]> Cypherpunks repositories - gostls13.git/commit
crypto/rc4: optimize generic implementation slightly
authorMichael Munday <mike.munday@ibm.com>
Fri, 23 Mar 2018 13:00:47 +0000 (13:00 +0000)
committerBrad Fitzpatrick <bradfitz@golang.org>
Fri, 23 Mar 2018 15:47:48 +0000 (15:47 +0000)
commit41402b59bd7401bf0b4ff3043dfe3d3fa712e05c
tree91f9c636f142e196101eba19866dbbbdef592866
parent58734039bdddbdfd3c7cd4f9fc232a75f5cef32a
crypto/rc4: optimize generic implementation slightly

The compiler can't currently figure out that it can eliminate both c.s
loads (using store to load forwarding) in the second line of the
following code:

...
c.s[i], c.s[j] = c.s[j], c.s[i]
x := c.s[j] + c.s[i]
...

The compiler eliminates the second load of c.s[j] (using the original
value of c.s[i]), however the load of c.s[i] remains because the compiler
doesn't know that c.s[i] and c.s[j] either overlap completely or not at
all.

Introducing temporaries to make this explicit improves the performance
of the generic code slightly, the goal being to remove the assembly in
this package in the future. This change also hoists a bounds check out
of the main loop which gives a slight performance boost and also makes
the behaviour identical to the assembly implementation when len(dst) <
len(src).

name       old speed     new speed     delta
RC4_128-4  491MB/s ± 3%  596MB/s ± 5%  +21.51%  (p=0.000 n=9+9)
RC4_1K-4   504MB/s ± 2%  616MB/s ± 1%  +22.33%  (p=0.000 n=10+10)
RC4_8K-4   509MB/s ± 1%  630MB/s ± 2%  +23.85%  (p=0.000 n=8+9)

Change-Id: I27adc775713b2e74a1a94e0c1de0909fb4379463
Reviewed-on: https://go-review.googlesource.com/102335
Run-TryBot: Michael Munday <mike.munday@ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
src/crypto/rc4/rc4.go