Unboxed tuple error when importing a module with `no-specialise`

[michaelpj]

I have modules A and B. A imports B, and B has OPTIONS_GHC -fno-specialise. ghcide gives me the following error on A (but not B). The error goes away if I remove the no-specialise on B.

  Program error: Error: bytecode compiler can't handle unboxed tuples and sums.
  Possibly due to foreign import/export decls in source.
  Workaround: use -fobject-code, or compile this module to .o separately.

I remembered this because I saw Ben Gamari report something similar on IRC, not sure if it's the same.

[michaelpj]

I think worker/wrapper can produce unboxed tuples, which will probably usually be specialized away. So I can see how this would happen if we're not compiling B with reduced enough optimization settings to prevent that.

[pepeiborra]

Just for my understanding, do either B or A contain any TH?

ghcide is certainly trying to disable optimisation, but maybe the setting is getting overridden later on:

https://github.com/digital-asset/ghcide/blob/21161389842a3141f82c82a6757d6eadfb264e8b/exe/Main.hs#L666

[michaelpj]

They both do. Not sure if it matters!

[pepeiborra]

They do. Not sure if it matters!

Why would ghcide go through the trouble of generating byte code otherwise?

[ndmitchell]

So hypothesis is we compile the module to byte code to generate TH. But something interferes and turns the optimisation up sufficiently high that the bytecode contains unboxed tuples? @michaelpj - do you have any -O2 or similar pragmas either in files, or in the Cabal/Stack from which you build? If so, does removing those and restarting fix it?

[michaelpj]

Not AFAICT. I'll experiment some more and see if anything affects it.

One thing I tried to try and pin it down:

• Set optimization: 0 in cabal.project
• Load the module up in cabal repl

That works fine, which I would have thought corresponds to the ghcide situation.

[ndmitchell]

Ghcide and Cabal Repl are about a million miles apart nowadays - entirely feasible for it to work with one but not the other. Is it likely we can get a minimal reproducer?

[michaelpj]

Sure, I just wanted to establish that it's not Totally Broken but only broken in ghcide (and therefore plausible that this is a bug rather than me making a mistake).

I'll try and boil this down, not sure when I'll get to it.

[michaelpj]

Okay, I definitely need to boil this down, it seems to be non-deterministic too :(

[michaelpj]

Which probably invalidates my claim that removing no-specialise helps!

[ndmitchell]

If you can reproduce this on the command line with ghcide running on a file without an IDE that would be easier to debug - might be easier to reduce as well.