[Missed optimization] `has_single_bit` when the value is known to be non-zero

[MitalAshok]

bool test1(unsigned n) {
    __builtin_assume(n);
    return __builtin_popcount(n) == 1;
}

bool test2(unsigned n) {
    __builtin_assume(n);
    return (n != 0) && ((n & (n-1u)) == 0);
}

Where test1 looks like libstdc++'s std::has_single_bit and test2 looks like libc++'s std::has_single_bit.
These are optimized differently with the non-zero assumption: https://godbolt.org/z/b8jTae91x

test1:                                  # @test1
        popcnt  eax, edi
        cmp     eax, 1
        sete    al
        ret
test2:                                  # @test2
        blsr    eax, edi
        sete    al
        ret

(without the assumption, they are both optimized to use popcnt)

return __builtin_popcount(n) <= 1; is optimized to blsr

[dtcxzyw]

As we don't know which form (ctpop(x) <u 2 or == 1) is better for this pattern (See #67268), do we need a new intrinsic function like i1 is_power_of_2(iN %x, i1 zero_is_poison)?

@goldsteinn @nikic @topperc @arsenm @RKSimon

[arsenm]

A new intrinsic is way too heavy of a solution for a very specific optimization

[dtcxzyw]

A new intrinsic is way too heavy of a solution for a very specific optimization

How about adding a new parameter i1 is_zero_poison to ctpop?

[efriedma-quic]

Isn't that the same as call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %x)?

[dtcxzyw]

Isn't that the same as call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %x)?

You are right :)