The final result of all this refactoring: instead of doing stat immediately

followed by an open for every source file we open, probe the file system with
'open' and then do an fstat when it succeeds.  open+fstat is faster than
stat+open because the kernel only has to perform the string->inode mapping
once.  Presumably it gets faster the deeper in your filesystem a lookup
happens.

For -Eonly on cocoa.h, this reduces system time from 0.042s to 0.039s on
my machine, a 7.7% speedup.

llvm-svn: 120066