ZJIT: Decouple gen_function_stub and gen_function_stub_hit_trampoline

[tenderlove]

Before this change, gen_function_stub and
gen_function_stub_hit_trampoline communicated via a scratch register. We would like gen_function_stub_hit_trampoline to have more freedom with regard to the registers it uses, especially for the CCall in to function_stub_hit. Instead of communicating via scratch register, we'll communicate via stack.

Practically speaking, this means:

• Stop using x15 (scratch reg) to communicate iseq call addr from call stub to function sub hit trampoline; use stack instead
• Don't try to CCall with x15 as first argument; can't use scratch reg in parallel move of arguments

Here is pseudo assembly of before this commit:

some_send_direct_in_a_ruby_method(JIT code):
  mov x15, gen_function_stub
  mov x0, self
  mov x1, 1
  blr x15

gen_function_stub:
  mov x15, 0xISEQADDR (the address of the ISEQ we _want_ to compile)
  jmp function_stub_hit_trampoline

function_stub_hit_trampoline:
  function prologue
  cpush ALL_JIT_REGS
  mov x0, x15 # currently x15 is 0xISEQADDR
  mov x1, CFP
  mov x2, SP
  blr function_stub_hit
  mov x15, x0 # write jump address to x15 (code pointer for compiled iseq)
  cpop ALL_JIT_REGS
  function epilogue
  jmp x15

Here is pseudo assembly of after this commit:

some_send_direct_in_a_ruby_method(JIT code):
  mov x15, gen_function_stub
  mov x0, self
  mov x1, 1
  blr x15

gen_function_stub:
  mov x15, 0xISEQADDR (the address of the ISEQ we _want_ to compile)
  push x15
  jmp function_stub_hit_trampoline

function_stub_hit_trampoline:
  pop x15 # get the ISEQ addr from gen_function_stub
  function prologue
  cpush ALL_JIT_REGS
  mov x0, x15 # currently x15 is 0xISEQADDR
  mov x1, CFP
  mov x2, SP
  blr function_stub_hit
  mov x15, x0 # write jump address to x15 (code pointer for compiled iseq)
  cpop ALL_JIT_REGS
  function epilogue
  jmp x15