X-Git-Url: https://wimlib.net/git/?a=blobdiff_plain;f=include%2Fwimlib%2Flz_extend.h;fp=include%2Fwimlib%2Flz_extend.h;h=ad16847d9a769ab17175a45dcc4f352366f3e1c4;hb=e9ab3bf39bf6689a62a94fc45ccc2b3f78e183c8;hp=858cb9a3b1f791edccbb6f2663f71fc8a3eb3983;hpb=a582ba68e1919a72066ae7e4d6c0c36d6db66675;p=wimlib

diff --git a/include/wimlib/lz_extend.h b/include/wimlib/lz_extend.h
index 858cb9a3..ad16847d 100644
--- a/include/wimlib/lz_extend.h
+++ b/include/wimlib/lz_extend.h
@@ -24,52 +24,58 @@
 #include "wimlib/bitops.h"
 #include "wimlib/unaligned.h"
 
+#include <smmintrin.h>
+
 /* Return the number of bytes at @matchptr that match the bytes at @strptr, up
  * to a maximum of @max_len.  Initially, @start_len bytes are matched.  */
 static inline u32
 lz_extend(const u8 * const strptr, const u8 * const matchptr,
 	  const u32 start_len, const u32 max_len)
 {
+#if 0
 	u32 len = start_len;
 	machine_word_t v_word;
 
-	if (UNALIGNED_ACCESS_IS_FAST) {
+	for (;;) {
+		v_word = load_word_unaligned(&matchptr[len]) ^
+			 load_word_unaligned(&strptr[len]);
+		if (v_word != 0 || len >= max_len)
+			break;
+		len += WORDSIZE;
+	}
+
+	return min(max_len, len + (ffsw(v_word) >> 3));
+#else
 
-		if (likely(max_len - len >= 4 * WORDSIZE)) {
+	const u8 *p1 = strptr + start_len;
+	const u8 *p2 = matchptr + start_len;
 
-		#define COMPARE_WORD_STEP					\
-			v_word = load_word_unaligned(&matchptr[len]) ^		\
-				 load_word_unaligned(&strptr[len]);		\
-			if (v_word != 0)					\
-				goto word_differs;				\
-			len += WORDSIZE;					\
+	u8 saved = strptr[max_len];
+	((u8 *)strptr)[max_len] = matchptr[max_len] + 1;
 
-			COMPARE_WORD_STEP
-			COMPARE_WORD_STEP
-			COMPARE_WORD_STEP
-			COMPARE_WORD_STEP
-		#undef COMPARE_WORD_STEP
-		}
+	__asm__(
+		"  movdqu (%[p1]), %%xmm0                    \n"
+		"  pcmpestri $0x18, (%[p2]), %%xmm0          \n"
+		"  jc 2f                                     \n"
+		"1:                                          \n"
+		"  add $0x10, %[p1]                          \n"
+		"  add $0x10, %[p2]                          \n"
+		"  movdqu (%[p1]), %%xmm0                    \n"
+		"  pcmpestri $0x18, (%[p2]), %%xmm0          \n"
+		"  jnc 1b                                    \n"
+		"2:                                          \n"
+		"  add %%rcx, %[p1]                          \n"
+		"  add %%rcx, %[p2]                          \n"
+		: [p1] "+r" (p1), [p2] "+r" (p2)
+		: "a" (16), "d" (16)
+		: "rcx", "cc", "xmm0", "memory"
+	       );
 
-		while (len + WORDSIZE <= max_len) {
-			v_word = load_word_unaligned(&matchptr[len]) ^
-				 load_word_unaligned(&strptr[len]);
-			if (v_word != 0)
-				goto word_differs;
-			len += WORDSIZE;
-		}
-	}
 
-	while (len < max_len && matchptr[len] == strptr[len])
-		len++;
-	return len;
+	((u8 *)strptr)[max_len] = saved;
 
-word_differs:
-	if (CPU_IS_LITTLE_ENDIAN)
-		len += (ffsw(v_word) >> 3);
-	else
-		len += (8 * WORDSIZE - 1 - flsw(v_word)) >> 3;
-	return len;
+	return p1 - strptr;
+#endif
 }
 
 #endif /* _WIMLIB_LZ_EXTEND_H */