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
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
|
//@ compile-flags: -Copt-level=3 -Z merge-functions=disabled
//@ only-x86_64
//@ min-llvm-version: 20
//@ ignore-std-debug-assertions (`ptr::swap_nonoverlapping` has one which blocks some optimizations)
#![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 [6 x i8], 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) {
// CHECK-NOT: alloca
// Swapping `i48` might be cleaner in LLVM-IR here, but `i32`+`i16` isn't bad,
// and is closer to the assembly it generates anyway.
// CHECK-NOT: load{{ }}
// CHECK: load i32{{.+}}align 2
// CHECK-NEXT: load i32{{.+}}align 2
// CHECK-NEXT: store i32{{.+}}align 2
// CHECK-NEXT: store i32{{.+}}align 2
// CHECK: load i16{{.+}}align 2
// CHECK-NEXT: load i16{{.+}}align 2
// CHECK-NEXT: store i16{{.+}}align 2
// CHECK-NEXT: store i16{{.+}}align 2
// CHECK-NOT: store{{ }}
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 i64, ptr %x, align 2
// CHECK-DAG: %[[YVAL:.+]] = load i64, ptr %y, align 2
// CHECK-DAG: store i64 %[[YVAL]], ptr %x, align 2
// CHECK-DAG: store i64 %[[XVAL]], ptr %y, align 2
swap(x, y)
}
// CHECK-LABEL: @swap_vecs
#[no_mangle]
pub fn swap_vecs(x: &mut Vec<u32>, y: &mut Vec<u32>) {
// CHECK-NOT: alloca
// There are plenty more loads and stores than just these,
// but at least one sure better be 64-bit (for size or capacity).
// CHECK: load i64
// CHECK: load i64
// CHECK: store i64
// CHECK: store i64
// CHECK: ret void
swap(x, y)
}
// CHECK-LABEL: @swap_slices
#[no_mangle]
pub fn swap_slices<'a>(x: &mut &'a [u32], y: &mut &'a [u32]) {
// CHECK-NOT: alloca
// CHECK: load ptr
// CHECK: load i64
// CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 16, i1 false)
// CHECK: store ptr
// CHECK: store i64
swap(x, y)
}
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: mul nuw nsw i64 %{{x|y}}.1, 3
// CHECK: load <{{[0-9]+}} x i64>
// CHECK: store <{{[0-9]+}} x i64>
// CHECK-COUNT-2: load i32
// CHECK-COUNT-2: store i32
// CHECK-COUNT-2: load i16
// CHECK-COUNT-2: store i16
// CHECK-COUNT-2: load i8
// CHECK-COUNT-2: store i8
if x.len() == y.len() {
x.swap_with_slice(y);
}
}
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
// Because the size in bytes in a multiple of 4, we can skip the smallest sizes.
// CHECK: load <{{[0-9]+}} x i64>
// CHECK: store <{{[0-9]+}} x i64>
// CHECK-COUNT-2: load i32
// CHECK-COUNT-2: store i32
// CHECK-NOT: load i16
// CHECK-NOT: store i16
// CHECK-NOT: load i8
// CHECK-NOT: store i8
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>{{.+}}, align 8,
// CHECK: store <{{[0-9]+}} x i64>{{.+}}, align 8,
if x.len() == y.len() {
x.swap_with_slice(y);
}
}
#[repr(C, packed)]
pub struct Packed {
pub first: bool,
pub second: usize,
}
// CHECK-LABEL: @swap_packed_structs
#[no_mangle]
pub fn swap_packed_structs(x: &mut Packed, y: &mut Packed) {
// CHECK-NOT: alloca
// CHECK-NOT: load
// CHECK-NOT: store
// CHECK: %[[A:.+]] = load i64, ptr %x, align 1,
// CHECK-NEXT: %[[B:.+]] = load i64, ptr %y, align 1,
// CHECK-NEXT: store i64 %[[B]], ptr %x, align 1,
// CHECK-NEXT: store i64 %[[A]], ptr %y, align 1,
// CHECK-NOT: load
// CHECK-NOT: store
// CHECK: %[[C:.+]] = load i8, ptr %[[X8:.+]], align 1,
// CHECK-NEXT: %[[D:.+]] = load i8, ptr %[[Y8:.+]], align 1,
// CHECK-NEXT: store i8 %[[D]], ptr %[[X8]], align 1,
// CHECK-NEXT: store i8 %[[C]], ptr %[[Y8]], align 1,
// CHECK-NOT: load
// CHECK-NOT: store
// CHECK: ret void
swap(x, y)
}
|
