| 1 | const std = @import("std"); |
| 2 | const testing = std.testing; |
| 3 | const assert = std.debug.assert; |
| 4 | const maxInt = std.math.maxInt; |
| 5 | const minInt = std.math.minInt; |
| 6 | |
| 7 | const builtin = @import("builtin"); |
| 8 | const compiler_rt = @import("../compiler_rt.zig"); |
| 9 | const symbol = compiler_rt.symbol; |
| 10 | |
| 11 | const endian = builtin.cpu.arch.endian(); |
| 12 | |
| 13 | inline fn limbGet(limbs: []const u64, i: usize) u64 { |
| 14 | return switch (endian) { |
| 15 | .little => limbs[i], |
| 16 | .big => limbs[limbs.len - 1 - i], |
| 17 | }; |
| 18 | } |
| 19 | |
| 20 | inline fn limbSet(limbs: []u64, i: usize, value: u64) void { |
| 21 | switch (endian) { |
| 22 | .little => limbs[i] = value, |
| 23 | .big => limbs[limbs.len - 1 - i] = value, |
| 24 | } |
| 25 | } |
| 26 | |
| 27 | fn usedLimbCount(bits: u16) u16 { |
| 28 | return @divCeil(bits, 64); |
| 29 | } |
| 30 | |
| 31 | fn limbCount(bits: u16) u16 { |
| 32 | return @divExact(std.zig.target.intByteSize(&builtin.target, bits), 8); |
| 33 | } |
| 34 | |
| 35 | fn varLimbs(ptr: [*]u64, bits: u16) []u64 { |
| 36 | const limb_cnt = usedLimbCount(bits); |
| 37 | const true_limb_cnt = limbCount(bits); |
| 38 | return switch (endian) { |
| 39 | .little => ptr[0..limb_cnt], |
| 40 | .big => ptr[true_limb_cnt - limb_cnt .. true_limb_cnt], |
| 41 | }; |
| 42 | } |
| 43 | |
| 44 | fn constLimbs(ptr: [*]const u64, bits: u16) []const u64 { |
| 45 | const limb_cnt = usedLimbCount(bits); |
| 46 | const true_limb_cnt = limbCount(bits); |
| 47 | return switch (endian) { |
| 48 | .little => ptr[0..limb_cnt], |
| 49 | .big => ptr[true_limb_cnt - limb_cnt .. true_limb_cnt], |
| 50 | }; |
| 51 | } |
| 52 | |
| 53 | fn fixLastLimb(out_ptr: [*]u64, is_signed: bool, bits: u16) void { |
| 54 | const limb_cnt = usedLimbCount(bits); |
| 55 | const true_limb_cnt = limbCount(bits); |
| 56 | if (limb_cnt == true_limb_cnt) return; |
| 57 | const true_out = out_ptr[0..true_limb_cnt]; |
| 58 | |
| 59 | const ms = limbGet(true_out, limb_cnt - 1); |
| 60 | const sign: u64 = if (!is_signed or @as(i64, @bitCast(ms)) >= 0) 0 else ~@as(u64, 0); |
| 61 | for (limb_cnt..true_limb_cnt) |i| { |
| 62 | limbSet(true_out, i, sign); |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | fn Limbs(T: type) type { |
| 67 | const int_info = @typeInfo(T).int; |
| 68 | const limb_cnt = comptime limbCount(int_info.bits); |
| 69 | return [limb_cnt]u64; |
| 70 | } |
| 71 | |
| 72 | fn asLimbs(v: anytype) Limbs(@TypeOf(v)) { |
| 73 | const T = @TypeOf(v); |
| 74 | const int_info = @typeInfo(T).int; |
| 75 | const limb_cnt = comptime limbCount(int_info.bits); |
| 76 | const ET = @Int(int_info.signedness, limb_cnt * 64); |
| 77 | const low_to_high: Limbs(T) = @bitCast(@as(ET, v)); |
| 78 | switch (endian) { |
| 79 | .little => return low_to_high, |
| 80 | .big => { |
| 81 | var swapped: Limbs(T) = undefined; |
| 82 | for (low_to_high, 0..) |x, i| { |
| 83 | swapped[limb_cnt - i - 1] = x; |
| 84 | } |
| 85 | return swapped; |
| 86 | }, |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | fn limbWrap(limb: u64, is_signed: bool, bits: u16) u64 { |
| 91 | assert(bits % 64 != 0); |
| 92 | const pad_bits: u6 = @intCast(64 - bits % 64); |
| 93 | if (!is_signed) { |
| 94 | const s = limb << pad_bits; |
| 95 | return s >> pad_bits; |
| 96 | } else { |
| 97 | const s = @as(i64, @bitCast(limb)) << pad_bits; |
| 98 | return @bitCast(s >> pad_bits); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | comptime { |
| 103 | symbol(&__addo_limb64, "__addo_limb64"); |
| 104 | } |
| 105 | |
| 106 | fn __addo_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, b_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) bool { |
| 107 | const limb_cnt = usedLimbCount(bits); |
| 108 | const out = varLimbs(out_ptr, bits); |
| 109 | const a = constLimbs(a_ptr, bits); |
| 110 | const b = constLimbs(b_ptr, bits); |
| 111 | |
| 112 | var carry: u1 = 0; |
| 113 | var i: usize = 0; |
| 114 | while (i < limb_cnt - 1) : (i += 1) { |
| 115 | const s1 = @addWithOverflow(limbGet(a, i), limbGet(b, i)); |
| 116 | const s2 = @addWithOverflow(s1[0], carry); |
| 117 | carry = s1[1] | s2[1]; |
| 118 | limbSet(out, i, s2[0]); |
| 119 | } |
| 120 | |
| 121 | const limb: u64 = b: { |
| 122 | if (!is_signed) { |
| 123 | const s1 = @addWithOverflow(limbGet(a, i), limbGet(b, i)); |
| 124 | const s2 = @addWithOverflow(s1[0], carry); |
| 125 | carry = s1[1] | s2[1]; |
| 126 | break :b s2[0]; |
| 127 | } else { |
| 128 | const as: i64 = @bitCast(limbGet(a, i)); |
| 129 | const bs: i64 = @bitCast(limbGet(b, i)); |
| 130 | const s1 = @addWithOverflow(as, bs); |
| 131 | const s2 = @addWithOverflow(s1[0], carry); |
| 132 | carry = s1[1] | s2[1]; |
| 133 | break :b @bitCast(s2[0]); |
| 134 | } |
| 135 | }; |
| 136 | |
| 137 | if (bits % 64 == 0) { |
| 138 | limbSet(out, i, limb); |
| 139 | fixLastLimb(out_ptr, is_signed, bits); |
| 140 | return carry != 0; |
| 141 | } else { |
| 142 | assert(carry == 0); |
| 143 | const wrapped_limb = limbWrap(limb, is_signed, bits); |
| 144 | limbSet(out, i, wrapped_limb); |
| 145 | fixLastLimb(out_ptr, is_signed, bits); |
| 146 | return wrapped_limb != limb; |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | fn test__addo_limb64(comptime T: type, a: T, b: T, expected: struct { T, bool }) !void { |
| 151 | const int_info = @typeInfo(T).int; |
| 152 | const is_signed = int_info.signedness == .signed; |
| 153 | |
| 154 | var a_limbs = asLimbs(a); |
| 155 | var b_limbs = asLimbs(b); |
| 156 | var out: Limbs(T) = undefined; |
| 157 | const overflow = __addo_limb64(&out, &a_limbs, &b_limbs, is_signed, int_info.bits); |
| 158 | |
| 159 | const expected_limbs = asLimbs(expected[0]); |
| 160 | try testing.expectEqual(expected_limbs, out); |
| 161 | try testing.expectEqual(expected[1], overflow); |
| 162 | } |
| 163 | |
| 164 | test __addo_limb64 { |
| 165 | try test__addo_limb64(u64, 1, 2, .{ 3, false }); |
| 166 | try test__addo_limb64(u64, maxInt(u64), 2, .{ 1, true }); |
| 167 | try test__addo_limb64(u65, maxInt(u65), 2, .{ 1, true }); |
| 168 | try test__addo_limb64(u255, 1, 2, .{ 3, false }); |
| 169 | |
| 170 | try test__addo_limb64(i64, 1, 2, .{ 3, false }); |
| 171 | try test__addo_limb64(i64, maxInt(i64), 1, .{ minInt(i64), true }); |
| 172 | try test__addo_limb64(i65, maxInt(i65), 1, .{ minInt(i65), true }); |
| 173 | try test__addo_limb64(i255, -3, 2, .{ -1, false }); |
| 174 | |
| 175 | try test__addo_limb64(u150, maxInt(u150), 2, .{ 1, true }); |
| 176 | try test__addo_limb64(i150, -3, 2, .{ -1, false }); |
| 177 | } |
| 178 | |
| 179 | comptime { |
| 180 | symbol(&__subo_limb64, "__subo_limb64"); |
| 181 | } |
| 182 | |
| 183 | fn __subo_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, b_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) bool { |
| 184 | const limb_cnt = usedLimbCount(bits); |
| 185 | const out = varLimbs(out_ptr, bits); |
| 186 | const a = constLimbs(a_ptr, bits); |
| 187 | const b = constLimbs(b_ptr, bits); |
| 188 | |
| 189 | var borrow: u1 = 0; |
| 190 | var i: usize = 0; |
| 191 | while (i < limb_cnt - 1) : (i += 1) { |
| 192 | const s1 = @subWithOverflow(limbGet(a, i), limbGet(b, i)); |
| 193 | const s2 = @subWithOverflow(s1[0], borrow); |
| 194 | borrow = s1[1] | s2[1]; |
| 195 | limbSet(out, i, s2[0]); |
| 196 | } |
| 197 | |
| 198 | const limb: u64 = b: { |
| 199 | if (!is_signed) { |
| 200 | const s1 = @subWithOverflow(limbGet(a, i), limbGet(b, i)); |
| 201 | const s2 = @subWithOverflow(s1[0], borrow); |
| 202 | borrow = s1[1] | s2[1]; |
| 203 | break :b s2[0]; |
| 204 | } else { |
| 205 | const as: i64 = @bitCast(limbGet(a, i)); |
| 206 | const bs: i64 = @bitCast(limbGet(b, i)); |
| 207 | const s1 = @subWithOverflow(as, bs); |
| 208 | const s2 = @subWithOverflow(s1[0], borrow); |
| 209 | borrow = s1[1] | s2[1]; |
| 210 | break :b @bitCast(s2[0]); |
| 211 | } |
| 212 | }; |
| 213 | |
| 214 | if (bits % 64 == 0) { |
| 215 | limbSet(out, i, limb); |
| 216 | fixLastLimb(out_ptr, is_signed, bits); |
| 217 | return borrow != 0; |
| 218 | } else { |
| 219 | const wrapped_limb = limbWrap(limb, is_signed, bits); |
| 220 | limbSet(out, i, wrapped_limb); |
| 221 | fixLastLimb(out_ptr, is_signed, bits); |
| 222 | return borrow != 0 or wrapped_limb != limb; |
| 223 | } |
| 224 | } |
| 225 | |
| 226 | fn test__subo_limb64(comptime T: type, a: T, b: T, expected: struct { T, bool }) !void { |
| 227 | const int_info = @typeInfo(T).int; |
| 228 | const is_signed = int_info.signedness == .signed; |
| 229 | |
| 230 | var a_limbs = asLimbs(a); |
| 231 | var b_limbs = asLimbs(b); |
| 232 | var out: Limbs(T) = undefined; |
| 233 | const overflow = __subo_limb64(&out, &a_limbs, &b_limbs, is_signed, int_info.bits); |
| 234 | |
| 235 | const expected_limbs = asLimbs(expected[0]); |
| 236 | try testing.expectEqual(expected_limbs, out); |
| 237 | try testing.expectEqual(expected[1], overflow); |
| 238 | } |
| 239 | |
| 240 | test __subo_limb64 { |
| 241 | try test__subo_limb64(u64, 3, 2, .{ 1, false }); |
| 242 | try test__subo_limb64(u64, 0, 1, .{ maxInt(u64), true }); |
| 243 | try test__subo_limb64(u65, 0, 1, .{ maxInt(u65), true }); |
| 244 | try test__subo_limb64(u255, 3, 2, .{ 1, false }); |
| 245 | |
| 246 | try test__subo_limb64(i64, 1, 2, .{ -1, false }); |
| 247 | try test__subo_limb64(i64, minInt(i64), 1, .{ maxInt(i64), true }); |
| 248 | try test__subo_limb64(i65, minInt(i65), 1, .{ maxInt(i65), true }); |
| 249 | try test__subo_limb64(i255, -1, 2, .{ -3, false }); |
| 250 | |
| 251 | try test__subo_limb64(u150, 2, maxInt(u150), .{ 3, true }); |
| 252 | try test__subo_limb64(i150, -3, 2, .{ -5, false }); |
| 253 | } |
| 254 | |
| 255 | comptime { |
| 256 | symbol(&__cmp_limb64, "__cmp_limb64"); |
| 257 | } |
| 258 | |
| 259 | // a < b -> -1 |
| 260 | // a == b -> 0 |
| 261 | // a > b -> 1 |
| 262 | fn __cmp_limb64(a_ptr: [*]const u64, b_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) i8 { |
| 263 | const limb_cnt = usedLimbCount(bits); |
| 264 | const a = constLimbs(a_ptr, bits); |
| 265 | const b = constLimbs(b_ptr, bits); |
| 266 | |
| 267 | var i: usize = 0; |
| 268 | if (is_signed) { |
| 269 | const sa: i64 = @bitCast(limbGet(a, limb_cnt - 1)); |
| 270 | const sb: i64 = @bitCast(limbGet(b, limb_cnt - 1)); |
| 271 | if (sa < sb) return -1; |
| 272 | if (sa > sb) return 1; |
| 273 | i += 1; |
| 274 | } |
| 275 | |
| 276 | while (i < limb_cnt) : (i += 1) { |
| 277 | const ai = limbGet(a, limb_cnt - 1 - i); |
| 278 | const bi = limbGet(b, limb_cnt - 1 - i); |
| 279 | if (ai < bi) return -1; |
| 280 | if (ai > bi) return 1; |
| 281 | } |
| 282 | |
| 283 | return 0; |
| 284 | } |
| 285 | |
| 286 | fn test__cmp_limb64(comptime T: type, a: T, b: T, expected: i8) !void { |
| 287 | const int_info = @typeInfo(T).int; |
| 288 | const is_signed = int_info.signedness == .signed; |
| 289 | |
| 290 | var a_limbs = asLimbs(a); |
| 291 | var b_limbs = asLimbs(b); |
| 292 | const actual = __cmp_limb64(&a_limbs, &b_limbs, is_signed, int_info.bits); |
| 293 | |
| 294 | try testing.expectEqual(expected, actual); |
| 295 | } |
| 296 | |
| 297 | test __cmp_limb64 { |
| 298 | try test__cmp_limb64(u64, 1, 2, -1); |
| 299 | try test__cmp_limb64(u64, 2, 2, 0); |
| 300 | try test__cmp_limb64(u64, 3, 2, 1); |
| 301 | |
| 302 | try test__cmp_limb64(u65, 1, 2, -1); |
| 303 | try test__cmp_limb64(u65, maxInt(u65), maxInt(u65), 0); |
| 304 | try test__cmp_limb64(u65, maxInt(u65), maxInt(u65) - 1, 1); |
| 305 | |
| 306 | try test__cmp_limb64(u255, 1, 2, -1); |
| 307 | try test__cmp_limb64(u255, 7, 7, 0); |
| 308 | try test__cmp_limb64(u255, maxInt(u255), maxInt(u255) - 1, 1); |
| 309 | |
| 310 | try test__cmp_limb64(i64, -1, 0, -1); |
| 311 | try test__cmp_limb64(i64, 0, 0, 0); |
| 312 | try test__cmp_limb64(i64, 1, 0, 1); |
| 313 | |
| 314 | try test__cmp_limb64(i65, minInt(i65), maxInt(i65), -1); |
| 315 | try test__cmp_limb64(i65, -1, -1, 0); |
| 316 | try test__cmp_limb64(i65, maxInt(i65), minInt(i65), 1); |
| 317 | |
| 318 | try test__cmp_limb64(i255, -3, 2, -1); |
| 319 | try test__cmp_limb64(i255, -5, -5, 0); |
| 320 | try test__cmp_limb64(i255, 2, -3, 1); |
| 321 | |
| 322 | try test__cmp_limb64(u150, maxInt(u150) - 5, maxInt(u150) - 5, 0); |
| 323 | try test__cmp_limb64(i150, minInt(i150), -5, -1); |
| 324 | } |
| 325 | |
| 326 | comptime { |
| 327 | symbol(&__and_limb64, "__and_limb64"); |
| 328 | } |
| 329 | |
| 330 | fn __and_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, b_ptr: [*]const u64, bits: u16) callconv(.c) void { |
| 331 | const limb_cnt = limbCount(bits); |
| 332 | const out = out_ptr[0..limb_cnt]; |
| 333 | const a = a_ptr[0..limb_cnt]; |
| 334 | const b = b_ptr[0..limb_cnt]; |
| 335 | |
| 336 | var i: usize = 0; |
| 337 | while (i < limb_cnt) : (i += 1) { |
| 338 | limbSet(out, i, limbGet(a, i) & limbGet(b, i)); |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | fn test__and_limb64(comptime T: type, a: T, b: T, expected: T) !void { |
| 343 | const int_info = @typeInfo(T).int; |
| 344 | |
| 345 | var a_limbs = asLimbs(a); |
| 346 | var b_limbs = asLimbs(b); |
| 347 | var out: Limbs(T) = undefined; |
| 348 | __and_limb64(&out, &a_limbs, &b_limbs, int_info.bits); |
| 349 | |
| 350 | const expected_limbs = asLimbs(expected); |
| 351 | try testing.expectEqual(expected_limbs, out); |
| 352 | } |
| 353 | |
| 354 | test __and_limb64 { |
| 355 | try test__and_limb64(u64, 1, 2, 0); |
| 356 | try test__and_limb64(u64, maxInt(u64), 2, 2); |
| 357 | try test__and_limb64(u65, maxInt(u65), 2, 2); |
| 358 | try test__and_limb64(u255, maxInt(u255), 7, 7); |
| 359 | |
| 360 | try test__and_limb64(i64, 1, 2, 0); |
| 361 | try test__and_limb64(i64, -1, 2, 2); |
| 362 | try test__and_limb64(i65, minInt(i65), -1, minInt(i65)); |
| 363 | try test__and_limb64(i255, -1, 2, 2); |
| 364 | |
| 365 | try test__and_limb64(u150, maxInt(u150), 7, 7); |
| 366 | try test__and_limb64(i150, -2, 3, 2); |
| 367 | } |
| 368 | |
| 369 | comptime { |
| 370 | symbol(&__or_limb64, "__or_limb64"); |
| 371 | } |
| 372 | |
| 373 | fn __or_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, b_ptr: [*]const u64, bits: u16) callconv(.c) void { |
| 374 | const limb_cnt = limbCount(bits); |
| 375 | const out = out_ptr[0..limb_cnt]; |
| 376 | const a = a_ptr[0..limb_cnt]; |
| 377 | const b = b_ptr[0..limb_cnt]; |
| 378 | |
| 379 | var i: usize = 0; |
| 380 | while (i < limb_cnt) : (i += 1) { |
| 381 | limbSet(out, i, limbGet(a, i) | limbGet(b, i)); |
| 382 | } |
| 383 | } |
| 384 | |
| 385 | fn test__or_limb64(comptime T: type, a: T, b: T, expected: T) !void { |
| 386 | const int_info = @typeInfo(T).int; |
| 387 | |
| 388 | var a_limbs = asLimbs(a); |
| 389 | var b_limbs = asLimbs(b); |
| 390 | var out: Limbs(T) = undefined; |
| 391 | __or_limb64(&out, &a_limbs, &b_limbs, int_info.bits); |
| 392 | |
| 393 | const expected_limbs = asLimbs(expected); |
| 394 | try testing.expectEqual(expected_limbs, out); |
| 395 | } |
| 396 | |
| 397 | test __or_limb64 { |
| 398 | try test__or_limb64(u64, 1, 2, 3); |
| 399 | try test__or_limb64(u64, maxInt(u64), 2, maxInt(u64)); |
| 400 | try test__or_limb64(u65, maxInt(u65), 2, maxInt(u65)); |
| 401 | try test__or_limb64(u255, 1, 2, 3); |
| 402 | |
| 403 | try test__or_limb64(i64, 1, 2, 3); |
| 404 | try test__or_limb64(i64, -1, 2, -1); |
| 405 | try test__or_limb64(i65, minInt(i65), 1, minInt(i65) + 1); |
| 406 | try test__or_limb64(i255, -3, 2, -1); |
| 407 | |
| 408 | try test__or_limb64(u150, maxInt(u150) - 1, 3, maxInt(u150)); |
| 409 | try test__or_limb64(i150, -2, 3, -1); |
| 410 | } |
| 411 | |
| 412 | comptime { |
| 413 | symbol(&__xor_limb64, "__xor_limb64"); |
| 414 | } |
| 415 | |
| 416 | fn __xor_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, b_ptr: [*]const u64, bits: u16) callconv(.c) void { |
| 417 | const limb_cnt = limbCount(bits); |
| 418 | const out = out_ptr[0..limb_cnt]; |
| 419 | const a = a_ptr[0..limb_cnt]; |
| 420 | const b = b_ptr[0..limb_cnt]; |
| 421 | |
| 422 | var i: usize = 0; |
| 423 | while (i < limb_cnt) : (i += 1) { |
| 424 | limbSet(out, i, limbGet(a, i) ^ limbGet(b, i)); |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | fn test__xor_limb64(comptime T: type, a: T, b: T, expected: T) !void { |
| 429 | const int_info = @typeInfo(T).int; |
| 430 | |
| 431 | var a_limbs = asLimbs(a); |
| 432 | var b_limbs = asLimbs(b); |
| 433 | var out: Limbs(T) = undefined; |
| 434 | __xor_limb64(&out, &a_limbs, &b_limbs, int_info.bits); |
| 435 | |
| 436 | const expected_limbs = asLimbs(expected); |
| 437 | try testing.expectEqual(expected_limbs, out); |
| 438 | } |
| 439 | |
| 440 | test __xor_limb64 { |
| 441 | try test__xor_limb64(u64, 1, 2, 3); |
| 442 | try test__xor_limb64(u64, 3, 2, 1); |
| 443 | try test__xor_limb64(u65, maxInt(u65), 2, maxInt(u65) - 2); |
| 444 | try test__xor_limb64(u255, 7, 3, 4); |
| 445 | |
| 446 | try test__xor_limb64(i64, 3, 2, 1); |
| 447 | try test__xor_limb64(i64, -1, 2, -3); |
| 448 | try test__xor_limb64(i65, minInt(i65), -1, maxInt(i65)); |
| 449 | try test__xor_limb64(i255, -3, 2, -1); |
| 450 | |
| 451 | try test__xor_limb64(u150, maxInt(u150) - 1, 3, maxInt(u150) - 2); |
| 452 | try test__xor_limb64(i150, -2, 3, -3); |
| 453 | } |
| 454 | |
| 455 | comptime { |
| 456 | symbol(&__not_limb64, "__not_limb64"); |
| 457 | } |
| 458 | |
| 459 | fn __not_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) void { |
| 460 | const limb_cnt = usedLimbCount(bits); |
| 461 | const out = varLimbs(out_ptr, bits); |
| 462 | const a = constLimbs(a_ptr, bits); |
| 463 | |
| 464 | var i: usize = 0; |
| 465 | while (i < limb_cnt - 1) : (i += 1) { |
| 466 | limbSet(out, i, ~limbGet(a, i)); |
| 467 | } |
| 468 | |
| 469 | var limb: u64 = ~limbGet(a, i); |
| 470 | if (!is_signed and bits % 64 != 0) { |
| 471 | limb = limbWrap(limb, is_signed, bits); |
| 472 | } |
| 473 | limbSet(out, i, limb); |
| 474 | fixLastLimb(out_ptr, is_signed, bits); |
| 475 | } |
| 476 | |
| 477 | fn test__not_limb64(comptime T: type, a: T, expected: T) !void { |
| 478 | const int_info = @typeInfo(T).int; |
| 479 | const is_signed = int_info.signedness == .signed; |
| 480 | |
| 481 | var a_limbs = asLimbs(a); |
| 482 | var out: Limbs(T) = undefined; |
| 483 | __not_limb64(&out, &a_limbs, is_signed, int_info.bits); |
| 484 | |
| 485 | const expected_limbs = asLimbs(expected); |
| 486 | try testing.expectEqual(expected_limbs, out); |
| 487 | } |
| 488 | |
| 489 | test __not_limb64 { |
| 490 | try test__not_limb64(u64, 1, maxInt(u64) - 1); |
| 491 | try test__not_limb64(u64, 3, maxInt(u64) - 3); |
| 492 | try test__not_limb64(u65, maxInt(u65), 0); |
| 493 | try test__not_limb64(u255, 7, maxInt(u255) - 7); |
| 494 | |
| 495 | try test__not_limb64(i64, 3, -4); |
| 496 | try test__not_limb64(i64, -1, 0); |
| 497 | try test__not_limb64(i65, minInt(i65), maxInt(i65)); |
| 498 | try test__not_limb64(i255, -3, 2); |
| 499 | |
| 500 | try test__not_limb64(u150, maxInt(u150), 0); |
| 501 | try test__not_limb64(i150, maxInt(i150), minInt(i150)); |
| 502 | } |
| 503 | |
| 504 | comptime { |
| 505 | symbol(&__shlo_limb64, "__shlo_limb64"); |
| 506 | } |
| 507 | |
| 508 | fn __shlo_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, shift: u16, is_signed: bool, bits: u16) callconv(.c) bool { |
| 509 | const limb_cnt = usedLimbCount(bits); |
| 510 | const out = varLimbs(out_ptr, bits); |
| 511 | const a = constLimbs(a_ptr, bits); |
| 512 | |
| 513 | assert(shift < bits); |
| 514 | |
| 515 | const limb_shift = shift / 64; |
| 516 | const bit_shift = shift % 64; |
| 517 | |
| 518 | var carry: u64 = 0; |
| 519 | var i: usize = 0; |
| 520 | while (i < limb_cnt - 1) : (i += 1) { |
| 521 | if (i < limb_shift) { |
| 522 | limbSet(out, i, 0); |
| 523 | } else { |
| 524 | const limb = limbGet(a, i - limb_shift); |
| 525 | limbSet(out, i, (limb << @intCast(bit_shift)) | carry); |
| 526 | carry = if (bit_shift != 0) (limb >> @intCast(64 - bit_shift)) else 0; |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | const limb = limbGet(a, i - limb_shift); |
| 531 | const raw_last = (limb << @intCast(bit_shift)) | carry; |
| 532 | carry = if (bit_shift != 0) (limb >> @intCast(64 - bit_shift)) else 0; |
| 533 | |
| 534 | const last = if (bits % 64 == 0) raw_last else limbWrap(raw_last, is_signed, bits); |
| 535 | limbSet(out, i, last); |
| 536 | |
| 537 | const sign_extend: u64 = if (is_signed and (last >> 63) == 1) ~@as(u64, 0) else 0; |
| 538 | const expected_carry: u64 = if (bit_shift == 0) 0 else sign_extend >> @intCast(64 - bit_shift); |
| 539 | |
| 540 | var overflow = carry != expected_carry; |
| 541 | if (bits % 64 != 0) { |
| 542 | overflow = overflow or raw_last != last; |
| 543 | } |
| 544 | |
| 545 | var j = limb_cnt - limb_shift; |
| 546 | while (j < limb_cnt) : (j += 1) { |
| 547 | overflow = overflow or limbGet(a, j) != sign_extend; |
| 548 | } |
| 549 | |
| 550 | fixLastLimb(out_ptr, is_signed, bits); |
| 551 | return overflow; |
| 552 | } |
| 553 | |
| 554 | fn test__shlo_limb64(comptime T: type, a: T, shift: u16, expected: struct { T, bool }) !void { |
| 555 | const int_info = @typeInfo(T).int; |
| 556 | const is_signed = int_info.signedness == .signed; |
| 557 | |
| 558 | var a_limbs = asLimbs(a); |
| 559 | var out: Limbs(T) = undefined; |
| 560 | const overflow = __shlo_limb64(&out, &a_limbs, shift, is_signed, int_info.bits); |
| 561 | |
| 562 | const expected_limbs = asLimbs(expected[0]); |
| 563 | try testing.expectEqual(expected_limbs, out); |
| 564 | try testing.expectEqual(expected[1], overflow); |
| 565 | } |
| 566 | |
| 567 | test __shlo_limb64 { |
| 568 | try test__shlo_limb64(u64, 0x1234_5678_9ABC_DEF0, 4, .{ 0x2345_6789_ABCD_EF00, true }); |
| 569 | try test__shlo_limb64(u64, 0x8000_0000_0000_0001, 63, .{ 0x8000_0000_0000_0000, true }); |
| 570 | try test__shlo_limb64(u65, 1, 64, .{ 0x1_0000_0000_0000_0000, false }); |
| 571 | try test__shlo_limb64(u65, 0x1_0000_0000_0000_0000, 1, .{ 0, true }); |
| 572 | try test__shlo_limb64(u128, 0x1234_5678_9ABC_DEF0_1234_5678_9ABC_DEF0, 4, .{ 0x2345_6789_ABCD_EF01_2345_6789_ABCD_EF00, true }); |
| 573 | try test__shlo_limb64(u255, maxInt(u255), 1, .{ maxInt(u255) - 1, true }); |
| 574 | try test__shlo_limb64(u633, 1 << 299, 333, .{ 1 << 632, false }); |
| 575 | try test__shlo_limb64(u633, 1 << 300, 333, .{ 0, true }); |
| 576 | try test__shlo_limb64(u633, 1 << 298, 333, .{ 1 << 631, false }); |
| 577 | |
| 578 | try test__shlo_limb64(i64, -2, 1, .{ -4, false }); |
| 579 | try test__shlo_limb64(i64, minInt(i64), 1, .{ 0, true }); |
| 580 | try test__shlo_limb64(i64, minInt(i64), 63, .{ 0, true }); |
| 581 | try test__shlo_limb64(i65, minInt(i63), 1, .{ minInt(i64), false }); |
| 582 | try test__shlo_limb64(i65, -1, 17, .{ -1 << 17, false }); |
| 583 | try test__shlo_limb64(i65, -3, 64, .{ -1 << 64, true }); |
| 584 | try test__shlo_limb64(i128, -0x1234_5678_9ABC_DEF0_1234_5678_9ABC_DEF0, 4, .{ -0x2345_6789_ABCD_EF01_2345_6789_ABCD_EF00, true }); |
| 585 | try test__shlo_limb64(i255, -3, 1, .{ -6, false }); |
| 586 | try test__shlo_limb64(i633, 1 << 298, 333, .{ 1 << 631, false }); |
| 587 | try test__shlo_limb64(i633, 1 << 299, 333, .{ minInt(i633), true }); |
| 588 | try test__shlo_limb64(i633, 1 << 300, 333, .{ 0, true }); |
| 589 | try test__shlo_limb64(i633, 1 << 297, 333, .{ 1 << 630, false }); |
| 590 | try test__shlo_limb64(i633, -1 << 299, 333, .{ -1 << 632, false }); |
| 591 | try test__shlo_limb64(i633, -1 << 300, 333, .{ 0, true }); |
| 592 | try test__shlo_limb64(i633, -1 << 298, 333, .{ -1 << 631, false }); |
| 593 | |
| 594 | try test__shlo_limb64(u150, maxInt(u150), 1, .{ maxInt(u150) - 1, true }); |
| 595 | try test__shlo_limb64(i150, -3, 1, .{ -6, false }); |
| 596 | } |
| 597 | |
| 598 | comptime { |
| 599 | symbol(&__shr_limb64, "__shr_limb64"); |
| 600 | } |
| 601 | |
| 602 | fn __shr_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, shift: u16, is_signed: bool, bits: u16) callconv(.c) void { |
| 603 | const limb_cnt = usedLimbCount(bits); |
| 604 | const out = varLimbs(out_ptr, bits); |
| 605 | const a = constLimbs(a_ptr, bits); |
| 606 | |
| 607 | assert(shift < bits); |
| 608 | |
| 609 | const limb_shift = shift / 64; |
| 610 | const bit_shift = shift % 64; |
| 611 | |
| 612 | const ms = limbGet(a, limb_cnt - 1); |
| 613 | const sign_extend: u64 = if (is_signed and (ms >> 63) == 1) ~@as(u64, 0) else 0; |
| 614 | |
| 615 | var carry: u64 = if (bit_shift != 0) (sign_extend << @intCast(64 - bit_shift)) else 0; |
| 616 | var i: usize = 0; |
| 617 | while (i < limb_cnt) : (i += 1) { |
| 618 | const j = limb_cnt - 1 - i; |
| 619 | if (i < limb_shift) { |
| 620 | limbSet(out, j, sign_extend); |
| 621 | } else { |
| 622 | const limb = limbGet(a, j + limb_shift); |
| 623 | limbSet(out, j, (limb >> @intCast(bit_shift)) | carry); |
| 624 | carry = if (bit_shift != 0) (limb << @intCast(64 - bit_shift)) else 0; |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | fixLastLimb(out_ptr, is_signed, bits); |
| 629 | } |
| 630 | |
| 631 | fn test__shr_limb64(comptime T: type, a: T, shift: u16, expected: T) !void { |
| 632 | const int_info = @typeInfo(T).int; |
| 633 | const is_signed = int_info.signedness == .signed; |
| 634 | |
| 635 | var a_limbs = asLimbs(a); |
| 636 | var out: Limbs(T) = undefined; |
| 637 | __shr_limb64(&out, &a_limbs, shift, is_signed, int_info.bits); |
| 638 | |
| 639 | const expected_limbs = asLimbs(expected); |
| 640 | try testing.expectEqual(expected_limbs, out); |
| 641 | } |
| 642 | |
| 643 | test __shr_limb64 { |
| 644 | try test__shr_limb64(u64, 0x1234_5678_9ABC_DEF0, 4, 0x0123_4567_89AB_CDEF); |
| 645 | try test__shr_limb64(u64, 0x8000_0000_0000_0001, 63, 1); |
| 646 | try test__shr_limb64(u65, 0x1_0000_0000_0000_0000, 64, 1); |
| 647 | try test__shr_limb64(u65, 0x1_0000_0000_0000_0001, 1, 0x0_8000_0000_0000_0000); |
| 648 | try test__shr_limb64(u128, 0x1234_5678_9ABC_DEF0_1234_5678_9ABC_DEF0, 4, 0x0123_4567_89AB_CDEF_0123_4567_89AB_CDEF); |
| 649 | try test__shr_limb64(u255, maxInt(u255), 1, maxInt(u254)); |
| 650 | try test__shr_limb64(u633, 1 << 333, 333, 1); |
| 651 | try test__shr_limb64(u633, 1 << 334, 333, 2); |
| 652 | try test__shr_limb64(u633, 1 << 332, 333, 0); |
| 653 | |
| 654 | try test__shr_limb64(i64, -2, 1, -1); |
| 655 | try test__shr_limb64(i64, minInt(i64), 63, -1); |
| 656 | try test__shr_limb64(i65, minInt(i65), 1, minInt(i65) | (1 << 63)); |
| 657 | try test__shr_limb64(i65, -1, 17, -1); |
| 658 | try test__shr_limb64(i128, -0x1234_5678_9ABC_DEF0_1234_5678_9ABC_DEF0, 4, -0x0123_4567_89AB_CDEF_0123_4567_89AB_CDEF); |
| 659 | try test__shr_limb64(i255, -3, 1, -2); |
| 660 | try test__shr_limb64(i633, 1 << 333, 333, 1); |
| 661 | try test__shr_limb64(i633, 1 << 334, 333, 2); |
| 662 | try test__shr_limb64(i633, 1 << 332, 333, 0); |
| 663 | try test__shr_limb64(i633, -1 << 333, 333, -1); |
| 664 | try test__shr_limb64(i633, -1 << 334, 333, -2); |
| 665 | try test__shr_limb64(i633, -1 << 332, 333, -1); |
| 666 | |
| 667 | try test__shr_limb64(u150, maxInt(u150), 1, maxInt(u149)); |
| 668 | try test__shr_limb64(i150, -3, 1, -2); |
| 669 | } |
| 670 | |
| 671 | comptime { |
| 672 | symbol(&__clz_limb64, "__clz_limb64"); |
| 673 | } |
| 674 | |
| 675 | fn __clz_limb64(a_ptr: [*]const u64, bits: u16) callconv(.c) u16 { |
| 676 | const limb_cnt = usedLimbCount(bits); |
| 677 | const a = constLimbs(a_ptr, bits); |
| 678 | |
| 679 | var res: u16 = 0; |
| 680 | var i: usize = 0; |
| 681 | |
| 682 | if (bits % 64 != 0) { |
| 683 | const limb = limbGet(a, limb_cnt - 1); |
| 684 | if (limb == 0) { |
| 685 | res += bits % 64; |
| 686 | } else { |
| 687 | return @clz(limb << @intCast(64 - bits % 64)); |
| 688 | } |
| 689 | i += 1; |
| 690 | } |
| 691 | |
| 692 | while (i < limb_cnt) : (i += 1) { |
| 693 | const j = limb_cnt - 1 - i; |
| 694 | const limb = limbGet(a, j); |
| 695 | if (limb == 0) { |
| 696 | res += 64; |
| 697 | } else { |
| 698 | res += @clz(limb); |
| 699 | break; |
| 700 | } |
| 701 | } |
| 702 | |
| 703 | return res; |
| 704 | } |
| 705 | |
| 706 | fn test__clz_limb64(comptime T: type, a: T, expected: u16) !void { |
| 707 | const int_info = @typeInfo(T).int; |
| 708 | |
| 709 | var a_limbs = asLimbs(a); |
| 710 | const out = __clz_limb64(&a_limbs, int_info.bits); |
| 711 | |
| 712 | try testing.expectEqual(expected, out); |
| 713 | } |
| 714 | |
| 715 | test __clz_limb64 { |
| 716 | try test__clz_limb64(u64, 0, 64); |
| 717 | try test__clz_limb64(u65, 1 << 64, 0); |
| 718 | try test__clz_limb64(u65, 1 << 9, 55); |
| 719 | try test__clz_limb64(u128, 1 << 31, 96); |
| 720 | try test__clz_limb64(u255, 1 << 62, 192); |
| 721 | |
| 722 | try test__clz_limb64(i64, -1, 0); |
| 723 | try test__clz_limb64(i65, minInt(i65), 0); |
| 724 | try test__clz_limb64(i65, 1 << 32, 32); |
| 725 | try test__clz_limb64(i128, 0, 128); |
| 726 | try test__clz_limb64(i255, 1 << 130, 124); |
| 727 | |
| 728 | try test__clz_limb64(u150, 1 << 31, 118); |
| 729 | try test__clz_limb64(i150, maxInt(u65) - 1, 85); |
| 730 | } |
| 731 | |
| 732 | comptime { |
| 733 | symbol(&__ctz_limb64, "__ctz_limb64"); |
| 734 | } |
| 735 | |
| 736 | fn __ctz_limb64(a_ptr: [*]const u64, bits: u16) callconv(.c) u16 { |
| 737 | const limb_cnt = usedLimbCount(bits); |
| 738 | const a = constLimbs(a_ptr, bits); |
| 739 | |
| 740 | var res: u16 = 0; |
| 741 | var i: usize = 0; |
| 742 | while (i < limb_cnt - 1) : (i += 1) { |
| 743 | const limb = limbGet(a, i); |
| 744 | if (limb == 0) { |
| 745 | res += 64; |
| 746 | } else { |
| 747 | res += @ctz(limb); |
| 748 | return res; |
| 749 | } |
| 750 | } |
| 751 | |
| 752 | const limb = limbGet(a, i); |
| 753 | if (bits % 64 != 0 and limb == 0) { |
| 754 | res += bits % 64; |
| 755 | } else { |
| 756 | res += @ctz(limb); |
| 757 | } |
| 758 | |
| 759 | return res; |
| 760 | } |
| 761 | |
| 762 | fn test__ctz_limb64(comptime T: type, a: T, expected: u16) !void { |
| 763 | const int_info = @typeInfo(T).int; |
| 764 | |
| 765 | var a_limbs = asLimbs(a); |
| 766 | const out = __ctz_limb64(&a_limbs, int_info.bits); |
| 767 | |
| 768 | try testing.expectEqual(expected, out); |
| 769 | } |
| 770 | |
| 771 | test __ctz_limb64 { |
| 772 | try test__ctz_limb64(u64, 1 << 17, 17); |
| 773 | try test__ctz_limb64(u65, 1 << 64, 64); |
| 774 | try test__ctz_limb64(u65, 0, 65); |
| 775 | try test__ctz_limb64(u128, 1 << 100, 100); |
| 776 | try test__ctz_limb64(u255, 1 << 200, 200); |
| 777 | |
| 778 | try test__ctz_limb64(i64, -1 << 9, 9); |
| 779 | try test__ctz_limb64(i65, minInt(i65), 64); |
| 780 | try test__ctz_limb64(i65, 0, 65); |
| 781 | try test__ctz_limb64(i128, -1 << 73, 73); |
| 782 | try test__ctz_limb64(i255, 1 << 130, 130); |
| 783 | |
| 784 | try test__ctz_limb64(u150, 1 << 101, 101); |
| 785 | try test__ctz_limb64(i150, -1 << 74, 74); |
| 786 | } |
| 787 | |
| 788 | comptime { |
| 789 | symbol(&__popcount_limb64, "__popcount_limb64"); |
| 790 | } |
| 791 | |
| 792 | fn __popcount_limb64(a_ptr: [*]const u64, bits: u16) callconv(.c) u16 { |
| 793 | const limb_cnt = usedLimbCount(bits); |
| 794 | const a = constLimbs(a_ptr, bits); |
| 795 | |
| 796 | var res: u16 = 0; |
| 797 | var i: usize = 0; |
| 798 | while (i < limb_cnt - 1) : (i += 1) { |
| 799 | res += @popCount(limbGet(a, i)); |
| 800 | } |
| 801 | |
| 802 | var limb = limbGet(a, i); |
| 803 | if (bits % 64 != 0) { |
| 804 | limb <<= @intCast(64 - bits % 64); |
| 805 | } |
| 806 | res += @popCount(limb); |
| 807 | |
| 808 | return res; |
| 809 | } |
| 810 | |
| 811 | fn test__popcount_limb64(comptime T: type, a: T, expected: u16) !void { |
| 812 | const int_info = @typeInfo(T).int; |
| 813 | |
| 814 | var a_limbs = asLimbs(a); |
| 815 | const out = __popcount_limb64(&a_limbs, int_info.bits); |
| 816 | |
| 817 | try testing.expectEqual(expected, out); |
| 818 | } |
| 819 | |
| 820 | test __popcount_limb64 { |
| 821 | try test__popcount_limb64(u64, 0xF0F0_0000_0000_0001, 9); |
| 822 | try test__popcount_limb64(u65, 1 << 64, 1); |
| 823 | try test__popcount_limb64(u65, maxInt(u65), 65); |
| 824 | try test__popcount_limb64(u128, (1 << 100) | (1 << 5) | 1, 3); |
| 825 | try test__popcount_limb64(u255, maxInt(u255), 255); |
| 826 | |
| 827 | try test__popcount_limb64(i64, -1, 64); |
| 828 | try test__popcount_limb64(i65, minInt(i65), 1); |
| 829 | try test__popcount_limb64(i65, -1, 65); |
| 830 | try test__popcount_limb64(i128, -1 << 7, 121); |
| 831 | try test__popcount_limb64(i255, -1 << 200, 55); |
| 832 | |
| 833 | try test__popcount_limb64(u150, (1 << 149) | (1 << 65) | 1, 3); |
| 834 | try test__popcount_limb64(i150, -1 << 7, 143); |
| 835 | } |
| 836 | |
| 837 | comptime { |
| 838 | symbol(&__bitreverse_limb64, "__bitreverse_limb64"); |
| 839 | } |
| 840 | |
| 841 | fn __bitreverse_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) void { |
| 842 | const limb_cnt = usedLimbCount(bits); |
| 843 | const out = varLimbs(out_ptr, bits); |
| 844 | const a = constLimbs(a_ptr, bits); |
| 845 | |
| 846 | var i: usize = 0; |
| 847 | while (i < limb_cnt) : (i += 1) { |
| 848 | const j = limb_cnt - 1 - i; |
| 849 | limbSet(out, j, @bitReverse(limbGet(a, i))); |
| 850 | } |
| 851 | |
| 852 | if (bits % 64 != 0) { |
| 853 | __shr_limb64(out_ptr, out_ptr, 64 - bits % 64, is_signed, bits); |
| 854 | } |
| 855 | fixLastLimb(out_ptr, is_signed, bits); |
| 856 | } |
| 857 | |
| 858 | fn test__bitreverse_limb64(comptime T: type, a: T, expected: T) !void { |
| 859 | const int_info = @typeInfo(T).int; |
| 860 | const is_signed = int_info.signedness == .signed; |
| 861 | |
| 862 | var a_limbs = asLimbs(a); |
| 863 | var out: Limbs(T) = undefined; |
| 864 | __bitreverse_limb64(&out, &a_limbs, is_signed, int_info.bits); |
| 865 | |
| 866 | const expected_limbs = asLimbs(expected); |
| 867 | try testing.expectEqual(expected_limbs, out); |
| 868 | } |
| 869 | |
| 870 | test __bitreverse_limb64 { |
| 871 | try test__bitreverse_limb64(u64, 1 << 7, 1 << 56); |
| 872 | try test__bitreverse_limb64(u65, 1 << 64, 1); |
| 873 | try test__bitreverse_limb64(u65, 1 << 9, 1 << 55); |
| 874 | try test__bitreverse_limb64(u128, 1 << 100, 1 << 27); |
| 875 | try test__bitreverse_limb64(u255, 1 << 200, 1 << 54); |
| 876 | |
| 877 | try test__bitreverse_limb64(i64, -1, -1); |
| 878 | try test__bitreverse_limb64(i65, 1 << 32, 1 << 32); |
| 879 | try test__bitreverse_limb64(i65, minInt(i65), 1); |
| 880 | try test__bitreverse_limb64(i128, 1 << 63, 1 << 64); |
| 881 | try test__bitreverse_limb64(i255, 1 << 130, 1 << 124); |
| 882 | |
| 883 | try test__bitreverse_limb64(u150, 1 << 9, 1 << 140); |
| 884 | try test__bitreverse_limb64(i150, minInt(i150), 1); |
| 885 | } |
| 886 | |
| 887 | comptime { |
| 888 | symbol(&__byteswap_limb64, "__byteswap_limb64"); |
| 889 | } |
| 890 | |
| 891 | fn __byteswap_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) void { |
| 892 | const limb_cnt = usedLimbCount(bits); |
| 893 | const out = varLimbs(out_ptr, bits); |
| 894 | const a = constLimbs(a_ptr, bits); |
| 895 | |
| 896 | assert(bits % 8 == 0); |
| 897 | |
| 898 | var i: usize = 0; |
| 899 | while (i < limb_cnt) : (i += 1) { |
| 900 | const j = limb_cnt - 1 - i; |
| 901 | limbSet(out, j, @byteSwap(limbGet(a, i))); |
| 902 | } |
| 903 | |
| 904 | if (bits % 64 != 0) { |
| 905 | __shr_limb64(out_ptr, out_ptr, 64 - bits % 64, is_signed, bits); |
| 906 | } |
| 907 | fixLastLimb(out_ptr, is_signed, bits); |
| 908 | } |
| 909 | |
| 910 | fn test__byteswap_limb64(comptime T: type, a: T, expected: T) !void { |
| 911 | const int_info = @typeInfo(T).int; |
| 912 | const is_signed = int_info.signedness == .signed; |
| 913 | |
| 914 | var a_limbs = asLimbs(a); |
| 915 | var out: Limbs(T) = undefined; |
| 916 | __byteswap_limb64(&out, &a_limbs, is_signed, int_info.bits); |
| 917 | |
| 918 | const expected_limbs = asLimbs(expected); |
| 919 | try testing.expectEqual(expected_limbs, out); |
| 920 | } |
| 921 | |
| 922 | test __byteswap_limb64 { |
| 923 | try test__byteswap_limb64(u64, 0x0123_4567_89AB_CDEF, 0xEFCD_AB89_6745_2301); |
| 924 | try test__byteswap_limb64(u72, 0x01_23_45_67_89_AB_CD_EF_11, 0x11_EF_CD_AB_89_67_45_23_01); |
| 925 | try test__byteswap_limb64(u128, 1 << 72, 1 << 48); |
| 926 | try test__byteswap_limb64(u248, 1, 1 << 240); |
| 927 | try test__byteswap_limb64(u256, 1 << 120, 1 << 128); |
| 928 | |
| 929 | try test__byteswap_limb64(i64, minInt(i64), 128); |
| 930 | try test__byteswap_limb64(i72, 1, 1 << 64); |
| 931 | try test__byteswap_limb64(i72, -1, -1); |
| 932 | try test__byteswap_limb64(i128, 1 << 56, 1 << 64); |
| 933 | try test__byteswap_limb64(i248, minInt(i248), 128); |
| 934 | |
| 935 | try test__byteswap_limb64(u152, 1, 1 << 144); |
| 936 | try test__byteswap_limb64(i152, 1 << 56, 1 << 88); |
| 937 | } |
| 938 | |
| 939 | comptime { |
| 940 | symbol(&__mulo_limb64, "__mulo_limb64"); |
| 941 | } |
| 942 | |
| 943 | inline fn add3(x: *[3]u64, start: usize, v0: u64) void { |
| 944 | var i = start; |
| 945 | var v = v0; |
| 946 | while (i < 3) : (i += 1) { |
| 947 | const s = @addWithOverflow(x[i], v); |
| 948 | x[i] = s[0]; |
| 949 | if (s[1] == 0) break; |
| 950 | v = 1; |
| 951 | } |
| 952 | } |
| 953 | |
| 954 | fn mulwide(a: u64, b: u64) [2]u64 { |
| 955 | const muldXi = @import("mulXi3.zig").muldXi; |
| 956 | return @bitCast(muldXi(u64, a, b)); |
| 957 | } |
| 958 | |
| 959 | fn __mulo_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, b_ptr: [*]const u64, is_signed: bool, bits: u16) callconv(.c) bool { |
| 960 | const limb_cnt = usedLimbCount(bits); |
| 961 | |
| 962 | const out = varLimbs(out_ptr, bits); |
| 963 | const a = constLimbs(a_ptr, bits); |
| 964 | const b = constLimbs(b_ptr, bits); |
| 965 | |
| 966 | @memset(out, 0); |
| 967 | |
| 968 | const all_ones = ~@as(u64, 0); |
| 969 | const a_neg = is_signed and ((limbGet(a, limb_cnt - 1) >> 63) != 0); |
| 970 | const b_neg = is_signed and ((limbGet(b, limb_cnt - 1) >> 63) != 0); |
| 971 | |
| 972 | var carry: [3]u64 = @splat(0); |
| 973 | var hi_zero = true; |
| 974 | var hi_ones = true; |
| 975 | var hi_borrow: u1 = 0; |
| 976 | var raw_last: u64 = 0; |
| 977 | |
| 978 | var k: usize = 0; |
| 979 | while (k < 2 * limb_cnt) : (k += 1) { |
| 980 | var acc = carry; |
| 981 | |
| 982 | var i: usize = if (k < limb_cnt) 0 else k - (limb_cnt - 1); |
| 983 | while (i < limb_cnt and i <= k) : (i += 1) { |
| 984 | const j = k - i; |
| 985 | if (j >= limb_cnt) continue; |
| 986 | |
| 987 | const p = mulwide(limbGet(a, i), limbGet(b, j)); |
| 988 | add3(&acc, 0, p[0]); |
| 989 | add3(&acc, 1, p[1]); |
| 990 | } |
| 991 | |
| 992 | var limb = acc[0]; |
| 993 | if (k < limb_cnt) { |
| 994 | limbSet(out, k, limb); |
| 995 | if (k == limb_cnt - 1) raw_last = limb; |
| 996 | } else { |
| 997 | if (is_signed) { |
| 998 | const h = k - limb_cnt; |
| 999 | |
| 1000 | const s0 = @subWithOverflow(limb, if (a_neg) limbGet(b, h) else 0); |
| 1001 | const s1 = @subWithOverflow(s0[0], if (b_neg) limbGet(a, h) else 0); |
| 1002 | const s2 = @subWithOverflow(s1[0], hi_borrow); |
| 1003 | |
| 1004 | limb = s2[0]; |
| 1005 | hi_borrow = @intFromBool(s0[1] != 0 or s1[1] != 0 or s2[1] != 0); |
| 1006 | } |
| 1007 | |
| 1008 | hi_zero = hi_zero and limb == 0; |
| 1009 | hi_ones = hi_ones and limb == all_ones; |
| 1010 | } |
| 1011 | |
| 1012 | carry = .{ acc[1], acc[2], 0 }; |
| 1013 | } |
| 1014 | |
| 1015 | const last = if (bits % 64 == 0) raw_last else limbWrap(raw_last, is_signed, bits); |
| 1016 | if (bits % 64 != 0) { |
| 1017 | limbSet(out, limb_cnt - 1, last); |
| 1018 | } |
| 1019 | |
| 1020 | fixLastLimb(out_ptr, is_signed, bits); |
| 1021 | |
| 1022 | if (!is_signed) { |
| 1023 | return !hi_zero or raw_last != last; |
| 1024 | } |
| 1025 | |
| 1026 | const sign_extend: u64 = if ((last >> 63) == 1) all_ones else 0; |
| 1027 | return (raw_last != last) or if (sign_extend == 0) !hi_zero else !hi_ones; |
| 1028 | } |
| 1029 | |
| 1030 | fn test__mulo_limb64(comptime T: type, a: T, b: T, expected: struct { T, bool }) !void { |
| 1031 | const int_info = @typeInfo(T).int; |
| 1032 | const is_signed = int_info.signedness == .signed; |
| 1033 | |
| 1034 | var a_limbs = asLimbs(a); |
| 1035 | var b_limbs = asLimbs(b); |
| 1036 | var out: Limbs(T) = undefined; |
| 1037 | const overflow = __mulo_limb64(&out, &a_limbs, &b_limbs, is_signed, int_info.bits); |
| 1038 | |
| 1039 | const expected_limbs = asLimbs(expected[0]); |
| 1040 | try testing.expectEqual(expected_limbs, out); |
| 1041 | try testing.expectEqual(expected[1], overflow); |
| 1042 | } |
| 1043 | |
| 1044 | test __mulo_limb64 { |
| 1045 | try test__mulo_limb64(u64, 3, 5, .{ 15, false }); |
| 1046 | try test__mulo_limb64(u64, maxInt(u64), 2, .{ maxInt(u64) - 1, true }); |
| 1047 | try test__mulo_limb64(u65, 1 << 32, 1 << 32, .{ 1 << 64, false }); |
| 1048 | try test__mulo_limb64(u65, 1 << 64, 2, .{ 0, true }); |
| 1049 | try test__mulo_limb64(u128, 1 << 80, 1 << 40, .{ 1 << 120, false }); |
| 1050 | try test__mulo_limb64(u128, 1 << 100, 1 << 40, .{ 0, true }); |
| 1051 | try test__mulo_limb64(u255, 7, 9, .{ 63, false }); |
| 1052 | try test__mulo_limb64(u255, maxInt(u255), 2, .{ maxInt(u255) - 1, true }); |
| 1053 | |
| 1054 | try test__mulo_limb64(i64, -3, 2, .{ -6, false }); |
| 1055 | try test__mulo_limb64(i64, maxInt(i64), 2, .{ -2, true }); |
| 1056 | try test__mulo_limb64(i65, 1 << 63, 2, .{ minInt(i65), true }); |
| 1057 | try test__mulo_limb64(i65, -1 << 32, 1 << 16, .{ -1 << 48, false }); |
| 1058 | try test__mulo_limb64(i128, 1 << 100, 1 << 27, .{ minInt(i128), true }); |
| 1059 | try test__mulo_limb64(i128, -1 << 80, 1 << 40, .{ -1 << 120, false }); |
| 1060 | try test__mulo_limb64(i255, -3, 2, .{ -6, false }); |
| 1061 | try test__mulo_limb64(i255, maxInt(i255), 2, .{ -2, true }); |
| 1062 | |
| 1063 | try test__mulo_limb64(u200, 0, maxInt(u200), .{ 0, false }); |
| 1064 | try test__mulo_limb64(u200, 1, maxInt(u200), .{ maxInt(u200), false }); |
| 1065 | try test__mulo_limb64(u200, 1 << 100, 1 << 99, .{ 1 << 199, false }); |
| 1066 | try test__mulo_limb64(u200, 1 << 100, 1 << 100, .{ 0, true }); |
| 1067 | try test__mulo_limb64(u200, maxInt(u200), maxInt(u200), .{ 1, true }); |
| 1068 | |
| 1069 | try test__mulo_limb64(i200, 0, -1, .{ 0, false }); |
| 1070 | try test__mulo_limb64(i200, -1, -1, .{ 1, false }); |
| 1071 | try test__mulo_limb64(i200, -1, minInt(i200), .{ minInt(i200), true }); |
| 1072 | try test__mulo_limb64(i200, maxInt(i200), 2, .{ -2, true }); |
| 1073 | try test__mulo_limb64(i200, 1 << 100, 1 << 98, .{ 1 << 198, false }); |
| 1074 | try test__mulo_limb64(i200, 1 << 100, 1 << 99, .{ minInt(i200), true }); |
| 1075 | try test__mulo_limb64(i200, maxInt(i200), maxInt(i200), .{ 1, true }); |
| 1076 | try test__mulo_limb64(i200, minInt(i200), minInt(i200), .{ 0, true }); |
| 1077 | |
| 1078 | try test__mulo_limb64(u150, maxInt(u150), 2, .{ maxInt(u150) - 1, true }); |
| 1079 | try test__mulo_limb64(i150, maxInt(i150), 2, .{ -2, true }); |
| 1080 | } |
| 1081 | |
| 1082 | comptime { |
| 1083 | symbol(&__abs_limb64, "__abs_limb64"); |
| 1084 | } |
| 1085 | |
| 1086 | fn __abs_limb64(out_ptr: [*]u64, a_ptr: [*]const u64, bits: u16) callconv(.c) void { |
| 1087 | const limb_cnt = limbCount(bits); |
| 1088 | const out = out_ptr[0..limb_cnt]; |
| 1089 | const a = a_ptr[0..limb_cnt]; |
| 1090 | |
| 1091 | const ms = limbGet(a, limb_cnt - 1); |
| 1092 | if ((ms >> 63) == 0) { |
| 1093 | @memcpy(out, a); |
| 1094 | return; |
| 1095 | } |
| 1096 | |
| 1097 | var carry: u1 = 1; |
| 1098 | var i: usize = 0; |
| 1099 | while (i < limb_cnt) : (i += 1) { |
| 1100 | const s = @addWithOverflow(~limbGet(a, i), carry); |
| 1101 | limbSet(out, i, s[0]); |
| 1102 | carry = s[1]; |
| 1103 | } |
| 1104 | } |
| 1105 | |
| 1106 | fn test__abs_limb64(comptime T: type, a: T, expected: @Int(.unsigned, @typeInfo(T).int.bits)) !void { |
| 1107 | const int_info = @typeInfo(T).int; |
| 1108 | comptime assert(int_info.signedness == .signed); |
| 1109 | |
| 1110 | var a_limbs = asLimbs(a); |
| 1111 | var out: Limbs(@TypeOf(expected)) = undefined; |
| 1112 | __abs_limb64(&out, &a_limbs, int_info.bits); |
| 1113 | |
| 1114 | const expected_limbs = asLimbs(expected); |
| 1115 | try testing.expectEqual(expected_limbs, out); |
| 1116 | } |
| 1117 | |
| 1118 | test __abs_limb64 { |
| 1119 | try test__abs_limb64(i64, 0, 0); |
| 1120 | try test__abs_limb64(i64, -1, 1); |
| 1121 | try test__abs_limb64(i64, minInt(i64), 1 << 63); |
| 1122 | try test__abs_limb64(i65, -1, 1); |
| 1123 | try test__abs_limb64(i65, minInt(i65), 1 << 64); |
| 1124 | try test__abs_limb64(i65, maxInt(i65), maxInt(i65)); |
| 1125 | try test__abs_limb64(i128, -1 << 80, 1 << 80); |
| 1126 | try test__abs_limb64(i128, 1 << 64, 1 << 64); |
| 1127 | try test__abs_limb64(i200, -1 << 198, 1 << 198); |
| 1128 | try test__abs_limb64(i255, -5, 5); |
| 1129 | try test__abs_limb64(i255, minInt(i255), 1 << 254); |
| 1130 | |
| 1131 | try test__abs_limb64(i150, -40, 40); |
| 1132 | } |