https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116754
Bug ID: 116754
Summary: libstdc++ std::ranges::copy performance issue
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: addmlbx at gmail dot com
Target Milestone: ---
Created attachment 59130
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=59130&action=edit
struct __memcpyable definition
libstdc++ std::ranges::copy implementation has bug which prevents calling
memmove in case iterators are pointers:
there is an `if constexpr (__memcpyable<_Iter, _Out>::__value)`
statement used to determine whether memmove should be called and _Iter and _Out
template parameters are swapped (__memcpyable expects output iterator to be
given first). Therefore this statement evaluates to false and memmove is not
called.
I noticed this bug when I was comparing the assembly produced for copying some
containers like std::vector, std::array and C-style array in the std::copy and
std::ranges::copy:
When calling std::copy on these containers, it's inlined down to the memmove
call. But for the std::ranges::copy the assembly is different: it fallback to
the generic cycle-based implementation.
Definition of the struct __memcpyable is shown on the attached picture in order
to demonstrate that it (and it's specializations) expect output iterator to be
passed first