beagleboard 500MHz

Calls to either x(500000000), y(500000000) or z(500000000) take 6 seconds. So it's 6 cycles per loop iteration everywhere.

Checking for condition and branching in the next instruction seems to pair well?

This agrees with Cortex-A8 TRM:

"Conditional branches

Conditional branches are executed differently than other conditional instructions. Most
conditional instructions take the destination register as an additional source and the
condition codes are resolved in E2. Branches do not require the destination register, PC,
as an additional source because they already use the PC as a source. They are also
different than normal conditional operations because the flags resolve the condition
codes in E3 rather than E2. This enables the pairing of a flag setting instruction and a
branch in the same cycle."

///////////////////////////////////////

    .text
    .global x
    .global y
    .global z
x:
    mov  r1, #1
1:
    add  r0, r0, #0
    add  r0, r0, #0
    cmp  r1, #0
    beq  unreachable_label
    add  r0, r0, #0
    add  r0, r0, #0
    subs r0, r0, #1
    bgt  1b
    bx   lr

y:
    mov  r1, #1
1:
    add  r0, r0, #0
    add  r0, r0, #0
    add  r0, r0, #0
    add  r0, r0, #0
    add  r0, r0, #0
    subs r0, r0, #1
    bgt  1b
    bx   lr

z:
    mov  r1, #1
1:
    add  r0, r0, #0
    add  r0, r0, #0
    subs r0, r0, #1
    add  r0, r0, #0
    add  r0, r0, #0
    add  r0, r0, #0
    bgt  1b
    bx   lr