Profiling showed that 'CheckImplicitConversions' was very slow because of the...

Profiling showed that 'CheckImplicitConversions' was very slow because of the call to getSpellingLoc().  On 'aes.c'
in the LLVM test suite, this function was consuming 7.4% of -fsyntax-only time.  This change fixes this issue
by delaying the check that the warning would be issued within a system macro by as long as possible.  The
main negative of this change is now the logic for this check is done in multiple places in this function instead
of just in one place up front.

llvm-svn: 127425