Re: Suggested improvement to std::list

[Loren James Rittle <rittle at latour dot rsch dot comm dot mot dot com>]

In article <20030212134535.5336e419.bkoz@redhat.com>,
Benjamin Kosnik<bkoz@redhat.com> writes:

>> Could you create a test case that demonstrates an actual performance
>> improvement?  We like to have those for proposed performance patches
>> (I hold myself to that standard and would like that information before
>> I consider it for approval).

> What do you think about checking these in to testsuite/performance? That
> way, when there is support in the testsuites for performance analysis
> and performance regression testing, there will be tests to use.

> Thoughts?

(a) At least one of us (i.e. I know I have ;-) promised to create such
a harness even if it were non-portable.  (b) If people contribute
performance test cases, then I agree that we should check them in
somewhere even if no public harness is available yet.  (b, part 2) I
and others have posted many cases over the years.  They get lost over
time when only archived in e-mail.  (c) I need to take a vacation just
to work on GCC.

I have thought about this a little bit.  If all performance tests were
written in a common form:

// Copyright and other text

// Standard includes, testsuite_hooks.h

// dg-like comment language, although perhaps much simpler to parse
// MAX_CONCURRENCY = C (C=1 for non-reentrant test)
// MAX_LOOP = L (L=1 for non-restartable test)
// FIND_CPU_PER_ITERATION | FIND_MAX_MEMORY | FIND_KERNEL_CALLS_PER_ITERATION

void test01 (void)
{
}

// no main

Then I could create a main driver for POSIX.  Someone else could do
windows.  More abstraction of things occurring in the test and to be
studied are possible, but frankly if we don't start simple, I don't
see this ever happening.

Regards,
Loren