GHC 8.6.5 takes minutes to compile Agda's Parser.hs

[andreasabel]

It is known that the .hs files generated by happy 2.0 can be 25% larger than the ones generated by happy 1.20. This is e.g. the case for Agda's Parser.hs generated from https://github.com/agda/agda/blob/7ea5bfa3d6bd374a2f4a0c40534a515b399787a9/src/full/Agda/Syntax/Parser/Parser.y .

However, we are also seeing a compile-time deterioration for some GHCs, e.g. 8.6.5.

Anecdotally, GHC 8.6.5 pauses for several minutes on my machine when type-checking the happy-2.0 generated Parser.hs. (Type-checking this is usually instantaneous.)

$ ./tc.sh -w ghc-8.6.5
Build profile: -w ghc-8.6.5 -O1
In order, the following will be built (use -v for more details):
 - Agda-2.8.0 (lib:Agda, exe:agda, exe:agda-mode) (first run)
...
[ 67 of 434] Compiling Agda.Utils.Monad ( src/full/Agda/Utils/Monad.hs, nothing )
[ 68 of 434] Compiling Agda.Utils.ListT ( src/full/Agda/Utils/ListT.hs, nothing ) [Agda.Utils.Monad changed]
[ 69 of 434] Compiling Agda.Utils.FileName ( src/full/Agda/Utils/FileName.hs, nothing ) [Agda.Utils.Monad changed]
[ 85 of 434] Compiling Agda.Utils.Benchmark ( src/full/Agda/Utils/Benchmark.hs, nothing ) [Agda.Utils.Monad changed]
[122 of 434] Compiling Agda.TypeChecking.Free.Lazy ( src/full/Agda/TypeChecking/Free/Lazy.hs, nothing ) [Agda.Utils.Monad changed]
[148 of 434] Compiling Agda.Syntax.Parser.Helpers ( src/full/Agda/Syntax/Parser/Helpers.hs, nothing ) [Agda.Utils.Monad changed]
[149 of 434] Compiling Agda.Syntax.Parser.Parser ( dist-no-code/build/x86_64-osx/ghc-8.6.5/Agda-2.8.0/build/Agda/Syntax/Parser/Parser.hs, nothing ) [Agda.Utils.Monad changed]
(5 min wait here)
```

[andreasabel]

I think I was too quick to blame 2.0 here. It is also slow with 1.20.1.1. Seems to be a GHC problem that was likely fixed in 8.8:

• https://gitlab.haskell.org/ghc/ghc/-/issues/14683

Similar issue:

• GHC takes a very long time to compile parser generated by Happy 1.19.8 compared to Happy 1.19.5 #109

[sgraf812]

Yes, that's the issue I had in mind as well. Good thing it's not entirely on 2.0

Out of curiousity: You say

It is known that the .hs files generated by happy 2.0 can be 25% larger than the ones generated by happy 1.20.

That seems surprising to me, because it's not like we changed the generation of the automaton. Do you mean .hs file size of .o file size? Could you perhaps look at a diff and give a qualitative assessment which parts have become larger?

[andreasabel]

So for -acg the generated .hs increased from 830K to 1050K, and for no options from 590K to 940K. I attach these products if you want to dig deeper.
Parser.zip

[sgraf812]

Thanks. I now recall why the parser is larger. It's the result of #266 (silly me) and I don't think we should go back.

Ironically, the biggest array in there is still happyExpList, which ends up as a dead binding because Agda doesn't use the expected token list feature. I have plans for getting rid of this array in #271 which will hopefully land with #272/#283 (which are essentially the same patch).