Use the conventional integer calling convention: $a0 holds the argument and $v0 holds the result. The function is recursive, so every call must preserve its own return address and original argument. The first recursive result must also survive the second call; a caller-saved register alone would not protect it.
This implementation reserves a 32-byte stack frame, keeping the stack pointer eight-byte aligned. Offsets 0–15 are available as the outgoing argument area, 16 stores the original argument, 20 stores the first result, 24 is padding, and 28 stores the return address. It changes no callee-saved register. Assemble in classic MIPS delayed-branch mode with instruction reordering disabled, such as .set noreorder.
foo:
addiu $sp,$sp,-32  # allocate 32-byte aligned frame
sw $ra,28($sp)  # save caller return address
sw $a0,16($sp)  # save original n
beq $a0,$zero,foo_base
sll $zero,$zero,0
addiu $t0,$zero,1
beq $a0,$t0,foo_base
sll $zero,$zero,0
addiu $t0,$zero,2
beq $a0,$t0,foo_base
sll $zero,$zero,0
addiu $a0,$a0,-1
jal foo
sll $zero,$zero,0
sw $v0,20($sp)  # preserve foo(n-1) across next call
lw $a0,16($sp)  # recover original n
sll $zero,$zero,0
addiu $a0,$a0,-3
jal foo
sll $zero,$zero,0
lw $t0,20($sp)
sll $zero,$zero,0
addu $v0,$v0,$t0  # combine the two recursive results
beq $zero,$zero,foo_return
sll $zero,$zero,0
foo_base:
addu $v0,$a0,$zero  # base case result is n
foo_return:
lw $ra,28($sp)
sll $zero,$zero,0
addiu $sp,$sp,32  # release the frame
jr $ra
sll $zero,$zero,0
All displayed operation mnemonics denote single machine instructions. In particular, sll $zero,$zero,0 is the actual harmless shift encoding often abbreviated by the pseudoinstruction nop; addu $v0,$a0,$zero performs a copy without the move pseudoinstruction. Each branch or jump has an explicit harmless branch delay slot, and harmless shifts also separate loads from their subsequent uses. The base cases return the original argument; other calls return the sum of the two recursive values. Every return restores the saved return address and exactly undoes the stack allocation.[ref]
The stated domain is nonnegative integer , with sufficient stack space and a result representable in the intended integer type. The non-trapping additions implement that mathematical result within range; beyond it, machine wraparound would require a separately specified integer semantics.