18 KiB
Exactprinting
Brittany uses the ghc-exactprint
library/wrapper around the GHC API to
parse haskell source code into a syntax tree and into "annotations". The
unannotated syntax tree would lose information, such as the exact (relative)
position in the source text, or any comments; this is what annotations provide.
Following that name, we'll call "exactprinting" the aspect of reproducing
comments and relative positions - most importantly additional newlines - while
round-tripping through brittany. The focus is not on the API of the
ghc-exactprint
library, but on the corresponding data-flow through brittany.
Take note that the --dump-bridoc-*
output filters out the constructors
responsible for comments and for applying DeltaPositions.
This is done to keep the output more readable, but could confuse you if you
try to understand how comments work.
TLDR - Practical Suggestions for Implementing Layouters
This advice does not explain how comments work, but if you are implementing a layouter it might cover most cases without requiring you to understand the details.
-
Ideally, we should wrap the
BriDoc
of any construct that as a location (i.e. has the form(L _ something)
) (and consequently has anAnnKey
) usingdocWrapNode
. As an example, look at thelayoutExpr
function and how it appliesdocWrapNode lexpr $ ..
right at the top. -
If we have not done the above, it is somewhat likely that comments "get eaten". For such cases:
-
Take a small reproduction case
-
Throw it at
brittany --dump-ast-full
and see where the comment is in the syntax tree. See where the corresponding syntax node is consumed/transformed by brittany and wrap it withdocWrapNode
. -
If it is unclear what alternative (of a
docAlt
orrunFilteredAlternative
) applies, try insertingdocDebug "myLabel"
nodes to track down which alternative applies.
-
-
For comments that do appear in the output but at the wrong location, there are two classes of problems: Firstly we have comments that move "down" past other stuff (even switching order of comments is possible). Use the steps from the last item to figure out which syntax tree constructor is relevant, and try inserting
docMoveToKWDP
or replacedocWrapNode
with a manually refined combination ofdocWrapNodePrior
anddocWrapNodeRest
. -
For comments that do appear in the output in roughly the right position, only with the wrong indentation, the cause most likely is a mis-interpretation of DPs that can be fixed by inserting a
docSetIndentLevel
at the right position - right before printing the thing that provides the "layouting rule" indentation, i.e. the body of ado
/let
/where
block. -
There is one other cause for off-by-one errors in comment position: Whitespace. In general, layouters should prefer to use
docSeparator
to insert space between syntax elements rather than including spaces in literal strings. As an example, usedocSeq [docLit "then", docSeparator]
or the equivalentappSep (docLit "then")
rather thandocLit "then "
. The reason is that comment positions are relative to the last non-whitespace, anddocSeparator
is interpreted in just the right fashion: It inserts a whitespace, but keeps track of the correct comment offset. (Also, subsequentdocSeparators
are merged into one.) -
If all of this fails, read below, bother the maintainers and/or make use of the more advanced debugging features (there is a
#define
inBackendUtils.hs
that you can turn on to insert all kinds of verbose output in-line with the actual output).
A Small Example
main = do
putStr "hello" -- very suspense
putStrLn " world" --nice
If you pass this to brittany --dump-ast-full
you'll see .. a 100 line syntax
tree. Yeah, raw syntax tree are a bit unwieldly.
(btw I'd use clipread | brittany --dump-ast-full
for that purpose, where
clipread
boils down to xclip -o -selection clipboard
. If you have not set
up that script on your system, you really should.)
To simplify this slightly, we will focus down on just the syntax tree of
the do
block, which is the HsDo
constructor.
---- ast ----
A Just (Ann (DP (0,0)) [] [] [((AnnComment (Comment "--nice" stdin:3:21-26 Nothing)),DP (0,1)),((G AnnEofPos),DP (1,0))] Nothing Nothing)
HsModule
..
[ A Just (Ann (DP (0,0)) [] [] [] Nothing Nothing)
ValD
FunBind
A Just (Ann (DP (0,0)) [] [] [((G AnnVal),DP (0,0))] Nothing Nothing)
Unqual {OccName: main}
MG
A Nothing
[ A Just (Ann (DP (0,0)) [] [] [((G AnnEqual),DP (0,1))] Nothing Nothing)
Match
FunRhs
..main..
Prefix
NoSrcStrict
[]
GRHSs
[ A Just (Ann (DP (0,-1)) [] [] [] Nothing Nothing)
GRHS
[]
A Just (Ann (DP (0,1)) [] [] [((G AnnDo),DP (0,0))] Nothing Nothing)
HsDo
DoExpr
A Nothing
[ A Just (Ann (DP (1,2)) [] [] [] Nothing Nothing)
BodyStmt
A Just (Ann (DP (0,0)) [] [] [] Nothing Nothing)
HsApp
..putStr..
.."hello"..
..
..
, A Just (Ann (DP (1,0)) [((Comment "-- very suspense" stdin:2:18-33 Nothing),DP (0,1))] [] [] Nothing Nothing)
BodyStmt
A Just (Ann (DP (0,0)) [] [] [] Nothing Nothing)
HsApp
..putStrLn..
.." world"..
..
..
]
]
A (Nothing) (EmptyLocalBinds)
]
FromSource
WpHole
[]
]
..
So this is a haskell module, HsModule
containing a function bind FunBind
containing a match group, containing a Match, containing some right-hand-side
expression which in this case is just a do block HsDo
which contains two
applications HsApp
of a function putStr(Ln)
plus some string literal.
There is no need to understand this, as long as you can roughly see how this representation corresponds to the input source code.
For the purpose of exactprinting, what we need to look at are the annotations.
The ghc-exactprint
library returns the syntax tree and annotations as two
different entities:
- You can start looking at the module level and work your way down to any syntactical construct from there;
- The Annotation type and its
Ann
constructor.
In the above --dump-ast-full
output these two are mixed together using the
fake A
constructor that is just a pair of a Maybe Annotation
and of one
node in the syntax tree. It was produced by recursively printing the syntax
tree, and for each node n
we print A (getAnnotation n) n
. So let's focus
on the Annotation
type.
The ghc-exactprint
Annotation Type
Ann
{ annEntryDelta :: !DeltaPos
, annPriorComments :: ![(Comment, DeltaPos)]
, annFollowingComments :: ![(Comment, DeltaPos)]
, annsDP :: ![(KeywordId, DeltaPos)]
, annSortKey :: !(Maybe [SrcSpan])
, annCapturedSpan :: !(Maybe AnnKey)
}
But please refer to the ghc-exactprint docs for the fully commented version.
A few things to point out:
-
There are three constructors that contain the
Comment
type in that constructor.annPriorComments
andannFollowingComments
are obvious, but a third hides behind theKeywordId
type. Source code comments may appear in one of these three locations. -
The
DeltaPos
type and itsDP
constructor can be seen in the above output everywhere. It contains information about relative positioning of both comments and syntax nodes. Please test what changes if you insert a newline beforeputStrLn
, or add spaces before one of the comments, and see how the--dump-ast-full
output changes. -
The exact semantics of the
DP
value, especially when it comes to indentation, are a source of constant joy. If the values don't make sense, you are on the right track. Just figure out what DP is connected to what change in the syntax tree for now. -
We have two comments in the source code, which appear in opposite order in the
--dump-ast-full
output. The reason is that comments mostly appear in the middle of two AST nodes, and it is somewhat arbitary whether we connected them as an "after" comment of the first or as an "before" comment of the second node. And keep in mind that we have a third constructor that can contain comments that are somewhere in the "middle" of a node, too. -
We have
DP
s with negative offsets. Did I mention how much funDP
s are? I have no idea where the above-1
comes from. -
The
annsDP
field may also contain theDP
s of syntax that is somewhere "in the middle" of a syntax node, e.g. the position of theelse
keyword.We will discuss the semantics of
DP
further down below.
Data-Flow of a Comment When Round-Tripping
Parsing with ghc-exactprint
returns both a syntax tree and a map of
annotations (Map AnnKey Annotation
). Let's consider just the comment
"-- very suspense" in the above example: The annotations map would contain
the following mapping:
AnnKey {stdin:3:3-19} (CN "BodyStmt")
-> Ann { annEntryDelta = DP (1,0)
, annPriorComments =
[((Comment "-- very suspense" stdin:2:18-33 Nothing),DP (0,1))]
, annFollowingComments = []
, annsDB = []
, annSortKey = Nothing
, annCapturedSpan = Nothing
}
where the AnnKey
is connected to the syntax node BodyStmt
with the given
source location.
Brittany keeps the annotations map around, and the BriDoc
structure contains
nodes that have AnnKey
values, i.e. the BriDoc
nested documented structure
similarly only contains references into the annotations map. The corresponding
constructors of the BriDoc(F)
type are:
data BriDoc
= ..
| BDAnnotationPrior AnnKey BriDoc
| BDAnnotationKW AnnKey (Maybe AnnKeywordId) BriDoc
| BDAnnotationRest AnnKey BriDoc
| BDMoveToKWDP AnnKey AnnKeywordId Bool BriDoc -- True if should respect x offset
| ..
when rendering a BriDoc
to text, the semantics of the above nodes can be
described roughly like this:
-
render (BDAnnotationPrior annkey bd)
extracts the "before" type comments under the givenannkey
from the annotations map (this is a stateful process - they are really removed from the map). It renders these comments. If we are in a new line, we respect theannEntryDelta :: DeltaPos
value to insert newlines. The "if in a new line" check prevents us from inserting newlines in the case that brittany chose to transform a multi-line layout into a single-line layout.Then we recursively process
render bd
. -
render (BDAnnotationsKW annkey mkwId bd)
similarly first renders the comments extracted from the annotations map under the givenannkey
before callingrender bd
. For example, this would allow us to print the comments before the closing bracket]
of an empty list literal e.g.[{-put numbers here to do X-}]
. -
render (BDMoveToKWDP annkey kwId xToo bd
moves to the relative position of the given keyword before continuing withrender bd
. It is used for example to insert newlines before awhere
keyword to match those of the original source code. -
render (BDAnnotationsRest annkey bd)
first callsrender bd
and then takes any remaining comments it can find in the annotations map under the givenannkey
and prints them.
Some Notes to This Design
-
We heavily rely on the
ghc-exactprint
library and its types and their semantics. We could define our own data structures to capture comments and whitespace offsets. While this could allow us to make the later steps of the process easier by more closely matching the information we need when rendering aBriDoc
document, it would involve a mildly complex extra transformation step fromghc-exactprint
annotations to hypotheticalbrittany
annotations. -
For those cases where we rely on
ghc-exactprint
to output syntax thatbrittany
does not know yet, it is mandatory that we keep the annotations around. -
We make the rendering stateful in the annotations. The main advantage to this is that we can keep track of any comments that have not yet been reproduced in the output, and as a last resort append them at the end. The effect of that is that comments "move" down in the document when brittany is not exact, but at least it does not "eat" comments. The latter can still happen though if we forget to include a given
AnnKey
at all in theBriDoc
document.Of course this is a bit yucky, but it seems to be a sensible measure for the long transitioning period where
brittany
is not perfect. -
It may be surprising to nest things like we do in the
BriDoc
type. The intuitive document representation for something like-- before foo -- after
might be
sequence [comment "-- before", text "foo", comment "-- after"]
but instead we use
BDAnnotationsPrior annkey1 -- yields "-- before" BDAnnotationsRest annkey1 -- yields "-- after" BDLit "foo"
which may seem unnecessarily nested. But this representation has certain advantages, most importantly rewriting/restructuring the tree is straigh-forward: consider how
BDAnnotationsPrior annkey (BDSeq [a, b])
can be transformed intoBDSeq [BDAnnotationsPrior annkey a, b]
. You can do the same transformation using the "flat" representation, but there are way more cases to consider.
DeltaPosition semantics
DeltaPositions (we'll just say DP
which matches the constructor name for this
type) are used to specify where to place comments and regular syntax (including
keywords). This covers both newlines and indentation, and for indentation
includes the case where indentation is mandatory ("layouting rule").
Let us look at this example, which was constructed so that each comment contains its own DP:
do -- DP (0, 1)
-- DP (2, 2) two newlines, two spaces indentation
abc
-- DB (1, 0) one newline, zero indentation relative to the do-indentation
def
The first comment is of the easy sort, because it occurs at the end of a non-empty line: There is no row offset, and the column offset matches the number of spaces (before the "--") after the last thing in the line.
The second comment does have a row offset: After the last comment, we have to insert two line-breaks, then apply the indentation (two spaces) and then insert the comment starting with "--". This is straight-forward so far.
The third comment however highlights how DPs are affected by the layouting rule.
Caveat One: Indentation relative to layouting rule indentation level
Following the first two cases, one would assume that the DP would be
(1, 2)
. However, for cases where the layouting rule applies
(do
, let
, where
) the indentation of the comments is expressed relative
to the current indentation according to the layouting rule. Unfortunately,
this current indentation is not known until the first construct after
the let, so in the above example, the comment between the do
and the first
construct (abc
) has an indentation relative to the enclosing indentation
level (which is 0 for this example). This applies even if the comment is
connected to the first construct (if the first comment is a "prior" comment
of the "abc" syntax node).
This applies not only to comments, but also to the DPs of all syntax nodes (including keywords).
This also means that it is possible to have negative indentation. Consider this comment:
do
abc
-- fancy comment
def
Caveat Two: Caveat one applies to more than the layouting rule
There are syntactic constructs, for example data declarations, where the
layouting rule does not apply, but for the purpose of DP
indentations
ghc-exactprint
pretends that it does. For example:
data Foo = Foo
{ myInt :: Int
-- DP (1, -7) relative to the `Foo` constructor (!)
}
The layouting rule does not apply in any way here. Still, we get a rather unexpected DP.
DeltaPositions of keywords and syntax nodes
We have mostly talked about comments, but DPs exist and work for keywords and syntax nodes just like they do for comments.
func = x
where
x = 42
here, the where
keyword has a DP of (2, 1)
and the x = 42
equation
has a DP of (2, 2)
. We make use of these DPs using the BDMoveToKWDP
or the
BDAnnotationPrior
constructors of the BriDoc
document. The former would be
used for the where
keyword, the latter would be applied to the equation
document.