First take on using AES-NI instructions

This first naive use of the new Intel AES-NI instructions boosts the
performance of AES on CPUs supporting this by 3 to 5 times.

Results from running
  ./benchmark --cipher-repetitions 10 --large-buffers  cipher aes
on a
  cpu family      : 6
  model           : 37
  model name      : Intel(R) Core(TM) i5 CPU         660  @ 3.33GHz
  stepping        : 2
  cpu MHz         : 3325.494
  cache size      : 4096 KB
  cpu cores       : 2

yields this:

    ECB/Stream         CBC             CFB             OFB             CTR
--------------- --------------- --------------- --------------- ---------------
  130ms   110ms   110ms   100ms   110ms   110ms   160ms   150ms   170ms   170ms
   40ms    40ms    20ms    30ms    30ms    20ms    70ms    70ms    80ms    80ms

The first line is with runtime switched off AES-NI instructions (don't
set use_aesni in do_setkey), the second with enabled AES-NI.  By
fixing the alignment, I hope to squeeze out a little more even with
this naive implementation.

1 files changed, 6 insertions, 0 deletions

diff --git a/README b/README
index 514464ab..6fe1cfb5 100644
--- a/README
+++ b/README
@@ -170,6 +170,12 @@
                      available.  Try this if you get problems with
                      assembler code.
 
+     --disable-aesni-support
+                     Disable support for the AES-NI instructions of
+                     newer Intel CPUs.  The default is to use AES-NI
+                     if available.  Try this if you get problems with
+                     assembler code.
+
      --disable-O-flag-munging
                      Some code is too complex for some compilers while
                      in higher optimization modes, thus the compiler