| author | |
| committer | |
| log | bf89482f23aae3524c92d385d6d8241d065d1031 |
| tree | 1acb589b414ea47fe5923216a97ce0ef9eee4490 |
| parent | 5cfcb015033864c769235726975ab9919c217ef9 |
3 files changed, 147 insertions(+), 138 deletions(-)
lib/compiler_rt.zig+1| ... | ... | @@ -233,6 +233,7 @@ comptime { |
| 233 | 233 | |
| 234 | 234 | _ = @import("compiler_rt/memcpy.zig"); |
| 235 | 235 | _ = @import("compiler_rt/memset.zig"); |
| 236 | _ = @import("compiler_rt/memmove.zig"); | |
| 236 | 237 | _ = @import("compiler_rt/memcmp.zig"); |
| 237 | 238 | _ = @import("compiler_rt/bcmp.zig"); |
| 238 | 239 | _ = @import("compiler_rt/ssp.zig"); |
lib/compiler_rt/memcpy.zig+2-138| ... | ... | @@ -5,145 +5,9 @@ const builtin = @import("builtin"); |
| 5 | 5 | comptime { |
| 6 | 6 | if (builtin.object_format != .c) { |
| 7 | 7 | @export(&memcpy, .{ .name = "memcpy", .linkage = common.linkage, .visibility = common.visibility }); |
| 8 | @export(&memmove, .{ .name = "memmove", .linkage = common.linkage, .visibility = common.visibility }); | |
| 9 | 8 | } |
| 10 | 9 | } |
| 11 | 10 | |
| 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); | |
| 14 | } | |
| 15 | ||
| 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 | } | |
| 22 | ||
| 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; | |
| 38 | } | |
| 39 | ||
| 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 | } | |
| 50 | ||
| 51 | if (len > 16) { | |
| 52 | @branchHint(.unlikely); | |
| 53 | blockCopy(dest, src, 16, len); | |
| 54 | return dest; | |
| 55 | } | |
| 56 | ||
| 57 | blockCopy(dest, src, 8, len); | |
| 58 | ||
| 59 | return dest; | |
| 60 | } | |
| 61 | ||
| 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 | } | |
| 68 | ||
| 69 | inline fn copyLong(dest: [*]u8, src: [*]const u8, len: usize) void { | |
| 70 | var array: [8]@Vector(32, u8) = undefined; | |
| 71 | ||
| 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].*; | |
| 75 | ||
| 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 | } | |
| 84 | } | |
| 85 | } | |
| 86 | ||
| 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); | |
| 96 | } | |
| 97 | } | |
| 98 | ||
| 99 | inline fn copyForward(dest: [*]u8, src: [*]const u8, len: usize) void { | |
| 100 | const tail: @Vector(32, u8) = src[len - 32 ..][0..32].*; | |
| 101 | ||
| 102 | const N: usize = len & ~@as(usize, 127); | |
| 103 | var i: usize = 0; | |
| 104 | ||
| 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 | } | |
| 111 | ||
| 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); | |
| 117 | } | |
| 118 | } | |
| 119 | ||
| 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 | } | |
| 146 | ||
| 147 | inline for (array[1..], 0..) |vec, i| dest[i * 32 ..][0..32].* = vec; | |
| 148 | dest[len - 32 ..][0..32].* = array[0]; | |
| 11 | fn memcpy(opt_dest: ?[*]u8, opt_src: ?[*]const u8, len: usize) callconv(.C) ?[*]u8 { | |
| 12 | return @call(.always_inline, @import("memmove.zig").memmove, .{ opt_dest, opt_src, len }); | |
| 149 | 13 | } |
lib/compiler_rt/memmove.zig created+144| ... | ... | @@ -0,0 +1,144 @@ |
| 1 | const std = @import("std"); | |
| 2 | const common = @import("./common.zig"); | |
| 3 | const builtin = @import("builtin"); | |
| 4 | ||
| 5 | comptime { | |
| 6 | if (builtin.object_format != .c) { | |
| 7 | @export(&memmove, .{ .name = "memmove", .linkage = common.linkage, .visibility = common.visibility }); | |
| 8 | } | |
| 9 | } | |
| 10 | ||
| 11 | pub fn memmove(opt_dest: ?[*]u8, opt_src: ?[*]const u8, len: usize) callconv(.C) ?[*]u8 { | |
| 12 | // a port of https://github.com/facebook/folly/blob/1c8bc50e88804e2a7361a57cd9b551dd10f6c5fd/folly/memcpy.S | |
| 13 | if (len == 0) { | |
| 14 | @branchHint(.unlikely); | |
| 15 | return opt_dest; | |
| 16 | } | |
| 17 | ||
| 18 | const dest = opt_dest.?; | |
| 19 | const src = opt_src.?; | |
| 20 | ||
| 21 | if (len < 8) { | |
| 22 | @branchHint(.unlikely); | |
| 23 | if (len == 1) { | |
| 24 | @branchHint(.unlikely); | |
| 25 | dest[0] = src[0]; | |
| 26 | } else if (len >= 4) { | |
| 27 | @branchHint(.unlikely); | |
| 28 | blockCopy(dest, src, 4, len); | |
| 29 | } else { | |
| 30 | blockCopy(dest, src, 2, len); | |
| 31 | } | |
| 32 | return dest; | |
| 33 | } | |
| 34 | ||
| 35 | if (len > 32) { | |
| 36 | @branchHint(.unlikely); | |
| 37 | if (len > 256) { | |
| 38 | @branchHint(.unlikely); | |
| 39 | copyMove(dest, src, len); | |
| 40 | return dest; | |
| 41 | } | |
| 42 | copyLong(dest, src, len); | |
| 43 | return dest; | |
| 44 | } | |
| 45 | ||
| 46 | if (len > 16) { | |
| 47 | @branchHint(.unlikely); | |
| 48 | blockCopy(dest, src, 16, len); | |
| 49 | return dest; | |
| 50 | } | |
| 51 | ||
| 52 | blockCopy(dest, src, 8, len); | |
| 53 | ||
| 54 | return dest; | |
| 55 | } | |
| 56 | ||
| 57 | inline fn blockCopy(dest: [*]u8, src: [*]const u8, block_size: comptime_int, len: usize) void { | |
| 58 | const first = @as(*align(1) const @Vector(block_size, u8), src[0..block_size]).*; | |
| 59 | const second = @as(*align(1) const @Vector(block_size, u8), src[len - block_size ..][0..block_size]).*; | |
| 60 | dest[0..block_size].* = first; | |
| 61 | dest[len - block_size ..][0..block_size].* = second; | |
| 62 | } | |
| 63 | ||
| 64 | inline fn copyLong(dest: [*]u8, src: [*]const u8, len: usize) void { | |
| 65 | var array: [8]@Vector(32, u8) = undefined; | |
| 66 | ||
| 67 | inline for (.{ 64, 128, 192, 256 }, 0..) |N, i| { | |
| 68 | array[i * 2] = src[(N / 2) - 32 ..][0..32].*; | |
| 69 | array[(i * 2) + 1] = src[len - N / 2 ..][0..32].*; | |
| 70 | ||
| 71 | if (len <= N) { | |
| 72 | @branchHint(.unlikely); | |
| 73 | for (0..i + 1) |j| { | |
| 74 | dest[j * 32 ..][0..32].* = array[j * 2]; | |
| 75 | dest[len - ((j * 32) + 32) ..][0..32].* = array[(j * 2) + 1]; | |
| 76 | } | |
| 77 | return; | |
| 78 | } | |
| 79 | } | |
| 80 | } | |
| 81 | ||
| 82 | inline fn copyMove(dest: [*]u8, src: [*]const u8, len: usize) void { | |
| 83 | if (@intFromPtr(src) >= @intFromPtr(dest)) { | |
| 84 | @branchHint(.unlikely); | |
| 85 | copyForward(dest, src, len); | |
| 86 | } else if (@intFromPtr(src) + len > @intFromPtr(dest)) { | |
| 87 | @branchHint(.unlikely); | |
| 88 | overlapBwd(dest, src, len); | |
| 89 | } else { | |
| 90 | copyForward(dest, src, len); | |
| 91 | } | |
| 92 | } | |
| 93 | ||
| 94 | inline fn copyForward(dest: [*]u8, src: [*]const u8, len: usize) void { | |
| 95 | const tail: @Vector(32, u8) = src[len - 32 ..][0..32].*; | |
| 96 | ||
| 97 | const N: usize = len & ~@as(usize, 127); | |
| 98 | var i: usize = 0; | |
| 99 | ||
| 100 | while (i < N) : (i += 128) { | |
| 101 | dest[i..][0..32].* = src[i..][0..32].*; | |
| 102 | dest[i + 32 ..][0..32].* = src[i + 32 ..][0..32].*; | |
| 103 | dest[i + 64 ..][0..32].* = src[i + 64 ..][0..32].*; | |
| 104 | dest[i + 96 ..][0..32].* = src[i + 96 ..][0..32].*; | |
| 105 | } | |
| 106 | ||
| 107 | if (len - i <= 32) { | |
| 108 | @branchHint(.unlikely); | |
| 109 | dest[len - 32 ..][0..32].* = tail; | |
| 110 | } else { | |
| 111 | copyLong(dest[i..], src[i..], len - i); | |
| 112 | } | |
| 113 | } | |
| 114 | ||
| 115 | inline fn overlapBwd(dest: [*]u8, src: [*]const u8, len: usize) void { | |
| 116 | var array: [5]@Vector(32, u8) = undefined; | |
| 117 | array[0] = src[len - 32 ..][0..32].*; | |
| 118 | inline for (1..5) |i| array[i] = src[(i - 1) << 5 ..][0..32].*; | |
| 119 | ||
| 120 | const end: usize = (@intFromPtr(dest) + len - 32) & 31; | |
| 121 | const range = len - end; | |
| 122 | var s = src + range; | |
| 123 | var d = dest + range; | |
| 124 | ||
| 125 | while (@intFromPtr(s) > @intFromPtr(src + 128)) { | |
| 126 | // zig fmt: off | |
| 127 | const first = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 32)).*; | |
| 128 | const second = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 64)).*; | |
| 129 | const third = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 96)).*; | |
| 130 | const fourth = @as(*align(1) const @Vector(32, u8), @ptrCast(s - 128)).*; | |
| 131 | ||
| 132 | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 32))).* = first; | |
| 133 | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 64))).* = second; | |
| 134 | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 96))).* = third; | |
| 135 | @as(*align(32) @Vector(32, u8), @alignCast(@ptrCast(d - 128))).* = fourth; | |
| 136 | // zig fmt: on | |
| 137 | ||
| 138 | s -= 128; | |
| 139 | d -= 128; | |
| 140 | } | |
| 141 | ||
| 142 | inline for (array[1..], 0..) |vec, i| dest[i * 32 ..][0..32].* = vec; | |
| 143 | dest[len - 32 ..][0..32].* = array[0]; | |
| 144 | } |