Replace the "clever trick" with the popcount intrinsic

[quad]

Both GCC and clang have long supported the __builtin_popcount family of bit operation builtins.

Let's use them!

(One C23 day, we could #include <stdbit.h> and stdc_count_ones.)

---

I ran this test to confirm identical behaviour:

#include <assert.h>
#include <stdint.h>
#include <stdio.h>

static int
old_count_set_bits(uint32_t x)
{
    register uint32_t i = x;	/* take no chances! */

    /* clever trick for adding bits together in parallel to count them */
    i = ((i & 0xAAAAAAAA) >> 1) + (i & ~0xAAAAAAAA);
    i = ((i & 0xCCCCCCCC) >> 2) + (i & ~0xCCCCCCCC);
    i = ((i & 0xF0F0F0F0) >> 4) + (i & ~0xF0F0F0F0);
    i = ((i & 0xFF00FF00) >> 8) + (i & ~0xFF00FF00);
    i = ((i & 0xFFFF0000) >> 16) + (i & ~0xFFFF0000);

    return i;
}

static int
new_count_set_bits(uint32_t x)
{
    return __builtin_popcountg(x);
}

int
main(void)
{
    for (uint32_t x = 0; x < UINT32_MAX; x++)
    {
        assert(old_count_set_bits(x) == new_count_set_bits(x));

        if (0 == (x & (UINT32_MAX >> 4)))
        {
            printf(
                "%llu%% (%x / %x)\n",
                ((uint64_t) x * 100) / UINT32_MAX,
                x,
                UINT32_MAX
            );
        }
    }
}

It's still magic that computers can now count to UINT32_MAX, in seconds!

[wrog]

It's still magic that computers can now count to UINT32_MAX, in seconds

Yeah, sometimes it's nice to be living in The Future.

[wrog]

So, for this, I would want the autoconf thingie that checks whether the compiler can actually do this (is this in C11?) on the platform in question -- currently we have something known to work everywhere and I'd like to preserve that state of affairs, even if we almost certainly want to use the builtin if it's available -- and then be #ifdeffing out the guts of count_set_bits (and maybe making it conditionally inline cf. how Unicode routines in utf.h/utf-ctype.h work) rather than replacing it everywhere.

Also now wondering about similar issues in bf_random() (see numbers.c) which might then make this worth a new fake system header...

[quad]

popcount isn’t in C11, sadly. Is that our minimum supported standard? Or are you thinking to maybe carve off a place to put functions like this that enable backwards compat? Regardless, I’ll give a crack at the m4 madness when I get a free moment!

…

On Wed, Oct 15, 2025, at 02:16, Roger Crew wrote: *wrog* left a comment (wrog/lambdamoo#19) <#19 (comment)> So, for this, I would want the autoconf thingie that checks whether the compiler can actually do this (is this in C11?) on the platform in question, and then be #ifdeffing out the guts of count_set_bits rather than replacing it everywhere. Also now wondering about similar issues in bf_random() (see numbers.c) which might then make this worth a new fake system header... — Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAF3WPGYLEZMDR2N7GTULL3XUHTDAVCNFSM6AAAAACILNSDNSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTIMBSGM3TKMZUGE>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[wrog]

Right now we're at C99. I think I could be talked into C11 (since that's been 15 years and we've actually already got one POSIX 2007ism that I know of [the %ms in scxnf]) but I at least need to see what's in there. Baby steps :-)

[quad]

I will never get around to taking a crack at the m4 madness because, as it turns out, becoming a father has made me bereft of free moments. (unsubscribing!)