Sample based profiling

[Joerg Sonnenberger <joerg%bec.de@localhost>]

Hello all,
the recent work on static PIE support have exposed some fundamental
issues in the way profiling support is currently implemented. When -pg
is used for linking, the binary will essentially get a copy of libc
linked in, but without actually being a static binary. This means that
all the functionality that is supposed to be provided by the dynamic
linker is pulled in from libc_pg.a and not from the dynamic linker. I
see three options going forward for this:

(1) Introduce a dynamic version of libc_pg.
(2) Stop pretending that -pg means dynamic linking and implicitly add
-static.
(3) Stop pretending that -pg makes sense on modern hardware.

Fundamentally, I think the third option is the correct way forward.
Sample based profiling with the default HZ setting just doesn't give
very useful results except for extremely long running processes -- you
get a sample every few million CPU cycles and want to extract something
useful from that? There are options for doing sample based profiling
without -pg. devel/gperftools provides one and doesn't require custom
builds of everything. Both GCC and Clang support basic block counting
options useful for coverage testing, removing the other case where gperf
might have been useful in the past. To fully support gperftools, we need
to provide some kernel glue to support stack unwinding across signals,
but that's something desirable anyway. dtrace can also be used for
profiling. I don't see the point of keeping another library archive for
everything in base around...

Joerg