lzx-decompress.c: Add SSE2 version of undo_call_insn_preprocessing()
authorEric Biggers <ebiggers3@gmail.com>
Tue, 27 May 2014 14:28:59 +0000 (09:28 -0500)
committerEric Biggers <ebiggers3@gmail.com>
Tue, 27 May 2014 14:32:39 +0000 (09:32 -0500)
src/lzx-decompress.c
src/resource.c

index 05093266ce4bb4ce8e2406ec9b1c86998d09341c..d3ba46af846de85ec23f9cd35743fddfda05e2a7 100644 (file)
@@ -7,7 +7,7 @@
  */
 
 /*
- * Copyright (C) 2012, 2013 Eric Biggers
+ * Copyright (C) 2012, 2013, 2014 Eric Biggers
  *
  * This file is part of wimlib, a library for working with WIM files.
  *
 
 #include <string.h>
 
+#ifdef __SSE2__
+#  include <emmintrin.h>
+#endif
+
 /* Huffman decoding tables and maps from symbols to code lengths. */
 struct lzx_tables {
 
@@ -711,20 +715,19 @@ lzx_decode_match(unsigned main_element, int block_type,
 }
 
 static void
-undo_call_insn_translation(u32 *call_insn_target, s32 input_pos,
-                          s32 file_size)
+undo_call_insn_translation(u32 *call_insn_target, s32 input_pos)
 {
        s32 abs_offset;
        s32 rel_offset;
 
        abs_offset = le32_to_cpu(*call_insn_target);
-       if (abs_offset >= -input_pos && abs_offset < file_size) {
+       if (abs_offset >= -input_pos && abs_offset < LZX_WIM_MAGIC_FILESIZE) {
                if (abs_offset >= 0) {
                        /* "good translation" */
                        rel_offset = abs_offset - input_pos;
                } else {
                        /* "compensating translation" */
-                       rel_offset = abs_offset + file_size;
+                       rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
                }
                *call_insn_target = cpu_to_le32(rel_offset);
        }
@@ -753,15 +756,99 @@ undo_call_insn_translation(u32 *call_insn_target, s32 input_pos,
  * as it is used in calculating the translated jump targets.  But in WIM files,
  * this file size is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).*/
 static void
-undo_call_insn_preprocessing(u8 *uncompressed_data, s32 uncompressed_size)
+undo_call_insn_preprocessing(u8 *uncompressed_data, size_t uncompressed_size)
 {
-       for (s32 i = 0; i < uncompressed_size - 10; i++) {
-               if (uncompressed_data[i] == 0xe8) {
-                       undo_call_insn_translation((u32*)&uncompressed_data[i + 1],
-                                                  i,
-                                                  LZX_WIM_MAGIC_FILESIZE);
-                       i += 4;
-               }
+#ifdef __SSE2__
+
+       /* SSE2 vectorized implementation for x86_64.  This speeds up LZX
+        * decompression by about 5-8% overall.  (Usually --- the performance
+        * actually regresses slightly in the degenerate case that the data
+        * consists entirely of 0xe8 bytes.)  */
+       __m128i *p128 = (__m128i *)uncompressed_data;
+       u32 valid_mask = 0xFFFFFFFF;
+
+       if (uncompressed_size >= 32 &&
+           ((uintptr_t)uncompressed_data % 16 == 0))
+       {
+               __m128i * const end128 = p128 + uncompressed_size / 16 - 1;
+
+               /* Create a vector of all 0xe8 bytes  */
+               const __m128i e8_bytes = _mm_set1_epi8(0xe8);
+
+               /* Iterate through the 16-byte vectors in the input.  */
+               do {
+                       /* Compare the current 16-byte vector with the vector of
+                        * all 0xe8 bytes.  This produces 0xff where the byte is
+                        * 0xe8 and 0x00 where it is not.  */
+                       __m128i cmpresult = _mm_cmpeq_epi8(*p128, e8_bytes);
+
+                       /* Map the comparison results into a single 16-bit
+                        * number.  It will contain a 1 bit when the
+                        * corresponding byte in the current 16-byte vector is
+                        * an e8 byte.  Note: the low-order bit corresponds to
+                        * the first (lowest address) byte.  */
+                       u32 e8_mask = _mm_movemask_epi8(cmpresult);
+
+                       if (!e8_mask) {
+                               /* If e8_mask is 0, then none of these 16 bytes
+                                * have value 0xe8.  No e8 translation is
+                                * needed, and there is no restriction that
+                                * carries over to the next 16 bytes.  */
+                               valid_mask = 0xFFFFFFFF;
+                       } else {
+                               /* At least one byte has value 0xe8.
+                                *
+                                * The AND with valid_mask accounts for the fact
+                                * that we can't start an e8 translation that
+                                * overlaps the previous one.  */
+                               while ((e8_mask &= valid_mask)) {
+
+                                       /* Count the number of trailing zeroes
+                                        * in e8_mask.  This will produce the
+                                        * index of the byte, within the 16, at
+                                        * which the next e8 translation should
+                                        * be done.  */
+                                       u32 bit = __builtin_ctz(e8_mask);
+
+                                       /* Do the e8 translation.  */
+                                       u8 *p8 = (u8 *)p128 + bit;
+                                       undo_call_insn_translation((s32 *)(p8 + 1),
+                                                                  p8 - uncompressed_data);
+
+                                       /* Don't start an e8 translation in the
+                                        * next 4 bytes.  */
+                                       valid_mask &= ~((u32)0x1F << bit);
+                               }
+                               /* Moving on to the next vector.  Shift and set
+                                * valid_mask accordingly.  */
+                               valid_mask >>= 16;
+                               valid_mask |= 0xFFFF0000;
+                       }
+               } while (++p128 < end128);
+       }
+
+       u8 *p8 = (u8 *)p128;
+       while (!(valid_mask & 1)) {
+               p8++;
+               valid_mask >>= 1;
+       }
+#else /* __SSE2__  */
+       u8 *p8 = uncompressed_data;
+#endif /* !__SSE2__  */
+
+       if (uncompressed_size > 10) {
+               /* Finish any bytes that weren't processed by the vectorized
+                * implementation.  */
+               u8 *p8_end = uncompressed_data + uncompressed_size - 10;
+               do {
+                       if (*p8 == 0xe8) {
+                               undo_call_insn_translation((s32 *)(p8 + 1),
+                                                          p8 - uncompressed_data);
+                               p8 += 5;
+                       } else {
+                               p8++;
+                       }
+               } while (p8 < p8_end);
        }
 }
 
index f7684b10baee9bab6fffaa109f0977e79a19d1ed..300aec8f1cbec29e0b919bba8243984e7e518271 100644 (file)
@@ -138,6 +138,7 @@ read_compressed_wim_resource(const struct wim_resource_spec * const rspec,
        int errno_save;
 
        u64 *chunk_offsets = NULL;
+       u8 *_ubuf = NULL;
        u8 *ubuf = NULL;
        void *cbuf = NULL;
        bool chunk_offsets_malloced = false;
@@ -368,13 +369,14 @@ read_compressed_wim_resource(const struct wim_resource_spec * const rspec,
 
        /* Allocate buffer for holding the uncompressed data of each chunk.  */
        if (chunk_size <= STACK_MAX) {
-               ubuf = alloca(chunk_size);
+               _ubuf = alloca(chunk_size + 15);
        } else {
-               ubuf = MALLOC(chunk_size);
-               if (ubuf == NULL)
+               _ubuf = MALLOC(chunk_size + 15);
+               if (_ubuf == NULL)
                        goto oom;
                ubuf_malloced = true;
        }
+       ubuf = (u8 *)(((uintptr_t)_ubuf + 15) & ~15);
 
        /* Allocate a temporary buffer for reading compressed chunks, each of
         * which can be at most @chunk_size - 1 bytes.  This excludes compressed
@@ -548,7 +550,7 @@ out_free_memory:
        if (chunk_offsets_malloced)
                FREE(chunk_offsets);
        if (ubuf_malloced)
-               FREE(ubuf);
+               FREE(_ubuf);
        if (cbuf_malloced)
                FREE(cbuf);
        errno = errno_save;