ghc 9.2.7, ghc 9.6.1 and replace old-ghc-nix with nixpkgs ghc

[hamishmack]

This PR started out adding GHC 9.6.1 (alpha). To make it now also switches the boot compilers from old-ghc-nix (which is unmaintained) to use the native compilers in nixpkgs. This means we can boot the haskell.nix versions of the compilers with more recent GHC versions. Hopefully this will allow us to stop maintaining support for old compilers (like 8.4.4 and 8.6.5). It should also reduce the depth of GHC recompilation that results from changes in haskell.nix.

Now 9.2.7 is available we are also adding that to this PR.

Once this is merged, let's merge

• gold only on linux, !android, !musl #1875
• ghc-lib-parser-9.6: add global unique counters patch #1876
• Add better support for armv7a cross compilation #1874
• Fetch the revised cabal files as part of elaborating hackage.nix #1860

[dhess]

Hi, thanks for all your hard work on this. There's a fix here for evaluating this branch with the nixpkgs-unstable pin: #1878

[amesgen]

Cross compilers (e.g. musl and mingw) are no longer be built after this change, probably due to the change in flake.nix:

diff --git a/flake.nix b/flake.nix
index 9b23b26f..f6df9ee8 100644
--- a/flake.nix
+++ b/flake.nix
@@ -68,7 +68,7 @@
   };
 
   outputs = { self, nixpkgs, nixpkgs-unstable, nixpkgs-2105, nixpkgs-2111, nixpkgs-2205, nixpkgs-2211, flake-utils, tullia, ... }@inputs:
-    let compiler = "ghc926";
+    let compiler = "ghc927";
       config = import ./config.nix;
 
       traceNames = prefix: builtins.mapAttrs (n: v:
@@ -163,13 +163,13 @@
           ci = import ./ci.nix { inherit (self.internal) compat; inherit system; };
         in stripAttrsForHydra (filterDerivations ci);
 
-      ciJobs =
+      requiredJobs =
         let
           inherit (legacyPackages) lib;
           names = x: lib.filter (n: n != "recurseForDerivations" && n != "meta")
               (builtins.attrNames x);
-          requiredJobs =
-            builtins.listToAttrs (
+        in
+          builtins.listToAttrs (
               lib.concatMap (nixpkgsVer:
                 let nixpkgsJobs = allJobs.${nixpkgsVer};
                 in lib.concatMap (compiler-nix-name:
@@ -185,9 +185,8 @@
                    }) (names ghcJobs))
                 ) (names nixpkgsJobs)
               ) (names allJobs));
-        in lib.optionalAttrs (system == "x86_64-linux") {
-          latest = allJobs.unstable.ghc8107.native or {};
-        } // requiredJobs;
+
+      ciJobs = allJobs;
 
       hydraJobs = ciJobs;

Before this PR:

 $ nix eval --json .#hydraJobs --apply __attrNames
["latest","required-R2205-ghc8107-native","required-R2205-ghc865-native","required-R2211-ghc8107-native","required-R2211-ghc865-native","required-R2211-ghc902-native","required-R2211-ghc926-native","required-R2211-ghc944-native","required-unstable-ghc8107-aarch64-multiplatform","required-unstable-ghc8107-ghcjs","required-unstable-ghc8107-mingwW64","required-unstable-ghc8107-musl64","required-unstable-ghc8107-native","required-unstable-ghc865-ghcjs","required-unstable-ghc865-native","required-unstable-ghc884-ghcjs","required-unstable-ghc884-native","required-unstable-ghc902-aarch64-multiplatform","required-unstable-ghc902-mingwW64","required-unstable-ghc902-musl64","required-unstable-ghc902-native","required-unstable-ghc926-aarch64-multiplatform","required-unstable-ghc926-mingwW64","required-unstable-ghc926-musl64","required-unstable-ghc926-native","required-unstable-ghc944-mingwW64","required-unstable-ghc944-native"]

Note that this contains entries such as "required-unstable-ghc926-musl64".

After this PR:

 $ nix eval --json .#hydraJobs --apply __attrNames
["R2205","R2211","meta","unstable"]
~/git/iog/haskell.nix on  master
 $ nix eval --json .#hydraJobs.unstable --apply __attrNames
["ghc8107","ghc902","ghc927","ghc944","ghc96020230302","ghc961","meta","recurseForDerivations"]
~/git/iog/haskell.nix on  master
 $ nix eval --json .#hydraJobs.unstable.ghc927 --apply __attrNames
["meta","native","recurseForDerivations"]

Only "native" GHCs are built.

[angerman]

@amesgen thabks for the report. This PR ended up being way too large :-( we will focus on speeding up CI so that this kinda of PRs don't end up taking 1-2days per iteration :-/