aboutsummaryrefslogtreecommitdiff
path: root/lib/lib.c
diff options
context:
space:
mode:
authorDaniel Verkamp <daniel@drv.nu>2014-05-15 19:05:16 -0500
committerDaniel Verkamp <daniel@drv.nu>2014-05-15 19:05:16 -0500
commitafe951b19c0f8313731e28a5f347304d82e7f7af (patch)
treebd4251ec4a504a2aeb2c8736dc8907e4aa16767b /lib/lib.c
parent97641f459a3e74045305b4ef9b80bb00bc56a29f (diff)
downloadtoybox-afe951b19c0f8313731e28a5f347304d82e7f7af.tar.gz
Here's a quick cleanup of md5sum. Executive summary: smaller and faster.
On my machine, for a 2.2 GB file of random bytes, the timings with warm cache are: toybox before: 11.4 seconds toybox after: 8.3 seconds GNU md5sum: 3.9 seconds openssl dgst -md5: 3.5 seconds This is clearly better than before (3x openssl), but still slow (2x openssl). I suspect there is more low-hanging fruit to be had by eliminating the memcpy in hash_update (maybe not too much - hash_update accounts for about 4% of total runtime versus 92% for md5_transform according to perf - but this would also help sha1sum). make bloatcheck on x86_64 gcc 4.8.2 -Os: name old new delta ----------------------------------------------------------------------- md5rot 0 64 64 md5_transform 365 223 -142 ----------------------------------------------------------------------- -78 total Rationale for the changes: Move definition of 'rol' up so it can be used in md5_transform. This is purely cosmetic; it expands to exactly the same code. Put rotation counts in a lookup table instead of calculating them on the fly. This is mostly a wash size-wise, +5 bytes total, but worthwhile for readability and speed. Instead of accessing the state array using a rotating index (the variable formerly known as 'a'), access the state with constant offsets and rotate the contents of the array instead. This is the big win - it eliminates all the crazy memory addressing math inside the loop.
Diffstat (limited to 'lib/lib.c')
0 files changed, 0 insertions, 0 deletions