what's the difference between __builtin_popcountll and_mm_popcnt_u64?

Question

I was trying to how many 1 in 512MB memory and I found two possible methods, _mm_popcnt_u64() and __builtin_popcountll() in the gcc builtins.

_mm_popcnt_u64() is said to use the CPU introduction SSE4.2，which seems to be the fastest, and __builtin_popcountll() is excepted to use table lookup.

So, I think __builtin_popcountll() should be little slower than _mm_popcnt_u64().

However I got a result like this:

It took almost the same time for two methods. I highly doubt that they used the same way to work.

I also got this in popcntintrin.h

/* Calculate a number of bits set to 1. */
extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial___))
_mm_popcnt_u32 (unsigned int __X)
{
  return __builtin_popcount (__X);
}

#ifdef __x86_64__
extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_popcnt_u64 (unsigned long long __X)
{
  return __builtin_popcountll (__X);
}
#endif

So, I'm confused how __builtin_popcountll() works on earth

Answer 1

_mm_popcnt_u64 is part of <nmmintrin.h>, a header devised by Intel for utility functions for accessing SSE 4.2 instructions.

__builtin_popcountll is a GCC extension.

_mm_popcnt_u64 is portable to non-GNU compilers, and __builtin_popcountll is portable to non-SSE-4.2 CPUs. But on systems where both are available, both should compile to the exact same code.

Answer 2

If You compile without march flag, so with x86_64 default, builtin should be slower because it needs to dispatch function selecting between different architectures. This will cause no inlining and additional condition.

Answer 3

Besides any other consideration you cannot just time a loop and use the raw number for any meaningful benchmark.

As a rule of thumb, if it doesn't have error bars (variance) it is not a proper measure of anything. Next time try running your benchmarks 10 times (or 1000) each, compute the average and standard deviations, and make sure one of the results is better/worse than the other with high statistical confidence, i.e. > 99.9%.

https://en.wikipedia.org/wiki/Standard_deviation#Estimation

And as a side note, a 0.1% difference in a benchmark should usually be considered statistical noise, especially if you are measuring CPU instrinsics or any other function that takes under 100 cycles to execute.