Performance regression in SHA1

[dwmw2]

I updated to the latest Cryptogams code, from whatever I'd lifted from OpenSSL last time I was playing with this: https://gitlab.com/openconnect/openconnect/-/commit/4abf01865dea0feb43c328c3a5311533fca8640f

The update reduces the performance of the stitched AES+SHA1 encryption code by about 10%, and decryption by even more.

Ryzen 5

• Before:

  $ tests/esptest ; tests/esptest  ; tests/esptest  ; tests/esptest
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 4580535 packets in 5s (5130 Mb/s)
  Encrypted+decrypted 2446172 packets in 5s (2739 Mb/s)
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 4553131 packets in 5s (5099 Mb/s)
  Encrypted+decrypted 2500763 packets in 5s (2800 Mb/s)
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 4537943 packets in 5s (5082 Mb/s)
  Encrypted+decrypted 2460366 packets in 5s (2755 Mb/s)
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 4530504 packets in 5s (5074 Mb/s)
  Encrypted+decrypted 2500993 packets in 5s (2801 Mb/s)

• After:

  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 3920203 packets in 5s (4390 Mb/s)
  Encrypted+decrypted 1384902 packets in 5s (1551 Mb/s)
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 3833687 packets in 5s (4293 Mb/s)
  Encrypted+decrypted 1404777 packets in 5s (1573 Mb/s)
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 3922032 packets in 5s (4392 Mb/s)
  Encrypted+decrypted 1415149 packets in 5s (1584 Mb/s)
  CPU capabilities: 078bffff 7ed8320b 209c01a9 00000000
  Encrypted 3889960 packets in 5s (4356 Mb/s)
  Encrypted+decrypted 1414968 packets in 5s (1584 Mb/s)

  Haswell Xeon

• Before:

  $ tests/esptest ; tests/esptest  ; tests/esptest  ; tests/esptest 
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2346782 packets in 5s (2628 Mb/s)
  Encrypted+decrypted 1265742 packets in 5s (1417 Mb/s)
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2377519 packets in 5s (2662 Mb/s)
  Encrypted+decrypted 1265845 packets in 5s (1417 Mb/s)
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2371691 packets in 5s (2656 Mb/s)
  Encrypted+decrypted 1265596 packets in 5s (1417 Mb/s)
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2406201 packets in 5s (2694 Mb/s)
  Encrypted+decrypted 1263781 packets in 5s (1415 Mb/s)

• After:

  $ tests/esptest ; tests/esptest  ; tests/esptest  ; tests/esptest 
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2193897 packets in 5s (2457 Mb/s)
  Encrypted+decrypted 952022 packets in 5s (1066 Mb/s)
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2269787 packets in 5s (2542 Mb/s)
  Encrypted+decrypted 951414 packets in 5s (1065 Mb/s)
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2243736 packets in 5s (2512 Mb/s)
  Encrypted+decrypted 951446 packets in 5s (1065 Mb/s)
  CPU capabilities: ffebffff 7ffef3ff 000037ab 00000000
  Encrypted 2238346 packets in 5s (2506 Mb/s)
  Encrypted+decrypted 949672 packets in 5s (1063 Mb/s)

The test case is part of the OpenConnect test suite in my 'hacks2' branch which (currently) contains the commit linked above.

[dwmw2]

On Haswell, it was using sha1_block_data_order_avx2 before. Now it's just sha1_block_data_order.

[dwmw2]

Oh, this is just because sha1-x86_64.pl hasn't been updated in cryptogams since the repository was created. The one in OpenSSL is much newer.