1.50* attempted to work around Glob's implicit searching of non-top-level dot dirs.
This was overzealous; it meant that journal's include completely
excluded paths involving a glob and a dot dir or dot file anywhere in the path.
Now, the pre-1.50 behaviour is restored:
`*` and `**` won't implicitly match dot files or top-level dot directories.
They will implicitly search non-top-level dot directories, as before (#Glob#49).
Before 1.50, journal's include directive's handling of glob patterns (*, **, ?, etc.)
had these limitations:
- ** always searched intermediate dot directories
- ** matched only directories, not files
In 1.50-1.50.3, it had different limitations, some unintended:
- it ignored all dot files, dot dirs, and symbolic links to dot dirs,
even when explicitly mentioned in the pattern (unless using --old-glob)
- it showed symbolic links dereferenced, eg in `hledger files` output
Now it has fewer limitations, mainly this:
- it ignores all dot files and dot dirs, even when explicitly mentioned (unless using --old-glob)
Ie it no longer ignores symbolic links to dot dirs, and it no longer shows symbolic links dereferenced.
Also: including the current file is now always harmless, whether using a glob pattern or not.
Internally, file paths in the "include file stack" (jincludefilestack) are now just absolute,
but not canonicalised; showing symbolic links un-dereferenced in output and error messages seems
generally more useful. This might affect output elsewhere also.
(Those paths are still canonicalised on the fly when checking for include cycles,
not so efficiently: each time an include directive is parsed, all the current parent files
and all the new glob-matched include files will be re-canonicalised.
Hopefully this is unnoticeable.)
Avoiding potentially confusing silent fallback. Also,
- Drop support for Ledger's legacy LEDGER environment variable;
we now support only LEDGER_FILE, for simplicity.
- Clarify the behaviour, eg when a glob pattern matches multiple files
or when the value is empty.
Since 1.50, sourceFilePath, which does IO operations, was being called for every item in the journal.
On my machine this was causing a ~40% slowdown,
but probably it could be more depending on storage system.
Now it's once again called only once per include directive.
Speed seems slightly better now than 1.43 for some reason
(eg: 13k txns/s -> 8k txns/s -> 14k txns/s).
PeriodData's use of Int keys caused wrong results with periodic
reports involving dates outside the machine-specific limits of Int.
Those were:
64 bits: -25252734927764696-04-22..25252734927768413-06-12
32 bits: -5877752-05-08..5881469-05-27
16 bits: 1769-02-28..1948-08-04
8 bits: 1858-07-12..1859-03-24
32 bits is supported by MicroHS; 16 and 8 bits aren't supported by
any known haskell version, but that could change in future.
For example, on 64 bit machines we got:
25252734927768413-06-12 PeriodData's max date
(expenses) 1
25252734927768414-01-01 next year past PeriodData's max date
(expenses) 2
$ hledger reg -O csv --yearly
"txnidx","date","code","description","account","amount","total"
"0","-25252734927764696-11-10","","","expenses","1","1"
Now it uses Integer (like the time package), fixing the bug.
And benchmarking shows memory and time usage slightly improved
(surprisingly; tested with up to 500 subperiods, eg
hledger -f examples/10ktxns-1kaccts.journal reg -1 cur:A -D >/dev/null)
This allows us to guarantee that the report periods are well-formed and
don't contain errors (e.g. empty spans, spans not contiguous, spans not
a partition).
Note the underlying representation is now for disjoint spans, whereas
previously the end date of a span was equal to the start date of the
next span, and then was adjusted backwards one day when needed.
7 is no longer considered a scary shell character.
(This appears to have been harmless, used only for formatting the
balance assertion failure message.)
We no longer attempt to process timeclock entries in time order - that
was a wrong requirement, probably given by me, that can't work. Now
we just process them in parse order. This plus a little tweaking of
error checking fixes several ordering bugs with overlapping sessions
and also allows same-named overlapping sessions.
More cleanup will follow. More testing might show that --old-timeclock
is no longer needed.
Now we process entries in a more careful order: time, then parse
order, like journal format. This fixes the original issue and another
one mentioned at #2417.
The default timeclock parser (ie when not using --old-timeclock) has
the following changes, related to issues such as
[#2141], [#2365], [#2400], [#2417]:
- semicolon now always starts a comment; timeclock account names can't include semicolons
(though journal account names still can)
- clock-in and clock-out entries now have different syntax
- clock-ins now require an account name
- clock-outs now can have a comment and tags
- the doc has been rewritten, and now mentions the --old-timeclock flag
- lib: accountnamep and modifiedaccountnamep now take a flag to allow semicolons or not
