YJIT: Decrease IVAR_MAX_DEPTH to 8

[k0kubun]

I tested 5, 8, and 10 as GET_IVAR_MAX_DEPTH and SET_IVAR_MAX_DEPTH on SFR. depth=8 showed the best latency, and it had almost the same max code size as depth=5, which is 1.6% smaller than depth=10's.

[maximecb]

Thanks for running this experiment 🙏

Would just like to see the perf impact of depth 8 vs 10 on headline benchmarks before we merge.

[k0kubun]

Here's the results:

before: ruby 3.3.0dev (2023-09-07T18:22:34Z master fcdedf7f47) +YJIT [x86_64-linux]
after: ruby 3.3.0dev (2023-09-07T20:49:45Z yjit-ivar-8 322ac79141) +YJIT [x86_64-linux]

--------------  -----------  ----------  ---------  ----------  ----------  ---------  -------------  ------------
bench           before (ms)  stddev (%)  RSS (MiB)  after (ms)  stddev (%)  RSS (MiB)  after 1st itr  before/after
activerecord    31.3         5.6         56.4       31.2        5.3         56.4       0.98           1.00
chunky-png      399.5        0.6         57.6       399.2       0.5         50.5       1.01           1.00
erubi-rails     10.5         28.2        103.7      10.5        30.8        97.2       1.03           1.00
hexapdf         1343.4       1.1         272.9      1381.8      1.2         264.3      1.01           0.97
liquid-c        38.8         0.7         39.6       38.6        0.7         39.7       1.00           1.00
liquid-compile  39.8         0.6         34.8       39.8        0.8         34.7       1.00           1.00
liquid-render   56.2         0.3         38.0       56.5        0.3         38.0       1.00           1.00
lobsters        615.9        5.1         331.5      611.3       3.8         331.1      1.00           1.01
mail            82.4         0.3         50.2       82.9        0.2         50.6       0.99           0.99
psych-load      1234.0       0.2         36.5       1237.1      0.2         36.3       1.00           1.00
railsbench      1112.7       1.4         104.4      1116.8      1.3         102.8      0.99           1.00
ruby-lsp        36.1         26.8        130.6      36.0        26.5        129.9      1.01           1.00
sequel          47.3         2.0         40.9       47.3        1.7         40.8       1.00           1.00
--------------  -----------  ----------  ---------  ----------  ----------  ---------  -------------  ------------

Most benchmarks seem not impacted, but hexapdf does slow down a little.