| 1 | const std = @import("std"); |
| 2 | const assert = std.debug.assert; |
| 3 | const compiler_rt = @import("../compiler_rt.zig"); |
| 4 | const builtin = @import("builtin"); |
| 5 | |
| 6 | comptime { |
| 7 | if (builtin.object_format != .c) { |
| 8 | const export_options: std.builtin.ExportOptions = .{ |
| 9 | .name = "memcpy", |
| 10 | .linkage = compiler_rt.linkage, |
| 11 | .visibility = compiler_rt.visibility, |
| 12 | }; |
| 13 | |
| 14 | if (builtin.mode == .small or builtin.zig_backend == .stage2_aarch64) |
| 15 | @export(&memcpySmall, export_options) |
| 16 | else |
| 17 | @export(&memcpyFast, export_options); |
| 18 | } |
| 19 | } |
| 20 | |
| 21 | const Element = compiler_rt.PreferredLoadStoreElement; |
| 22 | |
| 23 | comptime { |
| 24 | assert(std.math.isPowerOfTwo(@sizeOf(Element))); |
| 25 | } |
| 26 | |
| 27 | fn memcpySmall(noalias dest: ?[*]u8, noalias src: ?[*]const u8, len: usize) callconv(.c) ?[*]u8 { |
| 28 | @setRuntimeSafety(false); |
| 29 | |
| 30 | for (0..len) |i| { |
| 31 | dest.?[i] = src.?[i]; |
| 32 | } |
| 33 | |
| 34 | return dest; |
| 35 | } |
| 36 | |
| 37 | fn memcpyFast(noalias dest: ?[*]u8, noalias src: ?[*]const u8, len: usize) callconv(.c) ?[*]u8 { |
| 38 | @setRuntimeSafety(false); |
| 39 | |
| 40 | const small_limit = 2 * @sizeOf(Element); |
| 41 | |
| 42 | if (copySmallLength(small_limit, dest.?, src.?, len)) return dest; |
| 43 | |
| 44 | copyForwards(dest.?, src.?, len); |
| 45 | |
| 46 | return dest; |
| 47 | } |
| 48 | |
| 49 | inline fn copySmallLength( |
| 50 | comptime small_limit: comptime_int, |
| 51 | dest: [*]u8, |
| 52 | src: [*]const u8, |
| 53 | len: usize, |
| 54 | ) bool { |
| 55 | if (len < 16) { |
| 56 | copyLessThan16(dest, src, len); |
| 57 | return true; |
| 58 | } |
| 59 | |
| 60 | if (comptime 2 < (std.math.log2(small_limit) + 1) / 2) { |
| 61 | if (copy16ToSmallLimit(small_limit, dest, src, len)) return true; |
| 62 | } |
| 63 | |
| 64 | return false; |
| 65 | } |
| 66 | |
| 67 | inline fn copyLessThan16( |
| 68 | dest: [*]u8, |
| 69 | src: [*]const u8, |
| 70 | len: usize, |
| 71 | ) void { |
| 72 | @setRuntimeSafety(false); |
| 73 | if (len < 4) { |
| 74 | if (len == 0) return; |
| 75 | dest[0] = src[0]; |
| 76 | dest[len / 2] = src[len / 2]; |
| 77 | dest[len - 1] = src[len - 1]; |
| 78 | return; |
| 79 | } |
| 80 | copyRange4(4, dest, src, len); |
| 81 | } |
| 82 | |
| 83 | inline fn copy16ToSmallLimit( |
| 84 | comptime small_limit: comptime_int, |
| 85 | dest: [*]u8, |
| 86 | src: [*]const u8, |
| 87 | len: usize, |
| 88 | ) bool { |
| 89 | @setRuntimeSafety(false); |
| 90 | inline for (2..(std.math.log2(small_limit) + 1) / 2 + 1) |p| { |
| 91 | const limit = 1 << (2 * p); |
| 92 | if (len < limit) { |
| 93 | copyRange4(limit / 4, dest, src, len); |
| 94 | return true; |
| 95 | } |
| 96 | } |
| 97 | return false; |
| 98 | } |
| 99 | |
| 100 | inline fn copyForwards( |
| 101 | noalias dest: [*]u8, |
| 102 | noalias src: [*]const u8, |
| 103 | len: usize, |
| 104 | ) void { |
| 105 | @setRuntimeSafety(false); |
| 106 | |
| 107 | copyFixedLength(dest, src, @sizeOf(Element)); |
| 108 | const alignment_offset = @alignOf(Element) - @intFromPtr(src) % @alignOf(Element); |
| 109 | const n = len - alignment_offset; |
| 110 | const d = dest + alignment_offset; |
| 111 | const s = src + alignment_offset; |
| 112 | |
| 113 | copyBlocksAlignedSource(@ptrCast(d), @ptrCast(@alignCast(s)), n); |
| 114 | |
| 115 | // copy last `@sizeOf(Element)` bytes unconditionally, since block copy |
| 116 | // methods only copy a multiple of `@sizeOf(Element)` bytes. |
| 117 | const offset = len - @sizeOf(Element); |
| 118 | copyFixedLength(dest + offset, src + offset, @sizeOf(Element)); |
| 119 | } |
| 120 | |
| 121 | inline fn copyBlocksAlignedSource( |
| 122 | noalias dest: [*]align(1) Element, |
| 123 | noalias src: [*]const Element, |
| 124 | max_bytes: usize, |
| 125 | ) void { |
| 126 | copyBlocks(dest, src, max_bytes); |
| 127 | } |
| 128 | |
| 129 | /// Copies the largest multiple of `@sizeOf(T)` bytes from `src` to `dest`, |
| 130 | /// that is less than `max_bytes` where `T` is the child type of `src` and |
| 131 | /// `dest`. |
| 132 | inline fn copyBlocks( |
| 133 | noalias dest: anytype, |
| 134 | noalias src: anytype, |
| 135 | max_bytes: usize, |
| 136 | ) void { |
| 137 | @setRuntimeSafety(false); |
| 138 | |
| 139 | const T = @typeInfo(@TypeOf(dest)).pointer.child; |
| 140 | comptime assert(T == @typeInfo(@TypeOf(src)).pointer.child); |
| 141 | |
| 142 | const loop_count = max_bytes / @sizeOf(T); |
| 143 | |
| 144 | for (dest[0..loop_count], src[0..loop_count]) |*d, s| { |
| 145 | d.* = s; |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | inline fn copyFixedLength( |
| 150 | noalias dest: [*]u8, |
| 151 | noalias src: [*]const u8, |
| 152 | comptime len: comptime_int, |
| 153 | ) void { |
| 154 | @setRuntimeSafety(false); |
| 155 | comptime assert(std.math.isPowerOfTwo(len)); |
| 156 | |
| 157 | const T = if (len >= @sizeOf(Element)) |
| 158 | Element |
| 159 | else if (len > @sizeOf(usize)) |
| 160 | @Vector(len, u8) |
| 161 | else |
| 162 | @Int(.unsigned, len * 8); |
| 163 | |
| 164 | const loop_count = @divExact(len, @sizeOf(T)); |
| 165 | |
| 166 | const d: [*]align(1) T = @ptrCast(dest); |
| 167 | const s: [*]align(1) const T = @ptrCast(src); |
| 168 | |
| 169 | inline for (0..loop_count) |i| { |
| 170 | d[i] = s[i]; |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | /// copy `len` bytes from `src` to `dest`; `len` must be in the range |
| 175 | /// `[copy_len, 4 * copy_len)`. |
| 176 | inline fn copyRange4( |
| 177 | comptime copy_len: comptime_int, |
| 178 | noalias dest: [*]u8, |
| 179 | noalias src: [*]const u8, |
| 180 | len: usize, |
| 181 | ) void { |
| 182 | @setRuntimeSafety(false); |
| 183 | comptime assert(std.math.isPowerOfTwo(copy_len)); |
| 184 | |
| 185 | const a = len & (copy_len * 2); |
| 186 | const b = a / 2; |
| 187 | |
| 188 | const last = len - copy_len; |
| 189 | const pen = last - b; |
| 190 | |
| 191 | copyFixedLength(dest, src, copy_len); |
| 192 | copyFixedLength(dest + b, src + b, copy_len); |
| 193 | copyFixedLength(dest + pen, src + pen, copy_len); |
| 194 | copyFixedLength(dest + last, src + last, copy_len); |
| 195 | } |
| 196 | |
| 197 | fn testMemcpyImpl(comptime memcpyImpl: anytype) !void { |
| 198 | const max_len = 1024; |
| 199 | var buffer: [max_len + @alignOf(Element) - 1]u8 align(@alignOf(Element)) = undefined; |
| 200 | for (&buffer, 0..) |*b, i| { |
| 201 | b.* = @intCast(i % 97); |
| 202 | } |
| 203 | var dest: [max_len + @alignOf(Element) - 1]u8 align(@alignOf(Element)) = undefined; |
| 204 | |
| 205 | for (0..max_len) |copy_len| { |
| 206 | for (0..@alignOf(Element)) |s_offset| { |
| 207 | for (0..@alignOf(Element)) |d_offset| { |
| 208 | @memset(&dest, 0xff); |
| 209 | const s = buffer[s_offset..][0..copy_len]; |
| 210 | const d = dest[d_offset..][0..copy_len]; |
| 211 | _ = memcpyImpl(@ptrCast(d.ptr), @ptrCast(s.ptr), s.len); |
| 212 | std.testing.expectEqualSlices(u8, s, d) catch |e| { |
| 213 | std.debug.print("error encountered for length={d}, s_offset={d}, d_offset={d}\n", .{ |
| 214 | copy_len, s_offset, d_offset, |
| 215 | }); |
| 216 | return e; |
| 217 | }; |
| 218 | } |
| 219 | } |
| 220 | } |
| 221 | } |
| 222 | test memcpySmall { |
| 223 | if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest; |
| 224 | try testMemcpyImpl(memcpySmall); |
| 225 | } |
| 226 | test memcpyFast { |
| 227 | if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest; |
| 228 | try testMemcpyImpl(memcpyFast); |
| 229 | } |