Redundant load/store pass

[ltratt]

This PR adds a new redundant load_store elimination pass (2bdc42b): it's simple, and more-or-less a direct port from jitc_yk. The main reason I've done this now is to help me understand what's needed from the optimiser API: indeed, a change was necessary (675444e).

@Fidget-Spinner My guess is that the known_bits pass might need the extended inst_committed API too. And perhaps other bits I haven't considered yet.

[ltratt]

Oops, I should have added the tests in 410fa9e in the original PR. Will need squashing.

[Fidget-Spinner]

Do you have any benchmark results? The refactoring looks fine to me.

[ltratt]

I do:

 fannkuchredux/yklua/10     13979 ± 103   13005 ± 143   0.93  6.97% faster
 nbody/yklua/250000          4694 ±  93    4458 ±  65   0.95  5.03% faster
 Heightmap/yklua/2000        8338 ±  76    7983 ±  12   0.96  4.26% faster
 sieve/yklua/3000            5066 ±  47    4860 ±  28   0.96  4.06% faster
 spectralnorm/yklua/1000    12564 ±  31   12340 ±   2   0.98  1.79% faster
 cd/yklua/250               33563 ± 140   33176 ± 161   0.99  1.15% faster
 mandelbrot/yklua/500        2266 ±   4    2292 ±   3   1.01  1.16% slower

Everything else is indistinguistable (apart from list, whose results bounce around too much to be useful). The mandelbrot result is odd but replicates at scale: there's nothing obvious causing it that I've yet understood. It may be as simple as slightly different code layout or similar...

[Fidget-Spinner]

Ok I've taken a look at this twice now. Let's merge this.

[Fidget-Spinner]

Please squash.

[ltratt]

Squashed.

[Fidget-Spinner]

@ltratt the new tests are failing, happy holidays!

I'll move my branch to use the new API after this is merged.

[ltratt]

Oops, fixed the doc string in the force push.

[Fidget-Spinner]

lang tests are failing

[ltratt]

Ah, I'm being bitten by the lack of "proper" jit-pre-opt. I'll need to chew on the best solution to this.

[ltratt]

After playing for a little bit, I don't see a nice way for j2 to do jit-pre-opt and jit-post-opt: it seems that I end up either giving misleading jit-pre-opt or messing up inlining edge cases (or add lots of code to do both, which is messy). The new commit thus doesn't try to do this: it either makes the IR tests more general or sets YKD_OPT=0. Since the lang_tests are mostly testing aot_to_hir I think the latter is OK.

While doing this, it turned out -- to my surprise! -- that we're also encountering the const case in j2 (which we didn't in jitc_yk, I think because of the worse inlining). So I've had to implement that too.

@Fidget-Spinner If you're OK with this, it will need squashing.

[Fidget-Spinner]

Turning off opt for certain lang tests seem fine to me. Please squash. I will copy my known bits changes to the new branch after merging.

[ltratt]

Squashed.