A few tweaks.

diff --git a/md5p8.diff b/md5p8.diff
index d94e5c3764f389550a77e2c7c5eda37a9156f6da..c9110e8ba9dd9c41e8cccaa2f21159505fecdbee 100644 (file)
--- a/md5p8.diff
+++ b/md5p8.diff
@@ -18,7 +18,7 @@ xxHash is still preferred (and faster), but this provides a reasonably
 fast fallback for the case where xxHash libraries are not available at
 build time.
 
-based-on: 9dd99521381e022d81528696ca2457e17727200b
+based-on: d47a80c05ecb2d4e8356697e3f71c219a639fecc
 diff --git a/Makefile.in b/Makefile.in
 --- a/Makefile.in
 +++ b/Makefile.in
@@ -50,15 +50,7 @@ diff --git a/Makefile.in b/Makefile.in
 diff --git a/checksum.c b/checksum.c
 --- a/checksum.c
 +++ b/checksum.c
-@@ -45,6 +45,7 @@ extern const char *checksum_choice;
- #define CSUM_MD4 4
- #define CSUM_MD5 5
- #define CSUM_XXH64 6
-+#define CSUM_MD5P8 7
- 
- struct name_num_obj valid_checksums = {
-       "checksum", NULL, NULL, 0, 0, {
-@@ -52,6 +53,7 @@ struct name_num_obj valid_checksums = {
+@@ -44,6 +44,7 @@ struct name_num_obj valid_checksums = {
                { CSUM_XXH64, "xxh64", NULL },
                { CSUM_XXH64, "xxhash", NULL },
  #endif
@@ -66,7 +58,7 @@ diff --git a/checksum.c b/checksum.c
                { CSUM_MD5, "md5", NULL },
                { CSUM_MD4, "md4", NULL },
                { CSUM_NONE, "none", NULL },
-@@ -139,6 +141,7 @@ int csum_len_for_type(int cst, BOOL flist_csum)
+@@ -131,6 +132,7 @@ int csum_len_for_type(int cst, BOOL flist_csum)
          case CSUM_MD4_OLD:
          case CSUM_MD4_BUSTED:
                return MD4_DIGEST_LEN;
@@ -74,7 +66,7 @@ diff --git a/checksum.c b/checksum.c
          case CSUM_MD5:
                return MD5_DIGEST_LEN;
  #ifdef SUPPORT_XXHASH
-@@ -164,6 +167,7 @@ int canonical_checksum(int csum_type)
+@@ -156,6 +158,7 @@ int canonical_checksum(int csum_type)
          case CSUM_MD4_BUSTED:
                break;
          case CSUM_MD4:
@@ -82,7 +74,18 @@ diff --git a/checksum.c b/checksum.c
          case CSUM_MD5:
                return -1;
  #ifdef SUPPORT_XXHASH
-@@ -197,9 +201,19 @@ uint32 get_checksum1(char *buf1, int32 len)
+@@ -168,7 +171,9 @@ int canonical_checksum(int csum_type)
+       return 0;
+ }
+ 
+-#ifndef HAVE_SIMD /* See simd-checksum-*.cpp. */
++#ifdef HAVE_SIMD /* See simd-checksum-*.cpp. */
++#define get_checksum2 get_checksum2_nosimd
++#else
+ /*
+   a simple 32 bit checksum that can be updated from either end
+   (inspired by Mark Adler's Adler-32 checksum)
+@@ -189,9 +194,18 @@ uint32 get_checksum1(char *buf1, int32 len)
        }
        return (s1 & 0xffff) + (s2 << 16);
  }
@@ -94,17 +97,15 @@ diff --git a/checksum.c b/checksum.c
 +void checksum2_disable_prefetch()
 +{
 +}
-+
-+void get_checksum2(char *buf, int32 len, char *sum, UNUSED(OFF_T prefetch_offset))
-+#else
-+void get_checksum2_nosimd(char *buf, int32 len, char *sum, UNUSED(OFF_T prefetch_offset))
  #endif
--
+ 
 -void get_checksum2(char *buf, int32 len, char *sum)
++/* Renamed to get_checksum2_nosimd() with HAVE_SIMD */
++void get_checksum2(char *buf, int32 len, char *sum, UNUSED(OFF_T prefetch_offset))
  {
        switch (xfersum_type) {
  #ifdef SUPPORT_XXHASH
-@@ -207,6 +221,7 @@ void get_checksum2(char *buf, int32 len, char *sum)
+@@ -199,6 +213,7 @@ void get_checksum2(char *buf, int32 len, char *sum)
                SIVAL64(sum, 0, XXH64(buf, len, checksum_seed));
                break;
  #endif
@@ -112,7 +113,7 @@ diff --git a/checksum.c b/checksum.c
          case CSUM_MD5: {
                MD5_CTX m5;
                uchar seedbuf[4];
-@@ -322,6 +337,21 @@ void file_checksum(const char *fname, const STRUCT_STAT *st_p, char *sum)
+@@ -314,6 +329,21 @@ void file_checksum(const char *fname, const STRUCT_STAT *st_p, char *sum)
                break;
          }
  #endif
@@ -134,7 +135,7 @@ diff --git a/checksum.c b/checksum.c
          case CSUM_MD5: {
                MD5_CTX m5;
  
-@@ -397,6 +427,7 @@ static union {
+@@ -389,6 +419,7 @@ static union {
  #ifdef SUPPORT_XXHASH
  static XXH64_state_t* xxh64_state;
  #endif
@@ -142,7 +143,7 @@ diff --git a/checksum.c b/checksum.c
  static int cursum_type;
  
  void sum_init(int csum_type, int seed)
-@@ -415,6 +446,9 @@ void sum_init(int csum_type, int seed)
+@@ -407,6 +438,9 @@ void sum_init(int csum_type, int seed)
                XXH64_reset(xxh64_state, 0);
                break;
  #endif
@@ -152,7 +153,7 @@ diff --git a/checksum.c b/checksum.c
          case CSUM_MD5:
                MD5_Init(&ctx.m5);
                break;
-@@ -457,6 +491,9 @@ void sum_update(const char *p, int32 len)
+@@ -449,6 +483,9 @@ void sum_update(const char *p, int32 len)
                XXH64_update(xxh64_state, p, len);
                break;
  #endif
@@ -162,7 +163,7 @@ diff --git a/checksum.c b/checksum.c
          case CSUM_MD5:
                MD5_Update(&ctx.m5, (uchar *)p, len);
                break;
-@@ -511,6 +548,9 @@ int sum_end(char *sum)
+@@ -503,6 +540,9 @@ int sum_end(char *sum)
                SIVAL64(sum, 0, XXH64_digest(xxh64_state));
                break;
  #endif
@@ -349,7 +350,15 @@ new file mode 100644
 diff --git a/lib/mdigest.h b/lib/mdigest.h
 --- a/lib/mdigest.h
 +++ b/lib/mdigest.h
-@@ -32,3 +32,14 @@ void md5_begin(md_context *ctx);
+@@ -18,6 +18,7 @@
+ #define CSUM_MD4 4
+ #define CSUM_MD5 5
+ #define CSUM_XXH64 6
++#define CSUM_MD5P8 7
+ 
+ typedef struct {
+       uint32 A, B, C, D;
+@@ -40,3 +41,14 @@ void md5_begin(md_context *ctx);
  void md5_update(md_context *ctx, const uchar *input, uint32 length);
  void md5_result(md_context *ctx, uchar digest[MD5_DIGEST_LEN]);
  #endif
@@ -429,11 +438,11 @@ diff --git a/simd-checksum-x86_64.cpp b/simd-checksum-x86_64.cpp
   */
  
  #ifdef __x86_64__
-@@ -423,6 +439,176 @@ uint32 get_checksum1(char *buf1, int32 len)
+@@ -423,6 +439,171 @@ uint32 get_checksum1(char *buf1, int32 len)
      return (s1 & 0xffff) + (s2 << 16);
  }
  
-+#define PREFETCH_ENABLE // debugging
++#define PREFETCH_ENABLE 1 // debugging
 +
 +#if 0 // debugging
 +#define PREFETCH_PRINTF(f_, ...) printf((f_), ##__VA_ARGS__)
@@ -443,8 +452,6 @@ diff --git a/simd-checksum-x86_64.cpp b/simd-checksum-x86_64.cpp
 +
 +#define PREFETCH_MIN_LEN 1024 // the overhead is unlikely to be worth the gain for small blocks
 +#define PREFETCH_MAX_BLOCKS 8
-+#define CSUM_MD5 5
-+#define CSUM_MD5P8 7
 +
 +typedef struct {
 +    int in_use;
@@ -484,12 +491,13 @@ diff --git a/simd-checksum-x86_64.cpp b/simd-checksum-x86_64.cpp
 +    }
 +}
 +
-+void checksum2_enable_prefetch(struct map_struct *map, OFF_T len, int32 blocklen)
++void checksum2_enable_prefetch(UNUSED(struct map_struct *map), UNUSED(OFF_T len), UNUSED(int32 blocklen))
 +{
 +#ifdef PREFETCH_ENABLE
 +    checksum2_disable_prefetch();
 +    int slots = md5_parallel_slots();
-+    if (((xfersum_type == CSUM_MD5) || (xfersum_type == CSUM_MD5P8)) && (slots > 1) && (len >= blocklen * PREFETCH_MAX_BLOCKS) && (blocklen >= PREFETCH_MIN_LEN)) {
++    if ((xfersum_type == CSUM_MD5 || xfersum_type == CSUM_MD5P8)
++     && slots > 1 && len >= blocklen * PREFETCH_MAX_BLOCKS && blocklen >= PREFETCH_MIN_LEN) {
 +        prefetch = (prefetch_t*)malloc(sizeof(prefetch_t));
 +        memset(prefetch, 0, sizeof(prefetch_t));
 +        prefetch->map = map;
@@ -499,10 +507,6 @@ diff --git a/simd-checksum-x86_64.cpp b/simd-checksum-x86_64.cpp
 +        prefetch->blocks = MIN(PREFETCH_MAX_BLOCKS, slots);
 +        PREFETCH_PRINTF("checksum2_enable_prefetch len:%ld blocklen:%d blocks:%d\n", prefetch->len, prefetch->blocklen, prefetch->blocks);
 +    }
-+#else
-+    (void)map;
-+    (void)len;
-+    (void)blocklen;
 +#endif
 +}
 +
@@ -1419,7 +1423,7 @@ new file mode 100644
 +    int slots = md5_parallel_slots();
 +    if ((streams < 1) || (streams > slots)) return 0;
 +    if (pre4 && post4) return 0;
-+    
++
 +    if (slots == 1) {
 +        MD5_CTX ctx;
 +        MD5_Init(&ctx);