Fix initialization of unboxed numbers in static blocks

[TheNumbat]

Flambda2 was incorrectly initializing unboxed static closure slots and static constants.

• Closure slots were all assumed to be Word_val, which was interpreted as a pointer and hence passed to setfield. It would then call caml_initialize with a gc-unfriendly value, like an unboxed float. This also wouldn't fully copy a vec128.

• Static constant boxed int64s and int64 arrays were incorrectly calling setfield with the unboxed int64 value due to using Word_int.

• When Config.runtime5 was set, make_update unconditionally used setfield.

• Naked vec128s would be written off the end of the closure block (this won't actually occur until unboxed vectors are exposed).

I found this in #2519, which re-enabled the assertion that we do not move between float & int registers. That was triggered when attempting to pass an unboxed float to caml_initialize.

[TheNumbat]

Note - writing an unboxed int32 into the closure environment currently needs to write 64 bits, because Project_value_slot loads Naked_int32s with the Word_int memory chunk instead of the sign-extending Thirtytwo_signed. This relies on the incoming int32 to already have been sign extended, which seems a bit dangerous.

I will instead make Project_value_slot sign extend (this is free on amd64).

[mshinwell]

I will discuss this with @TheNumbat

[Gbury]

Note - writing an unboxed int32 into the closure environment currently needs to write 64 bits, because Project_value_slot loads Naked_int32s with the Word_int memory chunk instead of the sign-extending Thirtytwo_signed. This relies on the incoming int32 to already have been sign extended, which seems a bit dangerous.

I will instead make Project_value_slot sign extend (this is free on amd64).

Well, this was written at a time where everything was a word, so this was reasonable and safe at that time. Clearly, with the advent of non-scannable slots in closures, as well as mixed blocks, we might want to change that, and indeed do sign-extending loads.

[TheNumbat]

Sounds good; the change to memory_chunk_of_kind makes int32 loads sign-extend.

[TheNumbat]

Added some more edits / explanations after discussing with @mshinwell

[mshinwell]

Please don't rebase this yet, as I'm still reviewing/editing.

[mshinwell]

@TheNumbat I've finished. I suggest looking at my new changes commit by commit, but read them all before commenting...

We've discussed and the appropriate changes have been made

[TheNumbat]

New changes look good, merging