Optimizations for Apple Silicon?

[Michael Wojcik]

> From: openssl-users <openssl-users-bounces at openssl.org> On Behalf Of Blumenthal, Uri - 0553 - MITLL
> Sent: Monday, 20 March, 2023 08:28

> Naïve questions, driven my current use of Apple Silicon (includes AES, SHA1, SHA2, SHA3 extended instructions):
> 1. Does the current stable OpenSSL-3.1.0 include (assembly?) code to take advantage, aka - utilize, these CPU instructions?

For 3.0.8, a quick look at Configurations/15-ios.conf shows it uses the armv4 assembly configuration, and at crypto/aes/asm/aes-armv4.pl suggests it's not using any dedicated instruction. The comments at the top of the latter list various performance improvements but nothing about using an extended instruction if it's available.

Glancing at some search results it appears the dedicated instruction can get down to 0.9 cycles/byte, whereas the OpenSSL source states it reaches 21.5 cycles/byte, so using the dedicated instruction would be a big performance gain -- if those sources are comparing the same thing (one might be including some portion of overhead excluded by the other), and only when doing AES operations, of course. With TLS, for example, I/O will typically dominate so speeding up may not do much for many applications.

Now that said, there seem to be crypto/*/asm/*-armv8.pl files, but 1) they don't seem to be used by any configurations, and 2) the AES one (vpaes-armv8.pl) is a vectorized AES but doesn't seem to use any dedicated instruction -- though I'm not at all an ARM assembly programmer, so take that with a lot of salt.

> 2. How can I check whether openssl installation (binary and libraries) are compiled with Silicon optimizations (if I did not
> compile from source myself)?

If I wanted to do this, I'd probably disassemble libcrypto on the target platform and search for the symbol AES_encrypt, and then look at the implementation, or just search the disassembly for the instruction in question with a suitable regex search. There might be an easier way.

> 3. What's the current analog of rdrand engine? I.e., does OpenSSL take input from RDRAND and its analog on AARCH64,
> and how can I check that it does?

RDRAND, yes, if OpenSSL was not built with no-rdrand. I don't know what the analog might be on ARM.

Hopefully someone can provide more detailed and authoritative answers...

-- 
Michael Wojcik