This makes it work like work-on-multi from Reflex Platform. In
particular, rather than making `.env` from `shellFor`, we make `.env`
the primitive, and `shellFor` works by combining together the arguments
of all the packages to `generic-builder` and taking the `.env` of the
resulting mashup-package.
There are 2 benefits of this:
1. The dependency logic is deduplicated. generic builder just concatted
lists, whereas all the envs until now would sieve apart haskell and
system build inputs. Now, they both decide haskell vs system the same
way: according to the argument list and without reflection.
Consistency is good, especially because it mean that if the build
works, the shell is more likely to work.
2. Cross is handled better. For native builds, because the
`ghcWithPackages` calls would shadow, we through both the regular
component (lib, exe, test, bench) haskell deps and Setup.hs haskell
deps in the same `ghcWithPackages` call. But for cross builds we use
`buildPackages.ghcWithPackages` to get the setup deps. This ensures
everything works correctly.
This is a bit dubvious, but the alternative of making nodejs a
nativeBuildInput for node packages is worse. In general the cross story
for interpreted languages is murky, and this fits that pattern.
Before, we'd always use `cc = null`, and check for that. The problem is
this breaks for cross compilation to platforms that don't support a C
compiler.
It's a very subtle issue. One might think there is no problem because we
have `stdenvNoCC`, and presumably one would only build derivations that
use that. The problem is that one still wants to use tools at build-time
that are themselves built with a C compiler, and those are gotten via
"splicing". The runtime version of those deps will explode, but the
build time / `buildPackages` versions of those deps will be fine, and
splicing attempts to work this by using `builtins.tryEval` to filter out
any broken "higher priority" packages (runtime is the default and
highest priority) so that both `foo` and `foo.nativeDrv` works.
However, `tryEval` only catches certain evaluation failures (e.g.
exceptions), and not arbitrary failures (such as `cc.attr` when `cc` is
null). This means `tryEval` fails to let us use our build time deps, and
everything comes apart.
The right solution is, as usually, to get rid of splicing. Or, baring
that, to make it so `foo` never works and one has to explicitly do
`foo.*`. But that is a much larger change, and certaily one unsuitable
to be backported to stable.
Given that, we instead make an exception-throwing `cc` attribute, and
create a `hasCC` attribute for those derivations which wish to
condtionally use a C compiler: instead of doing `stdenv.cc or null ==
null` or something similar, one does `stdenv.hasCC`. This allows quering
without "tripping" the exception, while also allowing `tryEval` to work.
No platform without a C compiler is yet wired up by default. That will
be done in a following commit.
I think those deps could be made to build, but I didn't want to get
bogged down investigating further. "Use flags" are always a good thing,
so this is fine for now.
