purescript-backend-erl

_PureScript to Erlang optimizing compiler, based on backend-optimizer._

purescript-backend-erl aims to be source-compatible with purerl code.
In particular, the representation of datatypes remains unchanged and existing FFI files should work as-is.
(The generated Erlang files will not be compatible.)

The main visible differences are uncurrying (which may affect the public API of generated modules) and the details of what errors may be thrown.

Warning: bigint literals are currently unsupported and you will get JSON decoding errors if you use them in your source code.
You can encode them with arithmetic as a workaround until things are patched.

Also, due to inlining, inlining of div for Int will occur more often, and this optimization is unsafe since it is optimized to div in Erlang which has different semantics for negative numbers and dividing by zero.

Usage

purs-backend-erl compiles code from CoreFn in ./output (./output/Module.Name/corefn.json) to Erlang modules in ./output-erl (./output-erl/Module.Name/module_name@{ps,foreign}.erl).
This is the same convention as purerl, just in ./output-erl instead of ./output to avoid conflicts.

Note: purescript-backend-erl is the name of the repository, while purs-backend-erl is the name of the executable and its associated npm package.

$3

Run tests with spago test -- --accept --run to accept new golden output and compile and run the resulting Erlang tests.
(Not all tests have assertions (yet).)

Results

Like purescript-backend-optimizer, performance improvement can be around 30% or even 40%.
The size of the compiled BEAM files is comparable to that of purerl.

Optimizations

Optimizations are largely those provided by backend-optimizer, adapted as necessary for the Erlang code generation and purerl libraries.
(Ref and ST optimizations were yeeted.)

See the backend-optimizer repo for examples of how to inline RowList operations, Generics, lenses, and others.

The additional Erlang-specific optimizations take place at various stages of the pipeline:

- Complexity analysis to guide inlining during backend-optimizer's evaluation
- Evaluation of known foreign functions in PS semantics (they do not have to be foreign, strictly speaking)
- Specializing known foreign functions during Erlang codegen
- This is where Erlang Lists and Tuples get specialized, since they work with Erlang data structures that cannot be represented by backend-optimizer.
- Uncurrying during codegen
- After Erlang codegen
- This is Common Subexpression Elimination part 2, Erlang Edition.
- Pattern matches are also generated based on demand analysis, to avoid deeply nested accessors.

The cherry on top is that Erlang has weird scoping rules, so the final step of codegen is renaming local variables throughout the whole AST.
Final beta-reduction of applied lambdas plus re-scoping of macros also happens there.

These optimizations were guided by inspecting the optimizations that Erlang performs, see ./opts.

$3

Unlike purerl, which can access externs through the Haskell API of the compiler, backend-erl does not have access to the types of declarations.
This necessitates generating uncurrying based on the function code after optimization, not the types.
This makes it less predictable and may change the public API of PureScript modules in their Erlang form.

- The meaning of foreign imports remains the same: foreign f/0 declarations are taken as-is, and f/2 declarations are taken to be pure binary functions, and so on.
- Transitivity: If g calls f with two arguments and a f/4 uncurrying exists, then a g/2 uncurrying is also generated that calls f/4.
- Uncurrying is affected by backend-optimizer inlining directives (inlining f arity=2 in the above example will also uncurry g/2, but via inlining this time, not by calling f/4).

The calling convention is that generated _nullary_ Erlang functions are always _pure_ wrappers around the code corresponding to a PureScript declaration.
Non-nullary generated Erlang functions may correspond to any of the three types of functions (curried, uncurried, or uncurried effectful), based upon their definition.
Non-nullary FFI functions only correspond to curried functions (_ -> _ -> _), not FnX nor EffectFnX.

$3

- Foreign functions in core libraries are specialized and inlined when possible, especially Erlang BIFs.
- Erlang Tuple operations are always inlined.
Erlang List operations are mostly inlined (TODO: inline uncons).
They are both treated as constructors by backend-optimizer so they will be inlined more aggressively.
- MagicDo is optimized only where the backend-optimizer inlinings make it obvious: abstract functions that are later called with Effect dictionaries do not have specializations generated (TODO).
- Foreign declarations are called directly, instead of via their wrappers in the generated PS modules (although those are still generated for backwards compatibility).

$3

Calls to top-level nullary functions are hoisted, to ensure that they are only instantiated once in a scope.

This is one of the main downsides of the Erlang codegen: top-level declarations have to be functions, so even with the CSE that purs itself performs, typeclass dictionaries get re-instantiated a lot.
(TODO: pull out typeclass methods as their own declarations to avoid dictionary creation in the first place.)

$3

Some pattern matches are rewritten back to Erlang case trees for legibility (the backend-optimizer uses Boolean branching and unsafe accessors exclusively).
(TODO: do this for product-like datatypes.)

$3

- Pull out typeclass methods
- More uncurrying (e.g. for compareImpl LT EQ GT)
- Ensure generics always inline
- Test optimizations
- Make it extensible?