Robert Roessler wrote:
Neil Hodgson wrote:
Robert Roessler:
...
1) is line-count a good determinant in evaluating hash algos?
2) source code size/layout may not be a good indicator of the length
of runtime code paths
Complexity for the code reader should be avoided. My reaction to
the Hsieh code was:
1) What's this preprocessor stuff? Is it alignment trickery?
Not really "alignment" in the sense of processor architectures - it
just facilitates the processing of 2 16-bit chunks per cycle of the
main loop without blowing past the end.
Given the amount of hardware optimizations in mainstream
processors nowadays, this is a little pointless.
[snip]
But a potential show-stopper anyway is that, since you want Scintilla
to compile across the widest possible set of architectures, Hsieh's
code as it stands will fail on non-x86 CPUs where aligned memory
accesses are *required*. Or at least I *think* unaligned 16-bit
accesses are as evil as unaligned 32-bit access on these arches - if
16-bit accesses are even allowed in these cases!
This is a bug waiting to happen on any app that wants to be
portable. Rip it out. Write clear code.
And yes, I also abhor the algorithm's apparent complexity - I just
like what it does. ;)
So Hsieh's code uses unaligned 16-bit accesses? Leaving out all
the levels of caching, an unaligned access (i.e. 16-bit or bigger
memory access with an odd address) always ends up causing two
memory accesses on modern x86 processors. If it straddles a cache
line, big ouch. If both bytes are not in the caches, this causes
two cache lines (2 x 16 bytes or more) to be fetched.
Such optimizations has no place in modern code, at least, in an
app such as Scintilla or SciTE. If you understand modern
microprocessor architecture, you will understand that many
old-style optimizations are not worth putting in an app where
high-quality code is important.
HTH,
--
Cheers,
Kein-Hong Man (esq.)
Kuala Lumpur, Malaysia
_______________________________________________
Scite-interest mailing list
[email protected]
http://mailman.lyra.org/mailman/listinfo/scite-interest
