This patches vpaes-armv8.pl to add vpaes_ctr32_encrypt_blocks. CTR mode
is by far the most important mode these days. It should have access to
_vpaes_encrypt_2x, which gives a considerable speed boost. Also exclude
vpaes_ecb_* as they're not even used.
For iOS, this change is completely a no-op. iOS ARMv8 always has crypto
extensions, and we already statically drop all other AES
implementations.
Android ARMv8 is *not* required to have crypto extensions, but every
ARMv8 device I've seen has them. For those, it is a no-op
performance-wise and a win on size. vpaes appears to be about 5.6KiB
smaller than the tables. ARMv8 always makes SIMD (NEON) available, so we
can statically drop aes_nohw.
In theory, however, crypto-less Android ARMv8 is possible. Today such
chips get a variable-time AES. This CL fixes this, but the performance
story is complex.
The Raspberry Pi 3 is not Android but has a Cortex-A53 chip
without crypto extensions. (But the official images are 32-bit, so even
this is slightly artificial...) There, vpaes is a performance win.
Raspberry Pi 3, Model B+, Cortex-A53
Before:
Did 265000 AES-128-GCM (16 bytes) seal operations in 1003312us (264125.2 ops/sec): 4.2 MB/s
Did 44000 AES-128-GCM (256 bytes) seal operations in 1002141us (43906.0 ops/sec): 11.2 MB/s
Did 9394 AES-128-GCM (1350 bytes) seal operations in 1032104us (9101.8 ops/sec): 12.3 MB/s
Did 1562 AES-128-GCM (8192 bytes) seal operations in 1008982us (1548.1 ops/sec): 12.7 MB/s
After:
Did 277000 AES-128-GCM (16 bytes) seal operations in 1001884us (276479.1 ops/sec): 4.4 MB/s
Did 52000 AES-128-GCM (256 bytes) seal operations in 1001480us (51923.2 ops/sec): 13.3 MB/s
Did 11000 AES-128-GCM (1350 bytes) seal operations in 1007979us (10912.9 ops/sec): 14.7 MB/s
Did 2013 AES-128-GCM (8192 bytes) seal operations in 1085545us (1854.4 ops/sec): 15.2 MB/s
The Pixel 3 has a Cortex-A75 with crypto extensions, so it would never
run this code. However, artificially ignoring them gives another data
point (ARM documentation[*] suggests the extensions are still optional
on a Cortex-A75.) Sadly, vpaes no longer wins on perf over aes_nohw.
But, it is constant-time:
Pixel 3, AES/PMULL extensions ignored, Cortex-A75:
Before:
Did 2102000 AES-128-GCM (16 bytes) seal operations in 1000378us (2101205.7 ops/sec): 33.6 MB/s
Did 358000 AES-128-GCM (256 bytes) seal operations in 1002658us (357051.0 ops/sec): 91.4 MB/s
Did 75000 AES-128-GCM (1350 bytes) seal operations in 1012830us (74049.9 ops/sec): 100.0 MB/s
Did 13000 AES-128-GCM (8192 bytes) seal operations in 1036524us (12541.9 ops/sec): 102.7 MB/s
After:
Did 1453000 AES-128-GCM (16 bytes) seal operations in 1000213us (1452690.6 ops/sec): 23.2 MB/s
Did 285000 AES-128-GCM (256 bytes) seal operations in 1002227us (284366.7 ops/sec): 72.8 MB/s
Did 60000 AES-128-GCM (1350 bytes) seal operations in 1016106us (59049.0 ops/sec): 79.7 MB/s
Did 11000 AES-128-GCM (8192 bytes) seal operations in 1094184us (10053.2 ops/sec): 82.4 MB/s
Note the numbers above run with PMULL off, so the slow GHASH is
dampening the regression. If we test aes_nohw and vpaes paired with
PMULL on, the 20% perf hit becomes a 31% hit. The PMULL-less variant is
more likely to represent a real chip.
This is consistent with upstream's note in the comment, though it is
unclear if 20% is the right order of magnitude: "these results are worse
than scalar compiler-generated code, but it's constant-time and
therefore preferred".
[*] http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.100458_0301_00_en/lau1442495529696.html
Bug: 246
Change-Id: If1dc87f5131fce742052498295476fbae4628dbf
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/35026
Commit-Queue: David Benjamin <davidben@google.com>
Reviewed-by: Adam Langley <agl@google.com>
55db667c62