tree 4fd90bda42147bd9d07476834c81f0decdb5fef1
parent a220a6024f66c123019b5c080f6bd8bcaf75448c
author David Benjamin <davidben@google.com> 1716040059 -0400
committer Boringssl LUCI CQ <boringssl-scoped@luci-project-accounts.iam.gserviceaccount.com> 1718127971 +0000

Remove OPENSSL_ia32cap_P references from AES-NI assembly

The AES-NI key schedule functions have two versions, dating to OpenSSL's
23f6eec71dbd472044db7dc854599f1de14a1f48. This cites RT#3576.
Unfortunately, OpenSSL purged their old RT bugs, without any archives,
so this context is now lost. Some archives of openssl-dev discussion
(also predating OpenSSL's archives) give most of the context:
https://groups.google.com/g/mailing.openssl.dev/c/OuFXwW4NfO8/m/7d2ZXVjkxVkJ

Broadly, although AES-NI has an aeskeygenassist instruction for the key
schedule, apparently it's overall faster to ignore it and use aesenclast
instead. But it's slower on older processors, so the assembly would
check for AVX && !XOP as a proxy. (Note we always set XOP to false, even
though this likely wasn't a capability check but a proxy for pre-Xen AMD
chips.)

It is unclear if the aeskeygenassist version is still worthwhile.
However, the aesenclast version requires SSSE3. SSSE3 long predates
AES-NI, but it's not clear if AES-NI implies SSSE3. In OpenSSL, the CCM
AES-NI assembly seems to assume it does. For now, I've preserved the
pair of them.

There are now only two assembly files with OPENSSL_ia32cap_P references!

Bug: 673
Change-Id: I990b1393d780db4caf074c184ce8bbd182da6e29
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/68690
Reviewed-by: Bob Beck <bbe@google.com>
Commit-Queue: David Benjamin <davidben@google.com>