Fixing a FIXME turned out to be pretty involved. I added an io function
that returns a unique boxed string (for the contents of a file) rather than
a string, and went from there. Also made the src field of codemap a unique
boxed string. This doesn't seem to make that much difference in amount of
allocation according to valgrind (disappointingly), but I also had to introduce
a copy somewhere else pending a new snapshot, so maybe that's it.
Specifically box the string (to avoid unnecessary copies) and store it
in codemap::filemap.
Remove the hack in driver::diagnostic that rereads the source from the
file and instead just get the source from the filemap.
(This commit is also a prerequisite for issue #1612)
The former contain a codemap (which is per-crate), and the latter don't. This
will be useful in order to allow more than one crate to be compiled in one run
of the compiler.
This correctly fixes issue #1362.
chpos/byte_pos are now the offsets within a particular file, but
rather the offsets within a virtual file with is formed by combing all
of the modules within a crate. Thus, resetting them to 0 causes an
overlap and hence, bogus source locations.
Fix#1362 by moving chpos/byte_pos to parse_sess so that
new_parser_from_source_str has access to them and hence can chose an
initial value that is not already been used in the crate.
Note that the trigger for bug 1361 was that syntax/ext/expand.rs calls
parse_expr_from_source_str (which calls new_parser_from_source_str)
using the same codemap as the current crate (and hence causing overlap
with files in the crate as new_parser_from_source_str resets the
chpos/byte_pos to 0).
