| 2628 | // Force AVX2 generation for this function on GCC/Clang |
| 2629 | #if defined( COMPILER_GCC_CLANG ) |
| 2630 | __attribute__( ( target( "avx2" ) ) ) |
| 2631 | #endif |
| 2632 | void strip_ansi_avx2( std::string& str ) { |
| 2633 | if ( str.empty() ) |
| 2634 | return; |
| 2635 | |
| 2636 | char* data = &str[0]; |
| 2637 | const size_t len = str.size(); |
| 2638 | size_t read_idx = 0; |
| 2639 | size_t write_idx = 0; |
| 2640 | |
| 2641 | const __m256i esc_vec = _mm256_set1_epi8( '\x1B' ); |
| 2642 | |
| 2643 | // Process 32-byte chunks |
| 2644 | while ( read_idx + 32 <= len ) { |
| 2645 | __m256i chunk = _mm256_loadu_si256( reinterpret_cast<const __m256i*>( data + read_idx ) ); |
| 2646 | __m256i cmp = _mm256_cmpeq_epi8( chunk, esc_vec ); |
| 2647 | int mask = _mm256_movemask_epi8( cmp ); |
| 2648 | |
| 2649 | if ( likely( mask == 0 ) ) { |
| 2650 | if ( read_idx != write_idx ) { |
| 2651 | _mm256_storeu_si256( reinterpret_cast<__m256i*>( data + write_idx ), chunk ); |
| 2652 | } |
| 2653 | read_idx += 32; |
| 2654 | write_idx += 32; |
| 2655 | } else { |
| 2656 | // Found ESC. Let the scalar loop handle the complexity of finding EXACT location |
| 2657 | // and parsing the variable length ANSI code. |
| 2658 | break; |
| 2659 | } |
| 2660 | } |
| 2661 | |
| 2662 | // Finish remaining with Scalar Logic |
| 2663 | while ( read_idx < len ) { |
| 2664 | if ( unlikely( data[read_idx] == '\x1B' ) ) { |
| 2665 | if ( read_idx + 1 < len && data[read_idx + 1] == '[' ) { |
| 2666 | size_t scan = read_idx + 2; |
| 2667 | while ( scan < len ) { |
| 2668 | unsigned char c = static_cast<unsigned char>( data[scan] ); |
| 2669 | if ( c >= 0x40 && c <= 0x7E ) { |
| 2670 | scan++; |
| 2671 | break; |
| 2672 | } |
| 2673 | scan++; |
| 2674 | } |
| 2675 | read_idx = scan; |
| 2676 | continue; |
| 2677 | } |
| 2678 | } |
| 2679 | if ( read_idx != write_idx ) { |
| 2680 | data[write_idx] = data[read_idx]; |
| 2681 | } |
| 2682 | write_idx++; |
| 2683 | read_idx++; |
| 2684 | } |
| 2685 | str.resize( write_idx ); |
| 2686 | } |
| 2687 | #endif // ARCH_X86 |