| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We were creating a doc map, a declaration map and a list of entities
separately by going through the HsGroup. These structures were all used
to build the interface of a module.
Instead of doing this, we can start by creating a list of declarations
from the HsGroup, then collect the docs directly from this list
(instead of using the list of entities), creating a documentation map.
We no longer need the Entity data type, and we can store a single
map from names to declarations and docs in the interface, instead of
the declaration map and the doc map.
This way, there is only one place where we filter out the declarations
that we don't want, and we can remove a lot of code.
Another advantage of this is that we can create the exports directly
out of the list of declarations when we export the full module contents.
(Previously we did a look up for each name to find the declarations).
This is faster and removes another point where we depend on names to
identify exported declarations, which is good because it eliminates
problems with instances (which don't have names).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Ross Paterson reported a bug where links would point to the defining module
instead of the "best" module for an identifier (e.g Int pointing to GHC.Base
instead of Data.Int). This patch fixes this problem by refactoring the way
renamed names are represented. Instead of representing them by:
> data DocName = Link Name | NoLink Name
they are now represented as such:
> data DocName = Documented Name Module | Undocumented Name
and the the link-env looks like this:
> type LinkEnv = Map Name Module
There are several reasons for this. First of all, the bug was caused by
changing the module part of Names during the renaming process, without changing
the Unique field. This caused names to be overwritten during the loading of
.haddock files (which caches names using the NameCache of the GHC session).
So we might create new Uniques during renaming to fix this (but I'm not
sure that would be problem-free). Instead, we just keep the Name and add the
Module where the name is best documented, since it can be useful to keep
the original Name around (for e.g. source-code location info and for users of
the Haddock API).
Also, the names Link/NoLink don't really make sense, since wether to use
links or not is entirely up to the users of DocName.
In the process of following this change into H.Backends.Html I removed the
assumption that binder names are Undocumented (which was just an unnecessary
assumption, the OccName is the only thing needed to render these). This will
probably make it possible to get rid of the renamer and replace it with a
traversal from SYB or Uniplate.
Since DocName has changed, InterfaceFile has changed so this patch also
increments the file-format version. No backwards-compatibility is implemented.
|
