https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125525
Bug ID: 125525
Summary: equivalent signed zero detection produces different
assembler
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: manu at gcc dot gnu.org
Target Milestone: ---
When implementing radix sort for floating-point values, it is common practice
to canonicalize -0.0 and +0.0 to 0.0. There are at least two ways to do this:
#include <stddef.h>
#include <stdint.h>
#include <string.h>
static inline uint64_t
load_u64_from_double(double x)
{
uint64_t u;
memcpy(&u, &x, sizeof(u));
return u;
}
uint64_t
double_to_sort_key_asc(double x)
{
uint64_t u = load_u64_from_double(x);
if ((u << 1) == 0)
u = 0;
uint64_t mask = (uint64_t)-(int64_t)(u >> 63) |
UINT64_C(0x8000000000000000);
return u ^ mask;
}
uint64_t
double_to_sort_key_asc2(double x)
{
uint64_t u = load_u64_from_double(x);
uint64_t is_zero = ((u << 1) == 0);
u &= ~(is_zero << 63); // clear sign bit only for zero
uint64_t mask = (uint64_t)-(int64_t)(u >> 63) |
UINT64_C(0x8000000000000000);
return u ^ mask;
}
I expected that both functions would produce the same assembler, however, -O2
-march=x86-64-v2 produces:
"double_to_sort_key_asc":
movabs rcx, -9223372036854775808
movq rax, xmm0
cqo
or rdx, rcx
xor rdx, rax
add rax, rax
cmovne rcx, rdx
mov rax, rcx
ret
"double_to_sort_key_asc2":
movq rax, xmm0
xor edx, edx
mov rcx, rax
add rcx, rcx
sete dl
sal rdx, 63
not rdx
and rdx, rax
mov rax, rdx
sar rax, 63
bts rax, 63
xor rax, rdx
ret
Note that GCC < 16.1 produced a conditional jump for the first variant but it
now produces cmovne. Using __builtin_expect to tell GCC 16.1 that the condition
is likely false produces again a conditional jump. Clang also produces a
conditional jump.
