Slow codec automatic derivation

[andreisilviudragnea]

When comparing codec automatic derivation with semi-automatic derivation (explicit or using JsonCodec), automatic derivation is very slow.

In each file which contains import io.circe.generic.auto._, if there is an expression like Decoder.apply[SomeType] (or any circe call which needs an implicit Decoder[T] instance), an automatically derived decoder will be generated for each implicit call site (DerivedDecoder.deriveDecoder gets called for each Decoder.apply[SomeType] call).

I have reproduced the bug here:

• CirceGenericAutoSpec.scala typer phase takes 6.9 seconds
• CirceGenericJsonCodecSpec.scala typer phase takes 114 ms
• CirceGenericSemiAutoSpec.scala typer phase takes 102 ms

The number of class files generated by CirceGenericAutoSpec.scala is huge compared to the other two examples.

Given how inefficient automatic derivation can be in these cases, I think this behaviour should be documented here.