X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=include%2Fwimlib%2Fdecompress_common.h;h=b6927bed556dc442f3364d52a1fb2b5e3e718fa6;hp=3f3742a02ea750aad757febe523109b9441954d6;hb=40a690416a3951361ec77d33a723dd4497fb7585;hpb=0158ecdaa9e7475a75799f369468f9af5a8ce4ea diff --git a/include/wimlib/decompress_common.h b/include/wimlib/decompress_common.h index 3f3742a0..b6927bed 100644 --- a/include/wimlib/decompress_common.h +++ b/include/wimlib/decompress_common.h @@ -14,6 +14,7 @@ #include "wimlib/compiler.h" #include "wimlib/endianness.h" #include "wimlib/types.h" +#include "wimlib/unaligned.h" /* Structure that encapsulates a block of in-memory data being interpreted as a * stream of bits, optionally with interwoven literal bytes. Bits are assumed @@ -59,7 +60,6 @@ static inline void bitstream_ensure_bits(struct input_bitstream *is, const unsigned num_bits) { /* This currently works for at most 17 bits. */ - wimlib_assert2(num_bits <= 17); if (is->bitsleft >= num_bits) return; @@ -67,8 +67,7 @@ bitstream_ensure_bits(struct input_bitstream *is, const unsigned num_bits) if (unlikely(is->end - is->next < 2)) goto overflow; - is->bitbuf |= (u32)le16_to_cpu(*(const le16 *)is->next) - << (16 - is->bitsleft); + is->bitbuf |= (u32)get_unaligned_u16_le(is->next) << (16 - is->bitsleft); is->next += 2; is->bitsleft += 16; @@ -76,7 +75,7 @@ bitstream_ensure_bits(struct input_bitstream *is, const unsigned num_bits) if (unlikely(is->end - is->next < 2)) goto overflow; - is->bitbuf |= (u32)le16_to_cpu(*(const le16 *)is->next); + is->bitbuf |= (u32)get_unaligned_u16_le(is->next); is->next += 2; is->bitsleft = 32; } @@ -131,11 +130,24 @@ bitstream_read_bits(struct input_bitstream *is, unsigned num_bits) static inline u8 bitstream_read_byte(struct input_bitstream *is) { - if (unlikely(is->end - is->next < 1)) + if (unlikely(is->end == is->next)) return 0; return *is->next++; } +/* Read and return the next 16-bit integer embedded in the bitstream. */ +static inline u16 +bitstream_read_u16(struct input_bitstream *is) +{ + u16 v; + + if (unlikely(is->end - is->next < 2)) + return 0; + v = get_unaligned_u16_le(is->next); + is->next += 2; + return v; +} + /* Read and return the next 32-bit integer embedded in the bitstream. */ static inline u32 bitstream_read_u32(struct input_bitstream *is) @@ -144,7 +156,7 @@ bitstream_read_u32(struct input_bitstream *is) if (unlikely(is->end - is->next < 4)) return 0; - v = le32_to_cpu(*(const le32 *)is->next); + v = get_unaligned_u32_le(is->next); is->next += 4; return v; } @@ -202,14 +214,14 @@ bitstream_align(struct input_bitstream *is) * input data is exhausted, the Huffman symbol is decoded as if the missing bits * are all zeroes. * - * XXX: This is mostly duplicated in lzms_huffman_decode_symbol() in - * lzms-decompress.c. */ + * XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in + * lzms_decompress.c. */ static inline u16 read_huffsym(struct input_bitstream *istream, const u16 decode_table[], unsigned table_bits, unsigned max_codeword_len) { - u16 entry; - u16 key_bits; + unsigned entry; + unsigned key_bits; bitstream_ensure_bits(istream, max_codeword_len); @@ -243,7 +255,7 @@ make_huffman_decode_table(u16 decode_table[], unsigned num_syms, /* - * Copy a LZ77 match at (dst - offset) to dst. + * Copy an LZ77 match at (dst - offset) to dst. * * The length and offset must be already validated --- that is, (dst - offset) * can't underrun the output buffer, and (dst + length) can't overrun the output @@ -253,50 +265,88 @@ make_huffman_decode_table(u16 decode_table[], unsigned num_syms, * This function won't write any data beyond this position. */ static inline void -lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend) +lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length) { const u8 *src = dst - offset; -#if defined(__x86_64__) || defined(__i386__) - /* Copy one 'unsigned long' at a time. On i386 and x86_64 this is + const u8 * const end = dst + length; + + /* + * Try to copy one machine word at a time. On i386 and x86_64 this is * faster than copying one byte at a time, unless the data is * near-random and all the matches have very short lengths. Note that * since this requires unaligned memory accesses, it won't necessarily * be faster on every architecture. * * Also note that we might copy more than the length of the match. For - * example, if an 'unsigned long' is 8 bytes and the match is of length - * 5, then we'll simply copy 8 bytes. This is okay as long as we don't - * write beyond the end of the output buffer, hence the check for - * (winend - (dst + length) >= sizeof(unsigned long) - 1). */ - if (offset >= sizeof(unsigned long) && - winend - (dst + length) >= sizeof(unsigned long) - 1) + * example, if a word is 8 bytes and the match is of length 5, then + * we'll simply copy 8 bytes. This is okay as long as we don't write + * beyond the end of the output buffer, hence the check for (winend - + * end >= WORDSIZE - 1). + */ + if (UNALIGNED_ACCESS_IS_VERY_FAST && + likely(winend - end >= WORDSIZE - 1)) { - /* Access memory through a packed struct. This tricks the - * compiler into allowing unaligned memory accesses. */ - struct ulong_wrapper { - unsigned long v; - } _packed_attribute; - const u8 * const end = dst + length; - unsigned long v; - - v = ((struct ulong_wrapper *)src)->v; - ((struct ulong_wrapper *)dst)->v = v; - dst += sizeof(unsigned long); - src += sizeof(unsigned long); - - if (dst < end) { + if (offset >= WORDSIZE) { + /* The source and destination words don't overlap. */ + + /* To improve branch prediction, one iteration of this + * loop is unrolled. Most matches are short and will + * fail the first check. But if that check passes, then + * it becomes increasing likely that the match is long + * and we'll need to continue copying. */ + + copy_word_unaligned(src, dst); + src += WORDSIZE; + dst += WORDSIZE; + + if (dst < end) { + do { + copy_word_unaligned(src, dst); + src += WORDSIZE; + dst += WORDSIZE; + } while (dst < end); + } + return; + } else if (offset == 1) { + + /* Offset 1 matches are equivalent to run-length + * encoding of the previous byte. This case is common + * if the data contains many repeated bytes. */ + + machine_word_t v = repeat_byte(*(dst - 1)); do { - v = ((struct ulong_wrapper *)src)->v; - ((struct ulong_wrapper *)dst)->v = v; - dst += sizeof(unsigned long); - src += sizeof(unsigned long); + store_word_unaligned(v, dst); + src += WORDSIZE; + dst += WORDSIZE; } while (dst < end); + return; } + /* + * We don't bother with special cases for other 'offset < + * WORDSIZE', which are usually rarer than 'offset == 1'. Extra + * checks will just slow things down. Actually, it's possible + * to handle all the 'offset < WORDSIZE' cases using the same + * code, but it still becomes more complicated doesn't seem any + * faster overall; it definitely slows down the more common + * 'offset == 1' case. + */ + } - return; + /* Fall back to a bytewise copy. */ + + if (min_length >= 2) { + *dst++ = *src++; + length--; + } + if (min_length >= 3) { + *dst++ = *src++; + length--; + } + if (min_length >= 4) { + *dst++ = *src++; + length--; } -#endif do { *dst++ = *src++; } while (--length);