YJIT: Fallback Integer#<< if a shift amount varies

[k0kubun]

• Chain-guard Integer#<< once and fallback to a C call if a shift amount varies.
• Rename an exit reason amt_changed to amount_changed since the name seemed a bit cryptic.

On nqueens Shopify/yjit-bench#260 with --yjit-call-threshold=1, this change pushes ratio_in_yjit from 0.0% to 99.9% and makes it 3.60x faster.

before: ruby 3.4.0dev (2024-01-05T16:51:37Z master 4d03140009) +YJIT [x86_64-linux]
after: ruby 3.4.0dev (2024-01-08T17:07:08Z yjit-ltlt 43edc3419f) +YJIT [x86_64-linux]

-------  -----------  ----------  ----------  ----------  -------------  ------------
bench    before (ms)  stddev (%)  after (ms)  stddev (%)  after 1st itr  before/after
nqueens  581.8        0.7         161.8       0.1         3.55           3.60
-------  -----------  ----------  ----------  ----------  -------------  ------------

[XrXr]

Looks like this effectively disables the specialized path, with ./miniruby --yjit-dump-disasm --yjit-call-threshold=1 -e '23434 << 3' generating a call:

  # regenerate_branch
  # Block: <main>@-e:1 (chain_depth: 1)
  # reg_temps: 00000011
  # Insn: 0004 opt_ltlt (stack_size: 2)
  # call to Integer#<<
  # RUBY_VM_CHECK_INTS(ec)
  0x104d1803c: ldur w11, [x20, #0x20]
  0x104d18040: tst w11, w11
  <snip>

Should be > 1 probably, and it looks like the defer_compilation is from gen_send_general() and it's unconditional.

[k0kubun]

Ugh, thanks for catching it. 43edc34

I don't get the point of this increment by the way. If it's only for the "Double defer!" detection, we could assign one bit from chain_depth_return_landing to it instead of overloading chain_depth. It seems counter-intuitive when you look at *_MAX_DEPTH constants.

[maximecb]

For some reason I'm not seeing any speedup on nqueens:

interp: ruby 3.4.0dev (2024-01-12T18:04:23Z master 0462b1b350) [arm64-darwin23]
yjit: ruby 3.4.0dev (2024-01-12T18:04:23Z master 0462b1b350) +YJIT dev [arm64-darwin23]

-------  -----------  ----------  ---------  ----------  ------------  -----------
bench    interp (ms)  stddev (%)  yjit (ms)  stddev (%)  yjit 1st itr  interp/yjit
nqueens  729.3        0.5         728.8      0.5         1.01          1.00       
-------  -----------  ----------  ---------  ----------  ------------  -----------
Legend:
- yjit 1st itr: ratio of interp/yjit time for the first benchmarking iteration.
- interp/yjit: ratio of interp/yjit time. Higher is better for yjit. Above 1 represents a speedup.

Wondering if maybe the last commit made this not work?

[k0kubun]

Did you use --yjit-call-threshold=1? As I said in Shopify/yjit-bench#260 (comment), the script doesn't JIT the method by default, so it needs to be updated.

[maximecb]

kk. I'll tweak the benchmark parameters.