Tremendous performance regression in 1.1.2 -> mainline

[Mark Mitchell]

>>>>> "David" == David Edelsohn <dje@watson.ibm.com> writes:

    David> 	I believe that we need to make the performance of GCC
    David> itself and the performance of GCC's generated code on key
    David> architectures part of the requirements for the next major
    David> GCC release.  The lack of compiler regression test failures
    David> is necessary but not sufficient.

I've already agreed with you in private email.  But, I'll say so again
in public email.  Yes, this will be a release criteria.

However, it's unlikely that *all* programs will compile as fast, or
run as fast.  We'll have to evaluate examples to decide whether they
are ones for which we should do something, or not.

Also, regarding heuristics: you're right that having clamps is often
effective.  However, I've often found that clamps in GCC are used as
an alternative to figuring out how to really improve the bottlenecks.
So, I'm in favor of clamps -- but after we try to make sure that the
underlying algorithms have really been coded efficiently.

None of this is specific commentary on any particular clamp, or any
particular program; it's all just meant in general.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com