https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122294
Bug ID: 122294
Summary: Suboptimal code for several implementations of
memalignment
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: peter0x44 at disroot dot org
Target Milestone: ---
For the following code:
#include <cstdint>
#include <stdlib.h>
size_t memalignment1(void *ptr)
{
return (size_t)ptr & -(size_t)ptr;
}
size_t memalignment2(void *ptr)
{
if (ptr == NULL)
return 0;
return (size_t)ptr & -(size_t)ptr;
}
size_t memalignment3(void *ptr)
{
uintptr_t addr;
int trailing_zeros;
if (ptr == NULL)
return 0;
addr = (uintptr_t)ptr;
trailing_zeros = __builtin_ctzll((unsigned long long)addr);
return (size_t)1 << trailing_zeros;
}
gcc emits at -Os:
memalignment1(void*):
mov rax, rdi
neg rax
and rax, rdi
ret
memalignment2(void*):
mov rax, rdi
xor edx, edx
neg rax
and rax, rdi
test rdi, rdi
cmove rax, rdx
ret
memalignment3(void*):
bsf rcx, rdi
mov eax, 1
xor edx, edx
sal rax, cl
test rdi, rdi
cmove rax, rdx
ret
Only the oneliner implementation generates optimal code. The rest are bad.
clang generates identical code to gcc for the oneliner, for all 3
implementations.
https://gcc.godbolt.org/z/qKhr5T8aq
