Without the volatile attribute, compilers are entitled to optimize out
the same asm(). In the case of __udelay() in syscounter.c, it calls
`get_ticks()` twice, one for the starting time and the second in the
loop to check the current time. When compilers inline `get_ticks()`
they see the same `mrrc` instructions and optimize out the second one.
This leads to infinite loop since we don't get updated value from the
system counter.