Adam Ierymenko
0c498556d5
Unroll Salsa20 fully for a little more speed (non-SSE now almost as fast as SSE)
2015-10-09 09:39:27 -07:00
Adam Ierymenko
160278c489
Little bit of reorg in Salsa20 which seems to speed things up very slightly.
2015-10-08 17:42:53 -07:00
Adam Ierymenko
789046ca57
Speed up Salsa20 just a bit.
2015-09-30 14:35:05 -07:00
Adam Ierymenko
a297e4a5bf
Add build def ZT_NO_TYPE_PUNNING, which when defined disables type punning code that might cause unaligned access errors on architectures that care (e.g. Android/ARM)
2015-07-08 09:12:51 -07:00
Adam Ierymenko
f19c3c51d3
Revert slow non-SSE Salsa20 modification since it did not fix Android/ARM issue. Also update Salsa20 comments and clean up a bit.
2015-07-02 09:00:00 -07:00
Adam Ierymenko
7c9949eea3
For @glimberg -- a *possible* fix to the alignment headaches on Android/ARM. If this works we should find a define that can be used to enable it there since it will slow things down on non-x86 other architectures.
2015-07-01 12:29:23 -07:00
Adam Ierymenko
8d2e20ede6
Get rid of __align stuff in Salsa20 -- not portable, does not seem to help much on newer chips.
2014-09-16 08:53:18 -07:00
Adam Ierymenko
f2d372545a
Salsa20 SSE Windows build fix -- turns out you can't be as loose with SSE intrinsics in Visual Studio
2014-07-26 20:07:38 -07:00
Adam Ierymenko
12692c551e
SSE optimized Salsa20 -- anywhere from 20% to 50% faster than plain C version
2014-07-15 17:56:09 -07:00
Adam Ierymenko
8c9b73f67b
Make Salsa20 variable-round, allowing for Salsa20/12 to be used for Packet encrypt and decrypt. Profiling analysis found that Salsa20 encrypt was accounting for a nontrivial percentage of CPU time, so it makes sense to cut this load fundamentally. There are no published attacks against Salsa20/12, and DJB believes 20 rounds to be overkill. This should be more than enough for our needs. Obviously incorporating ASM Salsa20 is among the next steps for performance.
2013-10-18 17:39:48 -04:00
Adam Ierymenko
ef3e319c64
Several things:
...
(1) Probable fix for issue #7 and major cleanup of EthernetTap code with consolidation for all unix-like systems and specialization for different flavors only when needed.
(2) Refactor of Buffer<> to make its members private, and Packet to use Buffer's methods exclusively to access them. This improves clarity and means we're no longer lying about Buffer's role in the code's security posture.
(3) Add -fstack-protect to Makefile to bounds check stack variables.
2013-07-09 14:06:55 -04:00
Adam Ierymenko
150850b800
New git repository for release - version 0.2.0 tagged
2013-07-04 16:56:19 -04:00
