| ... | ... | @@ -1,149 +1,213 @@ |
| 1 | 1 | const std = @import("std"); |
| 2 | const assert = std.debug.assert; |
| 2 | 3 | const common = @import("./common.zig"); |
| 3 | 4 | const builtin = @import("builtin"); |
| 4 | 5 | |
| 5 | 6 | comptime { |
| 6 | 7 | if (builtin.object_format != .c) { |
| 7 | | @export(&memcpy, .{ .name = "memcpy", .linkage = common.linkage, .visibility = common.visibility }); |
| 8 | | @export(&memmove, .{ .name = "memmove", .linkage = common.linkage, .visibility = common.visibility }); |
| 8 | const export_options: std.builtin.ExportOptions = .{ |
| 9 | .name = "memcpy", |
| 10 | .linkage = common.linkage, |
| 11 | .visibility = common.visibility, |
| 12 | }; |
| 13 | |
| 14 | if (builtin.mode == .ReleaseSmall) |
| 15 | @export(&memcpySmall, export_options) |
| 16 | else |
| 17 | @export(&memcpyFast, export_options); |
| 9 | 18 | } |
| 10 | 19 | } |
| 11 | 20 | |
| 12 | | fn memcpy(noalias opt_dest: ?[*]u8, noalias opt_src: ?[*]const u8, len: usize) callconv(.C) ?[*]u8 { |
| 13 | | return memmove(opt_dest, opt_src, len); |
| 21 | const Element = if (std.simd.suggestVectorLength(u8)) |vec_size| |
| 22 | @Type(.{ .vector = .{ |
| 23 | .child = u8, |
| 24 | .len = vec_size, |
| 25 | } }) |
| 26 | else |
| 27 | usize; |
| 28 | |
| 29 | comptime { |
| 30 | assert(@sizeOf(Element) >= @alignOf(Element)); |
| 31 | assert(std.math.isPowerOfTwo(@sizeOf(Element))); |
| 14 | 32 | } |
| 15 | 33 | |
| 16 | | fn memmove(opt_dest: ?[*]u8, opt_src: ?[*]const u8, len: usize) callconv(.C) ?[*]u8 { |
| 17 | | // a port of https://github.com/facebook/folly/blob/1c8bc50e88804e2a7361a57cd9b551dd10f6c5fd/folly/memcpy.S |
| 18 | | if (len == 0) { |
| 19 | | @branchHint(.unlikely); |
| 20 | | return opt_dest; |
| 21 | | } |
| 34 | fn memcpySmall(noalias dest: ?[*]u8, noalias src: ?[*]const u8, len: usize) callconv(.C) ?[*]u8 { |
| 35 | @setRuntimeSafety(builtin.is_test); |
| 22 | 36 | |
| 23 | | const dest = opt_dest.?; |
| 24 | | const src = opt_src.?; |
| 25 | | |
| 26 | | if (len < 8) { |
| 27 | | @branchHint(.unlikely); |
| 28 | | if (len == 1) { |
| 29 | | @branchHint(.unlikely); |
| 30 | | dest[0] = src[0]; |
| 31 | | } else if (len >= 4) { |
| 32 | | @branchHint(.unlikely); |
| 33 | | blockCopy(dest, src, 4, len); |
| 34 | | } else { |
| 35 | | blockCopy(dest, src, 2, len); |
| 36 | | } |
| 37 | | return dest; |
| 37 | for (0..len) |i| { |
| 38 | dest.?[i] = src.?[i]; |
| 38 | 39 | } |
| 39 | 40 | |
| 40 | | if (len > 32) { |
| 41 | | @branchHint(.unlikely); |
| 42 | | if (len > 256) { |
| 43 | | @branchHint(.unlikely); |
| 44 | | copyMove(dest, src, len); |
| 45 | | return dest; |
| 46 | | } |
| 47 | | copyLong(dest, src, len); |
| 48 | | return dest; |
| 49 | | } |
| 41 | return dest; |
| 42 | } |
| 50 | 43 | |
| 51 | | if (len > 16) { |
| 52 | | @branchHint(.unlikely); |
| 53 | | blockCopy(dest, src, 16, len); |
| 54 | | return dest; |
| 55 | | } |
| 44 | fn memcpyFast(noalias dest: ?[*]u8, noalias src: ?[*]const u8, len: usize) callconv(.C) ?[*]u8 { |
| 45 | @setRuntimeSafety(builtin.is_test); |
| 46 | |
| 47 | const small_limit = 2 * @sizeOf(Element); |
| 56 | 48 | |
| 57 | | blockCopy(dest, src, 8, len); |
| 49 | if (copySmallLength(small_limit, dest.?, src.?, len)) return dest; |
| 50 | |
| 51 | copyForwards(dest.?, src.?, len); |
| 58 | 52 | |
| 59 | 53 | return dest; |
| 60 | 54 | } |
| 61 | 55 | |
| 62 | | inline fn blockCopy(dest: [*]u8, src: [*]const u8, block_size: comptime_int, len: usize) void { |
| 63 | | const first = @as(*align(1) const @Vector(block_size, u8), src[0..block_size]).*; |
| 64 | | const second = @as(*align(1) const @Vector(block_size, u8), src[len - block_size ..][0..block_size]).*; |
| 65 | | dest[0..block_size].* = first; |
| 66 | | dest[len - block_size ..][0..block_size].* = second; |
| 67 | | } |
| 56 | inline fn copySmallLength( |
| 57 | comptime small_limit: comptime_int, |
| 58 | dest: [*]u8, |
| 59 | src: [*]const u8, |
| 60 | len: usize, |
| 61 | ) bool { |
| 62 | if (len < 16) { |
| 63 | copyLessThan16(dest, src, len); |
| 64 | return true; |
| 65 | } |
| 68 | 66 | |
| 69 | | inline fn copyLong(dest: [*]u8, src: [*]const u8, len: usize) void { |
| 70 | | var array: [8]@Vector(32, u8) = undefined; |
| 67 | if (comptime 2 < (std.math.log2(small_limit) + 1) / 2) { |
| 68 | if (copy16ToSmallLimit(small_limit, dest, src, len)) return true; |
| 69 | } |
| 71 | 70 | |
| 72 | | inline for (.{ 64, 128, 192, 256 }, 0..) |N, i| { |
| 73 | | array[i * 2] = src[(N / 2) - 32 ..][0..32].*; |
| 74 | | array[(i * 2) + 1] = src[len - N / 2 ..][0..32].*; |
| 71 | return false; |
| 72 | } |
| 75 | 73 | |
| 76 | | if (len <= N) { |
| 77 | | @branchHint(.unlikely); |
| 78 | | for (0..i + 1) |j| { |
| 79 | | dest[j * 32 ..][0..32].* = array[j * 2]; |
| 80 | | dest[len - ((j * 32) + 32) ..][0..32].* = array[(j * 2) + 1]; |
| 81 | | } |
| 82 | | return; |
| 83 | | } |
| 74 | inline fn copyLessThan16( |
| 75 | dest: [*]u8, |
| 76 | src: [*]const u8, |
| 77 | len: usize, |
| 78 | ) void { |
| 79 | @setRuntimeSafety(builtin.is_test); |
| 80 | if (len < 4) { |
| 81 | if (len == 0) return; |
| 82 | dest[0] = src[0]; |
| 83 | dest[len / 2] = src[len / 2]; |
| 84 | dest[len - 1] = src[len - 1]; |
| 85 | return; |
| 84 | 86 | } |
| 87 | copyRange4(4, dest, src, len); |
| 85 | 88 | } |
| 86 | 89 | |
| 87 | | inline fn copyMove(dest: [*]u8, src: [*]const u8, len: usize) void { |
| 88 | | if (@intFromPtr(src) >= @intFromPtr(dest)) { |
| 89 | | @branchHint(.unlikely); |
| 90 | | copyForward(dest, src, len); |
| 91 | | } else if (@intFromPtr(src) + len > @intFromPtr(dest)) { |
| 92 | | @branchHint(.unlikely); |
| 93 | | overlapBwd(dest, src, len); |
| 94 | | } else { |
| 95 | | copyForward(dest, src, len); |
| 90 | inline fn copy16ToSmallLimit( |
| 91 | comptime small_limit: comptime_int, |
| 92 | dest: [*]u8, |
| 93 | src: [*]const u8, |
| 94 | len: usize, |
| 95 | ) bool { |
| 96 | @setRuntimeSafety(builtin.is_test); |
| 97 | inline for (2..(std.math.log2(small_limit) + 1) / 2 + 1) |p| { |
| 98 | const limit = 1 << (2 * p); |
| 99 | if (len < limit) { |
| 100 | copyRange4(limit / 4, dest, src, len); |
| 101 | return true; |
| 102 | } |
| 96 | 103 | } |
| 104 | return false; |
| 97 | 105 | } |
| 98 | 106 | |
| 99 | | inline fn copyForward(dest: [*]u8, src: [*]const u8, len: usize) void { |
| 100 | | const tail: @Vector(32, u8) = src[len - 32 ..][0..32].*; |
| 107 | inline fn copyForwards( |
| 108 | noalias dest: [*]u8, |
| 109 | noalias src: [*]const u8, |
| 110 | len: usize, |
| 111 | ) void { |
| 112 | @setRuntimeSafety(builtin.is_test); |
| 113 | assert(len >= 2 * @sizeOf(Element)); |
| 114 | |
| 115 | dest[0..@sizeOf(Element)].* = src[0..@sizeOf(Element)].*; |
| 116 | const alignment_offset = @alignOf(Element) - @intFromPtr(src) % @alignOf(Element); |
| 117 | const n = len - alignment_offset; |
| 118 | const d = dest + alignment_offset; |
| 119 | const s = src + alignment_offset; |
| 120 | |
| 121 | copyBlocksAlignedSource(@ptrCast(d), @alignCast(@ptrCast(s)), n); |
| 122 | |
| 123 | // copy last `@sizeOf(Element)` bytes unconditionally, since block copy |
| 124 | // methods only copy a multiple of `@sizeOf(Element)` bytes. |
| 125 | const offset = len - @sizeOf(Element); |
| 126 | dest[offset..][0..@sizeOf(Element)].* = src[offset..][0..@sizeOf(Element)].*; |
| 127 | } |
| 101 | 128 | |
| 102 | | const N: usize = len & ~@as(usize, 127); |
| 103 | | var i: usize = 0; |
| 129 | inline fn copyBlocksAlignedSource( |
| 130 | noalias dest: [*]align(1) Element, |
| 131 | noalias src: [*]const Element, |
| 132 | max_bytes: usize, |
| 133 | ) void { |
| 134 | copyBlocks(dest, src, max_bytes); |
| 135 | } |
| 104 | 136 | |
| 105 | | while (i < N) : (i += 128) { |
| 106 | | dest[i..][0..32].* = src[i..][0..32].*; |
| 107 | | dest[i + 32 ..][0..32].* = src[i + 32 ..][0..32].*; |
| 108 | | dest[i + 64 ..][0..32].* = src[i + 64 ..][0..32].*; |
| 109 | | dest[i + 96 ..][0..32].* = src[i + 96 ..][0..32].*; |
| 110 | | } |
| 137 | /// Copies the largest multiple of `@sizeOf(T)` bytes from `src` to `dest`, |
| 138 | /// that is less than `max_bytes` where `T` is the child type of `src` and |
| 139 | /// `dest`. |
| 140 | inline fn copyBlocks( |
| 141 | noalias dest: anytype, |
| 142 | noalias src: anytype, |
| 143 | max_bytes: usize, |
| 144 | ) void { |
| 145 | @setRuntimeSafety(builtin.is_test); |
| 146 | |
| 147 | const T = @typeInfo(@TypeOf(dest)).pointer.child; |
| 148 | comptime assert(T == @typeInfo(@TypeOf(src)).pointer.child); |
| 111 | 149 | |
| 112 | | if (len - i <= 32) { |
| 113 | | @branchHint(.unlikely); |
| 114 | | dest[len - 32 ..][0..32].* = tail; |
| 115 | | } else { |
| 116 | | copyLong(dest[i..], src[i..], len - i); |
| 150 | const loop_count = max_bytes / @sizeOf(T); |
| 151 | |
| 152 | for (dest[0..loop_count], src[0..loop_count]) |*d, s| { |
| 153 | d.* = s; |
| 117 | 154 | } |
| 118 | 155 | } |
| 119 | 156 | |
| 120 | | inline fn overlapBwd(dest: [*]u8, src: [*]const u8, len: usize) void { |
| 121 | | var array: [5]@Vector(32, u8) = undefined; |
| 122 | | array[0] = src[len - 32 ..][0..32].*; |
| 123 | | inline for (1..5) |i| array[i] = src[(i - 1) << 5 ..][0..32].*; |
| 124 | | |
| 125 | | const end: usize = (@intFromPtr(dest) + len - 32) & 31; |
| 126 | | const range = len - end; |
| 127 | | var s = src + range; |
| 128 | | var d = dest + range; |
| 129 | | |
| 130 | | while (@intFromPtr(s) > @intFromPtr(src + 128)) { |
| 131 | | // zig fmt: off |
| 132 | | const first = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 32)).*; |
| 133 | | const second = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 64)).*; |
| 134 | | const third = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 96)).*; |
| 135 | | const fourth = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 128)).*; |
| 136 | | |
| 137 | | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 32))).* = first; |
| 138 | | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 64))).* = second; |
| 139 | | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 96))).* = third; |
| 140 | | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 128))).* = fourth; |
| 141 | | // zig fmt: on |
| 142 | | |
| 143 | | s -= 128; |
| 144 | | d -= 128; |
| 145 | | } |
| 157 | /// copy `len` bytes from `src` to `dest`; `len` must be in the range |
| 158 | /// `[copy_len, 4 * copy_len)`. |
| 159 | inline fn copyRange4( |
| 160 | comptime copy_len: comptime_int, |
| 161 | noalias dest: [*]u8, |
| 162 | noalias src: [*]const u8, |
| 163 | len: usize, |
| 164 | ) void { |
| 165 | @setRuntimeSafety(builtin.is_test); |
| 166 | comptime assert(std.math.isPowerOfTwo(copy_len)); |
| 167 | assert(len >= copy_len); |
| 168 | assert(len < 4 * copy_len); |
| 169 | |
| 170 | const a = len & (copy_len * 2); |
| 171 | const b = a / 2; |
| 172 | |
| 173 | const last = len - copy_len; |
| 174 | const pen = last - b; |
| 175 | |
| 176 | dest[0..copy_len].* = src[0..copy_len].*; |
| 177 | dest[b..][0..copy_len].* = src[b..][0..copy_len].*; |
| 178 | dest[pen..][0..copy_len].* = src[pen..][0..copy_len].*; |
| 179 | dest[last..][0..copy_len].* = src[last..][0..copy_len].*; |
| 180 | } |
| 181 | |
| 182 | test { |
| 183 | const S = struct { |
| 184 | fn testFunc(comptime copy_func: anytype) !void { |
| 185 | const max_len = 1024; |
| 186 | var buffer: [max_len + @alignOf(Element) - 1]u8 align(@alignOf(Element)) = undefined; |
| 187 | for (&buffer, 0..) |*b, i| { |
| 188 | b.* = @intCast(i % 97); |
| 189 | } |
| 190 | var dest: [max_len + @alignOf(Element) - 1]u8 align(@alignOf(Element)) = undefined; |
| 191 | |
| 192 | for (0..max_len) |copy_len| { |
| 193 | for (0..@alignOf(Element)) |s_offset| { |
| 194 | for (0..@alignOf(Element)) |d_offset| { |
| 195 | @memset(&dest, 0xff); |
| 196 | const s = buffer[s_offset..][0..copy_len]; |
| 197 | const d = dest[d_offset..][0..copy_len]; |
| 198 | _ = copy_func(@ptrCast(d.ptr), @ptrCast(s.ptr), s.len); |
| 199 | std.testing.expectEqualSlices(u8, s, d) catch |e| { |
| 200 | std.debug.print("error encountered for length={d}, s_offset={d}, d_offset={d}\n", .{ |
| 201 | copy_len, s_offset, d_offset, |
| 202 | }); |
| 203 | return e; |
| 204 | }; |
| 205 | } |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | }; |
| 146 | 210 | |
| 147 | | inline for (array[1..], 0..) |vec, i| dest[i * 32 ..][0..32].* = vec; |
| 148 | | dest[len - 32 ..][0..32].* = array[0]; |
| 211 | try S.testFunc(memcpySmall); |
| 212 | try S.testFunc(memcpyFast); |
| 149 | 213 | } |