Dependencies between the compiler phases

Each phase is necessary for successfully compiling Scala code (except
for the optimization phases). If you want to compile only a subset of
the language, you could (probably) skip some, but I am very skeptical
this works well. You can try '-Yskip:' and see for yourself.

All phases assume their place in the sequence, and handle a (subset) of
the AST or types. After erasure, since types are erased, all phases will
assume the simpler language of types, and therefore won't work as expected.

If you need to write a new code generator you can insert it before any
phase, for instance erasure (all phases produce well-typed code), and
take advantage of the extra available information (but also handle a
richer language, like inner classes).

iulian

Valentin Wüstholz wrote:
> Hi,
>
> I'm working on a new Scala backend as part of my Master's thesis at
> ETH Zürich. I already asked the same question on the "scala" list and
> people there suggested that this list might be a better place to ask
> questions like these.
>
> I was wondering which compiler phases depend on each other. Is this
> documented anywhere? I use Scala 2.7.5 for my compiler plugin.
>
> I'm particularly interested in the phases that depend on the 'erasure'
> phase. Apparently at least the 'constructors' and the 'flatten' phase
> seem to depend on it. Is this correct?
>
> What about other phases, such as lazyVals, lambdaLift, flatten, mixer
> or cleanup?
>
> Cheers,
>
> Valentin
>

Thanks! In general, we don't really mind if we can only support a
subset of the language since this is just a proof of concept
implementation.

But putting the code generator before the erasure phase would mean
that we have to implement a lot of the later phases ourselves. This
would take a lot of time besides being a maintenance nightmare.
Therefore we had hoped that a phase like erasure would be more or less
right before the JVM code generator. This should make it a lot easier
to target different platforms. After all, erasure is something quite
JVM specific. C#/.Net for instance doesn't use erasure.

Cheers,

Valentin

On Tue, Sep 8, 2009 at 5:33 PM, Iulian Dragos wrote:
> Each phase is necessary for successfully compiling Scala code (except for
> the optimization phases). If you want to compile only a subset of the
> language, you could (probably) skip some, but I am very skeptical this works
> well. You can try '-Yskip:' and see for yourself.
>
> All phases assume their place in the sequence, and handle a (subset) of the
> AST or types. After erasure, since types are erased, all phases will assume
> the simpler language of types, and therefore won't work as expected.
>
> If you need to write a new code generator you can insert it before any
> phase, for instance erasure (all phases produce well-typed code), and take
> advantage of the extra available information (but also handle a richer
> language, like inner classes).
>
> iulian
>
> Valentin Wüstholz wrote:
>>
>> Hi,
>>
>> I'm working on a new Scala backend as part of my Master's thesis at
>> ETH Zürich. I already asked the same question on the "scala" list and
>> people there suggested that this list might be a better place to ask
>> questions like these.
>>
>> I was wondering which compiler phases depend on each other. Is this
>> documented anywhere? I use Scala 2.7.5 for my compiler plugin.
>>
>> I'm particularly interested in the phases that depend on the 'erasure'
>> phase. Apparently at least the 'constructors' and the 'flatten' phase
>> seem to depend on it. Is this correct?
>>
>> What about other phases, such as lazyVals, lambdaLift, flatten, mixer
>> or cleanup?
>>
>> Cheers,
>>
>> Valentin
>>
>
>
> --
> Iulian Dragos
> =============
> EPFL-IIF-LAMP 1
> INR 321 http://lamp.epfl.ch/~dragos
>
>