| ... | ... | @@ -953,14 +953,105 @@ test "len" { |
| 953 | 953 | try testing.expect(len(c_ptr) == 2); |
| 954 | 954 | } |
| 955 | 955 | |
| 956 | | pub fn indexOfSentinel(comptime Elem: type, comptime sentinel: Elem, ptr: [*:sentinel]const Elem) usize { |
| 956 | const backend_supports_vectors = switch (builtin.zig_backend) { |
| 957 | .stage2_llvm, .stage2_c => true, |
| 958 | else => false, |
| 959 | }; |
| 960 | |
| 961 | pub fn indexOfSentinel(comptime T: type, comptime sentinel: T, p: [*:sentinel]const T) usize { |
| 957 | 962 | var i: usize = 0; |
| 958 | | while (ptr[i] != sentinel) { |
| 963 | |
| 964 | if (backend_supports_vectors and |
| 965 | !@inComptime() and |
| 966 | (@typeInfo(T) == .Int or @typeInfo(T) == .Float) and std.math.isPowerOfTwo(@bitSizeOf(T))) |
| 967 | { |
| 968 | switch (@import("builtin").cpu.arch) { |
| 969 | // The below branch assumes that reading past the end of the buffer is valid, as long |
| 970 | // as we don't read into a new page. This should be the case for most architectures |
| 971 | // which use paged memory, however should be confirmed before adding a new arch below. |
| 972 | .aarch64, .x86, .x86_64 => if (comptime std.simd.suggestVectorSize(T)) |block_len| { |
| 973 | comptime std.debug.assert(std.mem.page_size % block_len == 0); |
| 974 | const Block = @Vector(block_len, T); |
| 975 | const mask: Block = @splat(sentinel); |
| 976 | |
| 977 | // First block may be unaligned |
| 978 | const start_addr = @intFromPtr(&p[i]); |
| 979 | const offset_in_page = start_addr & (std.mem.page_size - 1); |
| 980 | if (offset_in_page < std.mem.page_size - block_len) { |
| 981 | // Will not read past the end of a page, full block. |
| 982 | const block: Block = p[i..][0..block_len].*; |
| 983 | const matches = block == mask; |
| 984 | if (@reduce(.Or, matches)) { |
| 985 | return i + std.simd.firstTrue(matches).?; |
| 986 | } |
| 987 | |
| 988 | i += (std.mem.alignForward(usize, start_addr, @alignOf(Block)) - start_addr) / @sizeOf(T); |
| 989 | } else { |
| 990 | // Would read over a page boundary. Per-byte at a time until aligned or found. |
| 991 | // 0.39% chance this branch is taken for 4K pages at 16b block length. |
| 992 | // |
| 993 | // An alternate strategy is to do read a full block (the last in the page) and |
| 994 | // mask the entries before the pointer. |
| 995 | while ((@intFromPtr(&p[i]) & (@alignOf(Block) - 1)) != 0) : (i += 1) { |
| 996 | if (p[i] == sentinel) return i; |
| 997 | } |
| 998 | } |
| 999 | |
| 1000 | std.debug.assert(std.mem.isAligned(@intFromPtr(&p[i]), @alignOf(Block))); |
| 1001 | while (true) { |
| 1002 | const block: *const Block = @ptrCast(@alignCast(p[i..][0..block_len])); |
| 1003 | const matches = block.* == mask; |
| 1004 | if (@reduce(.Or, matches)) { |
| 1005 | return i + std.simd.firstTrue(matches).?; |
| 1006 | } |
| 1007 | i += block_len; |
| 1008 | } |
| 1009 | }, |
| 1010 | else => {}, |
| 1011 | } |
| 1012 | } |
| 1013 | |
| 1014 | while (p[i] != sentinel) { |
| 959 | 1015 | i += 1; |
| 960 | 1016 | } |
| 961 | 1017 | return i; |
| 962 | 1018 | } |
| 963 | 1019 | |
| 1020 | test "indexOfSentinel vector paths" { |
| 1021 | const Types = [_]type{ u8, u16, u32, u64 }; |
| 1022 | const allocator = std.testing.allocator; |
| 1023 | |
| 1024 | inline for (Types) |T| { |
| 1025 | const block_len = comptime std.simd.suggestVectorSize(T) orelse continue; |
| 1026 | |
| 1027 | // Allocate three pages so we guarantee a page-crossing address with a full page after |
| 1028 | const memory = try allocator.alloc(T, 3 * std.mem.page_size / @sizeOf(T)); |
| 1029 | defer allocator.free(memory); |
| 1030 | @memset(memory, 0xaa); |
| 1031 | |
| 1032 | // Find starting page-alignment = 0 |
| 1033 | var start: usize = 0; |
| 1034 | const start_addr = @intFromPtr(&memory); |
| 1035 | start += (std.mem.alignForward(usize, start_addr, std.mem.page_size) - start_addr) / @sizeOf(T); |
| 1036 | try testing.expect(start < std.mem.page_size / @sizeOf(T)); |
| 1037 | |
| 1038 | // Validate all sub-block alignments |
| 1039 | const search_len = std.mem.page_size / @sizeOf(T); |
| 1040 | memory[start + search_len] = 0; |
| 1041 | for (0..block_len) |offset| { |
| 1042 | try testing.expectEqual(search_len - offset, indexOfSentinel(T, 0, @ptrCast(&memory[start + offset]))); |
| 1043 | } |
| 1044 | memory[start + search_len] = 0xaa; |
| 1045 | |
| 1046 | // Validate page boundary crossing |
| 1047 | const start_page_boundary = start + (std.mem.page_size / @sizeOf(T)); |
| 1048 | memory[start_page_boundary + block_len] = 0; |
| 1049 | for (0..block_len) |offset| { |
| 1050 | try testing.expectEqual(2 * block_len - offset, indexOfSentinel(T, 0, @ptrCast(&memory[start_page_boundary - block_len + offset]))); |
| 1051 | } |
| 1052 | } |
| 1053 | } |
| 1054 | |
| 964 | 1055 | /// Returns true if all elements in a slice are equal to the scalar value provided |
| 965 | 1056 | pub fn allEqual(comptime T: type, slice: []const T, scalar: T) bool { |
| 966 | 1057 | for (slice) |item| { |
| ... | ... | @@ -1016,12 +1107,79 @@ pub fn lastIndexOfScalar(comptime T: type, slice: []const T, value: T) ?usize { |
| 1016 | 1107 | |
| 1017 | 1108 | pub fn indexOfScalarPos(comptime T: type, slice: []const T, start_index: usize, value: T) ?usize { |
| 1018 | 1109 | if (start_index >= slice.len) return null; |
| 1019 | | for (slice[start_index..], start_index..) |c, i| { |
| 1020 | | if (c == value) return i; |
| 1110 | |
| 1111 | var i: usize = start_index; |
| 1112 | if (backend_supports_vectors and |
| 1113 | !@inComptime() and |
| 1114 | (@typeInfo(T) == .Int or @typeInfo(T) == .Float) and std.math.isPowerOfTwo(@bitSizeOf(T))) |
| 1115 | { |
| 1116 | if (comptime std.simd.suggestVectorSize(T)) |block_len| { |
| 1117 | // For Intel Nehalem (2009) and AMD Bulldozer (2012) or later, unaligned loads on aligned data result |
| 1118 | // in the same execution as aligned loads. We ignore older arch's here and don't bother pre-aligning. |
| 1119 | // |
| 1120 | // Use `comptime std.simd.suggestVectorSize(T)` to get the same alignment as used in this function |
| 1121 | // however this usually isn't necessary unless your arch has a performance penalty due to this. |
| 1122 | // |
| 1123 | // This may differ for other arch's. Arm for example costs a cycle when loading across a cache |
| 1124 | // line so explicit alignment prologues may be worth exploration. |
| 1125 | |
| 1126 | // Unrolling here is ~10% improvement. We can then do one bounds check every 2 blocks |
| 1127 | // instead of one which adds up. |
| 1128 | const Block = @Vector(block_len, T); |
| 1129 | if (i + 2 * block_len < slice.len) { |
| 1130 | const mask: Block = @splat(value); |
| 1131 | while (true) { |
| 1132 | inline for (0..2) |_| { |
| 1133 | const block: Block = slice[i..][0..block_len].*; |
| 1134 | const matches = block == mask; |
| 1135 | if (@reduce(.Or, matches)) { |
| 1136 | return i + std.simd.firstTrue(matches).?; |
| 1137 | } |
| 1138 | i += block_len; |
| 1139 | } |
| 1140 | if (i + 2 * block_len >= slice.len) break; |
| 1141 | } |
| 1142 | } |
| 1143 | |
| 1144 | // {block_len, block_len / 2} check |
| 1145 | inline for (0..2) |j| { |
| 1146 | const block_x_len = block_len / (1 << j); |
| 1147 | comptime if (block_x_len < 4) break; |
| 1148 | |
| 1149 | const BlockX = @Vector(block_x_len, T); |
| 1150 | if (i + block_x_len < slice.len) { |
| 1151 | const mask: BlockX = @splat(value); |
| 1152 | const block: BlockX = slice[i..][0..block_x_len].*; |
| 1153 | const matches = block == mask; |
| 1154 | if (@reduce(.Or, matches)) { |
| 1155 | return i + std.simd.firstTrue(matches).?; |
| 1156 | } |
| 1157 | i += block_x_len; |
| 1158 | } |
| 1159 | } |
| 1160 | } |
| 1161 | } |
| 1162 | |
| 1163 | for (slice[i..], i..) |c, j| { |
| 1164 | if (c == value) return j; |
| 1021 | 1165 | } |
| 1022 | 1166 | return null; |
| 1023 | 1167 | } |
| 1024 | 1168 | |
| 1169 | test "indexOfScalarPos" { |
| 1170 | const Types = [_]type{ u8, u16, u32, u64 }; |
| 1171 | |
| 1172 | inline for (Types) |T| { |
| 1173 | var memory: [64 / @sizeOf(T)]T = undefined; |
| 1174 | @memset(&memory, 0xaa); |
| 1175 | memory[memory.len - 1] = 0; |
| 1176 | |
| 1177 | for (0..memory.len) |i| { |
| 1178 | try testing.expectEqual(memory.len - i - 1, indexOfScalarPos(T, memory[i..], 0, 0).?); |
| 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | |
| 1025 | 1183 | pub fn indexOfAny(comptime T: type, slice: []const T, values: []const T) ?usize { |
| 1026 | 1184 | return indexOfAnyPos(T, slice, 0, values); |
| 1027 | 1185 | } |
Andrew Kelley <andrew@ziglang.org> 2023-10-01 00:07:57-07:00
GitHub <noreply@github.com> 2023-10-01 00:07:57-07:00
Signed by PGP key 4AEE18F83AFDEB23
lib