Added 8.2.1-binary

enableShared in generic-builder.nix should default to what the GHC compiler was compiled with.

Why should it?

Why should it?

My understanding is that if your compiler wasn't compiled with "enableShared", setting "enableSharedLibraries" to true, will break things.

enableSharedLibraries ? (ghc.isGhcjs or false) || stdenv.lib.versionOlder "7.7" ghc.version)

is always true today. I gather this is because old GHC didn't support dynamic linking, so that reads "true where possible". If anyone has a custom compiler that doesn't support dynamic linking, they can do passthru = { enableShared = false; }; just like with the prebuilt ones. If we want to add additional conditions we can do

enableSharedLibraries ? ghc.enableShared or false && additionalCondition

The point being deduplicating the logic like this doesn't tie our hands.

My understanding is that if your compiler wasn't compiled with "enableShared", setting "enableSharedLibraries" to true, will break things.

I see an argument to be made that you'd want things to break in that case. The enableShared default in the generic builder can be viewed as expressing a policy that compilers ought to fulfill and if someone uses the generic builder with a compiler that does not fulfill that policy then you ought to pass a proper enableShared setting along your custom compiler. That approach makes these choices more explicit, whereas the new scheme that @Ericson2314 chose to merge without waiting for my review makes this choice implicit.

Feel free to revert if you disagree. This was motivated by the need to have "enableSharedLibraries" set to false in prebuilt Android & iOS (which is the default for their config) and they are broken otherwise. I felt like this approach would avoid the most duplication but there are certainly other ways to accomplish this. If you do want to revert I hope it is okay to add an extra condition to enableSharedLibraries in generic-builder.nix.

Yeah we can revert this in the interim, but let me continue to argue for this approach.

@peti so first of all, are you OK with us adding the cross exceptions to the generic builder default? If we do, then the default becomes a verbose condition that doesn't to me express a policy other than "build shared libraries where possible". (Any other interpretation would strike me as an "overfit" one.) If we don't, then users must manually instruct generic-builder even when they don't actually care and just want working builds.

And even in the leave-as-is case, I still read the default that way. It makes no sense that below a certain version most people just don't want shared libraries. It makes far more sense that most people always want shared libraries, but also have static libraries than none at all. This PR exactly formalizies that second generalization of people's desires with the new default.

Oh, a change we could make is flip the or false to or true in case people forget the passthru, or just remove it altogether to require people to so-annotate their custom compilers.

I do see how a missing attribute causing static builds is a subtle and not-declarative-in-spirit default.

I just would have liked to have this conversation before the merge, not after. Anyway, I don't feel strongly enough about this issue to bother reverting those changes now. Both schemes have their advantages, and the state we have now feels just fine.