string::find complexity.

[Benjamin Kosnik]

>Hmmm.  Code size matters.  Also testing on different architectures
>matters.

Yes, it would be nice to explicitly test on non-x86 architectures, and
include size information. It turns out that make check-performance runs
on powerpc-darwin now, so that's an option for some of us.

In my experience, code size fluctuates more on compiler changes than on
library changes, so testing without keeping each constant is less
interesting and useful.

>What we could really use, of course, is a sizable C++ program to
>optimize.  Once that one is working pretty well, pick another, and
>optimize both simultaneously.  But I don't know how to get that.

.... it also has to use the standard library bits in question....

For a while, I was using some of the MPEG recorders/encoders on
sourceforge for infomal io testing. These weren't super useful in
practice, because buffered speed has been acceptable for this purpose
since 3.1 days...

>There must be some other good ideas out there....

Some of the performance testsuite tests are not quite as trivial as the
one-line loops. I think the producer-consumer allocator bit is decent,
for instance.

Any synthetic benchmark, however, will be less interesting than real code.

-benjamin