[cbdfghh]

OpenSSL uses assembly not for speed but for security. You need to make sure algorithms don't "optimize" leaking data.

For example, a strcmp on a secret field is insecure because of timing attacks.

The only way to ensure the CPU takes a fixed amount of time is through assembler

[bastawhiz]

That also brings about the risk of doing bad things that the compiler could otherwise prevent. Simply using a non-optimizing compiler with a timing attack-safe algorithm (and using the resulting machine code) would have the same effect. There's little reason to actually write assembly by hand, in that case, unless you're trying to milk performance.