YJIT: Merge lea and mov on x86_64 when possible

[k0kubun]

I started to look into adding optimization passes to the backend. This was one of the inefficient code I was aware of, so I worked on this first. My plan is to add more peephole optimizations to x86_merge over time.

Generated code

Before

  # save SP to CFP
  0x564019e65142: lea rcx, [rbx - 0x10]
  0x564019e65146: mov rbx, rcx
  0x564019e65149: mov qword ptr [r13 + 8], rbx

After

  # save SP to CFP
  0x5634800c413f: lea rbx, [rbx - 0x10]
  0x5634800c4143: mov qword ptr [r13 + 8], rbx

Code size stats

On railsbench, inline code size went down by 0.4%. I see no significant impact on its speed.

Before

inline_code_size:          2,073,884
outlined_code_size:        2,073,528

After

inline_code_size:          2,065,752
outlined_code_size:        2,065,384

[maximecb]

I wrote some very minor comments. Seems ready to merge otherwise.

We do have to be somewhat careful about adding many passes to the compiler as this will slow down our compilation speed. For now seems like not a huge concern, but something to keep in mind, especially since we aren't tracking compilation speed right now.