1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
|
// compile-flags: -O -Z merge-functions=disabled
// only-x86_64
// ignore-debug: the debug assertions get in the way
#![crate_type = "lib"]
use std::mem::swap;
type RGB48 = [u16; 3];
// CHECK-LABEL: @swap_rgb48_manually(
#[no_mangle]
pub fn swap_rgb48_manually(x: &mut RGB48, y: &mut RGB48) {
// FIXME: See #115212 for why this has an alloca again
// CHECK: alloca [3 x i16], align 2
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 6, i1 false)
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 6, i1 false)
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 6, i1 false)
let temp = *x;
*x = *y;
*y = temp;
}
// CHECK-LABEL: @swap_rgb48
#[no_mangle]
pub fn swap_rgb48(x: &mut RGB48, y: &mut RGB48) {
// FIXME: See #115212 for why this has an alloca again
// CHECK: alloca [3 x i16], align 2
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 6, i1 false)
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 6, i1 false)
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 6, i1 false)
swap(x, y)
}
type RGBA64 = [u16; 4];
// CHECK-LABEL: @swap_rgba64
#[no_mangle]
pub fn swap_rgba64(x: &mut RGBA64, y: &mut RGBA64) {
// CHECK-NOT: alloca
// CHECK-DAG: %[[XVAL:.+]] = load <4 x i16>, ptr %x, align 2
// CHECK-DAG: %[[YVAL:.+]] = load <4 x i16>, ptr %y, align 2
// CHECK-DAG: store <4 x i16> %[[YVAL]], ptr %x, align 2
// CHECK-DAG: store <4 x i16> %[[XVAL]], ptr %y, align 2
swap(x, y)
}
// LLVM doesn't vectorize a loop over 3-byte elements,
// so we chunk it down to bytes and loop over those instead.
type RGB24 = [u8; 3];
// CHECK-LABEL: @swap_rgb24_slices
#[no_mangle]
pub fn swap_rgb24_slices(x: &mut [RGB24], y: &mut [RGB24]) {
// CHECK-NOT: alloca
// CHECK: load <{{[0-9]+}} x i8>
// CHECK: store <{{[0-9]+}} x i8>
if x.len() == y.len() {
x.swap_with_slice(y);
}
}
// This one has a power-of-two size, so we iterate over it directly
type RGBA32 = [u8; 4];
// CHECK-LABEL: @swap_rgba32_slices
#[no_mangle]
pub fn swap_rgba32_slices(x: &mut [RGBA32], y: &mut [RGBA32]) {
// CHECK-NOT: alloca
// CHECK: load <{{[0-9]+}} x i32>
// CHECK: store <{{[0-9]+}} x i32>
if x.len() == y.len() {
x.swap_with_slice(y);
}
}
// Strings have a non-power-of-two size, but have pointer alignment,
// so we swap usizes instead of dropping all the way down to bytes.
const _: () = assert!(!std::mem::size_of::<String>().is_power_of_two());
// CHECK-LABEL: @swap_string_slices
#[no_mangle]
pub fn swap_string_slices(x: &mut [String], y: &mut [String]) {
// CHECK-NOT: alloca
// CHECK: load <{{[0-9]+}} x i64>
// CHECK: store <{{[0-9]+}} x i64>
if x.len() == y.len() {
x.swap_with_slice(y);
}
}
|
