Fast anonymous records in Haskell, at the cost of using some unsafe constructs. Current features:
lens-compatible lens for freeTake a look at Data.FastRecord.Example to get some idea on how things work.
An anonymous record tracks its schema in its type. We use ts :: [(Symbol, Type)] to represent a schema. The kind of ts is simple: a list of fields, each field consists of a name and a type. So '[ '("foo", Bool), '("bar", String)] is similar to data FooBar = FooBar { foo :: Bool, bar :: String }. The ' prefix on type constructors is related to the DataKinds extension.
Code related to schemas is in Data.FastRecord.Schema. We use type families to do some calculations at the type level (like the total number of fields, or the index of a particular field), then we apply constraints like KnownNat (FieldIndex ts sym t) to our function types, which means we can use Proxy#s to connect type-level and value-level and obtain results of type-level computations at runtime.
Functions with names ending with # explicitly pass Proxy#s around and aren't recommended for use by end-users.
Rec ts is the actual record value. The common representation of anonymous records is type-indexed heterogeneous lists, which is elegant and type-safe. Here we use primitive SmallArray#s instead to improve performance. The memory is more compact (a constant overhead over normal records), and field-accessing is O(1).
We can get/set a field of a record via getField#/setField#. We can also initialize an empty record via nullRec#, then set its individual fields.
The recommended way to get/set a field is to use lens. Import Data.FastRecord.Field, turn on the OverloadedLabels extension, and simply use label #foo as a lens for the foo field. It works for any schema including a foo field.
There are cases when we wish to update several fields at once. With nested single-field updates, the record is copied for multiple times which is a huge waste.
Data.FastRecord.BulkSet is for bulk updating. A BulkSet ts value describes the action of updating one or more fields of a Rec ts. It can be generated and composed using the Monoid interface (if there's overlapping, the right side has higher priority). We can also use the ~= operator to describe assigning a single field.
So a typical BulkSet ts looks like Proxy @"foo" ~= True <> Proxy @"bar" ~= "233" (with TypeApplications turned on)
A BulkSet ts value can be used to update a record (via bulkSet) or initialize a new record (via newRec).
fast-record needs some more polishing before it's available on Hackage/Stackage. Possible improvements include:
Show, Eq/Ord, NFData, etc)