YJIT: Optimize local variables when EP == BP

[k0kubun]

This PR optimizes local variable access in getlocal and setlocal from REG_CFP->ep[offset] (2 instructions) to REG_SP[offset] (1 instruction) when EP == BP. When EP escapes, the block gets invalidated.

Because local variable access is now based off of SP, this will allow us to handle local variables using Opnd::Stack and allocate registers for them in a future PR.

Generated code

Before

  # Insn: 0000 getlocal_WC_0 (stack_size: 0)
  0x59e90a49a058: mov rax, qword ptr [r13 + 0x20]
  # reg_temps: 00000000 -> 00000001
  0x59e90a49a05c: mov rsi, qword ptr [rax - 0x18]

After

  # Insn: 0000 getlocal_WC_0 (stack_size: 0)
  # reg_temps: 00000000 -> 00000001
  0x6271819a1058: mov rsi, qword ptr [rbx - 0x20]

Benchmark

This seems to improve the performance of the following benchmarks.

before: ruby 3.4.0dev (2024-04-17T02:45:52Z master 0b630d6441) +YJIT [x86_64-linux]
after: ruby 3.4.0dev (2024-04-17T05:46:52Z yjit-ep-bp 3b316da2e3) +YJIT [x86_64-linux]

--------------  -----------  ----------  ----------  ----------  -------------  ------------
bench           before (ms)  stddev (%)  after (ms)  stddev (%)  after 1st itr  before/after
erubi-rails     1219.7       0.1         1160.5      0.1         1.04           1.05
liquid-compile  74.1         0.5         73.6        0.8         1.00           1.01
liquid-render   94.8         0.3         93.0        0.6         1.01           1.02
psych-load      1785.5       0.1         1743.8      0.1         1.02           1.02
protoboeuf      59.0         0.6         58.2        0.8         1.01           1.01
--------------  -----------  ----------  ----------  ----------  -------------  ------------

[maximecb]

Hi @k0kubun ! This looks good. Nice work.

For completeness, could you measure memory usage on lobsters? Just asking because of the extra invariants we need to care care of, might be a useful data point.

[k0kubun]

For completeness, could you measure memory usage on lobsters?

Sure.

With 25 itrs on lobsters, inline_code_size is reduced by 0.7%, and yjit_alloc_size is increased by 2.4%.

Before

inline_code_size:          7,612,421
outlined_code_size:        6,322,015
code_region_size:         15,269,888
code_region_overhead:      1,335,452 ( 8.7%)
yjit_alloc_size:          28,784,713

After

inline_code_size:          7,559,850
outlined_code_size:        6,229,344
code_region_size:         15,462,400
code_region_overhead:      1,673,206 (10.8%)
yjit_alloc_size:          29,475,775

[maximecb]

We may want to think about how to optimize memory usage for storing invariants in the future, but this PR seems like a clear step in the right direction 👍

[nobu]

This seems causing recent hang ups at make (check, --enable-yjit=dev, --yjit-call-threshold=1 --yjit-verify-ctx --yjit-code-gc).