbl-128/s_fmal.c have the same issue.
diff --git a/sysdeps/ieee754/dbl-64/s_fma.c b/sysdeps/ieee754/dbl-64/s_fma.c index bc3677d..a349243 100644 --- a/sysdeps/ieee754/dbl-64/s_fma.c +++ b/sysdeps/ieee754/dbl-64/s_fma.c @@ -175,7 +175,10 @@ __fma (double x, double y, double z) /* Ensure correct sign of exact 0 + 0. */ if (__glibc_unlikely ((x == 0 || y == 0) && z == 0)) - return x * y + z; + { + x = math_opt_barrier (x); + return x * y + z; + } fenv_t env; libc_feholdexcept_setround (&env, FE_TONEAREST);