veneer 0.2.3

A very thin std-like library that doesn't depend on libc
Documentation 
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

// On most modern Intel and AMD processors, "rep movsq" and "rep stosq" have
// been enhanced to perform better than an simple qword loop, making them ideal
// for implementing memcpy/memset. Note that "rep cmps" has received no such
// enhancement, so it is not used to implement memcmp.
//
// On certain recent Intel processors, "rep movsb" and "rep stosb" have been
// further enhanced to automatically select the best microarchitectural
// implementation based on length and alignment. See the following features from
// the "Intel® 64 and IA-32 Architectures Optimization Reference Manual":
//  - ERMSB - Enhanced REP MOVSB and STOSB (Ivy Bridge and later)
//  - FSRM - Fast Short REP MOV (Ice Lake and later)
//  - Fast Zero-Length MOVSB (On no current hardware)
//  - Fast Short STOSB (On no current hardware)
//
// To simplify things, we switch to using the byte-based variants if the "ermsb"
// feature is present at compile-time. We don't bother detecting other features.
// Note that ERMSB does not enhance the backwards (DF=1) "rep movsb".

use core::{arch::asm, intrinsics, mem};

#[inline(always)]
#[cfg(target_feature = "ermsb")]
pub unsafe fn copy_forward(dest: *mut u8, src: *const u8, count: usize) {
    // FIXME: Use the Intel syntax once we drop LLVM 9 support on rust-lang/rust.
    core::arch::asm!(
        "repe movsb (%rsi), (%rdi)",
        inout("rcx") count => _,
        inout("rdi") dest => _,
        inout("rsi") src => _,
        options(att_syntax, nostack, preserves_flags)
    );
}

#[inline(always)]
#[cfg(not(target_feature = "ermsb"))]
pub unsafe fn copy_forward(mut dest: *mut u8, mut src: *const u8, count: usize) {
    let (pre_byte_count, qword_count, byte_count) = rep_param(dest, count);
    // Separating the blocks gives the compiler more freedom to reorder instructions.
    asm!(
        "rep movsb",
        inout("ecx") pre_byte_count => _,
        inout("rdi") dest => dest,
        inout("rsi") src => src,
        options(att_syntax, nostack, preserves_flags)
    );
    asm!(
        "rep movsq",
        inout("rcx") qword_count => _,
        inout("rdi") dest => dest,
        inout("rsi") src => src,
        options(att_syntax, nostack, preserves_flags)
    );
    asm!(
        "rep movsb",
        inout("ecx") byte_count => _,
        inout("rdi") dest => _,
        inout("rsi") src => _,
        options(att_syntax, nostack, preserves_flags)
    );
}

#[inline(always)]
pub unsafe fn copy_backward(dest: *mut u8, src: *const u8, count: usize) {
    let (pre_byte_count, qword_count, byte_count) = rep_param(dest, count);
    // We can't separate this block due to std/cld
    asm!(
        "std",
        "rep movsb",
        "sub $7, %rsi",
        "sub $7, %rdi",
        "mov {qword_count}, %rcx",
        "rep movsq",
        "test {pre_byte_count:e}, {pre_byte_count:e}",
        "add $7, %rsi",
        "add $7, %rdi",
        "mov {pre_byte_count:e}, %ecx",
        "rep movsb",
        "cld",
        pre_byte_count = in(reg) pre_byte_count,
        qword_count = in(reg) qword_count,
        inout("ecx") byte_count => _,
        inout("rdi") dest.add(count - 1) => _,
        inout("rsi") src.add(count - 1) => _,
        // We modify flags, but we restore it afterwards
        options(att_syntax, nostack, preserves_flags)
    );
}

#[inline(always)]
#[cfg(target_feature = "ermsb")]
pub unsafe fn set_bytes(dest: *mut u8, c: u8, count: usize) {
    // FIXME: Use the Intel syntax once we drop LLVM 9 support on rust-lang/rust.
    core::arch::asm!(
        "repe stosb %al, (%rdi)",
        inout("rcx") count => _,
        inout("rdi") dest => _,
        inout("al") c => _,
        options(att_syntax, nostack, preserves_flags)
    )
}

#[inline(always)]
#[cfg(not(target_feature = "ermsb"))]
pub unsafe fn set_bytes(mut dest: *mut u8, c: u8, count: usize) {
    let c = c as u64 * 0x0101_0101_0101_0101;
    let (pre_byte_count, qword_count, byte_count) = rep_param(dest, count);
    // Separating the blocks gives the compiler more freedom to reorder instructions.
    asm!(
        "rep stosb",
        inout("ecx") pre_byte_count => _,
        inout("rdi") dest => dest,
        in("rax") c,
        options(att_syntax, nostack, preserves_flags)
    );
    asm!(
        "rep stosq",
        inout("rcx") qword_count => _,
        inout("rdi") dest => dest,
        in("rax") c,
        options(att_syntax, nostack, preserves_flags)
    );
    asm!(
        "rep stosb",
        inout("ecx") byte_count => _,
        inout("rdi") dest => _,
        in("rax") c,
        options(att_syntax, nostack, preserves_flags)
    );
}

#[inline(always)]
pub unsafe fn compare_bytes(a: *const u8, b: *const u8, n: usize) -> i32 {
    #[inline(always)]
    unsafe fn cmp<T, U, F>(mut a: *const T, mut b: *const T, n: usize, f: F) -> i32
    where
        T: Clone + Copy + Eq,
        U: Clone + Copy + Eq,
        F: FnOnce(*const U, *const U, usize) -> i32,
    {
        // Ensure T is not a ZST.
        const { assert!(mem::size_of::<T>() != 0) };

        let end = a.add(intrinsics::unchecked_div(n, mem::size_of::<T>()));
        while a != end {
            if a.read_unaligned() != b.read_unaligned() {
                return f(a.cast(), b.cast(), mem::size_of::<T>());
            }
            a = a.add(1);
            b = b.add(1);
        }
        f(
            a.cast(),
            b.cast(),
            intrinsics::unchecked_rem(n, mem::size_of::<T>()),
        )
    }
    let c1 = |mut a: *const u8, mut b: *const u8, n| {
        for _ in 0..n {
            if a.read() != b.read() {
                return i32::from(a.read()) - i32::from(b.read());
            }
            a = a.add(1);
            b = b.add(1);
        }
        0
    };
    let c2 = |a: *const u16, b, n| cmp(a, b, n, c1);
    let c4 = |a: *const u32, b, n| cmp(a, b, n, c2);
    let c8 = |a: *const u64, b, n| cmp(a, b, n, c4);
    let c16 = |a: *const u128, b, n| cmp(a, b, n, c8);
    c16(a.cast(), b.cast(), n)
}

/// Determine optimal parameters for a `rep` instruction.
fn rep_param(dest: *mut u8, mut count: usize) -> (usize, usize, usize) {
    // Unaligned writes are still slow on modern processors, so align the destination address.
    let pre_byte_count = ((8 - (dest as usize & 0b111)) & 0b111).min(count);
    count -= pre_byte_count;
    let qword_count = count >> 3;
    let byte_count = count & 0b111;
    (pre_byte_count, qword_count, byte_count)
}