https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113353
Bug ID: 113353
Summary: Wrong rounding in std::nearbyint when vectorized with
-funsafe-math-optimizations on PPC
Product: gcc
Version: 12.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: alexander.gr...@tu-dresden.de
Target Milestone: ---
According to the standard `std::nearbyint` and `std::rint` are supposed to give
the same results. However the former does not tie to even in halfway cases when
used in a tight loop.
The following code basically rounding "-8.5" produces `-9` when compiled with
`-funsafe-math-optimizations` but (as actually correct) `-8` when compiled
without it or when using `std::rint` instead. (Additional flags: `-O3
-mcpu=power9`)
This is at least unexpected given that
> The only difference between std::nearbyint and std::rint is that
> std::nearbyint never raises FE_INEXACT.
which in this case does not hold.
#include <iostream>
#include <cmath>
#include <vector>
extern "C" void kernel(const double* in_ptr,
double* out_ptr,
const long ks)
{
for(long i=0; i<ks; i+=1)
out_ptr[i] = std::nearbyint(in_ptr[i]);
}
int main()
{
std::vector<double> in(16, -8.5);
decltype(in) out(in.size());
kernel(in.data(), out.data(), in.size());
for(const auto t: out)
std::cout << t << ", ";
std::cout << '\n';
}
Is this an intended effect of -funsafe-math-optimizations? I'd expect to at
least be able to workaround this with -frounding-math or so.
