- Parse errors encountered in include files are treated as "final" parse
errors in the parent file, preventing backtracking and fixing an issue
in #853
We previously had another parser type, 'type ErroringJournalParser =
ExceptT String ...' for throwing parse errors without the possibility of
backtracking. This parser type was removed under the assumption that it
would be possible to write our parser without this capability. However,
after a hairy backtracking bug, we would now prefer to have the option
to prevent backtracking.
- Define a 'FinalParseError' type specifically for the 'ExceptT' layer
- Any parse error can be raised as a "final" parse error
- Tracks the stack of include files for parser errors, anticipating the
removal of the tracking of stacks of include files in megaparsec 7
- Although a stack of include files is also tracked in the 'StateT
Journal' layer of the parser, it seems easier to guarantee correct
error messages in the 'ExceptT FinalParserError' layer
- This does not make the 'StateT Journal' stack redundant because the
'ExceptT FinalParseError' stack cannot be used to detect cycles of
include files
- Don't immediately throw custom parse errors into 'ParsecT'; rather,
just construct and return them
- This anticipates the re-implementation of an 'ExceptT' layer of the
parser, which should be able throw custom parse errors
- In anticipation of megaparsec 7, which removes support for stacks of
include files (as far as I can tell)
- Intended for the 'StateT Journal' layer of the parser
- A stack of include files would be better in a 'ReaderT' layer, but I
don't want to add another layer to the parser
- Intended for detecting cycles of include files
- Potential issue: for proper error messages for include file cycles,
we must remember to provide the filepath of the root journal file via
the initial journal state passed to a 'JournalParser'; I imagine
that we may forget to do so because in all other cases it is okay
not to do so.
A bunch of account sorting changes that got intermingled.
First, account codes have been dropped. They can still be parsed and
will be ignored, for now. I don't know if anyone used them.
Instead, account display order is now controlled by the order of account
directives, if any. From the mail list:
I'd like to drop account codes, introduced in hledger 1.9 to control
the display order of accounts. In my experience,
- they are tedious to maintain
- they duplicate/compete with the natural tendency to arrange account
directives to match your mental chart of accounts
- they duplicate/compete with the tree structure created by account
names
and it gets worse if you think about using them more extensively,
eg to classify accounts by type.
Instead, I plan to just let the position (parse order) of account
directives determine the display order of those declared accounts.
Undeclared accounts will be displayed after declared accounts,
sorted alphabetically as usual.
Second, the various account sorting modes have been implemented more
widely and more correctly. All sorting modes (alphabetically, by account
declaration, by amount) should now work correctly in almost all commands
and modes (non-tabular and tabular balance reports, tree and flat modes,
the accounts command). Sorting bugs have been fixed, eg #875.
Only the budget report (balance --budget) does not yet support sorting.
Comprehensive functional tests for sorting in the accounts and balance
commands have been added. If you are confused by some sorting behaviour,
studying these tests is recommended, as sorting gets tricky.
Custom Show instances were obscuring important details in test failure
output again. The best policy seems to be: stick with default derived
Show instances as far as possible, but when necessary customize them
to conform to haskell syntax so pretty-show can do its thing (eg when
they contain Day values, cf https://github.com/haskell/time/issues/101).
This makes skipping/unskipping tests easier, and improves readability
a bit.
Note it's also possible to just write the test name with no preceding
function, when the type is constrained (see Journal.hs).
Amount's default show instance hid important details, making eg test
failures hard to understand. Showing full detail required increasing
the debug level which was inconvenient.
Now it has a single show instance which shows more information, is
fairly compact, and is pretty-printable with pretty-show.
Ellipses (..) in the output indicate where fields are
- not shown in full detail, and/or
- shown in pseudo syntax (double quoted) to work with pretty-show.
ghci> usd 1
OLD:
Amount {acommodity="$", aquantity=1.00, ..}
NEW:
Amount {acommodity = "$", aquantity = 1.00, aprice = NoPrice, astyle = AmountStyle "L False 2 Just '.' Nothing..", amultiplier = False}
MixedAmount's show instance is unchanged, but showMixedAmountDebug
is affected by this change:
ghci> putStrLn $ showMixedAmountDebug $ Mixed [usd 1]
OLD:
Mixed [Amount {acommodity="$", aquantity=1.00, aprice=, astyle=AmountStyle {ascommodityside = L, ascommodityspaced = False, asprecision = 2, asdecimalpoint = Just '.', asdigitgroups = Nothing}}]
NEW:
Mixed [Amount {acommodity="$", aquantity=1.00, aprice=, astyle=AmountStyle "L False 2 Just '.' Nothing.."}]
Same-line & next-line comments of transactions, postings, etc.
are now parsed a bit more precisely. Previously parsing no comment
gave the same result as an empty comment (a single newline); now
it gives an empty string.
Also, and perhaps as a consequence of the above, when there's no
same-line comment but there is a next-line comment, we'll insert an
empty first line, otherwise next-line comments would get moved up to
the same line when rendered.
Some doctests have been added.
This removes transactionModifierToFunction's extra query parameter;
the rewrite command sets it in the TransactionModifier instead, which
I think is equivalent. I had to change one functional test, but it
seems correct now, so perhaps it wasn't working right before ?
I was negligent and did not test enough. This should ignore
transaction comments in auto posting rules more safely.
It also adds support for trailing comments on the first line of auto
posting rules, which previously were misparsed as part of the query.
'fail' will just terminate the current parse branch, whereas here
we have encountered a definite error. Also bring the code to
get the current working directory inside 'getFilePaths', as it
logically belongs there.
Field names are supposed to be case insensitive, but a field assignment like
fields ...,Transaction_Date,...
date %Transaction_Date
was failing, because of the capitalised letters. Fixed now.
- expands the set of expected tokens when e.g. parsing the invalid
posting `account $1 a`
- whitespace can affect parse errors because of the longest match rule
where errors that occur later take precedence over those that occur
earlier
- inline `spaceamountormissingp` into `postingp`
- combine `rightsymbolamountp` and `nosymbolamountp`
- the multiplier symbol '*' for an amount must now always preceed a sign '-'
[breaking change]
- make amount parser labels more generic to simplify error messages
For Data/Dates.hs in particular:
- Changed `SimpleTextParser` to `TextParser m` for all parsers
- Changed `string` to the case-insensitive `string'` to match the
behaviour of `T.toLower` found in `parsePeriodExpr`
- export `periodexprp` for "direct" use
... for displaying the source line on which parse errors occured
Over the following set of commits, I will to refactor the parsers to
obviate the `ExceptT String` layer of the `ErroringJournalParser` type
so that all parse errors go through Megaparsec's parse error machinery.
base-compat-batteries provides the same API across more ghc versions
than base-compat does, at the cost of more dependencies. Eg it exports
Prelude.Compat ((<>)) with ghc 7.10/base 4.8, which we expect.
My belief is that several of our deps already require it so the added
cost is not too great. We should probably go back to base-compat when
possible though, eg when we stop supporting ghc 7.10.
The new version of our package set apparently contains both base-compat and
base-compat-batteries in its transitive closure. This breaks the doctest suite,
which just imports everything into scope when the tests are run, thereby making
module names like Prelude.Compat ambiguous.
We don't need to import Data.Monoid because Prelude.Compat exports "<>"
already. In fact, importing that module causes build failures:
Hledger/Read/Common.hs:725:62: error:
Ambiguous occurrence ‘<>’
It could refer to either ‘Sem.<>’,
imported from ‘Prelude.Compat’ at Hledger/Read/Common.hs:97:1-39
(and originally defined in ‘Data.Semigroup’)
or ‘Data.Monoid.<>’,
imported from ‘Data.Monoid’ at Hledger/Read/Common.hs:110:1-18
Fixes https://github.com/simonmichael/hledger/issues/794.
- Rationale:
- The information necessary for applying exponents to a number is more
explicitly represented in the inputs to `fromRawNumber` than in the outputs
- This way, `exponentp` may simply return an `Int`
- Purpose: to reduce the verbosity of the previous implementation
- Split off `AmbiguousNumber` into its own type
- Introduce a function `AmbiguousNumber -> RawNumber` explicitly capturing the
disambiguation logic
- Reduce the number of remaining constructors in `RawNumber` to just two,
`WithSeparator` and `NoSeparator`
- The choice to distinguish by the presence of digit separators is motivated
by the need for this information later on when disallowing exponents on
numbers with digit separators
- Extracts the handling of signs out of `fromRawNumber` and into `signp` itself
- Rationale: The sign can be applied independently from the logic in
`fromRawNumber`
It now works slightly differently. Eg:
- <unbudgeted>'s subaccounts are hidden by default
- --show-unbudgeted shows all unbudgeted accounts, including subaccounts of budgeted parents
- --show-unbudgeted doesn't affect the grouping under <unbudgeted>
IMHO it's a nice simplification and increase in consistency, while still meeting the original intent.
Accounts which have no budget goals within the report period are now
grouped under <unbudgeted> - not just accounts with no budget goals ever.
Haddocks have been clarified, especially for budgetRollup. In some
ways things are much clearer without this feature, but it remains
enabled by default for now.
Budgets were restricted to a single interval in 1.9, but this was
a mistake. This restores the 1.5 behaviour, where a budget can be built
up from multiple periodic transactions with different intervals.
A commodity directive that doesn't specify the decimal point character
increases ambiguity and the chance of misparsing numbers, especially
as it overrides all style information inferred from the journal amounts.
In some cases it caused amounts with a decimal point to be parsed as if
with a digit group separator so 1.234 became 1234.
We could augment it with extra info from the journal amounts, when available,
but it would still be possible to be ambiguous, and that won't be obvious.
A commodity directive is what we recommend to nail down the style.
It seems the simple and really only way to do this reliably is to require
an explicit decimal point character. Most folks probably do this already.
Unfortunately, it makes another potential incompatiblity with ledger and
beancount journals. But the error message will be clear and easy to
work around.
Inferred amounts now have the appropriate standard amount style applied.
And when checking for balanced transactions, amount styles declared with
commodity directives are also used (previously only inferred amount styles were).
Income, liability and equity balances, which until now we have
always displayed as negative numbers, are now shown as normally positive
by these reports.
Negative numbers now indicate a contra-balance (eg an overdrawn
checking account), a net loss, a negative net worth, etc.
This makes these reports more like conventional financial statements,
and easier to read and share with normal people.
