[jamesqo]

This is a revival of the PR at #6638 I had to close a few months earlier because it had too many comments. I basically re-wrote the code I had written earlier, except with a few changes:

• I removed the optimizations for ARM64. I know very little about that platform, so someone more qualified should probably optimize it.
• I kept things mostly the way they were for non-AMD64 platforms. The only major thing I changed was replacing branching at the beginning/end of the copy with unaligned writes.
• I very conservatively kept the P/Invoke threshold at 1024 bytes, because even though my benchmarks said that the managed implementation was faster up to 8192 bytes, @nietras' benchmarks were saying 1024.
  • After this is merged, I'll open up a new PR where I increase it back to 8192 and we can discuss more about the threshold size.
• I also renamed some things to make the code cleaner, e.g. so that variables have self-explanatory names.

Please note: I am looking to get this merged mostly as-is because it's a clear win over what we currently have. Once this gets in, then we can open the door for further discussion/experimentation/etc. over optimizations.

/cc @jkotas, @nietras @tannergooding @benaadams FYI

[jamesqo]

Again this is largely a cleaned-up version of the same code I had written earlier, but I'll post benchmarks regardless. I have to re-build CoreCLR and write the benchmarks however, so I won't be able to do so straightaway. I'll update this PR when I have them.

[benaadams]

What's the impact with #7198 ? Does it replace the pinvoke as suggested? (Would be in Win x64 only though)

[jamesqo]

Going to get back to this sometime later-- have been busy lately

[jkotas]

Superseded by #9786