xref uses wrong directory on multi-package cabal new-build projects

[rvl]

Thanks for the excellent emacs mode.

I have a project with the following structure:

/home/rodney/dev/myproject
├── .dir-locals.el
├── cabal.project
├── pkga
│   ├── pkga.cabal
│   ├── Module1.hs
│   └── Module2.hs
└── pkgb
    ├── pkgb.cabal
    └── Main.hs

Dante chooses the new-build method. I have configured dante-project-root to be nil, so it gets set automatically to the directory of cabal.project, which is /home/rodney/dev/myproject/.

I am editing pkga/Module1.hs and have set dante-target to pkga.

When I run xref-find-definitions on a symbol, I get taken to the non-existent file /home/rodney/dev/myproject/Module2.hs. The package directory pkga is missing from this path. It has got a path relative to pkga.cabal and appended it to the dante-project-root, which is the location of cabal.project.

My workaround is to customize the new-build entry of dante-methods-alist, and set the dominating file predicate to (lambda (d) (directory-files d t ".cabal$")). Cabal v2-build correctly locates the cabal.project and everything works. This means that the bare-cabal method will never be chosen, but that is not a problem for me.

[jyp]

Thanks for the report. Unfortunately it is unclear to me what should be changed on the emacs side, it seems to be more of a bug in the GHC side.

[sboosali]

(nvm)

[jyp]

Ok, I suppose the bug is in cabal, which loads GHC from another directory than the one it was started in, causing GHC to report wrong relative paths.

[bezirg]

I ran into the exact same problem, this time using new-impure-nix dante method.

Suggestion: wouldn't this problem be fixed if dante.el prepends the the paths returned from cabal usually beginning src/...

I tried this and it seems to work in my case of new-impure-nix:

(defun dante--match-src-span (string)
  "Extract a location from a ghc span STRING."
  ;; On external symbols, GHC may return a location such as integer-gmp-1.0.0.1:integer-gmp-1.0.0.1:GHC.Integer.Type
  (when (string-match "\\(.*?\\):(\\([0-9]+\\),\\([0-9]+\\))-(\\([0-9]+\\),\\([0-9]+\\))$" string)
    (let ((file (match-string 1 string))
          (line (string-to-number (match-string 2 string)))
          (col (string-to-number (match-string 3 string))))
      (xref-make-file-location
       (or (gethash file dante-original-buffer-map)
           (expand-file-name file (expand-file-name dante-target dante-project-root)))
       line (1- col)))))

I don't know however how it will affect people that do not rely on dante-target set in .dir-locals.el

[deepfire]

So, there are following entities:

1. Project root -- the absolute root that contains all packages with their .cabal files. Often contains the cabal.project file.
2. Packages -- the set of arbitrary subdirectories, each containing a .cabal file.

AIUI, what xref is handed over is something relative to №2, whereas what we want it to have is a project-root-relative path of a file, which really is:

projectRelativePath = packageSubpath </> packageRelativeSubpath

Next question is -- does Dante know the packageSubpath (i.e. the subdir that contains the .cabal file) for any given source file?