Re: CVS performance regression over 2.7.2 on SPECint95 on sparc

[Diego Novillo <dnovillo at redhat dot com>]

On Sat, 02 Dec 2000, Robert Dewar wrote:

> <<Oh. Well, yes. It comes and goes. I didn't read the comparison
> table you posted. How much of a regression are you observing?
> Bear in mind that you're comparing to a really ancient version of
> gcc.
> >>
> 
> But looking at a "really ancient" version of GCC and seeing a performance
> decrease is a rather worrisome think! We have seen similar performance
> degradations with gcc 2.9x, compared with 2.8.1, and are definitely
> worried that this may make our transition problematical.
>
I agree, of course. I was just pointing out that the impression I
have is that performance has not been a top priority in GCC
development.

I'm fairly new to GCC, but I have not seen any efforts at
monitoring the code quality using any kind of performance
benchmarks. Correctness and target-specific optimizations have
received more attention.

Given the range of targets GCC supports, keeping track of
performance regressions is a bit challenging. Typically, industry
standard benchmarks are not freely available and you need access
to lots of idle cycles. You also need to keep a running them
regularly and learn to distinguish ordinary fluctuations from
real performance regressions.