https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90285
Bug ID: 90285
Summary: Poor optimised codegen for memmove() back on top of
oneself
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: s_gccbugzilla at nedprod dot com
Target Milestone: ---
The following code produces poor optimised codegen on trunk GCC at the time of
writing (2019-04-29):
// Reinterprets a T into its array of bytes
// Currently defined behaviour in C++ 20 for
// trivially copyable types only. The proposal
// would be that this would become defined
// behaviour for most possible C++ types.
template<class T>
constexpr inline byte_array_ref<T> detach_cast(T &v) noexcept
{
byte_array_ref<T> ret = reinterpret_cast<byte_array_ref<T>>(v);
byte temp[sizeof(T)];
// Reinterpret bytes by copying (not UB for TC types)
memmove(temp, &v, sizeof(T));
// Put reinterpreted bytes back. This avoids the UB
// of reinterpret casting without creating new objects.
memmove(ret, temp, sizeof(T));
return ret;
}
You can see GCC's codegen here (it does two copies of 40Kb):
https://godbolt.org/z/sJWSc1
You can see clang's codegen here (which is optimal, nothing is copied):
https://godbolt.org/z/ou8VFT
I think GCC ought to not perform memory copies for the above code sequence.
Niall
