ZJIT: Experiment with going through memory for all local variable access

[XrXr]

ZJIT currently tries to lift local variables at environment level 0 into FrameState using the NoEPEscape invariant and the from-memory reloads at strategic places. This setup has had a few bugs in the past and a few outstanding ones.^{1 2}

We would like to know what performance looks like if all local variables went through EP to/from memory. It has the benefit of being close to interpreter semantics so easier to ensure correctness. With ZJIT: Optimize load store forwarding (ruby#16228) and ZJIT: Implement dead store elimination (ruby#16507) we may not be losing much speed by going through memory.

To give an estimate of speed impact, in YJIT, putting some local variables into registers gave about a 2% speedup (ruby#11157).

Footnotes

1. https://github.com/Shopify/ruby/issues/970 ↩

2. https://github.com/Shopify/ruby/issues/976 ↩

[XrXr]

Early result from a draft patch version ruby@c3bb308e47

master: ruby 4.1.0dev (2026-05-14T21:17:54Z master 9a55fc506d) +ZJIT +PRISM [x86_64-linux]
+patch: ruby 4.1.0dev (2026-05-14T21:45:27Z +patch c3bb308e47) +ZJIT +PRISM [x86_64-linux]
 
--------------  -------------  -------------  --------------  -------------
bench             master (ms)    +patch (ms)  +patch 1st itr  master/+patch
railsbench                N/A            N/A             N/A            N/A
shipit                    N/A            N/A             N/A            N/A
activerecord     103.2 ± 4.7%   106.5 ± 5.5%           0.918          0.969
chunky-png       280.2 ± 0.5%   292.9 ± 0.4%           0.953          0.957
erubi-rails      390.2 ± 0.4%   398.2 ± 0.7%           1.012          0.980
hexapdf         1090.2 ± 0.6%  1114.9 ± 0.7%           0.996          0.978
liquid-c          29.1 ± 7.9%    29.6 ± 9.0%           0.979          0.983
liquid-compile   21.3 ± 10.3%   21.5 ± 10.3%           1.031          0.992
liquid-render     59.7 ± 3.7%    61.4 ± 3.4%           0.992          0.972
lobsters         529.5 ± 2.0%   527.3 ± 2.2%           0.981          1.004
mail              65.4 ± 3.2%    66.1 ± 3.1%           1.034          0.989
psych-load       879.5 ± 1.2%   905.9 ± 1.4%           0.981          0.971
rubocop          95.4 ± 13.7%   94.6 ± 12.8%           1.011          1.009
ruby-lsp          71.5 ± 4.7%    71.9 ± 3.1%           1.010          0.995
sequel            27.7 ± 4.8%    28.5 ± 5.5%           1.018          0.973
--------------  -------------  -------------  --------------  -------------
 
Legend:
- +patch 1st itr: ratio of master/+patch time for the first benchmarking iteration.
- master/+patch: ratio of master/+patch time. Higher is better for +patch. Above 1 represents a speedup.

It can't run all the benchmarks yet and is unoptimized (setlocal always goes through a function call), but ain't looking bad.