On Thu, Aug 4, 2022 at 11:45 AM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Wed, Aug 3, 2022 at 9:11 AM Joanne Koong <joannelkoong@xxxxxxxxx> wrote: > > > > > > __builtin_memcpy() is best. When we write just "memcpy()" we still > > > rely on compiler to actually optimizing that to __builtin_memcpy(), > > > because there is no memcpy() (we'd get unrecognized extern error if > > > compiler actually emitted call to memcpy()). > > > > Ohh I see, thanks for the explanation! > > > > I am going to do some selftests cleanup this week, so I'll change the > > other usages of memcpy() to __builtin_memcpy() as part of that clean > > up. > > builtin_memcpy might be doing single byte copy when > alignment is not known which is often the case when > working with packets. > If we do this cleanup, let's copy-paste cilium's memcpy > helper that does 8 byte copy. > It's much better than builtin_memcpy. > https://github.com/cilium/cilium/blob/master/bpf/include/bpf/builtins.h Maybe we should have a separate patchset to changing to use cilium's other builtins as well (eg memzero, memcmp, memmove); I'll leave the memcpy() -> __builtin_memcpy() changes to be part of that patchset.
