libgslcblas link issue past GHC 9.6

[tmplt]

Some time during the last eight months, hmatrix-gsl seems to have broken in nixpkgs: if master or unstable is used, a undefined symbol: cblas_ctrmv will result during run-time. This is most likely an upstream issue, but I figured a first step was to pin a working derivation in my Haskell library, so that downstream users don't have to match an eight-month old nixpkgs. But I haven't had much luck with that.

I've created an MWE via https://github.com/tmplt/hmatrix-gsl-mwe

$ nix build github:tmplt/hmatrix-gsl-mwe # build should succeed

$ nix build github:tmplt/hmatrix-gsl-mwe --override-input nixpkgs github:nixos/nixpkgs/nixpkgs-unstable
warning: not writing modified lock file of flake 'github:tmplt/hmatrix-gsl-mwe':
• Updated input 'nixpkgs':
    'github:nixos/nixpkgs/d9b69c3ec2a2e2e971c534065bdd53374bd68b97?narHash=sha256-ATMHHrg3sG1KgpQA5x8I%2BzcYpp5Sf17FaFj/fN%2B8OoQ%3D' (2025-02-24)
  → 'github:nixos/nixpkgs/0bd3fe4080c5d2bc39810d05d9fd8780985e773b?narHash=sha256-5aLE9L0Q1ErEdNYCt4K1oHR7GnNglHVJJvfMTCDdLMQ%3D' (2025-10-13)
error: build of '/nix/store/jdsxzb3ybwql6nxznqr56a7gg5ajqh55-mwe-0.0.0.0.drv' on 'ssh://root@cpn1' failed: builder for '/nix/store/jdsxzb3ybwql6nxznqr56a7gg5ajqh55-mwe-0.0.0.0.drv' failed with exit code 1
error: builder for '/nix/store/jdsxzb3ybwql6nxznqr56a7gg5ajqh55-mwe-0.0.0.0.drv' failed with exit code 1;
       last 25 log lines:
       > Using strip version 2.44 found on system at:
       > /nix/store/dmypp1h4ldn0vfk3fi6yfyf5yxp9yz0k-gcc-wrapper-14.3.0/bin/strip
       > Using tar found on system at:
       > /nix/store/n09s57l86zxyjydkxagawxpkhh3d3dlh-gnutar-1.35/bin/tar
       > No uhc found
       > Running phase: buildPhase
       > Preprocessing library for mwe-0.0.0.0...
       > Building library for mwe-0.0.0.0...
       > [1 of 1] Compiling MWE              ( src/MWE.hs, dist/build/MWE.o, dist/build/MWE.dyn_o )
       > Preprocessing test suite 'doctests' for mwe-0.0.0.0...
       > Building test suite 'doctests' for mwe-0.0.0.0...
       > [1 of 1] Compiling Main             ( tests/doctests.hs, dist/build/doctests/doctests-tmp/Main.o )
       > [2 of 2] Linking dist/build/doctests/doctests
       > Running phase: checkPhase
       > Running 1 test suites...
       > Test suite doctests: RUNNING...
       > [ERROR  ] [ThreadId 5] /build/mwe-0.0.0.0/src/MWE.hs:9: failure in expression `mwe'
       > [ERROR  ] [ThreadId 5] expected: [10.0]
       > [ERROR  ] [ThreadId 5]  but got: /nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgsl.so: undefined symbol: cblas_ctrmv
       > [ERROR  ] [ThreadId 5]           ^
       >
       > [INFO   ] [ThreadId 4] Examples: 1  Tried: 1  Errors: 0  Unexpected output: 1
       > Test suite doctests: FAIL
       > Test suite logged to: dist/test/mwe-0.0.0.0-doctests.log
       > 0 of 1 test suites (0 of 1 test cases) passed.
       For full logs, run 'nix log /nix/store/jdsxzb3ybwql6nxznqr56a7gg5ajqh55-mwe-0.0.0.0.drv'.

In short: how can I override the gsl input to hmatrix-gsl so that it is pinned to a known working version?
Or is the root issue elsewhere? I'm bumping from GHC 9.6.6 to 9.10.3, after all.

This may not be a haskell-flake issue at all, but I figure I'll start here.

[tmplt]

In normal fashion I find a solution directly after asking for help: basePackages = pkgs.haskell.packages.ghc96; fixes the run-time error.

Is this the correct way to fix the issue?

Downstream (in a nix shell) I replace pkgs.haskellPackages.ghcWithPackages (_: [ mwe ]) with pkgs.haskell.packages.ghc96.ghcWithPackages (_: [ mwe ])

[srid]

Does this help? https://community.flake.parts/haskell-flake/dependency#path

[tmplt]

That results in the same error.

Applying the above broad-stroke fix gives us:

$ nix build && ldd result/lib/ghc-9.6.7/lib/x86_64-linux-ghc-9.6.7/libHSmwe-0.0.0.0-FGQA6RBdeWgBkoX90i0E3I-ghc9.6.7.so | grep libgsl
        libgsl.so.28 => /nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgsl.so.28 (0x00007fce27200000)
        libgslcblas.so.0 => /nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgslcblas.so.0 (0x00007fce293a5000)

but removing the basePackages override:

$ nix build
[snip]
> [ERROR  ] [ThreadId 6]  but got: /nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgsl.so: undefined symbol: cblas_ctrmv
[snip]

So a known working gsl (the very same derivation) is indeed used.
I wonder where this linking issue comes from.

It is reproducible with basePackages = pkgs.haskell.packages.ghc98;.

On default basePackages:

$ nix develop
$ cabal build
$ ldd dist-newstyle/build/x86_64-linux/ghc-9.10.3/mwe-0.0.0.0/t/doctests/build/doctests/doctests | grep libgsl
        libgsl.so.28 => /nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgsl.so.28 (0x00007ffff7a00000)
        libgslcblas.so.0 => /nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgslcblas.so.0 (0x00007ffff7e90000)

Just as with 9.6.

And just as the error says, the symbol is undefined in libgsl, but it is availabli in libgslcblas; thus:

$  LD_PRELOAD=/nix/store/ig9f31z5v8scrbfl56bl2kcl3f49x2j8-gsl-2.8/lib/libgslcblas.so  dist-newstyle/build/x86_64-linux/ghc-9.10.3/mwe-0.0.0.0/t/doctests/build/doctests/doctests
[INFO   ] [ThreadId 4] Examples: 1  Tried: 1  Errors: 0  Unexpected output: 0

Is this a cabal issue?

[tmplt]

I don't think this is a cabal issue. I've updated the MWE to include working-mwe and broken-mwe with their respective devShells:

haskellProjects = rec {
  broken = {
    projectRoot = builtins.toString (lib.fileset.toSource {
      root = ./.;
      fileset = lib.fileset.unions [ ./src ./tests ./mwe.cabal ];
    });

    autoWire = [ "packages" "devShells" ];
  };

  working = broken // { basePackages = pkgs.haskell.packages.ghc96; };
};

Both builds use cabal 3.16.0.0 (but different build hashes).
Via the Haskell Discourse I was recommended to explicitly toggle the onlygsl flag, but that did not yield different results.
This feels more like a Nix issue, but I'm unsure how do dig deeper.
Do you have any recommendations, @srid?

[tmplt]

Some experimentation led to tmplt/hmatrix-gsl-mwe@4e764e7: I cannot reproduce when building an executable rather than a test-suite.

I tried adding an explicit call to the MWE and see if it did anything: tmplt/hmatrix-gsl-mwe@ef7e3f5. Here, the explicit call works as expected, but the mainFromCabal call still fails.

Perhaps this is an issue with doctest-parallel?