The fundamental problem with the original code is that it accesses
p[-2] which is one before the beginning of the input buffer for
empty lines. rev.1.6 just moved the problem from failures when
p[-2] happens to be '\\' to failures when it happens to be '\0'.
rev.1.5 was confused about the trailing newline and other things.
I went back to rev.1.5 and fixed it. The result is the same as
Keith Bostic's final version in PR 1356 except it loses more
gracefully for excessively long input lines.
