This is essentially a partial revert of r105255, which fixed PR7272.
The test case it introduced[1] still passes:
test/Transforms/Inline/2010-05-31-ByvalTailcall.ll
I think this was misdiagnosed. The call to @ext in the test case should
never have been marked 'tail' by instcombine, tailcallelim, or any other
optimization because it captures a byval argument, which is using memory
from the caller. That's against the LangRef rules for tail calls. It
looks like the optimization that added the tail flag has been fixed in
the meantime, because the test still passes.