@ ARM version of trin.S, q.v.  The point here is to see how
        @ much of a slowdown dumbjit.c causes.  This takes 4.6 seconds
        @ on my Ryzen laptop, while the equivalent program compiled
        @ with dumbjit takes 14-15 seconds.
        .equ N, 100*1000
        .equ M, 10*1000
        .globl main
        .thumb_func
        .thumb
main:   push {r4, lr}
        ldr r4, =N

1:      ldr r0, =M
        bl tri
        sub r4, #1
        bne 1b

        mov r1, r0
        ldr r0, =format
        bl printf

        movs r1, #0
        pop {r4, pc}
        .ltorg
format: .asciz "%d\n"


tri:    movs r1, #0
        tst r0, r0
        beq 1f

2:      add r1, r0
        sub r0, #1
        bne 2b

1:      mov r0, r1
        bx lr