gistfile1.txt

/* Usage: gcc -DCPU_CLOCK_FREQUENCY=1008000000 test.S && time ./a.out
 */
	.cpu cortex-a53+fp+simd
	.file	"1.c"
	.text
	.align	2
	.global	main
	.type	main, %function
main:

#ifndef CPU_CLOCK_FREQUENCY
#error CPU_CLOCK_FREQUENCY must be defined
#endif

#define LOOP_UNROLL_FACTOR   30

        sub         sp, sp, #128
        add         x2, x2, #16
        bic         x2, x2, #15
        add         x3, x2, #64

        ldr         x1, =(CPU_CLOCK_FREQUENCY / LOOP_UNROLL_FACTOR)
        b           1f

    .balign 64
1:
    .rept LOOP_UNROLL_FACTOR
        add    x2, x2, x2
    .endr
        subs        x1, x1, #1
        bne         1b

        add         sp, sp, #128
        mov         w0, 0
        ret