Re: Role of profile/hashtable.h?!?

[Silvius Rus <rus at google dot com>]

On Sun, Jan 3, 2010 at 3:45 PM, Paolo Carlini <paolo.carlini@oracle.com> wrote:
>
> Hi,
> > If I remember well, we had to clone part of hashtable.h in order to
> > instrument decisions such as when to rehash. ?These decisions are
> > based on private data such as _M_num_elements. ?Other private data
> > member values may be collected by instrumentation code. ?In the case
> > of hashtable-based containers, profile instrumentation only cares
> > about hashtable details (distribution, rehashing), so instead of
> > instrumenting hash_map and hash_set, we decided to instrument only
> > hashtable. ?Thus you only see hashtable.h and not hash_map or
> > hash_set. ?Is there a better way to do this, which ideally would
> > eliminate or reduce greatly the amount of cloning?
> >
> Bah, if you ask me, I would just *not* instrument those legacy
> containers, which do not have an particular advantage compared to the
> tr1 and std unordered containers. Actually, have been moved already to
> the backward subdirectory and people still requiring those facilities
> certainly can't assume the availability of profile-mode, which is brand
> new. For comparison, we do *not* deliver debug-mode hash_* containers.
> What about just removing that code and concentrating on the rest?

Let me think about it a bit.  I still see tons of code using hash_* in
my benchmarks, so being able to test hash_* is important, at least to me.
However, I'd also much prefer not to carry it around any longer.

> > I will be revisiting the include/profile/ code early this coming week
> > and port a few needed fixes from the profile-stdlib branch and local
> > patches. ?In case you notice anything else questionable within
> > include/profile, this is a very good time to bring it up.
> >
> Ok thanks. I'm about to commit some changes which allows to cleanly
> regtest profile-mode. Some C++0x bits were missing and some tests were
> failing because we were forcing debug-mode in the testcases (which is
> not compatible with profile-mode)
>
> If I spot something else I will tell you, thanks for the good work!
>
> Paolo.

Thank you,
Silvius