Release calc version 2.11.0t8.3

CHANGES

@@ -44,6 +44,12 @@ Following is the change from calc version 2.11.0t8 to date:
 
     Misc source file cleanup for things such as } else { style consistency.
 
+    Fixed the basis for FNV-1 hashes.  Piror to this fix, the hash()
+    builtin produced FNV hash values that did not match the FNV-1
+    algorithm as specified in:
+
+	http://reality.sgi.com/chongo/tech/comp/fnv/index.html
+
 
 Following is the change from calc version 2.11.0t7 to 2.11.0t7.5:
 

assocfunc.c

@@ -57,7 +57,7 @@ associndex(ASSOC *ap, BOOL create, long dim, VALUE *indices)
 	 * so that we can first select the correct hash chain, and
 	 * also so we can quickly compare each element for a match.
 	 */
-	hash = (QCKHASH)0;
+	hash = FNV1_32_BASIS;
 	for (i = 0; i < dim; i++)
 		hash = hashvalue(&indices[i], hash);
 

func.c

@@ -1527,7 +1527,7 @@ f_hash(int count, VALUE **vals)
 	long lhash;
 	VALUE result;
 
-	hash = (QCKHASH)0;
+	hash = FNV1_32_BASIS;
 	while (count-- > 0)
 		hash = hashvalue(*vals++, hash);
 	lhash = (long) hash;

help/hash

@@ -12,10 +12,23 @@ TYPES
 DESCRIPTION
     Returns a hash value for one or more values of arbitrary types.
 
+    This function implements the Fowler/Noll/Vo hash-1 (FNV-1 hash).
+    The basis of the hash algorithm was taken from an idea sent
+    by Email to the IEEE POSIX P1003.2 mailing list from Phong Vo
+    (kpv@research.att.com) and Glenn Fowler (gsf@research.att.com).
+    Landon Curt Noll (http://reality.sgi.com/chongo) later improved on
+    their algorithm to come up with Fowler/Noll/Vo hash.
+
+    See:
+
+	http://reality.sgi.com/chongo/tech/comp/fnv/index.html
+
+    for more information in this hash.
+
 EXAMPLE
     > a = isqrt(2e1000); s = "xyz";
     > hash(a,s)
-	    870000771
+	    1916476840
 
 LIMITS
     The number of arguments is not to exceed 100.

lib/regress.cal

@@ -7115,6 +7115,8 @@ print '188: parsed test_natnumset()';
  */
 define test_somenew()
 {
+	local a, s;
+
 	print	'8200: Starting test_somenew';
 
 	vrfy(char(-1) == char(255),	'8201: char(-1) == char(255)');
@@ -7139,7 +7141,11 @@ define test_somenew()
 	vrfy(1/(1/0) == 0,		'8215: 1/(1/0) == 0');
 	vrfy(inverse(1/0) == 0,		'8216: inverse(1/0) == 0');
 
-	print '8217: Ending test_somenew';
+	a = isqrt(2e1000); s = "xyz";
+	print				'8217: a = isqrt(2e1000); s = "xyz";';
+	vrfy(hash(a,s) == 1916476840,	'8218: hash(a,s) == 1916476840');
+
+	print '8219: Ending test_somenew';
 }
 print '189: parsed test_somenew()';
 

seed.c

@@ -86,6 +86,50 @@ typedef struct s_hash64 hash64;
 #endif
 
 
+/*
+ * FNV-1 basis
+ *
+ * We start the hash at a non-zero value at the beginning so that
+ * hashing blocks of data with all 0 bits do not map onto the same
+ * 0 hash value.  The virgin value that we use below is the hash value
+ * that we would get from following 32 ASCII characters:
+ *
+ *              chongo <Landon Curt Noll> /\../\
+ *
+ * Note that the \'s above are not back-slashing escape characters.
+ * They are literal ASCII  backslash 0x5c characters.
+ *
+ * The effect of this virgin initial value is the same as starting
+ * with 0 and pre-pending those 32 characters onto the data being
+ * hashed.
+ *
+ * Yes, even with this non-zero virgin value there is a set of data
+ * that will result in a zero hash value.  Worse, appending any
+ * about of zero bytes will continue to produce a zero hash value.
+ * But that would happen with any initial value so long as the
+ * hash of the initial was the `inverse' of the virgin prefix string.
+ *
+ * But then again for any hash function, there exists sets of data
+ * which that the hash of every member is the same value.  That is
+ * life with many to few mapping functions.  All we do here is to
+ * prevent sets whose members consist of 0 or more bytes of 0's from
+ * being such an awkward set.
+ *
+ * And yes, someone can figure out what the magic 'inverse' of the
+ * 32 ASCII character are ... but this hash function is NOT intended
+ * to be a cryptographic hash function, just a fast and reasonably
+ * good hash function.
+ */
+#if defined(HAVE_B64)
+# define FNV1_64_BASIS ((hash64)(0xcbf29ce484222325ULL))
+#else
+# define FNV1_64_BASIS_0 ((USB32)0x2325)
+# define FNV1_64_BASIS_1 ((USB32)0x8422)
+# define FNV1_64_BASIS_2 ((USB32)0x9ce4)
+# define FNV1_64_BASIS_3 ((USB32)0xcbf2)
+#endif
+
+
 /*
  * hash_buf - perform a 64 bit Fowler/Noll/Vo hash on a buffer
  *
@@ -116,24 +160,16 @@ hash_buf(char *buf, unsigned len)
 	 * (gsf@research.att.com).
 	 *
 	 * See:
-	 *	http://reality.sgi.com/chongo/src/fnv/fnv_hash.tar.gz
+	 *	http://reality.sgi.com/chongo/tech/comp/fnv/index.html
-	 *	http://reality.sgi.com/chongo/src/fnv/h32.c
-	 *	http://reality.sgi.com/chongo/src/fnv/h64.c
 	 *
 	 * for information on 32bit and 64bit Fowler/Noll/Vo hashes.
 	 *
 	 * Landon Curt Noll (http://reality.sgi.com/chongo) later improved
 	 * on their algorithm to come up with Fowler/Noll/Vo hash.
-	 *
-	 * The 32 hash was able to process 234936 words from the web2 dictionary
-	 * without any 32 bit collisions using a constant of
-	 * 16777619 = 0x1000193.
-	 *
-	 * The 64 bit hash uses 1099511628211 = 0x100000001b3 instead.
 	 */
 #if defined(HAVE_B64)
 	/* hash each octet of the buffer */
-	for (hval = (hash64)0ULL; buf < buf_end; ++buf) {
+	for (hval = FNV1_64_BASIS; buf < buf_end; ++buf) {
 
 	    /* multiply by 1099511628211ULL mod 2^64 using 64 bit longs */
 	    hval *= (hash64)1099511628211ULL;
@@ -145,7 +181,11 @@ hash_buf(char *buf, unsigned len)
 #else /* HAVE_B64 */
 
 	/* hash each octet of the buffer */
-	for (val[0]=val[1]=val[2]=val[3]=0; buf < buf_end; ++buf) {
+	val[0] = FNV1_64_BASIS_0;
+	val[1] = FNV1_64_BASIS_1;
+	val[2] = FNV1_64_BASIS_2;
+	val[3] = FNV1_64_BASIS_3;
+	for (; buf < buf_end; ++buf) {
 
 	    /*
 	     * multiply by 1099511628211 mod 2^64 using 32 bit longs
@@ -167,7 +207,7 @@ hash_buf(char *buf, unsigned len)
 	    val[0] = tmp[0] & 0xffff;
 	    tmp[2] += (tmp[1] >> 16);
 	    val[1] = tmp[1] & 0xffff;
-	    val[3] += (tmp[2] >> 16);
+	    val[3] = tmp[3] + (tmp[2] >> 16);
 	    val[2] = tmp[2] & 0xffff;
 	    /*
 	     * Doing a val[3] &= 0xffff; is not really needed since it simply

version.c

@@ -12,7 +12,7 @@
 #define MAJOR_VER	2	/* major version */
 #define MINOR_VER	11	/* minor version */
 #define MAJOR_PATCH	0	/* patch level or 0 if no patch */
-#define MINOR_PATCH	"8.2"	/* test number or empty string if no patch */
+#define MINOR_PATCH	"8.3"	/* test number or empty string if no patch */
 
 /*
  * calc version constants

zmath.h

@@ -144,6 +144,43 @@ typedef SB32 LEN;			/* unit of length storage */
 #endif /* LONG_BITS == 64 */
 
 
+/*
+ * FNV-1 basis
+ *
+ * We start the hash at a non-zero value at the beginning so that
+ * hashing blocks of data with all 0 bits do not map onto the same
+ * 0 hash value.  The virgin value that we use below is the hash value
+ * that we would get from following 32 ASCII characters:
+ *
+ *              chongo <Landon Curt Noll> /\../\
+ *
+ * Note that the \'s above are not back-slashing escape characters.
+ * They are literal ASCII  backslash 0x5c characters.
+ *
+ * The effect of this virgin initial value is the same as starting
+ * with 0 and pre-pending those 32 characters onto the data being
+ * hashed.
+ *
+ * Yes, even with this non-zero virgin value there is a set of data
+ * that will result in a zero hash value.  Worse, appending any
+ * about of zero bytes will continue to produce a zero hash value.
+ * But that would happen with any initial value so long as the
+ * hash of the initial was the `inverse' of the virgin prefix string.
+ *
+ * But then again for any hash function, there exists sets of data
+ * which that the hash of every member is the same value.  That is
+ * life with many to few mapping functions.  All we do here is to
+ * prevent sets whose members consist of 0 or more bytes of 0's from
+ * being such an awkward set.
+ *
+ * And yes, someone can figure out what the magic 'inverse' of the
+ * 32 ASCII character are ... but this hash function is NOT intended
+ * to be a cryptographic hash function, just a fast and reasonably
+ * good hash function.
+ */
+#define FNV1_32_BASIS ((QCKHASH)(0x811c9dc5))
+
+
 /*
  * The largest power of 10 we will compute for our decimal conversion
  * internal constants is: 10^(2^TEN_MAX).