ZJIT: Write elements to an array at once on Insn::NewArray

[k0kubun]

Insn::NewArray calls rb_ary_push for each element, which is slower than what the interpreter does (rb_ec_ary_new_from_values).

For example, this code:

def test(a, b, c, d)
  [a, b, c, d]
end

test(1, 2, 3, 4)

results in:

  # Insn: v23 NewArray v15, v16, v17, v18
  # save PC to CFP
  0x5ff7238370fe: movabs r11, 0x5ff7463eae40
  0x5ff723837108: mov qword ptr [r13], r11
  # call rb_ary_new_capa
  0x5ff72383710c: push rsi
  0x5ff72383710d: push rdx
  0x5ff72383710e: push rcx
  0x5ff72383710f: push r8
  0x5ff723837111: mov edi, 4
  0x5ff723837116: call 0x5ff727c8f6e0
  0x5ff72383711b: pop r8
  0x5ff72383711d: pop rcx
  0x5ff72383711e: pop rdx
  0x5ff72383711f: pop rsi
  # call rb_ary_push
  0x5ff723837120: mov rdi, rax
  0x5ff723837123: push rdi
  0x5ff723837124: push rsi
  0x5ff723837125: push rdx
  0x5ff723837126: push rcx
  0x5ff723837127: push r8
  0x5ff723837129: push r8
  0x5ff72383712b: call 0x5ff727c8a510
  0x5ff723837130: pop r8
  0x5ff723837132: pop r8
  0x5ff723837134: pop rcx
  0x5ff723837135: pop rdx
  0x5ff723837136: pop rsi
  0x5ff723837137: pop rdi
  # call rb_ary_push
  0x5ff723837138: push rdi
  0x5ff723837139: push rsi
  0x5ff72383713a: push rdx
  0x5ff72383713b: push rcx
  0x5ff72383713c: push r8
  0x5ff72383713e: push r8
  0x5ff723837140: mov rsi, rdx
  0x5ff723837143: call 0x5ff727c8a510
  0x5ff723837148: pop r8
  0x5ff72383714a: pop r8
  0x5ff72383714c: pop rcx
  0x5ff72383714d: pop rdx
  0x5ff72383714e: pop rsi
  0x5ff72383714f: pop rdi
  # call rb_ary_push
  0x5ff723837150: push rdi
  0x5ff723837151: push rsi
  0x5ff723837152: push rdx
  0x5ff723837153: push rcx
  0x5ff723837154: push r8
  0x5ff723837156: push r8
  0x5ff723837158: mov rsi, rcx
  0x5ff72383715b: call 0x5ff727c8a510
  0x5ff723837160: pop r8
  0x5ff723837162: pop r8
  0x5ff723837164: pop rcx
  0x5ff723837165: pop rdx
  0x5ff723837166: pop rsi
  0x5ff723837167: pop rdi
  # call rb_ary_push
  0x5ff723837168: push rdi
  0x5ff723837169: push rsi
  0x5ff72383716a: push rdx
  0x5ff72383716b: push rcx
  0x5ff72383716c: push r8
  0x5ff72383716e: push r8
  0x5ff723837170: mov rsi, r8
  0x5ff723837173: call 0x5ff727c8a510
  0x5ff723837178: pop r8
  0x5ff72383717a: pop r8
  0x5ff72383717c: pop rcx
  0x5ff72383717d: pop rdx
  0x5ff72383717e: pop rsi
  0x5ff72383717f: pop rdi

but we shouldn't need to make 5 different C function calls.

Similarly to Insn::NewHash, gen_new_array should use gen_push_opnds and make a single C function call of rb_ec_ary_new_from_values, which newarray in insns.def uses. YJIT uses that function too.

[k0kubun]

context: Shopify/zjit#110 (comment)

We were not ready for using the stack as an argument at the time. We are, now.

[awdjob]

@k0kubun can I take a shot at this?

[k0kubun]

yes, thanks!

[tekknolagi]

I am going to un-assign. Will re-assign if you put up a PR.