All the passes now have their information (name, pre-requisites and post- properties) stored at the point where the pass is declared, which means the pass lists are just a simple list of pass functions.
The main consequence of this change was that the tests had to be changed. Now, instead of taking a "pass applied to data" item (type: PassM b), they take both the pass (type: Pass) and source data (type: b), and apply them later. This was the decision that involved the simplest changes to the existing tests (simply unbracketing the application of the pass to the source). I also had to include a few old-style versions though (testPass', testPassShouldFail') for where the functions were being used to test things that weren't actually passes (mainly StructureOccam).
Fixes#48
As part of this patch, I have also introduced a helper function that fiddles the type system for those passes that must run at the top-level (i.e. on A.AST) rather than on any Data t. They will give an error if not applied at the top-level.
This changes the Traversal API to the one that I've been working on in
the Polyplate branch, but implemented in terms of Data. The
performance isn't as good as the Polyplate version, but the code is a
lot simpler because it doesn't need all the type constraints (and it
doesn't make GHC struggle).
This also reworks all the passes in Tock to use the new API, including
those that previously used makeGeneric (which I've now removed) or
everywhereM. Most of the passes are simpler because of this, and I
suspect it's fixed a few subtle bugs resulting from missing recursion
in makeGeneric code.
I haven't yet profiled this, but subjectively it seems about the same
as the old Traversal (and thus faster for all the passes that didn't
yet use it).
This fixes ticket #47 from Trac, which explains how using a dependency graph for passes was a bit too over the top, and led to unexpected results. Under the "new" (the original!) system, the pass list is used as-is, but the dependencies are checked to make sure the pass list order isn't wrong. In future we should also add back running the properties at the appropriate point (currently disabled).
We now have three kinds of canned tree traversals, all of which are smart about
which types they're applied to: explicit-descent transformations,
implicit-descent transformations, and implicit-descent checks. I've only
provided depth-first application of the latter two, but we could do
breadth-first in the future if necessary.
This provides gmapMFor and gmapMFor2, which are like gmapM, but know what
they're looking for, and can therefore avoid going down branches of the tree
that won't contain any interesting types.
The basic approach is quite similar to Uniplate's PlateData: there's a function
(containsType) that'll tell you whether one type is contained somewhere within
another. However, unlike Uniplate, we build a static IntMap IntSet of the types
we need to know about, which allows rather quicker lookups. (I did try using
PlateData first.)
The result is that applyDepthM is now much quicker than it was before.
applyDepthM2 is a bit less impressive, which I assume is because it can't
really prune the tree much if it's looking for two types.
Future enhancements:
- convert more passes to use applyDepthM*;
- make gmapMFor* aware of constructors rather than just types, which should
allow a bit more pruning.
