Skip to content

Instantly share code, notes, and snippets.

@MateuszKubuszok
Last active August 7, 2023 00:39
Show Gist options
  • Save MateuszKubuszok/5c05ccc775ed81f8e102af3507d8e7ed to your computer and use it in GitHub Desktop.
Save MateuszKubuszok/5c05ccc775ed81f8e102af3507d8e7ed to your computer and use it in GitHub Desktop.
Chimney implementation ideas

Separate automatic derivation from semiautomatic

Directly solves:

General idea of the solution

Currently Chimney attempts to derive as much as possible with a single expression. It inline results on Scala 2 it uses an approach where when implicit is summoned it is expected that only user-provided implicit would be used. This is done because:

  • if user provided an implicit to override transformation, we have no choice, we would just use it

  • however, if it was an implicit provided by auto-derivation, then the resulting expr would be wrapped in

    new Transformer[From, To] { def transformer(src: From): To = derived expression }.transform(expr)

    while we are able to avoid all that unneessary allocation and return

    derived expression

    directly

  • in versions prior to 0.8.0-M1 it was achieved by summoning Transformer... and checking how it was created and discarding it if it was created by Transformer.derive implicit. This resulted in potentially exponential complexity as for each field and each subtype this transformer could be summonned, possibly derived... and then discarded.

So 0.8.0-M1 split Transformer and Transformer.AutoDerived to prevent this. However, it complicated the API in a way that doesn't appeach in other derivation libraries, so the proposed solution is to:

  • require users to import an implicit with automatic derivation themselves
  • remove distinction between Transformer and Transformer.AutoDerived
  • create a very detailed section in readme about automatic derivation and how it affects performance, because this change would quite often introduce tons of allocations which where not there before one users import auto._

Alternatively:

  • we could preserve distinction between those 2 types

  • introduce a separate method for summoning and using ONLY Transformer and not Transformer.AutoDerived limitting its usage to values explicitly provided by users

  • create some implicit returning Transformer rather than Transformer.AutoDerived which could be used when user would use something like:

    implicit def myType[A, B](implicit t: Transformer[A, B]): Transformer[X[A], Y[B]] = ...

Things to consider

If we managed to separate auto from semiautomatic - one way or another - we could consider separating:

  • dsl
  • internal.runtime
  • internal.compiletime

so that users could require in runtime only type classes and a few utilities used in generated code, and all macros could be required only for the compilation.

Custom constructor for target value

Directly solving:

Related:

General idea of the solution

  • we let used add a custom function with returns To type
  • this function is called instead of a constructor in ProductToProductRule
  • we make sure that all the existing modifiers and flags still work

Necessary steps

Adding custom function to DSL

To avoid gigantic changes to codebase, API, DSL we would assume that function HAS to have the same names of parameters used as values that appear in created type. Only then we would be able to call .withField* modifiers, with IDE support intact and making it working on Scala 2.12, 2.13 and 3. We can relax this requirement to say that each field with for which we use an override can only be overriden if there is such field defined for it (so that we could use it in DSL).

We would also have to check that types match between provided function and To fields. This would require us to store all inputs in type-level as some sort of Params.Value[B, "fieldNameB", Params.Value[A, "fieldNameA", Params.Empty]] type-level list.

Then we could implement a whitebox macro/transparent inline for .withConstructor((fieldNameA: A, fieldNameB: B) => ...) which would parse the AST, the value into runtimeDataStore and params types into TransformerCfg. This macro (if possible) should also parse things like .withConstructor(new Class(_, _, _)) (overriden constructor) or .withConstructor(Object.apply).

Using custom constructor in derivation

We can refactor TransformProductToProductRule so that matching Type[To] on Product.Constructor(arguments, parameters) could be skipped and we would use constructor extracted from RuntimeDataStore directl.

Testing

Virtually whole suite of TotalTransformerProductSpec would have to be copy pasted as TotalTransformerConstructorSpec and:

  • each .transformInto would have to be removed
  • each .into.transform would have to be replaced with .into.withConstructor().transform
  • each manually provided constructor would have to modify values to make sure that default constructor wasn't used instead
  • each of the above would have to be done for { () => ... }, new Target(_, _, _) and HelperObject.apply syntaxes to check that each of these works

Things to consider

  • how many helpers we want to add?
    • if we stay at a.into[B].withConstrcutor(...).transform things will be pretty simple
    • if we start considering thing like a.transformIntoVia[B](...) then each such feature would call for explicit helper. e.g. what if we wanted to use custom constructor with merging of 3 case classes? Where do we draw the line?
  • as a second step, later on in the future, we could allow .withConstructorPartial to enable support for smart constructors

Merge Transformers and recursive Patchers

Directly solving:

Related:

General idea of the solution

  • Patchers piggy-back on top of Transformers abstractions, so if Transformers supported (Foo, Bar) into Baz then Patchers could be implemented internally as (Patch, A) into A
  • currently Patchers check if all fields are used while Transformers doesn't, so if Transformers implemented policy checking, then Patchers could just use a default different policy than Transformers
  • making merge Transformers recursive would also made Patchers recursive
  • Patchers could have added extra rule for searching implicit Patchers

Necessary steps

Adding falback values to derivation

To avoid gigantic changes to codebase, API, DSL we would assume that Transformers and PartialTransformers have one main source value - which would be used in .withFieldComputed, .withFieldComputedPartial, .withFieldRenamed - and a list of "fallback" values which would be looked for a field if the main source value is missing it.

  1. TransformationContext would have to introduce val fallbacks: Vector[ExistentialExpr]
  2. as a intermediate fallback should be cleared when calling updateFromTo
  3. rule TransformProductToProductRule would need to expand its fallbacks section to:
    1. take each value from the list and use Product.Extraction to disect it (we might consider caching it somehow to avoid running it separately for each field) to see which fields each value could provide
    2. check if the sought fieldName is among available ones
    3. take first value that matches and use it as a source for deriveRecursiveTransformationExpr
  4. from this moment, if DSL could pass these extra values to Gateway
  5. TransformProductToProductRule, TransformTypeToValueClassRule, TransformValueClassToTypeRule TransformValueClassToValueClassRule needs to be extended to

From this moment on if DSL allowed adding fallback values and Gateway would pass it into TransformerContext then Chimney would support shallow case class merging.

Adding fallbacks to DSL

We have to keep in mind that Chimney already supports converting tuple to case class/another tuple, so we cannot just use (A, B).into[C] as the syntax for merging. Since fallback values would not be used in DSL in any other .with* method, we can just remember its type in some type-level list of TransformerCfg and allow adding them one by one e.g. like this:

val a: A
val b: B
val c: C

a.into[D]
  .withFallback(b)
  .withFallback(c)
  .transform

I'd suggest we didn't store these values in runtimeData: RuntimeDataStore - we can have another Vector[Any] in TransformerInto and in PartialTransformerInto, so that:

  • in cases when [Partial]TransformerInto is not used, then similarly to From expression the fallback values could be assigned to some val so that accessing them would be easier and without wrapping
  • in cases when [Partial]TransformerInto is used, then similarly to src: From expression the fallback values could be extracted from Vector[Any] and assigned to vals
  1. TransformerInto and PartialTransformerInto would have to be extended to contain val fallbacks: RuntimeDataStore and some methods for updating it (similarly to val runtimeData: RuntimeDataStore)

  2. to avoid conflicts and subtle bugs we should create third type-level list, next to TransformerCfg and TransformerFlags e.g.

    sealed trait TransformerFallbacks
    object TransformerFallbacks {
      final class Empty extends TransformerFallbacks
      final class Value[A, Fallbacks <: TransformerFallbacks] extends TransformerFallbacks
    }
  3. this list could be parsed before parsing configurations so that we could extract these values from val falbacks (if needed) and cache their value in vals

    1. each of fallbacks would have to have its not used warnings suppressed
  4. each val would have to create a reference to it and wrap it in ExistentialExpr to remember its type

  5. finally, these expressions would be used to initialize TransformationContext

Making fallbacks recursive

When calling deriveRecursiveTransformationExpr with a new context we are cleaning all fallbacks. From now on, we have to modify updateFromTo to take a new list of fallbacks, with .fieldName appended to each value which has such field, and removing from list each expression which doesn't have this value. Then we have to update each occurence of updateFromTo to update these values, or explicitly set it to empty (e.g. for collections and sealed traits).

Introducing policy checking

Because of Patchers existence and the way we want to implement them with merge transformers, we need at least 3 different policy options:

  • do not check fields usage (we need a catcher name) - the default value for Transformers and Partial Transformers
  • check if all fields in the main source were used - the default value for Patchers
  • check if all fields were used - would require using all filelds from source and its fallbacks

Later on we may need to add some option to add an exception, e.g. to not complain if field was unused in fallback because it existed in main source, or to implement #161 Allow to ignore specific field in patchers.

  1. we need to create a new Policy enum and add it to TransformerConfig
  2. since this is relevant only to TransformProductToProductRule, only this rule has to be updated to remember which values from source were used
  3. however we need to pay special attention because we are not distincting between vals from the class and vals from its parents and this will be relevant here, so we need to introduce that distinction in SourceType
  4. then we need to create a check working in parallel to generating each value, since we want to provide all errors at once, and if we know that not all values were used we should not prevent displaying that information if some field transformation couldn't be derived
  5. this check should only be performed if ctx.config contains the relevant policy

Rewriting Patchers

At this point Patchers implementation can be rewritten:

  1. config parsing could be left
  2. derivation for Patchers would transform PatchingContext into TransformationContext
    1. patch: Patch would become From, obj: A would become the only ExistentialExpr in fallback values, A would be used as To
    2. unless .ignoreRedundantPatcherFields was used the TransformationContext would set policy requiring using all fields from primary value
  3. PatcherMacros should override val rulesAvailableForPlatform to add:
    1. rule summoning implicit Patcher
    2. rule handling Options in Patcher in a Patcher-specific way
    3. probably remove some (all?) rules for Options used in Transformers derivation

Testing

We would have to add tests:

  • to check that uneeded fallback value doesn't produce warnings
  • to check that uneeded fallback value doesnt interfere with each field override flag
  • to check that used fallback doesn't interefere with overrides
  • to check that added policy options work and create compilation errors
  • to check that all of the above works with nested case class transformation with several values
    • including AnyVals on the path
  • to check that Patchers work predicatably in nested patching

Things to consider

  • should TransformImplicitRule attempt to summon for fallback values if it fail for main value?

  • should we add a flag for recursive fallback values usage? this would let us have a flag patching as default (keeping code backward compatible) and enabling recursive patching explictly

  • next step should we create specialized extension methods for various arities of tuples?:

    (a, b, c).mergeInto[D] // = a.into[D].withFallback(b).withFallback(c)
      .transform

    it looks useful but if we start adding such helpers for each such feature what will happen if we e.g. want to merge 3 case classes via custom constructor? where is the line?

Nested paths in overrides

General idea of the solution

Currently when using .withFieldConst, .withFieldComputed, .withFieldRenamed and their partial counterparts, we have a limitation that each path must be in form of either:

  • value => value.field
  • _.field

If the rename/value provision would have to happen in nesting, we are forced to create an implicit, so that it would be summoned inside the outer value transformation.

To address this we could:

  • replace FieldName <: String in DSL with some path type, e.g:

    sealed trait PathElement
    object PathElement {
      final class Field[Name <: String] extends PathElement
    }
    
    sealed trait Path
    object Path {
      final class Value[E <: PathElement] extends Path
      final class In[E <: PathElement, P <: Path] extends Path
    }

    (a breaking change!)

  • transformation context and recursive derivation would use non-empty list of paths and match on a content of a single-element path rather than String directly

Necessary steps

Modifying DSL

  1. we would have to introduct some path type e.g.

    sealed trait PathElement
    object PathElement {
      final class Field[Name <: String] extends PathElement
    }
    
    sealed trait Path
    object Path {
      final class Value[E <: PathElement] extends Path
      final class In[E <: PathElement, P <: Path] extends Path
    }
  2. we would have to replace FieldName <: String in TransformerCfg

  3. we would have to update whitebox macros in both Scala 2 and 3

Modifying derivation

  1. we would have to update TransformationContext to have a non-empty list of paths rather than String
  2. ProductToProductRule would have to compare not all values in fieldOverrides but only those which contain a single value (meaning we are at the same nesting level)
  3. new Context creation would have to drop outermost layer of overrides and filter out the remaining overrides checking if they override the field that is currently transformed

Testing

We could have to add tests:

  • to check that arbitrarily number of nesting works
  • to check that several overrides, sharing the same prefix work and that overrides don't propagate to where they shouldn't go

Things to consider

  • a future steps we could add similar operations like QuickLens have to support overriding a particular value in collecion/each value in a collection or in a particular subtype
    • this however, would require us to change the structure of field- and corproductOverrides because they could be interleaved
    • overrides of particular values in collection could only support build-in collections (too much of a pain to support everything) and they could reaaaally complicate the collections code
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment