https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113543
Bug ID: 113543
Summary: Poor codegen for bit-counting functions (countl_zero,
countl_one, countr_zero, countr_one)
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: janschultke at googlemail dot com
Target Milestone: ---
## Code to Reproduce (https://godbolt.org/z/qPeszhaPv)
#include <bit>
template <typename T>
T countr_zero(T x) {
return std::countr_zero(x);
}
template unsigned char countr_zero(unsigned char);
template unsigned short countr_zero(unsigned short);
template unsigned int countr_zero(unsigned int);
template unsigned long countr_zero(unsigned long);
template unsigned long long countr_zero(unsigned long long);
template <typename T>
T countr_one(T x) {
return std::countr_one(x);
}
template unsigned char countr_one(unsigned char);
template unsigned short countr_one(unsigned short);
template unsigned int countr_one(unsigned int);
template unsigned long countr_one(unsigned long);
template unsigned long long countr_one(unsigned long long);
template <typename T>
T countl_zero(T x) {
return std::countl_zero(x);
}
template unsigned char countl_zero(unsigned char);
template unsigned short countl_zero(unsigned short);
template unsigned int countl_zero(unsigned int);
template unsigned long countl_zero(unsigned long);
template unsigned long long countl_zero(unsigned long long);
template <typename T>
T countl_one(T x) {
return std::countl_zero(x);
}
template unsigned char countl_one(unsigned char);
template unsigned short countl_one(unsigned short);
template unsigned int countl_one(unsigned int);
template unsigned long countl_one(unsigned long);
template unsigned long long countl_one(unsigned long long);
## Summary
GCC consistently emits much more code for these function than clang.
For example, GCC:
> unsigned int countl_one<unsigned int>(unsigned int):
> xor eax, eax
> lzcnt eax, edi
> ret
Clang does not emit the extra xor instruction. I don't really know why. LZCNT
has a wide contract and should be equivalent to std::countl_zero.
It gets a lot worse though:
> unsigned short countl_zero<unsigned short>(unsigned short):
> mov eax, 16
> test di, di
> je .L23
> movzx edi, di
> lzcnt edi, edi
> lea eax, [rdi-16]
> .L23:
> ret
I don't really know what all of this schmutz is. Clang emits lzcnt and ret in
this case.
Another bit of disappointing codegen is this:
> unsigned char countr_zero<unsigned char>(unsigned char):
> movzx eax, dil
> xor edx, edx
> tzcnt edx, eax
> test dil, dil
> mov eax, 8
> cmovne eax, edx
> ret
Clang emits:
> or edi, 256
> tzcnt eax, edi
> ret
This clang codegen is very clever. It simply adds a bit on the left, so that
the 32-bit routine can be re-used with only one additional instruction.
