svn rev #23445: trunk/src/lib/crypto/ crypto_tests/ krb/crc32/

[Tom Yu]

ghudson at MIT.EDU writes:

> http://src.mit.edu/fisheye/changelog/krb5/?cs=23445
> Commit By: ghudson
> Log Message:
> Remove CRC32_SHIFT4 code as we are unlikely to ever need it.

I think someone (Bill Sommerfeld, maybe?) once told me that a shift-4
implementation of CRC32 could have a higher performance than a shift-8
implementation under some circumstances, because the entire CRC table
would remain in a single CPU cache line.  (and not need to reload
entries from main memory due to a cache miss) I don't remember being
able to demonstrate such a performance gain under any tests that I
ran, but perhaps I was not setting up the test conditions correctly.
It's also possible that modern CPU caches are large enough for this to
not be an issue.

Does anyone have additional information on this subject?  (It's
probably not very important as CRC32 is only really used as part of
the soon-to-be-deprecated single-DES cryptosystems.)