| author |  Daniele Cocca <daniele.cocca@gmail.com> 2022-03-13 00:43:51+00:00 |
| committer | Daniele Cocca <daniele.cocca@gmail.com> 2022-03-13 09:59:15+00:00 |
| log | d0277a3d17f738036a951a44f367839597dcaee4 |
| tree | 5415e8fcfc445b567fb2c0ba2e2f1b187f46d021 |
| parent | 87744a7ea9a2449764a110da4210d7750e3938ee |
CBE: implement popCount, byteSwap, bitReverse for ints <= 128 bits
This folds the airCountZeroes() code from
226fcd7c709ec664c5d883042cf7beb3026f66cb back into airBuiltinCall(),
since most of these builtins happen to require the same arguments and
can be unified under a common function signature.6 files changed, 107 insertions(+), 28 deletions(-)
src/codegen/c.zig+9-22
| ... | ... | @@ -1711,8 +1711,8 @@ fn genBody(f: *Function, body: []const Air.Inst.Index) error{ AnalysisFail, OutO |
| 1711 | 1711 | .memcpy => try airMemcpy(f, inst), |
| 1712 | 1712 | .set_union_tag => try airSetUnionTag(f, inst), |
| 1713 | 1713 | .get_union_tag => try airGetUnionTag(f, inst), |
| 1714 | .clz => try airCountZeroes(f, inst, "clz"), | |
| 1715 | .ctz => try airCountZeroes(f, inst, "ctz"), | |
| 1714 | .clz => try airBuiltinCall(f, inst, "clz"), | |
| 1715 | .ctz => try airBuiltinCall(f, inst, "ctz"), | |
| 1716 | 1716 | .popcount => try airBuiltinCall(f, inst, "popcount"), |
| 1717 | 1717 | .byte_swap => try airBuiltinCall(f, inst, "byte_swap"), |
| 1718 | 1718 | .bit_reverse => try airBuiltinCall(f, inst, "bit_reverse"), |
| ... | ... | @@ -3352,23 +3352,6 @@ fn airPtrToInt(f: *Function, inst: Air.Inst.Index) !CValue { |
| 3352 | 3352 | fn airBuiltinCall(f: *Function, inst: Air.Inst.Index, fn_name: [*:0]const u8) !CValue { |
| 3353 | 3353 | if (f.liveness.isUnused(inst)) return CValue.none; |
| 3354 | 3354 | |
| 3355 | const inst_ty = f.air.typeOfIndex(inst); | |
| 3356 | const local = try f.allocLocal(inst_ty, .Const); | |
| 3357 | const ty_op = f.air.instructions.items(.data)[inst].ty_op; | |
| 3358 | const writer = f.object.writer(); | |
| 3359 | const operand = try f.resolveInst(ty_op.operand); | |
| 3360 | ||
| 3361 | // TODO implement the function in zig.h and call it here | |
| 3362 | ||
| 3363 | try writer.print(" = {s}(", .{fn_name}); | |
| 3364 | try f.writeCValue(writer, operand); | |
| 3365 | try writer.writeAll(");\n"); | |
| 3366 | return local; | |
| 3367 | } | |
| 3368 | ||
| 3369 | fn airCountZeroes(f: *Function, inst: Air.Inst.Index, fn_name: [*:0]const u8) !CValue { | |
| 3370 | if (f.liveness.isUnused(inst)) return CValue.none; | |
| 3371 | ||
| 3372 | 3355 | const inst_ty = f.air.typeOfIndex(inst); |
| 3373 | 3356 | const local = try f.allocLocal(inst_ty, .Const); |
| 3374 | 3357 | const operand = f.air.instructions.items(.data)[inst].ty_op.operand; |
| ... | ... | @@ -3376,13 +3359,17 @@ fn airCountZeroes(f: *Function, inst: Air.Inst.Index, fn_name: [*:0]const u8) !C |
| 3376 | 3359 | const target = f.object.dg.module.getTarget(); |
| 3377 | 3360 | const writer = f.object.writer(); |
| 3378 | 3361 | |
| 3379 | const zig_bits = operand_ty.intInfo(target).bits; | |
| 3380 | _ = toCIntBits(zig_bits) orelse | |
| 3362 | const int_info = operand_ty.intInfo(target); | |
| 3363 | _ = toCIntBits(int_info.bits) orelse | |
| 3381 | 3364 | return f.fail("TODO: C backend: implement integer types larger than 128 bits", .{}); |
| 3365 | const signed_type = switch (int_info.signedness) { | |
| 3366 | .signed => "true", | |
| 3367 | .unsigned => "false", | |
| 3368 | }; | |
| 3382 | 3369 | |
| 3383 | 3370 | try writer.print(" = zig_{s}(", .{fn_name}); |
| 3384 | 3371 | try f.writeCValue(writer, try f.resolveInst(operand)); |
| 3385 | try writer.print(", {d});\n", .{zig_bits}); | |
| 3372 | try writer.print(", {d}, {s});\n", .{ int_info.bits, signed_type }); | |
| 3386 | 3373 | return local; |
| 3387 | 3374 | } |
| 3388 | 3375 |
src/link/C/zig.h+94-2
| ... | ... | @@ -502,8 +502,20 @@ zig_shl_sat_s(int, int, ((sizeof(int )) * CHAR_BIT - 1)) |
| 502 | 502 | zig_shl_sat_s(long, long, ((sizeof(long )) * CHAR_BIT - 1)) |
| 503 | 503 | |
| 504 | 504 | #define zig_bitsizeof(T) (CHAR_BIT * sizeof(T)) |
| 505 | #define zig_bit_mask(T, bit_width) \ | |
| 506 | ((bit_width) == 128 \ | |
| 507 | ? ((T)(((uint128_t) 0xffffffffffffffff) << 64 | \ | |
| 508 | (uint128_t) 0xffffffffffffffff)) \ | |
| 509 | : (((T)1 << (T)(bit_width)) - 1)) | |
| 510 | ||
| 511 | static inline uint128_t zig_sign_extend(uint128_t value, uint128_t zig_type_bit_width) { | |
| 512 | const uint128_t m = (uint128_t)1 << (zig_type_bit_width - 1); | |
| 513 | return (value ^ m) - m; | |
| 514 | } | |
| 515 | ||
| 516 | static inline int zig_clz(unsigned long long value, uint8_t zig_type_bit_width, bool signed_type) { | |
| 517 | (void)signed_type; // unused | |
| 505 | 518 | |
| 506 | static inline int zig_clz(uint64_t value, uint8_t zig_type_bit_width) { | |
| 507 | 519 | if (value == 0) return zig_type_bit_width; |
| 508 | 520 | if (zig_type_bit_width <= zig_bitsizeof(unsigned int)) |
| 509 | 521 | return (__builtin_clz(value) - zig_bitsizeof(unsigned int) + zig_type_bit_width); |
| ... | ... | @@ -512,9 +524,89 @@ static inline int zig_clz(uint64_t value, uint8_t zig_type_bit_width) { |
| 512 | 524 | return (__builtin_clzll(value) - zig_bitsizeof(unsigned long long) + zig_type_bit_width); |
| 513 | 525 | } |
| 514 | 526 | |
| 515 | static inline int zig_ctz(uint64_t value, uint8_t zig_type_bit_width) { | |
| 527 | static inline int zig_ctz(unsigned long long value, uint8_t zig_type_bit_width, bool signed_type) { | |
| 528 | (void)signed_type; // unused | |
| 529 | ||
| 516 | 530 | if (value == 0) return zig_type_bit_width; |
| 517 | 531 | if (zig_type_bit_width <= zig_bitsizeof(unsigned int)) return __builtin_ctz(value); |
| 518 | 532 | if (zig_type_bit_width <= zig_bitsizeof(unsigned long)) return __builtin_ctzl(value); |
| 519 | 533 | return __builtin_ctzll(value); |
| 520 | 534 | } |
| 535 | ||
| 536 | static inline int zig_popcount(unsigned long long value, uint8_t zig_type_bit_width, bool signed_type) { | |
| 537 | (void)signed_type; // unused | |
| 538 | ||
| 539 | const unsigned long long mask = zig_bit_mask(unsigned long long, zig_type_bit_width); | |
| 540 | if (zig_type_bit_width <= zig_bitsizeof(unsigned int)) | |
| 541 | return __builtin_popcount(value & mask); | |
| 542 | if (zig_type_bit_width <= zig_bitsizeof(unsigned long)) | |
| 543 | return __builtin_popcountl(value & mask); | |
| 544 | return __builtin_popcountll(value & mask); | |
| 545 | } | |
| 546 | ||
| 547 | static inline uint128_t zig_byte_swap(uint128_t value, uint8_t zig_type_bit_width, bool signed_type) { | |
| 548 | if (zig_type_bit_width <= 8) return value; | |
| 549 | if (zig_type_bit_width <= 16) return __builtin_bswap16(value); | |
| 550 | if (zig_type_bit_width <= 32) { | |
| 551 | const uint32_t swapped = __builtin_bswap32(value) >> (32 - zig_type_bit_width); | |
| 552 | return (signed_type ? zig_sign_extend(swapped, zig_type_bit_width) : swapped); | |
| 553 | } | |
| 554 | if (zig_type_bit_width <= 64) { | |
| 555 | const uint64_t swapped = __builtin_bswap64(value) >> (64 - zig_type_bit_width); | |
| 556 | return (signed_type ? zig_sign_extend(swapped, zig_type_bit_width) : swapped); | |
| 557 | } | |
| 558 | const uint128_t mask = zig_bit_mask(uint128_t, zig_type_bit_width); | |
| 559 | const uint128_t hi = __builtin_bswap64((uint64_t)(value >> 64)); | |
| 560 | const uint128_t lo = __builtin_bswap64((uint64_t)value); | |
| 561 | const uint128_t swapped = (((lo << 64 | hi) >> (128 - zig_type_bit_width))) & mask; | |
| 562 | return (signed_type ? zig_sign_extend(swapped, zig_type_bit_width) : swapped); | |
| 563 | } | |
| 564 | ||
| 565 | static inline uint128_t zig_bit_reverse(uint128_t value, uint8_t zig_type_bit_width, bool signed_type) { | |
| 566 | static const uint128_t reverse_lut[256] = { | |
| 567 | 0x00, 0x80, 0x40, 0xc0, 0x20, 0xa0, 0x60, 0xe0, 0x10, 0x90, 0x50, 0xd0, | |
| 568 | 0x30, 0xb0, 0x70, 0xf0, 0x08, 0x88, 0x48, 0xc8, 0x28, 0xa8, 0x68, 0xe8, | |
| 569 | 0x18, 0x98, 0x58, 0xd8, 0x38, 0xb8, 0x78, 0xf8, 0x04, 0x84, 0x44, 0xc4, | |
| 570 | 0x24, 0xa4, 0x64, 0xe4, 0x14, 0x94, 0x54, 0xd4, 0x34, 0xb4, 0x74, 0xf4, | |
| 571 | 0x0c, 0x8c, 0x4c, 0xcc, 0x2c, 0xac, 0x6c, 0xec, 0x1c, 0x9c, 0x5c, 0xdc, | |
| 572 | 0x3c, 0xbc, 0x7c, 0xfc, 0x02, 0x82, 0x42, 0xc2, 0x22, 0xa2, 0x62, 0xe2, | |
| 573 | 0x12, 0x92, 0x52, 0xd2, 0x32, 0xb2, 0x72, 0xf2, 0x0a, 0x8a, 0x4a, 0xca, | |
| 574 | 0x2a, 0xaa, 0x6a, 0xea, 0x1a, 0x9a, 0x5a, 0xda, 0x3a, 0xba, 0x7a, 0xfa, | |
| 575 | 0x06, 0x86, 0x46, 0xc6, 0x26, 0xa6, 0x66, 0xe6, 0x16, 0x96, 0x56, 0xd6, | |
| 576 | 0x36, 0xb6, 0x76, 0xf6, 0x0e, 0x8e, 0x4e, 0xce, 0x2e, 0xae, 0x6e, 0xee, | |
| 577 | 0x1e, 0x9e, 0x5e, 0xde, 0x3e, 0xbe, 0x7e, 0xfe, 0x01, 0x81, 0x41, 0xc1, | |
| 578 | 0x21, 0xa1, 0x61, 0xe1, 0x11, 0x91, 0x51, 0xd1, 0x31, 0xb1, 0x71, 0xf1, | |
| 579 | 0x09, 0x89, 0x49, 0xc9, 0x29, 0xa9, 0x69, 0xe9, 0x19, 0x99, 0x59, 0xd9, | |
| 580 | 0x39, 0xb9, 0x79, 0xf9, 0x05, 0x85, 0x45, 0xc5, 0x25, 0xa5, 0x65, 0xe5, | |
| 581 | 0x15, 0x95, 0x55, 0xd5, 0x35, 0xb5, 0x75, 0xf5, 0x0d, 0x8d, 0x4d, 0xcd, | |
| 582 | 0x2d, 0xad, 0x6d, 0xed, 0x1d, 0x9d, 0x5d, 0xdd, 0x3d, 0xbd, 0x7d, 0xfd, | |
| 583 | 0x03, 0x83, 0x43, 0xc3, 0x23, 0xa3, 0x63, 0xe3, 0x13, 0x93, 0x53, 0xd3, | |
| 584 | 0x33, 0xb3, 0x73, 0xf3, 0x0b, 0x8b, 0x4b, 0xcb, 0x2b, 0xab, 0x6b, 0xeb, | |
| 585 | 0x1b, 0x9b, 0x5b, 0xdb, 0x3b, 0xbb, 0x7b, 0xfb, 0x07, 0x87, 0x47, 0xc7, | |
| 586 | 0x27, 0xa7, 0x67, 0xe7, 0x17, 0x97, 0x57, 0xd7, 0x37, 0xb7, 0x77, 0xf7, | |
| 587 | 0x0f, 0x8f, 0x4f, 0xcf, 0x2f, 0xaf, 0x6f, 0xef, 0x1f, 0x9f, 0x5f, 0xdf, | |
| 588 | 0x3f, 0xbf, 0x7f, 0xff | |
| 589 | }; | |
| 590 | const uint128_t mask = zig_bit_mask(uint128_t, zig_type_bit_width); | |
| 591 | const uint128_t swapped = zig_byte_swap(value, zig_type_bit_width, signed_type); | |
| 592 | const uint128_t reversed = ( | |
| 593 | (reverse_lut[(swapped >> 0x78) & 0xff] << 0x78) | | |
| 594 | (reverse_lut[(swapped >> 0x70) & 0xff] << 0x70) | | |
| 595 | (reverse_lut[(swapped >> 0x68) & 0xff] << 0x68) | | |
| 596 | (reverse_lut[(swapped >> 0x60) & 0xff] << 0x60) | | |
| 597 | (reverse_lut[(swapped >> 0x58) & 0xff] << 0x58) | | |
| 598 | (reverse_lut[(swapped >> 0x50) & 0xff] << 0x50) | | |
| 599 | (reverse_lut[(swapped >> 0x48) & 0xff] << 0x48) | | |
| 600 | (reverse_lut[(swapped >> 0x40) & 0xff] << 0x40) | | |
| 601 | (reverse_lut[(swapped >> 0x38) & 0xff] << 0x38) | | |
| 602 | (reverse_lut[(swapped >> 0x30) & 0xff] << 0x30) | | |
| 603 | (reverse_lut[(swapped >> 0x28) & 0xff] << 0x28) | | |
| 604 | (reverse_lut[(swapped >> 0x20) & 0xff] << 0x20) | | |
| 605 | (reverse_lut[(swapped >> 0x18) & 0xff] << 0x18) | | |
| 606 | (reverse_lut[(swapped >> 0x10) & 0xff] << 0x10) | | |
| 607 | (reverse_lut[(swapped >> 0x08) & 0xff] << 0x08) | | |
| 608 | (reverse_lut[(swapped >> 0x00) & 0xff] << 0x00)); | |
| 609 | const uint128_t result = | |
| 610 | (reversed >> (zig_type_bit_width < 8 ? 8 - zig_type_bit_width : 0)) & mask; | |
| 611 | return (signed_type ? zig_sign_extend(result, zig_type_bit_width) : result); | |
| 612 | } |
test/behavior.zig+1-1
| ... | ... | @@ -139,6 +139,7 @@ test { |
| 139 | 139 | _ = @import("behavior/eval.zig"); |
| 140 | 140 | _ = @import("behavior/export_self_referential_type_info.zig"); |
| 141 | 141 | _ = @import("behavior/int128.zig"); |
| 142 | _ = @import("behavior/popcount.zig"); | |
| 142 | 143 | _ = @import("behavior/translate_c_macros.zig"); |
| 143 | 144 | _ = @import("behavior/union_with_members.zig"); |
| 144 | 145 | |
| ... | ... | @@ -147,7 +148,6 @@ test { |
| 147 | 148 | _ = @import("behavior/atomics.zig"); |
| 148 | 149 | _ = @import("behavior/export.zig"); |
| 149 | 150 | _ = @import("behavior/maximum_minimum.zig"); |
| 150 | _ = @import("behavior/popcount.zig"); | |
| 151 | 151 | _ = @import("behavior/saturating_arithmetic.zig"); |
| 152 | 152 | _ = @import("behavior/widening.zig"); |
| 153 | 153 | _ = @import("behavior/bugs/2114.zig"); |
test/behavior/bitreverse.zig-2
| ... | ... | @@ -5,7 +5,6 @@ const minInt = std.math.minInt; |
| 5 | 5 | |
| 6 | 6 | test "@bitReverse large exotic integer" { |
| 7 | 7 | if (builtin.zig_backend == .stage2_wasm) return error.SkipZigTest; |
| 8 | if (builtin.zig_backend == .stage2_c) return error.SkipZigTest; | |
| 9 | 8 | // Currently failing on stage1 for big-endian targets |
| 10 | 9 | if (builtin.zig_backend == .stage1) return error.SkipZigTest; |
| 11 | 10 | |
| ... | ... | @@ -14,7 +13,6 @@ test "@bitReverse large exotic integer" { |
| 14 | 13 | |
| 15 | 14 | test "@bitReverse" { |
| 16 | 15 | if (builtin.zig_backend == .stage2_wasm) return error.SkipZigTest; |
| 17 | if (builtin.zig_backend == .stage2_c) return error.SkipZigTest; | |
| 18 | 16 | if (builtin.zig_backend == .stage2_x86_64) return error.SkipZigTest; |
| 19 | 17 | if (builtin.zig_backend == .stage2_arm) return error.SkipZigTest; |
| 20 | 18 | if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest; |
test/behavior/byteswap.zig-1
| ... | ... | @@ -4,7 +4,6 @@ const expect = std.testing.expect; |
| 4 | 4 | |
| 5 | 5 | test "@byteSwap integers" { |
| 6 | 6 | if (builtin.zig_backend == .stage2_wasm) return error.SkipZigTest; |
| 7 | if (builtin.zig_backend == .stage2_c) return error.SkipZigTest; | |
| 8 | 7 | if (builtin.zig_backend == .stage2_x86_64) return error.SkipZigTest; |
| 9 | 8 | if (builtin.zig_backend == .stage2_arm) return error.SkipZigTest; |
| 10 | 9 | if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest; |
test/behavior/popcount.zig+3
| ... | ... | @@ -1,3 +1,4 @@ |
| 1 | const builtin = @import("builtin"); | |
| 1 | 2 | const std = @import("std"); |
| 2 | 3 | const expect = std.testing.expect; |
| 3 | 4 | const expectEqual = std.testing.expectEqual; |
| ... | ... | @@ -45,6 +46,8 @@ fn testPopCountIntegers() !void { |
| 45 | 46 | } |
| 46 | 47 | |
| 47 | 48 | test "@popCount vectors" { |
| 49 | if (builtin.zig_backend == .stage2_c) return error.SkipZigTest; // TODO | |
| 50 | ||
| 48 | 51 | comptime try testPopCountVectors(); |
| 49 | 52 | try testPopCountVectors(); |
| 50 | 53 | } |