Without this, the primop chokes on any thunks on attributes passed in
the attrset. It even is the reason why the test contained `builtins.seq`
to work around this. Supposedly, this might have been an intentional
restriction and changing this might break things in ways I cannot forsee
due to not knowing much about Flakes, however the status quo is equally
broken:
- The error message looks like an internal error and not like some
explicitly forbidden invariant violation.
- Seemingly simple syntax literals like "-1" compile to "__sub 0 1" and
thus create a thunk which then fails, which is utterly confusing ("why
does 1 work but not -1?")
- This is a stark violation of the principle of least surprise.
- Thunking relies on maybeThunk and thunk inlining optimizations, thus
not forcing thunks turns operational details of the evaluator into
language-observable behavior. That's bad.
I am changing this now regardless of the risk of breakage, because the
bytecode evaluator will have different thunk inlining optimizations and
thus inevitably cause mismatches in behavior anyways.
Change-Id: Ifc45c4d2900e40822383670b28e4e50ab8af317a
piegames
c1f75860d6