profile was calling vlong run-time support directly
which is counter-productive when those operations
are done by the compiler and might not even exist
in the library.  i've used A = A+B rather
than A += B to ensure the generated code is the same
(ie, calls to _addv and _subv) on those platforms
that still use them.  (those functions are handled specially for
profiling purposes in vlop.s on some platforms.)

it's probably time to put a bit more intelligent vlong
support into the portable part of cc, but this change
will allow alpha, amd64 and others to cope until i have
a chance to look at that (eg, for powerpc).