In particular if I have the following code:
func sum(n: Int, acc: Int) -> Int {
if n == 0 { return acc }
else { return sum(n - 1, acc + n) }
}
Will Swift compiler optimize it to a loop? And does it so in a more interesting case below?
func isOdd(n: Int) -> Bool {
if n == 0 { return false; }
else { return isEven(n - 1) }
}
func isEven(n: Int) -> Bool {
if n == 0 { return true }
else { return isOdd(n - 1) }
}
The best way to check is to examine the assembly language code generated by the compiler. I took the code above and compiled it with:
swift -O3 -S tco.swift >tco.asm
The relevant part of the output
.globl __TF3tco3sumFTSiSi_Si
.align 4, 0x90
__TF3tco3sumFTSiSi_Si:
pushq %rbp
movq %rsp, %rbp
testq %rdi, %rdi
je LBB0_4
.align 4, 0x90
LBB0_1:
movq %rdi, %rax
decq %rax
jo LBB0_5
addq %rdi, %rsi
jo LBB0_5
testq %rax, %rax
movq %rax, %rdi
jne LBB0_1
LBB0_4:
movq %rsi, %rax
popq %rbp
retq
LBB0_5:
ud2
.globl __TF3tco5isOddFSiSb
.align 4, 0x90
__TF3tco5isOddFSiSb:
pushq %rbp
movq %rsp, %rbp
testq %rdi, %rdi
je LBB1_1
decq %rdi
jo LBB1_9
movb $1, %al
LBB1_5:
testq %rdi, %rdi
je LBB1_2
decq %rdi
jo LBB1_9
testq %rdi, %rdi
je LBB1_1
decq %rdi
jno LBB1_5
LBB1_9:
ud2
LBB1_1:
xorl %eax, %eax
LBB1_2:
popq %rbp
retq
.globl __TF3tco6isEvenFSiSb
.align 4, 0x90
__TF3tco6isEvenFSiSb:
pushq %rbp
movq %rsp, %rbp
movb $1, %al
LBB2_1:
testq %rdi, %rdi
je LBB2_5
decq %rdi
jo LBB2_7
testq %rdi, %rdi
je LBB2_4
decq %rdi
jno LBB2_1
LBB2_7:
ud2
LBB2_4:
xorl %eax, %eax
LBB2_5:
popq %rbp
retq
There are no call instructions in the generated code, only conditional jumps (je
/ jne
/ jo
/ jno
). This clearly suggests that Swift does do tail call optimizations in both cases.
In addition, the isOdd
/isEven
functions are interesting in that the compiler not only seems to perform TCO but also inlines the other function in each case.