llvm/llvm-project
View on GitHubImplement `clz` using floating point math if `clz` instruction is not available
Open
#161,746 opened on Oct 2, 2025
good first issuellvm:SelectionDAGmissed-optimization
Repository metrics
- Stars
- (26,378 stars)
- PR merge metrics
- (Avg merge 1d 2h) (1,000 merged PRs in 30d)
Description
In https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogIEEE64Float we can find a neat trick to calculate the ilog2 of a u32 with floating point math:
u32 ilog2_u32(u32 x) {
// zexto extend to 64 bits
u64 x_u64 = x;
// set exponent to 52
x_u64 |= (52 + 1023ULL) << 52;
// move into f64 register
double f = __builtin_bit_cast(double, x_u64);
// subtract 2^52
f -= 0x1.0p52;
// move back into u64 register
x_u64 = __builtin_bit_cast(u64, f);
// extract exponent
return (x_u64 >> 52) - 1023;
}
Rearrange the equation:
ilog2(x) == 31 - clz(x)
clz(x) + ilog2(x) == 31
clz(x) == 31 - ilog2(x)
and we get:
u32 src32(u32 x) { return __builtin_clzg(x); }
u32 tgt32(u32 x) { return 31 - ilog2_u32(x); }
This generalizes to u8 and u16 using float instead of double
https://godbolt.org/z/cW8js3MK9 https://alive2.llvm.org/ce/z/-bK_nC