| ... | ... | @@ -502,19 +502,249 @@ 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) == zig_bitsizeof(T) \ |
| 507 | ? ((T)-1) \ |
| 508 | : (((T)1 << (T)(bit_width)) - 1)) |
| 505 | 509 | |
| 506 | | static inline int zig_clz(uint64_t value, uint8_t zig_type_bit_width) { |
| 510 | static inline int zig_clz(unsigned int value, uint8_t zig_type_bit_width) { |
| 507 | 511 | if (value == 0) return zig_type_bit_width; |
| 508 | | if (zig_type_bit_width <= zig_bitsizeof(unsigned int)) |
| 509 | | return (__builtin_clz(value) - zig_bitsizeof(unsigned int) + zig_type_bit_width); |
| 510 | | if (zig_type_bit_width <= zig_bitsizeof(unsigned long)) |
| 511 | | return (__builtin_clzl(value) - zig_bitsizeof(unsigned long) + zig_type_bit_width); |
| 512 | | return (__builtin_clzll(value) - zig_bitsizeof(unsigned long long) + zig_type_bit_width); |
| 512 | return __builtin_clz(value) - zig_bitsizeof(unsigned int) + zig_type_bit_width; |
| 513 | 513 | } |
| 514 | 514 | |
| 515 | | static inline int zig_ctz(uint64_t value, uint8_t zig_type_bit_width) { |
| 515 | static inline int zig_clzl(unsigned long value, uint8_t zig_type_bit_width) { |
| 516 | 516 | if (value == 0) return zig_type_bit_width; |
| 517 | | if (zig_type_bit_width <= zig_bitsizeof(unsigned int)) return __builtin_ctz(value); |
| 518 | | if (zig_type_bit_width <= zig_bitsizeof(unsigned long)) return __builtin_ctzl(value); |
| 519 | | return __builtin_ctzll(value); |
| 517 | return __builtin_clzl(value) - zig_bitsizeof(unsigned long) + zig_type_bit_width; |
| 520 | 518 | } |
| 519 | |
| 520 | static inline int zig_clzll(unsigned long long value, uint8_t zig_type_bit_width) { |
| 521 | if (value == 0) return zig_type_bit_width; |
| 522 | return __builtin_clzll(value) - zig_bitsizeof(unsigned long long) + zig_type_bit_width; |
| 523 | } |
| 524 | |
| 525 | #define zig_clz_u8 zig_clz |
| 526 | #define zig_clz_i8 zig_clz |
| 527 | #define zig_clz_u16 zig_clz |
| 528 | #define zig_clz_i16 zig_clz |
| 529 | #define zig_clz_u32 zig_clzl |
| 530 | #define zig_clz_i32 zig_clzl |
| 531 | #define zig_clz_u64 zig_clzll |
| 532 | #define zig_clz_i64 zig_clzll |
| 533 | |
| 534 | static inline int zig_clz_u128(uint128_t value, uint8_t zig_type_bit_width) { |
| 535 | if (value == 0) return zig_type_bit_width; |
| 536 | const uint128_t mask = zig_bit_mask(uint128_t, zig_type_bit_width); |
| 537 | const uint64_t hi = (value & mask) >> 64; |
| 538 | const uint64_t lo = (value & mask); |
| 539 | const int leading_zeroes = ( |
| 540 | hi != 0 ? __builtin_clzll(hi) : 64 + (lo != 0 ? __builtin_clzll(lo) : 64)); |
| 541 | return leading_zeroes - zig_bitsizeof(uint128_t) + zig_type_bit_width; |
| 542 | } |
| 543 | |
| 544 | #define zig_clz_i128 zig_clz_u128 |
| 545 | |
| 546 | static inline int zig_ctz(unsigned int value, uint8_t zig_type_bit_width) { |
| 547 | if (value == 0) return zig_type_bit_width; |
| 548 | return __builtin_ctz(value & zig_bit_mask(unsigned int, zig_type_bit_width)); |
| 549 | } |
| 550 | |
| 551 | static inline int zig_ctzl(unsigned long value, uint8_t zig_type_bit_width) { |
| 552 | if (value == 0) return zig_type_bit_width; |
| 553 | return __builtin_ctzl(value & zig_bit_mask(unsigned long, zig_type_bit_width)); |
| 554 | } |
| 555 | |
| 556 | static inline int zig_ctzll(unsigned long value, uint8_t zig_type_bit_width) { |
| 557 | if (value == 0) return zig_type_bit_width; |
| 558 | return __builtin_ctzll(value & zig_bit_mask(unsigned long, zig_type_bit_width)); |
| 559 | } |
| 560 | |
| 561 | #define zig_ctz_u8 zig_ctz |
| 562 | #define zig_ctz_i8 zig_ctz |
| 563 | #define zig_ctz_u16 zig_ctz |
| 564 | #define zig_ctz_i16 zig_ctz |
| 565 | #define zig_ctz_u32 zig_ctzl |
| 566 | #define zig_ctz_i32 zig_ctzl |
| 567 | #define zig_ctz_u64 zig_ctzll |
| 568 | #define zig_ctz_i64 zig_ctzll |
| 569 | |
| 570 | static inline int zig_ctz_u128(uint128_t value, uint8_t zig_type_bit_width) { |
| 571 | const uint128_t mask = zig_bit_mask(uint128_t, zig_type_bit_width); |
| 572 | const uint64_t hi = (value & mask) >> 64; |
| 573 | const uint64_t lo = (value & mask); |
| 574 | return (lo != 0 ? __builtin_ctzll(lo) : 64 + (hi != 0 ? __builtin_ctzll(hi) : 64)); |
| 575 | } |
| 576 | |
| 577 | #define zig_ctz_i128 zig_ctz_u128 |
| 578 | |
| 579 | static inline int zig_popcount(unsigned int value, uint8_t zig_type_bit_width) { |
| 580 | return __builtin_popcount(value & zig_bit_mask(unsigned int, zig_type_bit_width)); |
| 581 | } |
| 582 | |
| 583 | static inline int zig_popcountl(unsigned long value, uint8_t zig_type_bit_width) { |
| 584 | return __builtin_popcountl(value & zig_bit_mask(unsigned long, zig_type_bit_width)); |
| 585 | } |
| 586 | |
| 587 | static inline int zig_popcountll(unsigned long value, uint8_t zig_type_bit_width) { |
| 588 | return __builtin_popcountll(value & zig_bit_mask(unsigned long, zig_type_bit_width)); |
| 589 | } |
| 590 | |
| 591 | #define zig_popcount_u8 zig_popcount |
| 592 | #define zig_popcount_i8 zig_popcount |
| 593 | #define zig_popcount_u16 zig_popcount |
| 594 | #define zig_popcount_i16 zig_popcount |
| 595 | #define zig_popcount_u32 zig_popcountl |
| 596 | #define zig_popcount_i32 zig_popcountl |
| 597 | #define zig_popcount_u64 zig_popcountll |
| 598 | #define zig_popcount_i64 zig_popcountll |
| 599 | |
| 600 | static inline int zig_popcount_u128(uint128_t value, uint8_t zig_type_bit_width) { |
| 601 | const uint128_t mask = zig_bit_mask(uint128_t, zig_type_bit_width); |
| 602 | const uint64_t hi = (value & mask) >> 64; |
| 603 | const uint64_t lo = (value & mask); |
| 604 | return __builtin_popcountll(hi) + __builtin_popcountll(lo); |
| 605 | } |
| 606 | |
| 607 | #define zig_popcount_i128 zig_popcount_u128 |
| 608 | |
| 609 | #define zig_sign_extend(T) \ |
| 610 | static inline T zig_sign_extend_##T(T value, uint8_t zig_type_bit_width) { \ |
| 611 | const T m = (T)1 << (T)(zig_type_bit_width - 1); \ |
| 612 | return (value ^ m) - m; \ |
| 613 | } |
| 614 | |
| 615 | zig_sign_extend(uint8_t) |
| 616 | zig_sign_extend(uint16_t) |
| 617 | zig_sign_extend(uint32_t) |
| 618 | zig_sign_extend(uint64_t) |
| 619 | zig_sign_extend(uint128_t) |
| 620 | |
| 621 | #define zig_byte_swap_u(ZigTypeBits, CTypeBits) \ |
| 622 | static inline uint##CTypeBits##_t zig_byte_swap_u##ZigTypeBits(uint##CTypeBits##_t value, uint8_t zig_type_bit_width) { \ |
| 623 | return __builtin_bswap##CTypeBits(value) >> (CTypeBits - zig_type_bit_width); \ |
| 624 | } |
| 625 | |
| 626 | #define zig_byte_swap_s(ZigTypeBits, CTypeBits) \ |
| 627 | static inline int##CTypeBits##_t zig_byte_swap_i##ZigTypeBits(int##CTypeBits##_t value, uint8_t zig_type_bit_width) { \ |
| 628 | const uint##CTypeBits##_t swapped = zig_byte_swap_u##ZigTypeBits(value, zig_type_bit_width); \ |
| 629 | return zig_sign_extend_uint##CTypeBits##_t(swapped, zig_type_bit_width); \ |
| 630 | } |
| 631 | |
| 632 | #define zig_byte_swap(ZigTypeBits, CTypeBits) \ |
| 633 | zig_byte_swap_u(ZigTypeBits, CTypeBits) \ |
| 634 | zig_byte_swap_s(ZigTypeBits, CTypeBits) |
| 635 | |
| 636 | zig_byte_swap( 8, 16) |
| 637 | zig_byte_swap(16, 16) |
| 638 | zig_byte_swap(32, 32) |
| 639 | zig_byte_swap(64, 64) |
| 640 | |
| 641 | static inline uint128_t zig_byte_swap_u128(uint128_t value, uint8_t zig_type_bit_width) { |
| 642 | const uint128_t mask = zig_bit_mask(uint128_t, zig_type_bit_width); |
| 643 | const uint128_t hi = __builtin_bswap64((uint64_t)(value >> 64)); |
| 644 | const uint128_t lo = __builtin_bswap64((uint64_t)value); |
| 645 | return (((lo << 64 | hi) >> (128 - zig_type_bit_width))) & mask; |
| 646 | } |
| 647 | |
| 648 | zig_byte_swap_s(128, 128) |
| 649 | |
| 650 | static const uint8_t zig_bit_reverse_lut[256] = { |
| 651 | 0x00, 0x80, 0x40, 0xc0, 0x20, 0xa0, 0x60, 0xe0, 0x10, 0x90, 0x50, 0xd0, |
| 652 | 0x30, 0xb0, 0x70, 0xf0, 0x08, 0x88, 0x48, 0xc8, 0x28, 0xa8, 0x68, 0xe8, |
| 653 | 0x18, 0x98, 0x58, 0xd8, 0x38, 0xb8, 0x78, 0xf8, 0x04, 0x84, 0x44, 0xc4, |
| 654 | 0x24, 0xa4, 0x64, 0xe4, 0x14, 0x94, 0x54, 0xd4, 0x34, 0xb4, 0x74, 0xf4, |
| 655 | 0x0c, 0x8c, 0x4c, 0xcc, 0x2c, 0xac, 0x6c, 0xec, 0x1c, 0x9c, 0x5c, 0xdc, |
| 656 | 0x3c, 0xbc, 0x7c, 0xfc, 0x02, 0x82, 0x42, 0xc2, 0x22, 0xa2, 0x62, 0xe2, |
| 657 | 0x12, 0x92, 0x52, 0xd2, 0x32, 0xb2, 0x72, 0xf2, 0x0a, 0x8a, 0x4a, 0xca, |
| 658 | 0x2a, 0xaa, 0x6a, 0xea, 0x1a, 0x9a, 0x5a, 0xda, 0x3a, 0xba, 0x7a, 0xfa, |
| 659 | 0x06, 0x86, 0x46, 0xc6, 0x26, 0xa6, 0x66, 0xe6, 0x16, 0x96, 0x56, 0xd6, |
| 660 | 0x36, 0xb6, 0x76, 0xf6, 0x0e, 0x8e, 0x4e, 0xce, 0x2e, 0xae, 0x6e, 0xee, |
| 661 | 0x1e, 0x9e, 0x5e, 0xde, 0x3e, 0xbe, 0x7e, 0xfe, 0x01, 0x81, 0x41, 0xc1, |
| 662 | 0x21, 0xa1, 0x61, 0xe1, 0x11, 0x91, 0x51, 0xd1, 0x31, 0xb1, 0x71, 0xf1, |
| 663 | 0x09, 0x89, 0x49, 0xc9, 0x29, 0xa9, 0x69, 0xe9, 0x19, 0x99, 0x59, 0xd9, |
| 664 | 0x39, 0xb9, 0x79, 0xf9, 0x05, 0x85, 0x45, 0xc5, 0x25, 0xa5, 0x65, 0xe5, |
| 665 | 0x15, 0x95, 0x55, 0xd5, 0x35, 0xb5, 0x75, 0xf5, 0x0d, 0x8d, 0x4d, 0xcd, |
| 666 | 0x2d, 0xad, 0x6d, 0xed, 0x1d, 0x9d, 0x5d, 0xdd, 0x3d, 0xbd, 0x7d, 0xfd, |
| 667 | 0x03, 0x83, 0x43, 0xc3, 0x23, 0xa3, 0x63, 0xe3, 0x13, 0x93, 0x53, 0xd3, |
| 668 | 0x33, 0xb3, 0x73, 0xf3, 0x0b, 0x8b, 0x4b, 0xcb, 0x2b, 0xab, 0x6b, 0xeb, |
| 669 | 0x1b, 0x9b, 0x5b, 0xdb, 0x3b, 0xbb, 0x7b, 0xfb, 0x07, 0x87, 0x47, 0xc7, |
| 670 | 0x27, 0xa7, 0x67, 0xe7, 0x17, 0x97, 0x57, 0xd7, 0x37, 0xb7, 0x77, 0xf7, |
| 671 | 0x0f, 0x8f, 0x4f, 0xcf, 0x2f, 0xaf, 0x6f, 0xef, 0x1f, 0x9f, 0x5f, 0xdf, |
| 672 | 0x3f, 0xbf, 0x7f, 0xff |
| 673 | }; |
| 674 | |
| 675 | static inline uint8_t zig_bit_reverse_u8(uint8_t value, uint8_t zig_type_bit_width) { |
| 676 | const uint8_t reversed = zig_bit_reverse_lut[value] >> (8 - zig_type_bit_width); |
| 677 | return zig_sign_extend_uint8_t(reversed, zig_type_bit_width); |
| 678 | } |
| 679 | |
| 680 | #define zig_bit_reverse_i8 zig_bit_reverse_u8 |
| 681 | |
| 682 | static inline uint16_t zig_bit_reverse_u16(uint16_t value, uint8_t zig_type_bit_width) { |
| 683 | const uint16_t swapped = zig_byte_swap_u16(value, zig_type_bit_width); |
| 684 | const uint16_t reversed = ( |
| 685 | ((uint16_t)zig_bit_reverse_lut[(swapped >> 0x08) & 0xff] << 0x08) | |
| 686 | ((uint16_t)zig_bit_reverse_lut[(swapped >> 0x00) & 0xff] << 0x00)); |
| 687 | return zig_sign_extend_uint16_t( |
| 688 | reversed & zig_bit_mask(uint16_t, zig_type_bit_width), |
| 689 | zig_type_bit_width); |
| 690 | } |
| 691 | |
| 692 | #define zig_bit_reverse_i16 zig_bit_reverse_u16 |
| 693 | |
| 694 | static inline uint32_t zig_bit_reverse_u32(uint32_t value, uint8_t zig_type_bit_width) { |
| 695 | const uint32_t swapped = zig_byte_swap_u32(value, zig_type_bit_width); |
| 696 | const uint32_t reversed = ( |
| 697 | ((uint32_t)zig_bit_reverse_lut[(swapped >> 0x18) & 0xff] << 0x18) | |
| 698 | ((uint32_t)zig_bit_reverse_lut[(swapped >> 0x10) & 0xff] << 0x10) | |
| 699 | ((uint32_t)zig_bit_reverse_lut[(swapped >> 0x08) & 0xff] << 0x08) | |
| 700 | ((uint32_t)zig_bit_reverse_lut[(swapped >> 0x00) & 0xff] << 0x00)); |
| 701 | return zig_sign_extend_uint32_t( |
| 702 | reversed & zig_bit_mask(uint32_t, zig_type_bit_width), |
| 703 | zig_type_bit_width); |
| 704 | } |
| 705 | |
| 706 | #define zig_bit_reverse_i32 zig_bit_reverse_u32 |
| 707 | |
| 708 | static inline uint64_t zig_bit_reverse_u64(uint64_t value, uint8_t zig_type_bit_width) { |
| 709 | const uint64_t swapped = zig_byte_swap_u64(value, zig_type_bit_width); |
| 710 | const uint64_t reversed = ( |
| 711 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x38) & 0xff] << 0x38) | |
| 712 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x30) & 0xff] << 0x30) | |
| 713 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x28) & 0xff] << 0x28) | |
| 714 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x20) & 0xff] << 0x20) | |
| 715 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x18) & 0xff] << 0x18) | |
| 716 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x10) & 0xff] << 0x10) | |
| 717 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x08) & 0xff] << 0x08) | |
| 718 | ((uint64_t)zig_bit_reverse_lut[(swapped >> 0x00) & 0xff] << 0x00)); |
| 719 | return zig_sign_extend_uint64_t( |
| 720 | reversed & zig_bit_mask(uint64_t, zig_type_bit_width), |
| 721 | zig_type_bit_width); |
| 722 | } |
| 723 | |
| 724 | #define zig_bit_reverse_i64 zig_bit_reverse_u64 |
| 725 | |
| 726 | static inline uint128_t zig_bit_reverse_u128(uint128_t value, uint8_t zig_type_bit_width) { |
| 727 | const uint128_t swapped = zig_byte_swap_u128(value, zig_type_bit_width); |
| 728 | const uint128_t reversed = ( |
| 729 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x78) & 0xff] << 0x78) | |
| 730 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x70) & 0xff] << 0x70) | |
| 731 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x68) & 0xff] << 0x68) | |
| 732 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x60) & 0xff] << 0x60) | |
| 733 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x58) & 0xff] << 0x58) | |
| 734 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x50) & 0xff] << 0x50) | |
| 735 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x48) & 0xff] << 0x48) | |
| 736 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x40) & 0xff] << 0x40) | |
| 737 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x38) & 0xff] << 0x38) | |
| 738 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x30) & 0xff] << 0x30) | |
| 739 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x28) & 0xff] << 0x28) | |
| 740 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x20) & 0xff] << 0x20) | |
| 741 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x18) & 0xff] << 0x18) | |
| 742 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x10) & 0xff] << 0x10) | |
| 743 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x08) & 0xff] << 0x08) | |
| 744 | ((uint128_t)zig_bit_reverse_lut[(swapped >> 0x00) & 0xff] << 0x00)); |
| 745 | return zig_sign_extend_uint128_t( |
| 746 | reversed & zig_bit_mask(uint128_t, zig_type_bit_width), |
| 747 | zig_type_bit_width); |
| 748 | } |
| 749 | |
| 750 | #define zig_bit_reverse_i128 zig_bit_reverse_u128 |