libexpr: Reduce the size of Value down to 16 bytes (on 64 bit systems)

[xokdvium]

Motivation

This shaves off a very significant (~20% percent reduction in maximum heap size and ~17% in total bytes) amount of memory used for evaluation as well as reduces the GC-managed heap. See the results below. The fraction of collected memory is not affected.

Memory usage comparison

Below are the comparisons for NIX_SHOW_STATS=1 nix-env --query --available --out-path --file ../nixpkgs --eval-system x86_64-linux > /dev/null

Before

{
  ....
  "gc": {
    "cycles": 14,
    "heapSize": 14818668544,
    "totalBytes": 27457473440
  },
  ....
  "sizes": {
    "Attr": 16,
    "Bindings": 16,
    "Env": 8,
    "Value": 24
  },
  ....
  "values": {
    "bytes": 7345779024,
    "number": 306074126
  }
}

After

{
  ....
  "gc": {
    "cycles": 14,
    "heapSize": 11916210176,
    "totalBytes": 22553098160
  },
  ....
  "sizes": {
    "Attr": 16,
    "Bindings": 16,
    "Env": 8,
    "Value": 16
  },
  ....
  "values": {
    "bytes": 4897186016,
    "number": 306074126
  }
}

So the saving are around 2.45 GiB on a full nixpkgs eval.

Previously the union discriminator (InternalType) was
stored as a separate field in the Value, which takes up
whole 8 bytes due to padding needed for member alignment.
This effectively wasted 7 whole bytes of memory. Instead
of doing that InternalType is instead packed into pointer
alignment niches. As it turns out, there's more than enough
unused bits there for the bit packing to be effective.

See the doxygen comment in the ValueStorage specialization
for more details.

This does not add any performance overhead, even though
we now consistently assert the InternalType in all getters.

This can also be made atomic with a double width compare-and-swap
instruction on x86_64 (CMPXCHG16B instruction) for parallel evaluation.

Context

This is the best we can do with the current data model to address #8621.
The improvement is very significant and should give more leeway to the ever growing nixpkgs.

---

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[Mic92]

Cool stuff. Have only went over briefly. But I am interested in a review from @NaN-git

[xokdvium]

Previosly this code used a pointer to the listElements (either stored inline or in a big list on the heap). Since now inline values are tagged we have to copy them to the listView, so the pointer to elements would point to the stack.

For now I've changed this to a Value * instead (that is actually the same pointer, since the smallList is the first element in a Value structure). This should work fine, but maybe I'm missing something. @roberth do you think this is ok to do?

If there are no objections I'll remove the FIXME.

[xokdvium]

Also fixed i686-linux builds. Needed some SFINAE trickery to make the compiler not instantiate ValueStorage member functions that would be ill-formed on a non-64 bit system.