Missing Inferred Type in Tooltip when using Visible Type Applications

[julmb]

Your environment

Which OS do you use?
ArchLinux

Which version of GHC do you use and how did you install it?
9.6.4 from ghcup

How is your project built (alternative: link to the project)?
cabal

Which LSP client (editor/plugin) do you use?
Code - OSS + vscode-haskell

Which version of HLS do you use and how did you install it?
2.6.0.0 from ghcup

Have you configured HLS in any way (especially: a hie.yaml file)?

"haskell.manageHLS": "PATH",
"haskell.plugin.importLens.codeActionsOn": false,
"haskell.plugin.class.codeActionsOn": false,
"haskell.plugin.semanticTokens.globalOn": true,

Steps to reproduce

Hovering over print in this example:

main = (print :: Int -> IO ()) 0

Hovering over print in this example:

main = print @Int 0

Expected behaviour

The inferred type for print is shown in the tooltip in both examples.

Actual behaviour

The inferred type for print is only shown with type annotations via ::, but not with visible type applications via @. The inferred type is shown correctly without any type annotations (together with a warning about defaulting the type variable to Integer).

Debug information

I could not find anything relevant in the HLS output log either when typing the expression or when hovering to produce the tooltip.

[jhrcek]

I find this curious and managed to reproduce the issue (with latest hls + ghc 9.4.8 on fedora)

I narrowed down somewhat that the issue is observable in the vicinity of the following code:
The types info is generated from the list of types retrieved from HieAST here:

haskell-language-server/ghcide/src/Development/IDE/Spans/AtPoint.hs

Line 293 in a29d8e8

types = nodeType info

When type application is not used, this usually contains 2 values (most general type and currently inferred type(?))
But when type application is present, this list of types is empty.
But I have no idea why. If only there was an easy way to pretty-print hie ast around the point of interest to see what's going on. I might look into this over the weekend just to learn more deeply how this stuff works..

[soulomoon]

They are pretty dependent on the hieAst. Since we are hover only on the print.
The print @Int as a whole has types: [5]. Actually, the type follows the evidence veriable $dShow_xxx in the hieAst

we can see it from the HIeAst.

main = (print :: Int -> IO ()) 0 we have

[(0, ()), (1, IO), (2, Int), (3, Show), (4, 'Many),
 (5, HFunTy 4 2 1), (6, Type), (7, HTyVarTy a), (8, Show),
 (9, HFunTy 4 7 1), (10, HQualTy 8 9),
 (11, HForAllTy ((a, 6), [spec]) 10)]
 
            Node@hello.hs:5:9-13: Source: From source
                                  {(annotations: {(HsVar, HsExpr), (XExpr, HsExpr)}), 
                                   (types: [5, 11]), 
                                   (identifier info: {(name print,  Details:  Just 11 {usage}), 
                                                      (name $dShow_aRb,  Details:  Just 3 {usage of evidence variable})})}

main = print @Int 0 we have

[(0, ()), (1, IO), (2, Int), (3, Show), (4, 'Many),
 (5, HFunTy 4 2 1), (6, Type), (7, HTyVarTy a), (8, Show),
 (9, HFunTy 4 7 1), (10, HQualTy 8 9),
 (11, HForAllTy ((a, 6), [spec]) 10)]

        Node@hello.hs:5:8-17: Source: From source
                              {(annotations: {(HsAppType, HsExpr), (XExpr, HsExpr)}), 
                               (types: [5]), 
                               (identifier info: {(name $dShow_aGC,  Details:  Just 3 {usage of evidence variable})})}
                              
          Node@hello.hs:5:8-12: Source: From source
                                {(annotations: {(HsVar, HsExpr)}),  (types: [11]), 
                                 (identifier info: {(name print,  Details:  Just 11 {usage})})}
                                
          Node@hello.hs:5:15-17: Source: From source
                                 {(annotations: {(HsTyVar, HsType)}),  (types: []), 
                                  (identifier info: {(name Int,  Details:  Nothing {usage})})}

[jhrcek]

Nice investigation @soulomoon I'd be interested how you got the printout of the relevant ranges.
And do you have to recompile/reinstall hls just to get the "print" statements in?

[soulomoon]

You can setup ghc session and build the HieAst for a simple file @jhrcek , without the HLS.
It should speed up investigation on hieAst. I hope it helps.

    let filePath = "hello.hs"
    content <- Prelude.readFile filePath
    runGhc (Just libdir) $ do
        dflags <- getSessionDynFlags
        let dflags' = gopt_set dflags Opt_Haddock
        setSessionDynFlags dflags'
        target <- guessTarget filePath Nothing Nothing
        setTargets [target]
        load LoadAllTargets
        modSum <- getModSummary $ mkModuleName "Hello"
        parsedModule <- parseModule modSum
        typecheckedModule <- typecheckModule parsedModule
        let rms = fromJust $ renamedSource typecheckedModule
        let tcGblEnv = fst $ tm_internals_ typecheckedModule
        file <- mkHieFile modSum tcGblEnv rms
        let ast = snd $ List.head $ Map.toList $ getAsts $ hie_asts file
        let types = Arr.elems $ hie_types file
        liftIO $ Prelude.writeFile "hieAst.txt" $ showSDoc dflags' (ppr (Prelude.zip [0::Int ..] types)) ++ "\n" ++ showSDoc dflags' (ppr ast)

[jhrcek]

Awesome, thanks!
I'll play around with this snippet.
I turned your snipped into runnable script for local poking around.
It required me to declare Outputable instance for HieType to make it work.
https://gist.github.com/jhrcek/9785e1e4ae909995d06a8810e1d6d98c

This seems pretty tedious. I wonder what do the hls devs do when they want to investigate "what's in the hie ast"? Isn't there a way to pretty print it for debugging purposes? You know, to debug things like "the info I was expecting is not in this ast node, where can I get it instead?"..

[soulomoon]

Awesome, thanks! I'll play around with this snippet. I turned your snipped into runnable script for local poking around. It required me to declare Outputable instance for HieType to make it work. https://gist.github.com/jhrcek/9785e1e4ae909995d06a8810e1d6d98c

This seems pretty tedious. I wonder what do the hls devs do when they want to investigate "what's in the hie ast"? Isn't there a way to pretty print it for debugging purposes? You know, to debug things like "the info I was expecting is not in this ast node, where can I get it instead?"..

I am not sure the others, I just going through the source code of hls or ghc. it's a bit off topic here, maybe we can discuss this in matrix.

[michaelpj]

So it seems like this is a bug in HIE file generation? I would expect the print node to still have a type even with the type application.

[jhrcek]

I figured out a nicer way to show hie file (with full types reconstructed instead of with TypeIndexes) and it seems that in both cases the type info is associated with the print AST node. So not sure what leads to it not being rendered.
Will try to dig deeper over weekend.

[soulomoon]

I figured out a nicer way to show hie file (with full types reconstructed instead of with TypeIndexes) and it seems that in both cases the type info is associated with the print AST node. So not sure what leads to it not being rendered. Will try to dig deeper over weekend.

This does look better.

Take a look at the evidence variables. They are the ones with more concrete types.

[jhrcek]

Ok, I got to the bottom of this (what a fun learning experience this was 😄)

TL;DR: not showing types when TypeApplications is used seems like deliberate design - the type is not shown at the bottom of markdown block because it is already shown at the top (so having it repeated at the bottom would seem redundant)

More details:
As I commented above (refer to the screenshots) the HieAST does contain type information in both cases:

1. when explicit type annotation is used there are 2 types in the print's AST node:

• forall a. Show a => a -> IO ()
• Int -> IO ()

2. when TypeApplications is used there is only one type:

• forall a. Show a => a -> IO ()

Then there is this dropEnd1 logic which basically drops the last type from the list in cases when there's only single identifier associated with the hovered-over symbol. That logic has been implemented back in 2020 in this PR: 5dfee4c#diff-91f773bae93bbf86035573f4dbffa3035c1689b1bb7e17bde5937bb1271c0131R104
I think the reason it's there is because that last type is actually rendered at the top of the markdown block, so it seems it just "works as designed".

I hope the following screenshots from my debugging sessions (where I just Debug.Traced the relevant expressions in the linked code) help to clarify the difference in behavior more then lengthy explanations:

[michaelpj]

Shame there's no comment explaining, maybe @wz1000 remembers and could write one?

It seems like the logic we want is to not reproduce the "original" type signature. It's not clear that we're really achieving that, though; in that case we should be filtering out the original type signature from the list rather than what we're doing now. I'd be somewhat in favour of just displaying everything (ideally with a title for that section of the hover explaining what it means!), at least that's consistent.

[julmb]

This explains how the HIE AST is shown in the user interface, but I am still wondering why the AST contains only the general type forall a. Show a => a -> IO () and not the instantiated one Int -> IO ().

I was expecting HLS to behave similarly to GHCI in this case:

ghci> :type print
print :: Show a => a -> IO ()
ghci> :type print @Int
print @Int :: Int -> IO ()

In fact, I would argue that showing the instantiated signature is more important than showing the original one (which can always be looked up in the documentation or source).

Additionally, the fact that the user is using visible type applications indicates that the type inference situation is tricky and these are exactly the situations in which the user really benefits from insight into which types their expressions ultimately got assigned.

[wz1000]

This is just how the language works.

for print (1 :: Int), the typechecker has to elaborate print to print @Int. Since @Int wasn't written explicitly by the user, it is attached to the print.

If the user explicitly writes print @Int, then the typechecker doesn't need to do any elaboration, and the type of the span that print is used in is forall a. .... It would be incorrect to give that span the type Int -> ... because you can't apply a function Int -> ... to a type argument @Int.

I suppose what you want is the type of the expression print @Int, which is the feature request in #709, currently blocked on both GHC not providing us information in an easily accessible way and LSP not really supporting a reasonable way to expose this to the user.