Split compare258 into compare256 for longest_match and compare258 for deflate_quick.

1 files changed, 68 insertions, 36 deletions

diff --git a/compare258.c b/compare258.c
index d32e1de..db212d1 100644
--- a/compare258.c
+++ b/compare258.c
@@ -9,9 +9,9 @@
 #include "fallback_builtins.h"
 
 /* ALIGNED, byte comparison */
-static inline int32_t compare258_c_static(const unsigned char *src0, const unsigned char *src1) {
+static inline int32_t compare256_c_static(const unsigned char *src0, const unsigned char *src1) {
     const unsigned char *src0start = src0;
-    const unsigned char *src0end = src0 + 258;
+    const unsigned char *src0end = src0 + 256;
 
     do {
         if (*src0 != *src1)
@@ -26,31 +26,39 @@ static inline int32_t compare258_c_static(const unsigned char *src0, const unsig
         if (*src0 != *src1)
             break;
         src0 += 1, src1 += 1;
-        if (*src0 != *src1)
-            break;
-        src0 += 1, src1 += 1;
-        if (*src0 != *src1)
-            break;
-        src0 += 1, src1 += 1;
     } while (src0 < src0end);
- 
+
     return (int32_t)(src0 - src0start);
 }
 
+static inline int32_t compare258_c_static(const unsigned char *src0, const unsigned char *src1) {
+    if (*src0 != *src1)
+        return 0;
+    src0 += 1, src1 += 1;
+    if (*src0 != *src1)
+        return 1;
+    src0 += 1, src1 += 1;
+    if (*src0 != *src1)
+        return 2;
+
+    return compare256_c_static(src0, src1) + 2;
+}
+
 int32_t compare258_c(const unsigned char *src0, const unsigned char *src1) {
     return compare258_c_static(src0, src1);
 }
 
 #define LONGEST_MATCH   longest_match_c
+#define COMPARE256      compare256_c_static
 #define COMPARE258      compare258_c_static
 
 #include "match_p.h"
 
 #ifdef UNALIGNED_OK
 /* UNALIGNED_OK, 16-bit integer comparison */
-static inline int32_t compare258_unaligned_16_static(const unsigned char *src0, const unsigned char *src1) {
+static inline int32_t compare256_unaligned_16_static(const unsigned char *src0, const unsigned char *src1) {
     const unsigned char *src0start = src0;
-    const unsigned char *src0end = src0 + 258;
+    const unsigned char *src0end = src0 + 256;
 
     do {
         if (*(uint16_t *)src0 != *(uint16_t *)src1)
@@ -62,12 +70,19 @@ static inline int32_t compare258_unaligned_16_static(const unsigned char *src0,
         if (*(uint16_t *)src0 != *(uint16_t *)src1)
             break;
         src0 += 2, src1 += 2;
+        if (*(uint16_t *)src0 != *(uint16_t *)src1)
+            break;
+        src0 += 2, src1 += 2;
     } while (src0 < src0end);
 
-    if (*src0 == *src1)
-        src0 += 1;
+    return (int32_t)(src0 - src0start) + (src0 < src0end && *src0 == *src1);
+}
 
-    return (int32_t)(src0 - src0start);
+static inline int32_t compare258_unaligned_16_static(const unsigned char *src0, const unsigned char *src1) {
+    if (*(uint16_t *)src0 != *(uint16_t *)src1)
+        return (*src0 == *src1);
+
+    return compare256_unaligned_16_static(src0+2, src1+2) + 2;
 }
 
 int32_t compare258_unaligned_16(const unsigned char *src0, const unsigned char *src1) {
@@ -75,15 +90,15 @@ int32_t compare258_unaligned_16(const unsigned char *src0, const unsigned char *
 }
 
 #define LONGEST_MATCH   longest_match_unaligned_16
+#define COMPARE256      compare256_unaligned_16_static
 #define COMPARE258      compare258_unaligned_16_static
 
 #include "match_p.h"
 
 #ifdef HAVE_BUILTIN_CTZ
 /* UNALIGNED_OK, 32-bit integer comparison */
-static inline int32_t compare258_unaligned_32_static(const unsigned char *src0, const unsigned char *src1) {
-    const unsigned char *src0start = src0;
-    const unsigned char *src0end = src0 + 256;
+static inline int32_t compare256_unaligned_32_static(const unsigned char *src0, const unsigned char *src1) {
+    int32_t len = 0;
 
     do {
         uint32_t sv = *(uint32_t *)src0;
@@ -92,18 +107,26 @@ static inline int32_t compare258_unaligned_32_static(const unsigned char *src0,
 
         if (xor) {
             uint32_t match_byte = __builtin_ctz(xor) / 8;
-            return (int32_t)(src0 - src0start + match_byte);
+            return (int32_t)(len + match_byte);
         }
 
-        src0 += 4, src1 += 4;
-    } while (src0 < src0end);
+        src0 += 4, src1 += 4, len += 4;
+    } while (len < 256);
 
-    if (*(uint16_t *)src0 == *(uint16_t *)src1)
-        src0 += 2, src1 += 2;
-    else if (*src0 == *src1)
-        src0 += 1, src1 += 1;
+    return len;
+}
 
-    return (int32_t)(src0 - src0start);
+static inline int32_t compare258_unaligned_32_static(const unsigned char *src0, const unsigned char *src1) {
+    if (*(uint16_t *)src0 != *(uint16_t *)src1)
+        return (*src0 == *src1);
+
+    src0 += 2, src1 += 2;
+    if (*src0 != *src1)
+        return 2;
+    if (src0[1] != src1[1])
+        return 3;
+
+    return compare256_unaligned_32_static(src0, src1) + 2;
 }
 
 int32_t compare258_unaligned_32(const unsigned char *src0, const unsigned char *src1) {
@@ -111,6 +134,7 @@ int32_t compare258_unaligned_32(const unsigned char *src0, const unsigned char *
 }
 
 #define LONGEST_MATCH   longest_match_unaligned_32
+#define COMPARE256      compare256_unaligned_32_static
 #define COMPARE258      compare258_unaligned_32_static
 
 #include "match_p.h"
@@ -119,9 +143,8 @@ int32_t compare258_unaligned_32(const unsigned char *src0, const unsigned char *
 
 #ifdef HAVE_BUILTIN_CTZLL
 /* UNALIGNED_OK, 64-bit integer comparison */
-static inline int32_t compare258_unaligned_64_static(const unsigned char *src0, const unsigned char *src1) {
-    const unsigned char *src0start = src0;
-    const unsigned char *src0end = src0 + 256;
+static inline int32_t compare256_unaligned_64_static(const unsigned char *src0, const unsigned char *src1) {
+    int32_t len = 0;
 
     do {
         uint64_t sv = *(uint64_t *)src0;
@@ -130,18 +153,26 @@ static inline int32_t compare258_unaligned_64_static(const unsigned char *src0,
 
         if (xor) {
             uint64_t match_byte = __builtin_ctzll(xor) / 8;
-            return (int32_t)(src0 - src0start + match_byte);
+            return (int32_t)(len + match_byte);
         }
 
-        src0 += 8, src1 += 8;
-    } while (src0 < src0end);
+        src0 += 8, src1 += 8, len += 8;
+    } while (len < 256);
 
-    if (*(uint16_t *)src0 == *(uint16_t *)src1)
-        src0 += 2, src1 += 2;
-    else if (*src0 == *src1)
-        src0 += 1, src1 += 1;
+    return len;
+}
 
-    return (int32_t)(src0 - src0start);
+static inline int32_t compare258_unaligned_64_static(const unsigned char *src0, const unsigned char *src1) {
+    if (*(uint16_t *)src0 != *(uint16_t *)src1)
+        return (*src0 == *src1);
+
+    src0 += 2, src1 += 2;
+    if (*src0 != *src1)
+        return 2;
+    if (src0[1] != src1[1])
+        return 3;
+
+    return compare256_unaligned_64_static(src0, src1) + 2;
 }
 
 int32_t compare258_unaligned_64(const unsigned char *src0, const unsigned char *src1) {
@@ -149,6 +180,7 @@ int32_t compare258_unaligned_64(const unsigned char *src0, const unsigned char *
 }
 
 #define LONGEST_MATCH   longest_match_unaligned_64
+#define COMPARE256      compare256_unaligned_64_static
 #define COMPARE258      compare258_unaligned_64_static
 
 #include "match_p.h"