compiler.h: remove concept of different unaligned access speeds

[wimlib] / include / wimlib / decompress_common.h

diff --git a/include/wimlib/decompress_common.h b/include/wimlib/decompress_common.h
index 6762043c8b0ab30081c51b83454fd76820d27ecb..54241b36592495106ff7d2efc5970e3d84b52c20 100644 (file)
--- a/include/wimlib/decompress_common.h
+++ b/include/wimlib/decompress_common.h
@@ -10,9 +10,9 @@
 #ifndef _WIMLIB_DECOMPRESS_COMMON_H
 #define _WIMLIB_DECOMPRESS_COMMON_H
 
-#include "wimlib/assert.h"
+#include <string.h>
+
 #include "wimlib/compiler.h"
-#include "wimlib/endianness.h"
 #include "wimlib/types.h"
 #include "wimlib/unaligned.h"
 
@@ -60,7 +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;
@@ -68,7 +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)get_unaligned_u16_le(is->next) << (16 - is->bitsleft);
+       is->bitbuf |= (u32)get_unaligned_le16(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)get_unaligned_u16_le(is->next);
+               is->bitbuf |= (u32)get_unaligned_le16(is->next);
                is->next += 2;
                is->bitsleft = 32;
        }
@@ -93,9 +92,7 @@ overflow:
 static inline u32
 bitstream_peek_bits(const struct input_bitstream *is, const unsigned num_bits)
 {
-       if (unlikely(num_bits == 0))
-               return 0;
-       return is->bitbuf >> (32 - num_bits);
+       return (is->bitbuf >> 1) >> (sizeof(is->bitbuf) * 8 - num_bits - 1);
 }
 
 /* Remove @num_bits from the bitstream.  There must be at least @num_bits
@@ -144,7 +141,7 @@ bitstream_read_u16(struct input_bitstream *is)
 
        if (unlikely(is->end - is->next < 2))
                return 0;
-       v = get_unaligned_u16_le(is->next);
+       v = get_unaligned_le16(is->next);
        is->next += 2;
        return v;
 }
@@ -157,23 +154,22 @@ bitstream_read_u32(struct input_bitstream *is)
 
        if (unlikely(is->end - is->next < 4))
                return 0;
-       v = get_unaligned_u32_le(is->next);
+       v = get_unaligned_le32(is->next);
        is->next += 4;
        return v;
 }
 
-/* Read an array of literal bytes embedded in the bitstream.  Return a pointer
- * to the resulting array, or NULL if the read overflows the input buffer.  */
-static inline const u8 *
-bitstream_read_bytes(struct input_bitstream *is, size_t count)
+/* Read into @dst_buffer an array of literal bytes embedded in the bitstream.
+ * Return either a pointer to the byte past the last written, or NULL if the
+ * read overflows the input buffer.  */
+static inline void *
+bitstream_read_bytes(struct input_bitstream *is, void *dst_buffer, size_t count)
 {
-       const u8 *p;
-
        if (unlikely(is->end - is->next < count))
                return NULL;
-       p = is->next;
+       memcpy(dst_buffer, is->next, count);
        is->next += count;
-       return p;
+       return (u8 *)dst_buffer + count;
 }
 
 /* Align the input bitstream on a coding-unit boundary.  */
@@ -215,9 +211,9 @@ 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.  */
-static inline u16
+ * XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in
+ * lzms_decompress.c.  */
+static inline unsigned
 read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
             unsigned table_bits, unsigned max_codeword_len)
 {
@@ -254,6 +250,25 @@ make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
                          unsigned num_bits, const u8 lens[],
                          unsigned max_codeword_len);
 
+static inline void
+copy_word_unaligned(const void *src, void *dst)
+{
+       store_word_unaligned(load_word_unaligned(src), dst);
+}
+
+static inline machine_word_t
+repeat_byte(u8 b)
+{
+       machine_word_t v;
+
+       STATIC_ASSERT(WORDSIZE == 4 || WORDSIZE == 8);
+
+       v = b;
+       v |= v << 8;
+       v |= v << 16;
+       v |= v << ((WORDSIZE == 8) ? 32 : 0);
+       return v;
+}
 
 /*
  * Copy an LZ77 match at (dst - offset) to dst.
@@ -284,7 +299,7 @@ lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length)
         * beyond the end of the output buffer, hence the check for (winend -
         * end >= WORDSIZE - 1).
         */
-       if (UNALIGNED_ACCESS_IS_VERY_FAST &&
+       if (UNALIGNED_ACCESS_IS_FAST &&
            likely(winend - end >= WORDSIZE - 1))
        {