Call rb_ivar_get and rb_ivar_set instead of exiting for many ivars

[eileencodes]

Previously, when we have a lot of ivars defined, we would exit via jit_chain_guard for megamorphic ivars. Now if we have more than the max depth of ivars we can call rb_ivar_get and rb_ivar_set respectively.

Using the following script:

class A
  def initialize
    @a = 1
  end

  def a
    @a
  end
end

N = 30
N.times do |i|
  eval <<-eorb
class A#{i} < A
  def initialize
    @a#{i} = 1
    super
  end
end
  eorb
end

klasses = N.times.map { Object.const_get(:"A#{_1}") }

1000.times do
  klasses.each do |k|
    k.new.a
  end
end

Exits before this change show exits for getinstancevariable and setinstancevariable:

***YJIT: Printing YJIT statistics on exit***
method call exit reasons:
    klass_megamorphic:     24,975 (100.0%)
invokeblock exit reasons:
    (all relevant counters are zero)
invokesuper exit reasons:
    (all relevant counters are zero)
leave exit reasons:
    interp_return:     26,948 (100.0%)
     se_interrupt:          1 ( 0.0%)
getblockparamproxy exit reasons:
    (all relevant counters are zero)
getinstancevariable exit reasons:
    megamorphic:     13,986 (100.0%)
setinstancevariable exit reasons:
    megamorphic:     19,980 (100.0%)
opt_aref exit reasons:
    (all relevant counters are zero)
expandarray exit reasons:
    (all relevant counters are zero)
opt_getinlinecache exit reasons:
    (all relevant counters are zero)
invalidation reasons:
    (all relevant counters are zero)
num_send:                    155,823
num_send_known_class:              0 ( 0.0%)
num_send_polymorphic:        119,880 (76.9%)
bindings_allocations:              0
bindings_set:                      0
compiled_iseq_count:              36
compiled_block_count:            158
compiled_branch_count:           240
block_next_count:                 10
defer_count:                      70
freed_iseq_count:                  0
invalidation_count:                0
constant_state_bumps:              0
inline_code_size:             29,216
outlined_code_size:           27,948
freed_code_size:                   0
code_region_size:             65,536
live_context_size:             8,322
live_context_count:              219
live_page_count:                   4
freed_page_count:                  0
code_gc_count:                     0
num_gc_obj_refs:                 130
object_shape_count:              295
side_exit_count:              58,942
total_exit_count:             85,890
yjit_insns_count:          1,023,581
avg_len_in_yjit:                11.2
Top-4 most frequent exit ops (100.0% of exits):
    opt_send_without_block:     24,975 (42.4%)
       setinstancevariable:     19,980 (33.9%)
       getinstancevariable:     13,986 (23.7%)
		     leave:          1 ( 0.0%)

Exits after this change show we have no exits for getinstanevariable and setinstancevariable.

***YJIT: Printing YJIT statistics on exit***
method call exit reasons:
    klass_megamorphic:     24,975 (100.0%)
invokeblock exit reasons:
    (all relevant counters are zero)
invokesuper exit reasons:
    (all relevant counters are zero)
leave exit reasons:
    interp_return:     60,912 (100.0%)
     se_interrupt:          3 ( 0.0%)
getblockparamproxy exit reasons:
    (all relevant counters are zero)
getinstancevariable exit reasons:
    (all relevant counters are zero)
setinstancevariable exit reasons:
    (all relevant counters are zero)
opt_aref exit reasons:
    (all relevant counters are zero)
expandarray exit reasons:
    (all relevant counters are zero)
opt_getinlinecache exit reasons:
    (all relevant counters are zero)
invalidation reasons:
    (all relevant counters are zero)
num_send:                    155,823
num_send_known_class:              0 ( 0.0%)
num_send_polymorphic:        119,880 (76.9%)
bindings_allocations:              0
bindings_set:                      0
compiled_iseq_count:              36
compiled_block_count:            179
compiled_branch_count:           240
block_next_count:                 11
defer_count:                      70
freed_iseq_count:                  0
invalidation_count:                0
constant_state_bumps:              0
inline_code_size:             31,032
outlined_code_size:           29,708
freed_code_size:                   0
code_region_size:             65,536
live_context_size:             8,360
live_context_count:              220
live_page_count:                   4
freed_page_count:                  0
code_gc_count:                     0
num_gc_obj_refs:                 130
object_shape_count:              295
side_exit_count:              24,978
total_exit_count:             85,890
yjit_insns_count:          1,076,966
avg_len_in_yjit:                12.2
Top-2 most frequent exit ops (100.0% of exits):
    opt_send_without_block:     24,975 (100.0%)
		     leave:          3 ( 0.0%)

[eileencodes]

cc/ @tenderlove

[XrXr]

Looks like for gets this conflicts with #7334

[eileencodes]

Oh oops 🤦🏼‍♀️

[maximecb]

Kokubun did only implement the change for getivar though, so worth looking at his PR and implementing the change for set.

[k0kubun]

Yeah, I worked only on setivar because SFR was like:

getinstancevariable exit reasons:
    megamorphic: 85,342,998 (100.0%)
setinstancevariable exit reasons:
    megamorphic:          3 (100.0%)

but supporting setivar in this PR seems helpful for other benchmarks, looking at your description.

I confirmed that the CI of my PR was passing and merged it. Please rebase your branch against master 🙇

[k0kubun]

In #7334, I generated this at == max_chain_depth - 1 instead of == max_chain_depth (or >=) because I thought the C call block should also be counted as part of the number of chains. But if we think about the limit as the number of jit_chain_guard calls, then doing this at == max_chain_depth (or >=) makes sense 👍