| Message ID | 20220822020419.137775-1-michael.hudson@canonical.com |
|---|---|
| State | New |
| Headers | show |
| Series | [v2] Avoid undefined behaviour in ibm128 implementation of llroundl | expand |
* Michael Hudson-Doyle via Libc-alpha: > Detecting an overflow edge case depended on signed overflow of a long > long. Replace the signed long long with unsigned and cast it back to > unsigned before comparisons (which is implementation defined behaviour, > but I guess glibc does not support any one's complement > architectures...). > > BZ #29488 > --- > v2: added casts to some references to 'res' I missed in v1. This > version passes all tests on ppc64el with gcc 12 with both -O2 and -O3. > --- > sysdeps/ieee754/ldbl-128ibm/s_llroundl.c | 11 ++++++----- > 1 file changed, 6 insertions(+), 5 deletions(-) > > diff --git a/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c b/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c > index d85154e73a..5f54f92767 100644 > --- a/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c > +++ b/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c > @@ -28,7 +28,8 @@ long long > __llroundl (long double x) > { > double xh, xl; > - long long res, hi, lo; > + unsigned long long res; > + long long hi, lo; > > ldbl_unpack (x, &xh, &xl); > > @@ -69,7 +70,7 @@ __llroundl (long double x) > res = hi + lo; > > /* This is just sign(hi) == sign(lo) && sign(res) != sign(hi). */ > - if (__glibc_unlikely (((~(hi ^ lo) & (res ^ hi)) < 0))) > + if (__glibc_unlikely (((~(hi ^ lo) & (((long long)res) ^ hi)) < 0))) > goto overflow; Maybe it's clearer to use __builtin_add_overflow (or INT_ADD_WRAPV from <intprops.h>)? Thanks, Florian
On Aug 22 2022, Michael Hudson-Doyle via Libc-alpha wrote: > @@ -69,7 +70,7 @@ __llroundl (long double x) > res = hi + lo; The overflow happens here. Changing the type of the LHS does not change that.
On Mon, 22 Aug 2022, 20:34 Andreas Schwab, <schwab@suse.de> wrote: > On Aug 22 2022, Michael Hudson-Doyle via Libc-alpha wrote: > > > @@ -69,7 +70,7 @@ __llroundl (long double x) > > res = hi + lo; > > The overflow happens here. Changing the type of the LHS does not change > that. > IIRC it was one of the "res += 1"s lower down in the test case that was failing for me. But maybe a more thorough rewrite of this function is called for (if people still care about ibm128 I guess). Cheers, mwh
diff --git a/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c b/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c index d85154e73a..5f54f92767 100644 --- a/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c +++ b/sysdeps/ieee754/ldbl-128ibm/s_llroundl.c @@ -28,7 +28,8 @@ long long __llroundl (long double x) { double xh, xl; - long long res, hi, lo; + unsigned long long res; + long long hi, lo; ldbl_unpack (x, &xh, &xl); @@ -69,7 +70,7 @@ __llroundl (long double x) res = hi + lo; /* This is just sign(hi) == sign(lo) && sign(res) != sign(hi). */ - if (__glibc_unlikely (((~(hi ^ lo) & (res ^ hi)) < 0))) + if (__glibc_unlikely (((~(hi ^ lo) & (((long long)res) ^ hi)) < 0))) goto overflow; xh -= lo; @@ -82,7 +83,7 @@ __llroundl (long double x) } else if (xh == 0.5) { - if (xl > 0.0 || (xl == 0.0 && res >= 0)) + if (xl > 0.0 || (xl == 0.0 && ((long long)res) >= 0)) res += 1; } else if (-xh > 0.5) @@ -91,11 +92,11 @@ __llroundl (long double x) } else if (-xh == 0.5) { - if (xl < 0.0 || (xl == 0.0 && res <= 0)) + if (xl < 0.0 || (xl == 0.0 && ((long long)res) <= 0)) res -= 1; } - if (__glibc_unlikely (((~(hi ^ (res - hi)) & (res ^ hi)) < 0))) + if (__glibc_unlikely (((~(hi ^ (((long long)res) - hi)) & (((long long)res) ^ hi)) < 0))) goto overflow; return res;