rustbuild: Build jemalloc and libbacktrace only once (take 2)
This is a rebase of https://github.com/rust-lang/rust/pull/38583 without any additions, but with implemented @alexcrichton's suggestions.
~~This includes `exists(Makefile)` => `cfg(stage0)` suggestion... but it will break cross-compilation, no? Are `libstd/liballoc_jemalloc` cross-compiled for `target != host` built during `stage0`?~~
r? @alexcrichton
Extend Cell to work with non-Copy types
I'm not sure that I did this right but all of the tests pass.
I also had to move the `new()` function so that `Cell`s with non-`Copy` `T`s could be created. That wasn't in the RFC but I assume it needed to be done?
The previous way was not friendly to incremental compilation. The new
plan is to compute, for each body, a set of trait imports used in that
body (slightly subtle: for a closure, we assign the trait imports to the
enclosing fn). Then we walk all bodies and union these sets. The reason
we do this is that we can save the individual sets in the incremental
state, and then recompute only those sets that are needed. Before we
were planning to save only the final union, but in that case if some
components are invalidated we have to recompute *all* of them since we
don't have enough information to "partly" invalidate a result.
In truth, this set probably ought to be part of the `TypeckTables`;
however, I opted not to do that because I don't want to have to
save/restore the entire tables in the incremental state yet (since it
contains a lot of `NodeId` references, and removing those is a
significant refactoring).
Although signatures with anonymous parameters may not be deprecated or
removed at this point, the team seems to agree that the ability to have
an anonymous parameter is unfortunate historical baggage, and that we
shouldn't create new code that uses it.
Context: https://github.com/rust-lang/rust/issues/33417#issuecomment-276933861
