[compiler-rt] [scudo] Update header without read-modify-write operation (PR #66955)

[Kostya Kortchinsky via llvm-commits]

cryptoad wrote:

> What do you think?

I mean, it's fair. If the performance gains are significant, the protection might not be worth it. I initially went the side of security. I agree that the window for the race makes it difficult, and failure to win the race would mean an abort() which would prevent further attempts, or at least make them very noisy.

You might want to run that through the Fuchsia people as well, make sure that everyone knows it's coming. I do think that the race will happen in the wild (another example @ https://github.com/tensorflow/tensorflow/issues/58143), and will end up manifesting in some weird way that will be hard to debug - which is probably not worth the performance cost either.

I'll leave it to you and Christopher to make the decision.

https://github.com/llvm/llvm-project/pull/66955