popcount with clang

[michael-roe]

Modern compilers are clever; they can recognize some of the algorithms for popcount and replace them with a popcnt (Intel) or cpop (RISC-V) instruction.

int count(long x) { int v = 0; while(x != 0) { x &= x - 1; v++; } return v; }

This will get turned into popcntq %rdi, %rax by clang (with -O3 -march=x86-64-v2)

This would suggest that in order to get good performance, VOLK's popcount implementation ought to be one of the ones that is recognized by popular compilers.

[jdemel]

Thanks for the hint. Do you have a reference to read? And would you be willing to create a PR with this change?

[michael-roe]

I'm thinking about creating a pull request for this, but I'm trying to get cpu_features working on non-x86 architectures first, so I can test on a CPU that has popcount but isn't x86.