Promote instructions.

[ltratt]

This allows yk to constant fold a lot of the resulting instruction decoding. The benefits of doing so can be significant: it speeds big_loop.lua up by about 38%!

Needs ykjit/yk#1441 to be merged first.

[ltratt]

This will need squashing as I realised I could simplify things mildly in 26294d4.

[vext01]

please squash

[ltratt]

Squashed.

[ltratt]

Ah, this might only work after ykjit/yk#1442 (and maybe ykjit/yk#1443) have merged. I don't think I fully tested it pre-1442.

[vext01]

10:37:21 .thread '<unnamed>' panicked at ykrt/src/compile/jitc_yk/opt/mod.rs:74:41:
10:37:21 attempt to add with overflow

Is that the address space overflowing when adding the offset?

LLVM says about GEP:

The offsets are then added to the low bits of the base address up to the index type width, with silently-wrapping two’s complement arithmetic

[vext01]

(ah we raced)

[ltratt]

@vext01 With luck this one should now work.

[ltratt]

@vext01 We can retry this now.