Code Generation

Lower IR to assembly. For each IR instruction, emit corresponding asm.

Stack-based codegen (simplest):

• Every value lives on the stack initially.
• Operations pop operands, push results.

For a function:

Function calls:

gen_expr(arg1) -> push rax (etc., last)
mov rdi, [rsp+...]   ; load args from stack into reg
mov rsi, [rsp+...]
call <func>
add rsp, <args*8>    ; clean up
push rax             ; result

Stack frame layout (x86-64 ABI):

[caller's frame]
[return address]
[old rbp]              <- rbp
[local var 0]          <- rbp-8
[local var 1]          <- rbp-16
...
[temp slot]            <- rsp

Naive stack-machine codegen is suboptimal: every operation hits memory. Real compilers use register allocation to keep values in regs as long as possible.

For our compiler:

• Skip register allocation.
• Stack everything.
• Resulting code is correct but slow.
• Modern compilers' optimization output is 5-10x faster.