std.mem.swap is slow #25334
Description
Zig Version
0.15.1
Steps to Reproduce and Observed Behavior
Zig
const std = @import("std");
const Huge = struct {
data: [128]u128,
};
export fn f(x: *Huge, y: *Huge) void {
std.mem.swap(Huge, x, y);
}f:
push rbp
mov rbp, rsp
push r15
push r14
push rbx
sub rsp, 2056
lea r15, [rbp - 2080]
mov edx, 2048
mov r14, rdi
mov rbx, rsi
mov rsi, r14
mov rdi, r15
call memcpy@PLT
mov edx, 2048
mov rdi, r14
mov rsi, rbx
call memcpy@PLT
mov edx, 2048
mov rdi, rbx
mov rsi, r15
call memcpy@PLT
add rsp, 2056
pop rbx
pop r14
pop r15
pop rbp
retRust
https://godbolt.org/z/a7dv8jx7a
struct Huge([u128; 128]);
pub extern fn f(x: &mut Huge, y: &mut Huge) {
std::mem::swap(x, y);
}example::f::h4bf74e253b89ab56:
xor eax, eax
.LBB0_1:
movaps xmm0, xmmword ptr [rdi + 8*rax]
movaps xmm1, xmmword ptr [rdi + 8*rax + 16]
movaps xmm2, xmmword ptr [rsi + 8*rax]
movaps xmm3, xmmword ptr [rsi + 8*rax + 16]
movaps xmmword ptr [rdi + 8*rax], xmm2
movaps xmmword ptr [rdi + 8*rax + 16], xmm3
movaps xmmword ptr [rsi + 8*rax], xmm0
movaps xmmword ptr [rsi + 8*rax + 16], xmm1
movaps xmm0, xmmword ptr [rdi + 8*rax + 32]
movaps xmm1, xmmword ptr [rdi + 8*rax + 48]
movaps xmm2, xmmword ptr [rsi + 8*rax + 32]
movaps xmm3, xmmword ptr [rsi + 8*rax + 48]
movaps xmmword ptr [rdi + 8*rax + 32], xmm2
movaps xmmword ptr [rdi + 8*rax + 48], xmm3
movaps xmmword ptr [rsi + 8*rax + 32], xmm0
movaps xmmword ptr [rsi + 8*rax + 48], xmm1
add rax, 8
cmp rax, 256
jne .LBB0_1
retZig uses three memcpys, while Rust just shuffles some bytes in registers. I think the two implementations actually differ. They give different results if two Huge structs overlap! Zig assumes aliasing, and gives suboptimal code, but it rather should assert that no aliasing happens (adding noalias annotations doesn't change anything).
Expected Behavior
Zig should use O(1) extra memory to swap two structs of size N, with a no-alias safety check in std.mem.swap