Re: While we're at it

On Mon, 5 Apr 2010, Colm MacCárthaigh wrote:

> The very next paragraph explicitly mentions pipelining, so I don't
> agree with you that it has been omitted. 

Ok.

> I certainly don't call out that the L2 cache would contain the whole
> word, but this seems like an odd thing to call out. A register would
> contain the whole word, never mind any of the caches. The L1 and L2
> caches are both several orders of magnitude larger than a word.

In total, the L1 & L2 cache are of course larger. And a register _could_
contain a word, but doesn't do so in this case. In this case, it
contains a byte, which is compared to 0. It doesn't make the comparison
4 times on the same byte and there are no unnecessary comparisons, which
is sort of implied in your description.

But a complete word is brought into L2 cache to read one byte. When
aligned the next 3 tests access L2 cache at high speed and are
pipelined.  A conditional jump to do one test at a time would slow it
down because the pipeline is flushed when the branch is taken.

> Loop-unrolling is a very separate thing from caching optimisation,and
> is usually in opposition to it. Note that unrolled instructions
> require storage too, and so compete for cache line space.

Agree. Some middle ground is appropriate: not too much unrolled.  But
zero unrolling isn't ideal in many cases.

> > Show me a memory allocator that gives back unaligned memory, and I'll
> > show you how to fix that. Most of the time, we can be expect the memory
> > to be word-aligned. In the rare case it isn't, it won't matter too much.
> 
> The issue is not that an allocator may return unaligned memory, but
> that you may wish to count the length of a string that does not start
> at the alignment boundary. For one trivial, but realistic, example;
> 
> char * string = "Hello World"; // probably a word-aligned pointer
> int firstwordlength = strlen(string) - strlen(strtok(string, " "));
> // second call is unaligned

Sure. But one wants code that is fast in the most common case, and works 
in the less common case.


-- 
Av8 Internet   Prepared to pay a premium for better service?
www.av8.net         faster, more reliable, better service
617 256 5494