[Bug target/106484] Failure to optimize uint64_t/constant division on ARM32

["rsaxvc at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106484

--- Comment #5 from rsaxvc at gmail dot com ---
Related ticket requested with Clang:
https://github.com/llvm/llvm-project/issues/63731

latest umulh() function is a little shorter:

static uint64_t umulh(uint64_t a, uint64_t b)
{
const uint32_t a_lo = a;
const uint32_t a_hi = a >> 32;
const uint32_t b_lo = b;
const uint32_t b_hi = b >> 32;

/* FOIL method of multiplication
See https://en.wikipedia.org/wiki/FOIL_method,
but instead of binomials with constants a,b,c,d variables x,y: (ax+b) * (cy +
d),
consider it with variables a,b,c,d, constants x,y = 1<<32
Results in one UMULL or UMLAL(when merged with accumulate below) per multiply*/
const uint64_t acc0 = (uint64_t)a_lo * b_lo;
const uint64_t acc1 = (uint64_t)a_hi * b_lo;
const uint64_t acc2 = (uint64_t)a_lo * b_hi;
const uint64_t acc3 = (uint64_t)a_hi * b_hi;

/* Accumulate the results, keeping only the top 64-bits of the 128-bit result*/
uint64_t acc = acc0;
acc >>= 32;
acc += acc1;
acc += acc2;
acc >>= 32;
acc += acc3;

return acc;
}