`java.sql.Driver` missing from classpath in macro in presentation compiler

[kasiaMarek]

Original issue: https://github.com/scalameta/metals/issues/6494

Compiler version

3.4.2

Minimized code

//> using scala "3.4.2"
//> using dep "com.h2database:h2:2.2.224"

import scala.quoted._

class LoadStuff extends Selectable:
  def selectDynamic(name: String): Any = name

object LoadStuff:
  transparent inline def make = ${ makeImpl }

  private def makeImpl(using Quotes): Expr[Any] =
    import quotes.reflect.*

    Class.forName("org.h2.Driver") // problematic line

    val refinement = Refinement(TypeRepr.of[LoadStuff], "name", TypeRepr.of[String])

    refinement.asType match
      case '[refined] => '{ LoadStuff().asInstanceOf[refined] }

Output

Interactive compiler reports error diagnostics:

java.lang.ClassNotFoundException: java.sql.Driver
        at java.base/java.net.URLClassLoader.findClass(URLClassLoader.java:445)
        at java.base/java.lang.ClassLoader.loadClass(ClassLoader.java:593)
        at java.base/java.lang.ClassLoader.loadClass(ClassLoader.java:526)
        at java.base/java.lang.Class.forName0(Native Method)
        at java.base/java.lang.Class.forName(Class.java:421)
        at java.base/java.lang.Class.forName(Class.java:412)
        at a.LoadStuff$.makeImpl(LoadStuff.scala:13)
        at a.LoadStuff$.inline$makeImpl(LoadStuff.scala:10)

Expectation

Should compile fine and diagnostics should be empty.

[ncreep]

Hi,

Just adding some further info from the original issue.

Thinking about it, this can be further minimized, as this has nothing to do with Selectable. Apparently just being a transparent inline with class loading is enough to trigger this behavior:

object LoadStuff:
  transparent inline def make: Unit = ${ makeImpl }

  private def makeImpl(using Quotes): Expr[Unit] =
    Class.forName("org.h2.Driver")

    '{()}

This too triggers a loss of type-information at the call-site of make.

Removing transparent brings back type information.

Thanks

[ncreep]

Hi,

I'd like to take a stab at figuring out this issue. Can anyone please point me where to start looking in the repo?

Thanks

[jchyb]

Hello! Macro methods (the previously compiled .class code run on runtime) are called from the compiler/src/dotty/tools/dotc/quoted/Interpreter.scala class. There, a separately constructed classLoader is used to run that previously compiled code (the Interpreter name might be misleading, the only thing that is being interpreted is the ${ makeImpl } part, then makeImpl is run via invoking a method from a classfile). Since that class loader is constructed separately I suspect it might be the thing causing the issue. Previously we already had a problem with it in the cli, caused by the fact that calling java -cp SomeClass:compilerClasses -classPath OtherClasses replMain would not join SomeClass with OtherClasses in that class loader (but worked for non macro compilation). I do not know much about the presentation compiler, so I cannot help with that aspect unfortunately. Making any progress here or observation would be of great help, I recall taking a stab at this some time ago but not being able to make any progress

[ncreep]

Thank you for the detailed response!

Is there a guide on how to develop and diagnose the presentation compiler with Metals? E.g.:

• How do I see the exception in the issue description? It doesn't seem to be visible as a regular Metals user.
• How do I plug my local Metals setup to use a local version of the compiler?

Thanks

[tgodzik]

• How do I see the exception in the issue description? It doesn't seem to be visible as a regular Metals user.

That should be visible in .metals/reports or in .metals/metals.log

• How do I plug my local Metals setup to use a local version of the compiler?

You only have to publish locally the presentation compiler module, it should be picked up automatically.

[ncreep]

Thanks for the prompt response.

Trying to reproduce the bug now, and it's still missing type information, but also, I'm not seeing anything in the logs at all (only successful compilation).

In the original issue I reported in the Metals repository, I did get an exception visible somewhere, though I don't recall now where I got it from.

But even back then, as can be seen in the discussion there, I wasn't seeing the actual exception thrown by the presentation compiler, but rather some subsequent exception coming from Metals. It was mentioned there by Kasia, that it should be possible to get diagnostics directly from the presentation compiler. Can you please point me to where these are visible?

Thanks

[tgodzik]

Curious, I can't seem to reproduce either:

https://github.com/user-attachments/assets/2a02d862-c70f-4b2f-948e-25f398cbbe8e

The only issue I see is that no hover information is shown in some places, but not actual exception. We could try to fix that and use Hover tests for that.

[kasiaMarek]

Maybe it got somehow fixed or maybe there was some fallback in metals added, so there is no error visible. The error I posted was only visible when logging diagnostics from pc compile run, it did not surface.

[ncreep]

Although I've no idea where the exception has gone, the full issue at the usage site is still present:

• No hover
• No autocompletion
• No "insert type annotation"

Which I would imagine indicates that the presentation compiler decided to quit at this point.

[ncreep]

Also, how can I enable diagnostics from the presentation compiler?

Thanks

[ncreep]

Playing a bit more with this issue. It seems that there's a definite improvement in 3.5.x compared to 3.4.2.

In 3.4.2 the whole usage file completely loses type information in Metals. In 3.5.x it's only the results of the transparent inline that lose information, the rest of the file continues to function.

Maybe best effort compilation catches and suppresses errors mid way?

[kasiaMarek]

Also, how can I enable diagnostics from the presentation compiler?

If you are expecting Metals to display those diagnostics it won't. You have to change the code. Basically after type checking in run on compilation unit in pc you can call diagnostics (or something similar) method, and log that information. This is as far as I remember from the top of my head, I'd have to check for details.

[ncreep]

@kasiaMarek, got it, thanks.