calc/CHANGES

Following is the change from calc version 2.11.0t1 to date:

    Removed the makefile symbol MAIN.  Now forcing all funcions to correctly
    be reclared main.  To passify some old broken compilers, a return 0;
    (instead of an exit(0);) is used at the end of main().

    A few of files that were added to calc used 4 character indentation
    whereas most of calc uses 8 character indentation.  These imported
    sources have been changed to conform better with the calc style.

    Fixed misc compile warnings.

    Added the program calc_errno.c and the Makefile symbol ERRNO_DECL.
    If ERRNO_DECL is empty, calc_errno.c will try various ways to
    declare errno, sys_errlist and sys_nerr.  On success or when
    it gives up, calc_errno will output the middle of the calc_errno.h
    header file.  If ERRNO_DECL is  -DERRNO_NO_DECL, or -DERRNO_STD_DECL
    or -DERRNO_OLD_DECL then the Makefile will build the middle
    of the calc_errno.h header file without calc_errno.c's help.

    The func.c file now includes the constructed header file calc_errno.h
    to ensure that errno, sys_errlist and sys_nerr are declared correctly.


Following is the change from calc version 2.10.3t5.38 to 2.11.0t0:

    Fixed a few compile problems found under Red Hat 6.0 Linux.


Following is the change from calc version 2.10.3t5.38 to 2.11.3t5.46:

    Fixed a bug discovered by Ernest Bowen related to matrix-to-matrix copies.

    Bitwise operations on integers have been extended so that negative
    integers are treated in the same way as the integer types in C.

    Some changes have been made to lib/regress.cal and lib/natnumset.cal.

    Removed V_STRLITERAL and V_STRALLOC string type constants and
    renumbered the V_protection types.

    Added popcnt(x, bitval) builtin which counts the number of
    bits in x that match bitval.

    Misc compiler warning fixes.

    Fixed improper use of putchar() and printf() when printing rationals
    (inside qio.c).

    Fixed previously reported bug in popcnt() in relation to . values.

    Calc man page changes per suggestion from Martin Buck
    <Martin-2.Buck@student.uni-ulm.de>.  The calc man page is
    edited with a few more parameters from the Makefile.

    Misc Makefile changes per Martin Buck <Martin-2.Buck@student.uni-ulm.de>.

    Removed trailing blanks from files.

    Consolidated in the Makefile, where the debug and check rules are found.
    Fixed the regress.cal dependency list.

    Make chk and check will exit with an error if check.awk detects
    a problem in the regression output.  (Martin Buck)

    Fixed print line for test #4404.

    Moved custom.c and custom.h to the upper level to fix unresolved symbols.

    Moved help function processing into help.c.

    Moved nearly everything into libcalc.a to allow programs access to
    higher level calc objects (e.g., list, assoc, matrix, block, ...).

    Renamed PATCH_LEVEL to MAJOR_PATCH and SUB_PATCH_LEVEL to MINOR_PATCH.
    Added integers calc_major_ver, calc_minor_ver, calc_major_patch
    and string calc_minor_patch to libcalc.a.  Added CALC_TITLE to hold
    the "C-style arbitrary precision calculator" string.

    The function version(), now returns a malloced version string
    without the title.

    Consolidated multiple SGI IRIX -n32 sections (for r4k, r5k and r10k)
    into a single section.


Following is the change from calc version 2.10.3t5.34 to 2.10.3t5.37:

    Per request from David I Bell, the README line:

      I am allowing this calculator to be freely distributed for personal uses

    to:

      I am allowing this calculator to be freely distributed for your enjoyment

    Added help files for:

	address agd arrow dereference free freeglobals freeredc freestatics
	gd isptr mattrace oldvalue saveval & * -> and .

    Fixed blkcpy() and copy() arg order and processing.  Now:

	A = blk() = {1,2,3,4}
	B = blk()
	blkcpy(B,A)
	blkcpy(B,A)

    will result in B being twice as long as A.

    Since "make chk" pipes the regression output to awk, we cannot
    assume that stdout and stderr are ttys.  Tests #5985 and #5986
    have been removed for this reason.  (thanks to Martin Buck
    <Martin-2.Buck@student.uni-ulm.de> for this report)

    Fixed the order of prints in regress.cal.  By convention, a print
    of a test line happens after the test.  This is because function
    parsed messages occur after the function is parsed.  Also the
    boolean tesrt of vrfy happens before any print statement.
    Therefore a non-test line is tested and printed as follows:

	y = sha();
	print '7125: y = sha()';

    The perm(a,b) and comb(a,b) have been extented to arbitrary real a and
    integer b.

    Fixed a bug in minv().

    Moved string.c into libcalc.a.

    The NUMBER union was converted back into a flat structure.  Changes
    where 'num' and 'next' symbols were changed to avoid #define conflicts
    were reverse since the #define's needed to support the union went away.

    Removed trailing blanks from files.

    Ernest Bowen <ernie@turing.une.edu.au> sent in the following patch
    which is described in the next 34 points:

    (0) In the past:

		    A = B = strcat("abc", "def");

	would store "abc" and "def" as literal strings never to be freed, and
	store "abcdef" once each for both A and  B.  Now the "abc" and "bcd"
	are freed immediately after they are concatenated and "abcdef" is stored
	only once, just as the number 47 would be stored only once for

		    A = B = 47;

	The new STRING structure that achieves this stores not only the
	address of the first character in the string, but also the "length"
	with which the string was created, the current "links" count, and
	when links == 0 (which indicates the string has been freed) the
	address of the next freed STRING.  Except for the null string "",
	all string values are "allocated"; the concept of literal string
	remains for names of variables, object types and elements, etc.

    (1) strings may now include '\0', as in A = "abc\0def".  In normal printing
	this prints as "abc" and strlen(A) returns 3, but its "real" length
	of 7 is given by size(A). (As before there is an 8th zero character
	and sizeof(A) returns 8.)

    (2) If A is an lvalue whose current value is a string of size n, then
	for 0 <= i < n, A[i] returns the character with index i as an addressed
	octet using the same structure as for blocks, i.e. there is no
	distinction between a string-octet and a block-octet.  The same
	operations and functions can be used for both, and as before, an octet
	is in some respects a number in [0,256) and in others a one-character
	string.  For example, for A = "abc\0def" one will have both A[0] == "a"
	and A[0] == 97.  Assignments to octets can be used to change
	characters in the string, e.g. A[0] = "A", A[1] = 0, A[2] -= 32,
	A[3] = " " will change the above A to "A\0C def".

    (3) "show strings" now displays the indices, links, length, and some or all
	of the early and late characters in all unfreed strings which are values
	of lvalues or occur as "constants" in function definitions,
	using "\n", "\t", "\0", "\252", etc. when appropriate.  For example,
	the string A in (1) would be displayed as in the definition there.
	Only one line is used for each string.  I've also changed the
	analogous "show numbers" so that only some digits of numbers that
	would require more than one line are displayed.

    (4) "show literals" is analogous to "show constants" for number "constants"
	in that it displays only the strings that have been introduced by
	literal strings as in A = "abc".  There is a major difference between
	strings and numbers in that there are operations by which characters
	in any string may be changed.  For example, after A = "abc",
	A[0] = "X" changes A to "Xbc".  It follows that if a literal string
	is to be constant in the sense of never changing, such a character-
	changing operation should never be applied to that string.

	In this connection, it should be noted that if B is string-valued, then

			    A = B

	results in A referring to exactly the same string as B rather than to
	a copy of what is in B.  Thie is like the use of character-pointers in
	C, as in

			    char *s1, *s2;
			    s1 = "abc";
			    s2 = s1;

	To achieve the effect of

			    s2 = (char *) malloc(4);
			    strcpy(s2, s1);

	I have extended the str() function to accept a string as argument.  Then

			    A = str(B);

	will create a new string at a different location from that of B but
	with the same length and characters.  One will then have A == B,
	*A == *B, but &*A != &*B, &A[0] != &B[0].

	To assist in analyzing this sort of thing, I have defined a links()
	function which for number or string valued argument returns the number
	of links to the occurrence of that argument that is being referred to.
	For example, supposing "abc" has not been used earlier:

			    > A = "abc"
			    > links(A)
				    2
			    > links(A)
				    1

	The two links in the first call are to A and the current "oldvalue";
	in the second call, the only link is to A, the oldvalue now being 2.


    (5) strcat(S1, S2, ...) works as before; contribution of a string stops when
	'\0' is encountered.  E.g.

		    strcat("abc\0def", "ghi")

	will return "abcghi".

    (6) For concatenation of full strings I have chosen to follow
	some other languages (like Java, but not Mathematica which uses "<>")
	and use "+" so that, e.g.

		    "abc\0def" + "ghi"

	returns the string "abc\0defghi".  This immediately gives obvious
	meanings to multiplication by positive integers as in

		    2 * "abc" = "abc" + "abc" = "abcabc",

	to negation to reverse as string as in

		    - "abc" = "cba",

	to multiplication by fractions as in

		    0.5 * "abcd" = "ab",

	(where the length is int(0.5 * size("abcd")), and finally, by combining
	these to

		     k * A    and      A * k

	for any real number k and any string A.   In the case of k == 1, these
	return a new string rather than A itself.  (This differs from
	"" + A and A + "" which return A.)

    (7) char(x) has been changed so that it will accept any integer x or octet
	as argument and return a string of size one with character value
	x % 256.  In the past calc has required 0 <= x < 256; now negative
	x is acceptable; for example, 1000 * char(-1) will now produce the
	same as 1000 * "\377" or 1000 * "\xff".

    (8) For a string s, test(s) now returns zero not only for the null string
	"" but also for a string all of whose characters are '\0'.

    (9) Similarly <, <=, etc. now compare all characters including occurrences
	of '\0' until a difference is encountered or the end of a string is
	reached.  If no difference is encountered but one string is longer than
	the other, the longer string is considered as greater even if the
	remaining characters are all '\0'.

    (10) To retain the C sense of comparison of null-terminated strings I have
	 defined strcmp(S1, S2), and then, for completeness, strncmp(S1, S2, n).
	 For similar reasons, strcpy(S1, S2) and strncpy(S1, S2, n) have been
	 defined.

    (11) For strings, I have defined | and & as bitwise "or" and "and"
	 functions, with S1 | S2 having the size of the larger of S1 and S2,
	 S1 & S2 having the size of the smaller of S1 and S2.  By using, say,
	 4-character strings, one can simulate a C integral type so far as the
	 | and & operations are concerned.   It then seemed appropriate to
	 use the operator ~ for a "bitwise complement" as in C.  Thus I have
	 defined ~s for a string s to be the string of the same size as s
	 with each character being complemented by the C ~ operation.

    (12) For boolean algebra work on strings it is convenient also to have
	 the bitwise xor and setminus binary operations.  Using C's '^' for xor
	 would be confusing when this is used elsewhere for powers, so I
	 decided to use ~.  For setminus, I adopted the commonly used '\'.
	 Strings of fixed size n can now be used for a boolean algebra
	 structure with 8 * n elements.  The zero element is n * char(0),
	 the unity is n * char(-1), and one have all of the usual laws like
	 A & (B | C) == A & B | A * C,  A \ B = A & ~B, etc.

    (13) Having extended the bitwise operations for strings, it was appropriate
	 to do the same for integers.  Definitions of the binary ~ and \
	 operations for non-negative integers are straightforward.  For
	 the unary ~ operation, I decided to do what C does with integer
	 types, and defined ~N to be -N - 1.  With the appropriate extensions of
	 |, &, \ and the binary ~, one gets in effect the boolean algebra of
	 finite sets of natural numbers and their complements, by identifying
	 the set with distinct integer elements i_1, i_2, ... with the integer

		    2^i_1 + 2^i_2 + ...

	 For ~N for non-integer real N, I have simply used -N.  There is some
	 logic in this and it is certainly better than an error value.
	 I have not defined the binary operations |, &, ~, \ for non-integral
	 arguments.

	 The use of ~N in this way conflicts with calc's method of displaying
	 a number when it has to be rounded to config("display") decimals.
	 To resolve this, my preference would be to replace the printing of
	 "~" as a prefix by a trailing ellipsis "...", the rounding always
	 being towards zero.  E.g. with config("display", 5), 1/7 would print
	 as ".14285..." rather than "~.14285".   The config("outround")
	 parameter would determine the type of rounding only for the
	 equivalent of config("tilde", 0).

    (14) For objects, users may create their own definitions for binary |,
	 &, ~ and \ with xx_or, xx_and, xx_xor, xx_setminus functions.
	 For unary ~ and \ operations, I have used the names xx_comp and
	 xx_backslash.

    (15) For the obviously useful feature corresponding to cardinality of a
	 set, I have defined #S for a string S to be the number of nonzero bits
	 in S.   For a degree of consistency, it was then appropriate to
	 define #N for a nonnegative integer N to be the number of nonzero bits
	 in the binary representation of N.  I've extended this to arbitrary
	 real N by using in effect #(abs(num(N))).  I feel it is better to make
	 this available to users rather than having #N invoke an error message
	 or return an error value.  For defining #X for an xx-object X, I
	 have used the name xx_content to suggest that it is appropriate for
	 something which has the sense of a content (like number of members of,
	 area, etc.).

    (16) Having recognized # as a token, it seemed appropriate to permit its
	 use for a binary operation.  For real numbers x and y I have defined
	 x # y to be abs(x - y).  (This is often symbolized by x ~ y, but it
	 would be confusing to have x ~ y meaning xor(x,y) for strings and
	 abs(x-y) for numbers.)  Because '#' is commonly called the hash symbol,
	 I have used xx_hashop to permit definition of x # y for xx-objects.

    (17) For a similar reason I've added one line of code to codegen.c so that
	 /A returns the inverse of A.

    (18) Also for a list L, +L now returns the sum of the elements of L.  For
	 an xx object A, +A requires and uses the definition of xx_plus.

    (19) I have given the unary operators ~, #, /, \, and except at the
	 beginning of an expression + and -, the same precedence with
	 right-to-left associativity.  This precedence is now weaker than
	 unary * and &, but stronger than binary & and the shift and power
	 operators.  One difference from before is that now

			    a ^ - b ^ c

	 evaluates as a ^ (- (b ^ c)) rather than a ^ ((- b) ^ c).


    (20) For octets o1, o2, I've defined o1 | o2, o1 & o2, o1 ~ o2, ~o1 so
	 that they return 1-character strings.  #o for an octet o returns the
	 number of nonzero bits in o.

    (21) For substrings I've left substr() essentially as before, but
	 for consistency with the normal block/matrix indexing, I've extended
	 the segment function to accept a string as first argument.  Then

		    segment(A, m, n)

	 returns essentially the string formed from the character with index m
	 to the character with index n, ignoring indices < 0 and indices >=
	 len(A); thus, if m and n are both in [0, size(A))
	 the string is of length abs(m - n) + 1, the order of the characters
	 being reversed if n < m.  Here the indices for a list of size len are
	 0, 1, ..., len - 1.  As it makes some sense, if 0 <= n < size(A),

		    segment(A, n)

	 now returns the one-character string with its character being that with
	 index n in A.  (I've made a corresponding modification to the segment
	 function for lists.)  Some examples, if A = "abcdef",

		    segment(A,2,4) = "cde",

		    segment(A,4,2) = "edc",

		    segment(A,3) = "d",

		    segment(A, -2, 8) = "abcdef",

		    segment(A,7,8) = "".

    (22) As essentially particular cases of segment(), I've defined
	 head(A, n) and tail(A, n) to be the strings formed by the first
	 or last abs(n) characters of A, the strings being4]5O~? reversed '
	 if n is negative.   I've changed the definitions of head and tail for
	 lists to be consistent with this interpretation of negative n.

    (23) Similarly I've left strpos ezsentially as at present, but search
	 and rsearch have been extended to strings.  For example,

		    search(A, B, m, n)

	 returns the index i of the first occurrence of the string B in A
	 if m <= i < n, or the null value if there is no such occurrence.
	 As for other uses of search, negative m is interpreted as
	 size(A) + m, negative n as size(A) + n.  For a match in this
	 search, all size(B) characters, including occurrences of '\0',
	 in B must match successive characters in A.

	 The function rsearch() behaves similarly but searches in reverse order
	 of the indices.

    (24) A string A of length N determines in obvious ways arrays of M = 8 * N
	 bits.  If the characters in increasing index order are c_0, c_1, ...
	 and the bits in increasing order in c_i are b_j, b_j+1, ..., b_j+7
	 where j = 8 * i, I've taken the array of bits determined by A to be

		    b_0, b_1, ..., b_M-1

	 For example, since "a" = char(97) and 97 = 0b01100001, and
	 "b" = char(98) = 0b01100010, the string "ab" determines the 16-bit
	 array

		    1000011001000110

	 in which the bits in the binary representations of "a" and "b" have
	 been reversed.

	 bit with index n in this array.   This is consistent with the use of
	 bit for a number ch in [0,256), i.e. bit(char(ch), n) = bit(ch, n).
	 For n < 0 or n >= size(A), bit(A,n) returns the null value.

    (25) For assigning values to specified bits in a string, I've defined
	 setbit(A, n) and setbit(A, n, v).  The first assigns the value 1 to
	 bit(A, n), the second assigns test(v) to bit(A, n).

    (26) For consistency with the corresponding number operations, the shift
	 operations A << n and A >> n have been defined to give what look
	 like right- and left-shifts, respectively.  For example, "ab" << 2
	 returns the 16-bit array

		    0010000110010001

	 in which the array for "ab" has been moved 2 bits to the right.

    (27) To achieve much the same as the C strcpy and strncpy functions for
	 null-terminated strings, strcpy(S1, S2) and strncpy(S1, S2, n) have
	 been defined.  Unlike the blkcpy() and copy() functions, the copying
	 for these is only from the beginning of the strings.  Also, unlike C,
	 no memory overflow can occur as the copying ceases when size(S1) is
	 reached.  Note that these overwrite the content of S1 (which affects
	 all strings linked to it) as well as returning S1.  Examples:

	    S = strcpy(6 * "x", "abc")      <=>  S = "abc\0xx"

	    S = strcpy(3 * "x", "abcdef")   <=>  S = "abc"

	    S = strncpy(6 * "x", "abcd", 2) <=>  S = "ab\0xxx"

	    S = strncpy(6 * "x", "ab", 4)   <=>  S = "ab\0\0xx"

	    S = strncpy(6 * "x", "ab", 20)  <=>  S = "ab\0\0\0\0"

	 If a new string S not linked to S1 is to be created, this can be
	 achieved by using str(S1) in place of S1.  For example, the strcpy in

	    A = "xxxxxx"
	    S = strcpy(str("xxxxxx"), "abc")

	 would not change the value of A.

    (28) I've extended the definitions of copy(A, B, ssi, num, dsi) and
	 blkcpy(B, A, num, ssi, dsi) to allow for string-to-string copying
	 and block-to-string copying, but num is now an upper bound for the
	 number of characters to be copied - copying will cease before num
	 characters are copied if the end of the data in the source A or the
	 end of the destination B is reached.  As with other character-changing
	 operations, copying to a string B will not change the locations of
	 B[0], B[1], ... or the size of B.

	 In the case of copying a string to itself, characters are copied in
	 order of increasing index, which is different from block-to-block
	 copying where a memmove is used.  This affects only copy from a
	 string to itself.  For example,

		    A = "abcdefg";
		    copy(A, A, , , 2);

	 will result in A == "abababa".  If the overwriting that occurs here
	 is not wanted, one may use

		    A = "abcdefg";
		    copy(str(A), A, , , 2);

	  which results in A == "ababcde".

    (29) perm(a,b) and comb(a,b) have been extended to accept any real a and
	 any integer b except for perm(a, b) with integer a such that b <= a < 0
	 which gives a "division by zero" error.  For positive b, both functions
	 are polynomials in a of degree b;  for negative b, perm(a,b) is a
	 rational function (1/((a + 1) * (a+2) ...) with abs(b) factors in the
	 denominator), and comb(a,b) = 0.  (An obvious "todo" is to extend this
	 to complex or other types of a.)

    (30) Although it is not illegal, it seems pointless to use a comma operator
	 with a constant or simple variable as in

		    > 2 * 3,14159
			    14159
		    > a = 4; b = 5;
		    > A = (a , b + 2);
		    > A
			    7

	 I have added a few lines to addop.c so that when this occurs a
	 "unused value ignored" message and the relevant line number are
	 displayed.  I have found this useful as I occasionally type ','
	 when I mean '.'.

	 There may be one or two other changes resulting from the way I have
	 rewritten the optimization code in addop.c.  I think there was a bug
	 that assumed that PTR_SIZE would be the same as sizeof(long).  By
	 the way, the new OP_STRING is now of index rather than pointer type.
	 It follows that pointers are now used in opcodes only for global
	 variables.  By introducing a table of addresses of global variables
	 like those used for "constants" and "literal strings", the use of
	 pointers in opcodes could be eliminated.

    (31) When calc has executed a quit (or exit) statement in a function or
	 eval evaluation, it has invoked a call to math_error() which causes
	 a long jump to an initial state without freeing any data on the
	 stack, etc.  Maybe more detail should be added to math_error(), but
	 to achieve the freeing of memory for a quit statement and at the same
	 time give more information about its occurrence I have changed the
	 way opcodes.c handles OP_QUIT.  Now it should free the local variables
	 and whatever is on the stack, and display the name and line-number,
	 for each of the functions currently being evaluated.  The last
	 function listed should be the "top-level" one with name "*".
	 Strings being eval-ed will have name "**".

	 Here is a demo:

	    > global a;
	    >
	    > define f(x) {local i = x^2; a++;
	    >> if (x > 5) quit "Too large!"; return i;}
	    f() defined
	    > define g(x) = f(x) + f(2*x);
	    g() defined
	    > g(2)
		    20
	    > g(3)
	    Too large!
		    "f": line 3
		    "g": line 0
		    "*": line 6
	    > eval("g(3)")
	    Too large!
		    "f": line 3
		    "g": line 0
		    "**": line 1
		    "*": line 7
	    > a
		    6

    (32) I've made several small changes like removing

		    if (vp->v_type == V_NUM) {
			    q = qinv(vp->v_num);
			    if (stack->v_type == V_NUM)
				    qfree(stack->v_num);
			    stack->v_num = q;
			    stack->v_type = V_NUM;
			    return;
		    }

	 from the definition of o_invert.  Presumably these lines were intended
	 to speed up execution for the common case of numerical argument.
	 Comparing the runtimes with and without these lines for inverting
	 thousands of large random numbers in a matrix suggest that execution
	 for real numbers is slightly faster without these lines.

	 Maybe this and other similar treatment of "special cases" should be
	 looked at more closely.

    (33) The new lib script lib/natnumset.cal demonstrates how the new
	 string operators and functions may be used for defining and
	 working with sets of natural numbers not exceeding a
	 user-specified bound.


Following is the change from calc version 2.10.3t5.28 to 2.10.3t5.33:

    Added hnrmod(v, h, n, r) builtin to compute:

	v % (h * 2^n + r), h>0, n>0, r = -1, 0 or 1

    Changed lucas.cal and mersenne.cal to make use of hnrmod().

    A number of changes from Ernest Bowen:

	(1) introduction of unary & and * analogous to those in C;

	    For an lvalue var, &var returns what I call a
	    value-pointer; this is a constant which may be assigned to
	    a variable as in p = &var, and then *p in expressions has
	    the same effect as var.  Here is a simple example of their use:

		> define s(L) {local v=0; while (size(L)) v+= *pop(L);return v;}
		s() defined
		> global a = 1, b = 2;
		> L = list(&a, &b);
		> print s(L)
		3
		> b = 3;
		> print s(L)
		4

	    Octet-pointers, number-pointers, and string-pointers in
	    much the same way, but have not attempted to do much with
	    the latter two.

	    To print a pointer, use the "%p" specifier.

	    Some arithmetic operations has been defined for corresponding
	    C operations.  For example:

		> A = mat[4];
		> p = &A[0];
		> *(p+2) == A[2]
		> ++p
		> *p == A[1]

	    There is at present no protection against "illegal" use of &
	    and *, e.g. if one attempts here to assign a value to *(p+5),
	    or to use p after assigning another value to A.

	    NOTE: Unlike C, in calc &A[0] and A are quite different things.

	    NOTE: If the current value of a variable X is an octet,
	    number or string, *X may be used to to return the value of
	    X; in effect X is an address and *X is the value at X.

	    Added isptr(p) builtin to return 0 is p is not a pointer,
	    >0 if it is a pointer.  The value of isptr(p) comes from the
	    V_XYZ #define (see the top of value.h) of the value to which
	    p points.

	    To allow & to be used as a C-like address operator, use of it
	    has been dropped in calls to user-defined functions.  For the
	    time being I have replaced it by the back-quote `.  For example:

		> global a
		> define f(a,b) = a = b
		> f(&a,5)
		> print a
		0
		> f(`a,5)
		> print a
		5

	   However, one may use & in a similar way as in:

		> define g(a,b) = *a = b
		> g(&a, 7)
		> print a
		7

	   There is no hashvalue for pointers. Thus, like error values,
	   they cannot be used as indices in an association.

	   The -> also works in calc. For example:

		> obj xy {x,y}
		> obj uvw {u, v, w}
		> obj xy A = {1,2}
		> obj uvw B = {3,4,5}
		> p = &A
		> q = &B
		> p->x
			1
		> p->y = 6
		> A
			obj xy {1, 6}
		> q -> u
			3
		> p->y = q
		> A
			obj xy {1, v-ptr: 1400474c0}
		> p->y->u
			3
		> p->y->u = 7
		> B
			obj uvw {7, 4, 5}
		> p -> y = p
		> A
			obj xy {1, v-ptr: 140047490}
		> p -> y -> x
			1
		> p->y->y
			v-ptr: 140047490
		> p->y->y-> x
			1
		> p->y->y->x = 8
		> A
			obj xy {8, v-ptr: 140047490}


	(2) a method of "protecting" variables;

	    For the various kinds of "protection", of an l_value var,
	    bits of var->v_subtype, of which only bits 0 and 1 have been
	    used in the past to indicate literal and allocated strings.
	    This has meant initialization of var->v_subtype when a new var
	    is introduced, and for assignments, etc., examination of the
	    appropriate bits to confirm that the operation is to be permitted.

	    See help/protect for details.

	(3) automatic "freeing" of constants that are no longer required.

	    For the "freeing" of constants, the definition of a NUMBER
	    structure so that a NUMBER * q could be regarded as a
	    pointing to a "freed number" if q->links = 0.

	    The old q->num was changed to a union q->nu which had a pointer
	    to the old q->num if q->links > 0 and to the next freed number
	    if q->links = 0.  The old "num" is #defined to "nu->n_num".

	    The prior method calc has used for handling "constants" amounted
	    to leakage.  After:

		> define f(x) = 27 + x;
		> a = 27;

	    It is of course necessary for the constant 27 to be stored, but
	    if one now redefines f and a by:

		> define f(x) = 45 + x;
		> a = 45;

	    There seems little point in retaining 27 as a constant and
	    therefore using up memory.  If this example seems trivial,
	    replace 27 with a few larger numbers like 2e12345, or better,
	    -2e12345, for which calc needs memory for both 2e12345 and
	    -2e12345!

	    Constants are automatically freed a definition when a
	    function is re- or un-defined.

	    The qalloc(q) and qfree(q) functions have been changed so
	    that that q->links = 0 is permitted and indicates that q
	    has been freed.  If a number has been introduced as a
	    constant, i.e. by a literal numeral as in the above
	    examples, its links becoming zero indicates that it is no
	    longer required and its position in the table of constants
	    becomes available for a later new constant.

	(4) extension of transcendental functions like tan, tanh, etc.
	    to complex arguments

	(5) definition of gd(z) and agd(z), i.e. the gudermannian and
	    inverse gudermannian

	(6) introduction of show options for displaying information about
	    current constants, global variables, static variables, and cached
	    redc moduli.

	    To help you follow what is going on, the following show
	    items have been introduced:

		show constants ==> display the currently stored constants
		show numbers   ==> display the currently stored numbers
		show redcdata  ==> display the currently stored redc moduli
		show statics   ==> display info about static variables
		show real      ==> display only real-valued variables

	    The constants are automatically initialized as constants and
	    should always appear, with links >= 1, in in the list of constants.

	    The show command:

		show globals

	    has been redefined so that it gives information about all
	    current global and still active static variables.

	(7) definition of functions for freeing globals, statics, redc values

	    To free memory used by different kinds of variable, the following
	    builtins have been added:

		freeglobals();		/* free all globals */
		freestatics();		/* free all statics */
		freeredc();		/* free redc moduli */
		free(a, b, ...);	/* free specific variables */

	   NOTE: These functions do not "undefine" the variables, but
	   have the effect of assigning the null value to them, and so
	   frees the memory used for elements of a list, matrix or object.

	   See 10) below for info about "undefine *".

	(8) enhancement of handling of "old value": having it return an
	    lvalue and giving option of disabling updating.

	    Now, by default, "." return an lvalue with the appropriate
	    value instead of copying the old value.

	    So that a string of commands may be given without changing
	    the "oldvalue", the new builtin:

		saveval(0)

	    function simply disables the updating of the "." value.
	    The default updating can be resumed by calling:

		saveval(1)

	    The "." value:

		> 2 + 2
		4
		> .
		4

	    can now be treated as an unnamed variable.  For example:

		> mat x[3,3]={1,2,3,4,5,6,7,8,9}
		> x
		> print .[1,2]
		6

	(9) for a list L defining L[i] to be same as L[[i]]

	(10) extending undefine to permit its application to all user-defined
	     functions by using "undefine *".

	     The command:

		undefine *

	     undefines all current user-defined functions.  After
	     executing all the above freeing functions (and if
	     necessary free(.) to free the current "old value"), the
	     only remaining numbers as displayed by:

		show numbers

	     should be those associated with epsilon(), and if it has been
	     called, qpi().

	(11) storing the most recently calculated value of qpi(epsilon)i and
	     epsilon so that when called again with the same epsilon it
	     is copied rather than recalculateed.

	(12) defining trace() for square matrices

	(13) expression in parentheses may now be followed by a qualifier
	     computible with its type

	     When an expression in parentheses evaluates to an lvalue
	     whose current value is a matrix, list or object, it may
	     now be followed by a qualifier computible with its type.

	     For example:

		> A = list(1,2,4);
		> B = mat[2,2] = {5,6,7,8};
		> define f(x) = (x ? A : B)[[1]];
		> print f(1), f(0)
		2 6

		> obj xy {x,y}
		> C = obj xy = {4,5}
		> p = &C
		> *p.x
		Not indexing matrix or object
		> (*p).x
		4

	(14) swap(a,b) now permits swapping of octets in the same or different
	     blocks.

	     For example:

		> A = blk() = {1,2,3}
		> B = blk() = {4,5,6}
		> swap(A[0], B[2])
		> A
			chunksize = 256, maxsize = 256, datalen = 3
			060203

    A few bug fixes from Ernest Bowen:

	B1: qcmpi(q, n) in qmath.c sometimes gave the wrong result if
		LONG_BITS > BASEB, len = 1 and nf = 0, since it then
		reduces to the value of (nf != q->num.v[1]) in which
		q->num.v[1] is not part of the size-1 array of HALFs for
		q->num.  At present this is used only for changing opcodes
		for ^2 and ^4 from sequences involving OP_POWER to
		sequences using OP_SQUARE, which has no effect on the
		results of calculations.

	B2: in matdet(m) in matfunc.c, a copy of the matrix m was not freed
		when the determinant turned out have zero value.

	B3: in f_search() in func.c, a qlinking of the NUMBER * storing the
		the size of a file was not qfreed.

	B4: in comalloc() in commath.c the initial zero values for real and
		imag parts are qlinked but not qfreed when nonzero values are
		assigned to them.  Rather than changing
		the definition of comalloc(), I have included any relevant
		qfrees with the calls to comalloc() as in
			c = comalloc();
			qfree(c->real);
			c->real = ...

	B5: in calls to matsum(), zeros are qlinked but not qfreed.  Rather
		than changing addnumeric(), I have changed the definition
		of matsum(m) so that it simply adds the components of m,
		which requires only that the relevant additions be defined,
		not that all components of m be numbers.


    Simple arithmetic expressions with literal numbers are evaluated
    during compilation rather than execution.  So:

	define f(x) = 2 + 3 + x;

    will be stored as if defined by:

	define f(x) = 5 + x;

    Fixed bug with lowhex2bin converstion in lib_util.c.  It did not
    correctly convert from hex ASCII to binary values due to a table
    loading error.

    Fixed porting problem for NetBSD and FreeBSD by renaming the
    qdiv() function in qmath.c to qqdiv().

    Improved the speed of mfactor (from mfactor.cal library) for
    long Mersenne factorizations.  The default reporting loop
    is now 10000 cycles.

    SGI Mips r10k compile set is speced for IRIX6.5 with v7.2
    compilers.  A note for pre-IRIX6.5 and/or pre-v7.2 compilers
    is given in the compile set.

    Added regression tests related to saveval(), dot and pointers.


Following is the change from calc version 2.10.3t5.11 to 2.10.3t5.27:

    The todo help file as been updated with the in-progress items:

	XXX - block print function is not written yet ...

    Expanded the role of blk() to produce unnamed blocks as in:

		B = blk(len, chunk)

    and named blocks as in:

		B = blk(str, len, chunk)

    A block may be changed (with possible loss of data only if len is less
    than the old len) by:

		C = blk(B, len, chunk)

    For an unnamed block B, this creates a new block C and copies
    min(len, oldlen) octets to it, B remaining unchanged.   For a named
    block, the block B is changed and C refers to the same block as B,
    so that for example, C[i] = x will result in B[i] == x.  Thus, for a
    named block, "B = " does nothing (other than B = B) in:

		B = blk(B, len, chunk)

    but is necessary for changing an unnamed block.

    Renamed rmblk() to blkfree().

    The builtin function blkfree(val) will free memory allocated to block.
    If val is a named block, or the name of a named block, or the
    identifying index for a named block, blkfree(val) frees the
    memory block allocated to this named block.  The block remains
    in existence with the same name, identifying index, and chunksize,
    but its size and maxsize becomes zero and the pointer for the start
    of its data block null.

    The builtin function blocks() returns the number of blocks that
    have been created but not freed by the blkfree() function.  When called
    as blocks(id) and the argument id less than the number of named
    blocks that have been created, blocks(id) returns the named block
    with identifying index id.

    Removed the artifical limit of 20 named blocks.

    Added name() builtin to return the name of a type of value
    as a string.

    Added isdefined() to determine of a value is defined.

    Added isobjtype() to determine the type of an object.

    The isatty(v) builtin will return 1 if v is a file that associated
    with a tty (terminal, xterm, etc.) and 0 otherwise.  The isatty(v)
    builtin will no longer return an error if v is not a file or
    is a closed file.

    The isident(m) builtin will return 1 if m is a  identity matrix
    and 0 otherwise.  The isident(m) builtin will no longer return an
    error if m is not a matrix.

    Added extensive testing of isxxx() builtins and their operations
    on various types.

    Added md5() builtin to perform the MD5 Message-Digest Algorithm.

    Renamed isset() to bit().

    Blocks will expand when required by the copy() builtin function:

	> f = fopen("help/full", "r")
	> B = blk()
	> B
		chunksize = 256, maxsize = 256, datalen = 0
	> copy(B, f)
	> B
		chunksize = 256, maxsize = 310272, datalen = 310084
		2a2a2a2a2a2a2a2a2a2a2a2a2a0a2a20696e74726f0a2a2a2a2a2a2a2a2a...

	NOTE: Your results will differ because changes to help/full.

    The blkcpy() builtin args now more closely match that
    of memcpy(), strncpy:

	blkcpy(dst, src [, num [, dsi [, ssi]]])

    The copy() builtin args now more closely match that the cp command:

	copy(src, dst [, num [, ssi [, dsi]]])

    but otherwise does the same thing as blkcpy.

    Fixed lint problems for SunOS.

    Added have_memmv.c and HAVE_MEMMOVE Makefile variable to control
    use of memmove().  If empty, then memmove() is tested for and if
    not found, or if HAVE_MEMMOVE= -DHAVE_NO_MEMMOVE then an internal
    version of memmove() is used instead.

    Added regression tests for sha, sha1 and md5 builtin hash functions.

    Added xx_print to to the list of object routines are definable.
    Added xx_print.cal to the library to demo this feature.

    Moved blkcpy() routines have been moved to blkcpy.[ch].

    The blkcpy() & copy() builtings can not copy to/from numbers.
    For purposes of the copy, only the numerator is ignored.

    Resolved a number of missing symbols for libcalc users.

    Added lib_util.{c,h} to the calc source to support users of
    libcalc.a.  These utility routines are not directly used by
    calc but are otherwise have utility to those programmers who
    directly use libcalc.a instead.

    Added sample sub-directory.  This sub-directory contains a few
    sample programs that use libcalc.a.  These sample programs are
    built via the all rule because they will help check to see that
    libcalc.a library does not contain external references that
    cannot be resolved.  At the current time none of these sample
    programs are installed.

    Added a libcalc_call_me_last() call to return storage created
    by the libcalc_call_me_first() call.  This allows users of libcalc.a
    to free up a small amount of storage.

    Fixed some memory leaks associated with the random() Blum generator.

    Fixed fseek() file operations for SunOS.

    Fixed convz2hex() fencepost error.  It also removes leading 0's.

    Plugged a memory leak relating to pmod.  The following calculation:

	pmod(2, x, something)

    where x was not 2^n-1 would leak memory.  This has been fixed.


Following is the change from calc version 2.10.3t5.1 to 2.10.3t5.10:

    Misc printf warning bug fixes.

    Calc now permits chains of initializations as in:

		obj point {x,y} P = {1,2} = {3,4} = {5,6}

    Here the initializations are applied from left to right.  It may
    look silly, but the 1, 2, ... could be replaced by expressions with
    side effects.  As an example of its use suppose A and B are
    expressions with side effects:

		P = {A, B}

    has the effect of P.x = A; P.y = B.  Sometimes one might want these in
    the reverse order: P.y = B; P.x = A.  This is achieved by:

		P = { , B} = {A}

    Another example of its use:

		obj point Q = {P, P} = {{1, 2}, {3, 4}}

    which results in Q having Q.x.x = 1, Q.x.y = 2, etc.

    The role of the comma in has been changed.  Expressions such as:

                mat A[2], B[3]

    are equivalent to:

                (mat A[2]), (mat B[3])

    Now, expr1, expr2  returns type of expr2 rather than EXPR_RVALUE.  This
    permits expressions such as:

		(a = 2, b) = 3

    Also, expr1 ? expr2 : expr3  returns type(expr2) | type(expr3).
    This will make the result an lvalue (i.e. EXPR_RVALUE bit not set)
    For example, if both expr2 and expr3 are lvalues.  Then:

		a ? b : c = d

    has the effect of b = d if a is "nonzero", otherwise c = d.

    This may be compared with

		d = a ? b : c

    which does d = b if a is "nonzero", otherwise d = c.

    And now, expr1 || expr2 and expr1 && expr2 each return
    htype(expr1)| type(expr2).  So for example:

		a || b = c

    has the effect of a = c if a is "nonzero", otherwise b = c.
    And for example:

		a && b = c

    has the effect of a = c if a is "zero", otherwise b = c.

    At top level, newlines are neglected between '(' and the matching
    ')' in expressions and function calls.  For example, if f() has been
    already defined, then:


		a = (
			2
			+
			f
			(
			3
			)
		    )

    and

		b = sqrt (
			20
			,
			1
		    )

    will be accepted, and in interactive mode the continue-line prompt
    will be displayed.

    When calc sees a "for", "while", "do", or "switch", newlines will be
    ignored (and the line-continuation prompt displayed in interactive mode)
    until the expected conditions and statements are completed.
    For example:

	s = 0;
	for (i = 0; i < 5; i++)
	{
		s += i;
	}
	print s;

    Now 's' will print '10' instead of '5'.

    Added more regression tests to regress.cal.  Changed the error
    counter from 'err' to 'prob'.  The errmax() is set very high and
    the expected value of errcount() is kept in ecnt.

    Added the 'unexpected' help file which gives some unexpected
    surprises that C programmers may encounter.

    Updated the 'help', 'intro' and 'overview' to reflect the
    full ilst of non-builtin function help files.  Reorered the
    'full' help file.

    The blkalloc() builtin has been renamed blk().

    Only a "fixed" type of BLOCK will be used.  Other types of
    blocks in the future will be different VALUE types.

    Introduced an undefine command so that

		    undefine f, g, ...

    frees the memory used to store code for user-defined functions f,
    g, ..., effectively removing them from the list of defined
    functions.

    When working from a terminal or when config("lib_debug") > 0 advice
    that a function has been defined, undefined, or redefined is
    displayed in format "f() defined".

    Some experimental changes to block and octet handling, so that after:

		    B = blk(N)

    B[i] for 0 <= i < N behaves for some operations like an lvalue for
    a USB8 in B.

    xx_assign added to object functions to permit the possibility of
    specifying what A = B will do if A is an xx-object.  Normal
    assignment use of = is restored by the command: undefine
    xx_assign.

    For error-value err, errno(err) returns the error-code for err and
    stores this in calc_errno;  error(err) returns err as if
    error(errno(err)) were called.

    Anticipating probable future use, names have been introduced for
    the four characters @, #, $, `.  This completes the coverage of
    printable characters on a standard keyboard.

    Added sha() builtin to perform the old Secure Hash Algorithm
    (SHS FIPS Pub 180).

    Added sha1() builtin to perform the new Secure Hash Standard-1
    (SHS-1 FIPS Pub 180-1).

    Added ${LD_DEBUG} Makefile variable to allow for additional
    libraries to be compiled into calc ... for debugging purposes.
    In most cases, LD_DEBUG= is sufficent.

    Added ${CALC_ENV} makefile variable to allow for particular
    environment variables to be supplied for make {check,chk,debug}.
    In most cases, CALC_ENV= CALCPATH=./lib is sufficent.

    Added ${CALC_LIBS} to list the libaraies created and used to
    build calc.  The CALC_LIBS= custom/libcustcalc.a libcalc.a
    is standard for everyone.

    Improved how 'make calc' and 'make all' rules work with respect
    to building .h files.

    Added 'make run' to only run calc interactively with the
    ${CALC_ENV} calc environment.  Added 'make cvd', 'make dbx'
    and 'make gdb' rules to run debug calc with the respective
    debugger with the ${CALC_ENV} calc environment.

    Added cvmalloc_error() function to lib_calc.c as a hook for
    users of the SGI Workshop malloc debugging library.

    Cut down on places where *.h files include system files.
    The *.c should do that instead where it is reasonable.

    To avoid symbol conflicts, *.h files produced and shipped
    with calc are inclosed that as similar to the following:

	#if !defined(__CALC_H__)
	#define __CALC_H__
	..
	#endif /* !__CALC_H__ */

    Added memsize(x) builtin to print the best aproximation of the
    size of 'x' including overhead.  The sizeof(x) builtin attempts
    to cover just the storage of the value and not the overhead.
    Because -1, 0 and 1 ZVALUES are static common values, sizeof(x)
    ignores their storage.  Also sizeof(x) ignores the denominator of
    integers, and the imaginary parts of pure real numbers.  Added
    regression tests for memsize(), sizeof() and size().


Following is the change from calc version 2.10.3t4.16 to 2.10.3t5.0:

    The calc source now comes with a custom sub-directory which
    contains the custom interface code.  The main Makefile now
    drives the building and installing of this code in a similar
    way that it drives the lib and help sub-directories.  (see below)

    Made minor edits to most help files beginning with a thru e.

    The errno(n) sets a C-like errno to the value n; errno() returns
    the current errno value.  The argument for strerror() and error()
    defaults to this errno.

    Added more error() and errno() regression tests.

    The convention of using the global variable lib_debug at the
    end of calc librar scripts has been replaced with config("lib_debug").
    The "lib_debug" is reserved by convention for calc library scripts.
    This config parameter takes the place of the lib_debug global variable.
    By convention, "lib_debug" has the following meanings:

	<-1	no debug messages are printed though some internal
		    debug actions and information may be collected

	-1	no debug messages are printed, no debug actions will be taken

	0	only usage message regarding each important object are
		    printed at the time of the read (default)

	>0	messages regarding each important object are
		    printed at the time of the read in addition
		    to other debug messages

    The "calc_debug" is reserved by convention for internal calc routines.
    The output of "calc_debug" will change from release to release.
    Generally this value is used by calc wizards and by the regress.cal
    routine (make check).  By convention, "calc_debug" has the following
    meanings:

	<-1	reserved for future use

	-1	no debug messages are printed, no debug actions will be taken

	0	very little, if any debugging is performed (and then mostly
		    in alpha test code).  The only output is as a result of
		    internal fatal errors (typically either math_error() or
		    exit() will be called). (default)

	>0	a greater degree of debugging is performed and more
		    verbose messages are printed (regress.cal uses 1).

    The "user_debug" is provided for use by users.  Calc ignores this value
    other than to set it to 0 by default (for both "oldstd" and "newstd").
    No calc code or shipped library will change this value other than
    during startup or during a config("all", xyz) call.

    The following is suggested as a convention for use of "user_debug".
    These are only suggestions: feel free to use it as you like:

	<-1	no debug messages are printed though some internal
		    debug actions and information may be collected

	-1	no debug messages are printed, no debug actions will be taken

	0	very little, if any debugging is performed.  The only output
		    are from fatal errors. (default)

	>0	a greater degree of debugging is performed and more
		    verbose messages are printed

    Added more code that is deading with the BLOCK type.

    Added blkalloc() builtin.

    Split NAMETYPE definition out into nametype.h.

    Added OCTET type for use in processing block[i].

    Added free, copy, cmp, quickhash and print functions for
    HASH, BLOCK and OCTET.

    Added notes to config.h about what needs to be looked at when
    new configuration items are added.

    The null() builtin now takes arguments.

    Given the following:

	obj point {x,y}
	obj point P, Q

    will will now create P and Q as obj point objects.

    Added xx_or, xx_and, xx_not and xx_fact objfuncs.

    Added the custom() builtin function.  The custom() builtin
    interface is designed to make it easier for local custom
    modification to be added to calc.  Custom functions are
    non-standard or non-portable code.  For these reasons, one must can
    only execute custom() code by way of an explicit action.

    By default, custom() returns an error.  A new calc command line
    option of '-C' is required (as well as ALLOW_CUSTOM= -DCUSTOM
    Makefile variable set) to enable it.

    Added -C as a calc command line option.  This permits the
    custom() interface to be used.

    Added ALLOW_CUSTOM Makefile variable to permanently disable
    or selective enable the custom builtin interface.

    The rm() builtin now takes multiple filenames.  If the first
    arg is "-f", then 'no-such-file' errors are ignored.

    Added errcount([count]) builtin to return or set the error
    counter.  Added errmax([limit]) to rturn or set the error
    count limiter.

    Added -n as a calc command line option.  This has the effect
    of calling config("all", "newstd") at startup time.

    Added -e as a calc command line option to ignore all environment
    varialbes at startup time.  The getenv() builtin function will
    still return values, however.

    Added -i as a calc command line option.  This has the effect
    ignoring when errcount() exceeds errmax().

    Changed the config("maxerr") name to config("maxscan").  The
    old name of "maxerr" is kept for backward compatibility.

    Using an unknown -flag on the calc command like will
    generate a short usage message.

    Doing a 'help calc' displays the same info as 'help usage'.

    The 'make check' rule now uses the -i calc command line flag
    so that regress.cal can continue beyond when errcount exceeds
    errmax.  In regress.cal, vrfy() reports when errcount exceeds
    errmax and resets errmax to match errcount.  This check
    and report is independent of the test success of failure.

    Fixed missing or out of order tests in regress.cal.

    Misc Makefile cleanup in lib/Makefile and help/Makefile.

    The default errmax() value on startup is now 20.

    The custom() interface is now complete.  See help/custom and
    custom/HOW_TO_ADD files, which show up as the custom and new_custom
    help files, for more information.

    The help command will search ${LIBDIR}/custhelp if it fails to find
    a file in ${LIBDIR}.  This allows the help command to also print
    help for a custom function.  However if a standard help file and a
    custom help file share the same name, help will only print the
    standard help file.  One can skip the standard help file and print
    the custom help file by:

	help custhelp/name

    or by:

	custom("help", "name")

    Added minor sanity checks the help command's filename.

    Added show custom to display custom function information.

    Added the contrib help page to give information on how
    and where to submit new calc code, modes or custom functions.

    Added comment information to value.h about what needs to be
    checked or modified when a new value type is added.

    Both size(x) and sizeof(x) return information on all value types.
    Moved size and sizeof information from func.c and into new file: size.c.

    Added custom("devnull") to serve as a do-nothing interface tester.

    Added custom("argv" [,arg ...]) to print information about args.

    Added custom("sysinfo", "item") to print an internal calc #define
    parameter.

    The make depend rule also processes the custom/Makefile.

    Added xx_max and xx_min for objfuncs.

    The max(), min() builtins work for lists.


Following is the change from calc version 2.10.3t3 to 2.10.3t4.15:

    The priority of unary + and - to that of binary + and - when they are
    applied to a first or only term.  Thus:

	-16^-2 == -1/256
	-7^2 == -49
	-3! == -6

    Running ranlib is no longer the default.  Systems that need RANLIB
    should edit the Makefile and comment back in:

	RANLIB=ranlib

    Dropped support of SGI r8k.

    Added support for the SGI r5k.

    Added support for SGI Mips compiler version 7.1 or later.

    Removed "random" as a config() option.

    Removed CCZPRIME Makefile variable.

    Added zsquaremod() back into zmod.c to be used by the Blum-Blum-Shub
    generator for the special case of needing x^2 mod y.

    Moved the Blum-Blum-Shub code and defines from zrand.c and zrand.h
    into zrandom.c and zrandom.h.  Now only the a55 generator resides
    in zrand.c and zrand.h.

    Added random, srandom and randombit help files.

    Added random(), srandom() and randombit() builtin functions.  The
    cryptographically strong random number generator is code complete!

    Removed cryrand.cal now that a Blum-Blum-Shub generator is builtin.

    Improved the speed of seedrandom.cal.  It now uses the 13th
    builtin Blum-Blum-Shub seed.

    The randmprime.cal script makes use of the Blum-Blum-Shub generator.

    Added randombitrun.cal and randomrun.cal calc library files.
    These are the Blum-Blum-Shub analogs to the randbitrun.cal
    and randrun.cal a55 tests.

    Improved hash.c interface to lower level hash functions.  The hash
    interface does not yet have a func.c interface ...  it is still
    under test.

    Added randombitrun.cal to test the Blum-Blum-Shub generator.

    Added calc.h, hash.h, shs.h and value.h to LIB_H_SRC because some
    of the libcalc.a files need them.

    In the original version, each call to newerror(str) created a new
    error-value.  Now a new value will be created only if str has not
    been used in a previous call to newerror().  In effect, the string
    serves to identify the error-value; for example:

	    return newerror("Non-integer argument");

    can be used in one or more functions, some of which may be
    repeatedly called, but after it has been called once, it will
    always return the same value as if one had initially used the
    assignment:

	    non_integer_argument_error = newerror("Non-integer argument")

    and then in each function used:

	    return non_integer_argument_error;

    The new definition of newerror() permits its freer use in cases like:

	    define foo(a) {

		    if (!isint(a))
			    return newerror("Non-integer argument");
		    ...
	    }

    One might say that "new" in "newerror" used to mean being different
    from any earlier error-value.  Now it means being not one of the
    "original" or "old" error-values defined internally by calc.

    As newerror() and newerror("") specify no non-null string, it has
    been arranged that they return the same as newerror("???").

    Added "show errors" command analogous to "show functions" for
    user-defined functions.  One difference is that whereas the
    functions are created by explicit definitions, a new described
    error is created only when a newerror(...) is executed.

    Fixed macro symbol substitution problem uncovered by HPUX cpp bug in
    HVAL and related zrand.h macros.

    Added +e to CCMISC for HP-UX users.

    Fixed the prompt bug.

    Eliminated the hash_init() initialization function.

    The 'struct block' has been moved from value.c to a new file: block.h.

    Added "blkmaxprint" config value, which limits the octets to print
    for a block.  A "blkmaxprint" of 0 means to print all octets of a
    block, regardless of size.  The default is to print only the first
    256 octets.

    The "blkverbose" determines if all lines, including duplicates
    should be printed.  If TRUE, then all lines are printed.  If false,
    duplicate lines are skipped and only a "*" is printed in a sytle
    similar to od.  This config value has not meaning if "blkfmt" is
    "str".  The default value for "blkverbose" is FALSE: duplicate
    lines are not printed.

    The "blkbase" determines the base in which octets of a block
    are printed.  Possible values are:

        "hexadecimal"           Octets printed in 2 digit hex
        "hex"

        "octal"                 Octets printed in 3 digit octal
        "oct"

        "character"             Octets printed as chars with non-printing
        "char"                      chars as \123 or \n, \t, \r

        "binary"                Octets printed as 0 or 1 chars
        "bin"

        "raw"                   Octets printed as is, i.e. raw binary
        "none"

    The default "blkbase" is "hex".

    The "blkfmt" determines for format of how block are printed:

        "line"          print in lines of up to 79 chars + newline
        "lines"

        "str"           print as one long string
        "string"
        "strings"

        "od"            print in od-like format, with leading offset,
        "odstyle"          followed by octets in the given base
        "od_style"

        "hd"            print in hex dump format, with leading offset,
        "hdstyle"          followed by octets in the given base, followed
        "hd_style"         by chars or '.' if no-printable or blank

    The default "blkfmt" is "hd".

    Fixed a bug in coth() when testing acoth using coth(acoth(x)) == x
    within the rounding error.

    Assignments to matrices and objects has been changed.  The assignments in:

	A = list(1,2,3,4);
	B = makelist(4) = {1,2,3,4};

    will result in A == B.  Then:

	A = {,,5}

    will result in A == list(1,2,5,4).

    Made minor edits to most help files beginning with a thru d.

    Fixed error in using cmdbuf("").


Following is the change from calc version 2.10.3t0 to 2.10.3t2:

    Bumped to version 2.10.3 due to the amount of changes.

    Renamed qabs() to qqabs() to avoid conflicts with stdlib.h.

    Fixed a casting problem in label.c.

    A lot of work was performed on the code generation by Ernest Bowen
    <ernie@neumann.une.edu.au>.  Declarations no longer need to precese code:

	define f(x) {
		local i = x^2;
		print "i = ":i;
		local j = i;
		...
	}

    The scope of a variable extends from the end of the declaration (including
    initialization code for the variable) at which it is first created
    to the limit given by the following rules:

	local variable: to the end of the function being defined

	global variable: to the end of the session with calc

	static within a function definition: to the the first of:

	    an end of a global, static or local declaration (including
	    initialization code) with the same identifier

	    the end of the definition

	static at top level within a file: to the first of:

	    the next static declaration of the identifier at top level
	    in the file,

	    the next global declaration of the identifier at top level
	    in the file or in any function definition in the file,

	    the next global declaration of the identifier at any level
	    in a file being read as a result of a "read" command,

	    the end of the file.

    The scope of a top-level global or static variable may be
    interrupted by the use of the identifier as a parameter or local or
    static variable within a function definition in the file being
    read; it is restored (without change of value) after the definition.

    For example, The two static variables a and b are created,
    with zero value, when the definition is read; a is initialized
    with the value x if and when f(x) is first called with a positive
    even x, b is similarly initialized if and when f(x) is first called
    positive odd x.  Each time f(x) is called with positive integer x,
    a or b is incremented.  Finally the values of the static variables
    are assigned to the global variables a and b, and the resulting
    values displayed.  Immediately after the last of several calls to
    f(x), a = 0 if none of the x's have been positive even, otherwise
    a = the first positive even x + the number of positive even x's,
    and b = 0 if none of the x's have been positive odd, otherwise
    b = the first positive odd x + the number of positive odd x's:

	define f(x) {
		if (isint(x) && x > 0) {
			if (iseven(x)) {
				static a = x;
				a++;
			} else {
				static b = x;
				b++;
			}
		}
		global a = a, b = b;
		print "a =",a,"b =",b;
	}

    Fixed some faults in the handling of syntax errors for the matrix
    and object creation operators mat and obj.  In previous versions of calc:

	mat;				<- Bad dimension 0 for matrix
	mat A;				<- Bad dimension 0 for matrix
	global mat A;			<- Bad dimension 0 for matrix
	mat A[2], mat B[3]		<- Semicolon expected
	global mat A[2], mat B[3]	<- Bad syntax in declaration statement

    Now:

	this statement			has the same effect as
	--------------			----------------------
	mat A[2], B[3]			(A = mat[2]), B[3]

	global mat A[2], B[3]		global A, B; A = mat[2]; B = mat[3];

    Initialization remains essentially as before except that for objects,
    spaces between identifiers indicate assignments as in simple variable
    declarations.  Thus, after:

	obj point {x,y};
	obj point P, Q R = {1,2}

    P has {0,0}, Q and R have {1,2}.  In the corresponding expression with
    matrices commas between identifiers before the initialization are ignored.
    For example:

	this statement			has the same effect as
	--------------			----------------------
	mat A, B C [2] = {1,2}		A = B = C = (mat[2] = {1,2})

    One can also do things like:

	L = list(mat[2] = {1,2}, obj point = {3,4}, mat[2] = {5,6})
	A = mat[2,2] = {1,2,3,4}^2
	B = mat[2,2] = {1,2,3,4} * mat[2,2] = {5,6,7,8}

    where the initialization = has stronger binding than the assignment = and
    the * sign.

    Matrices and objects can be mixed in declarations after any simple
    variables as in:

	global a, b, mat A, B[2] = {3,4}, C[2] = {4,5}, obj point P = {5,6}, Q

    Fixed some bugs related to global and static scoping.  See the the
    5200 regress test and lib/test5200.cal for details.

    Optimized opcode generator so that functions defined using '=' do not
    have two unreached opcodes.  I.e.,:

	define f(x) = x^2
	show opcodes f

    Also unreachable opcodes UNDEF and RETURN are now not included at
    the end of any user-defined function.

    Changed the "no offset" indicator in label.c from 0 to -1; this
    permits goto jumps to the zero opcode position.

    Changed the opcode generation for "if (...)" followed by
    "break", "continue", or "goto", so that only one jump opcode is
    required.

    A label can now be immediately by a rightbrace.  For example:

	define test_newop3(x) {if (x < 0) goto l132; ++x; l132: return x;}

    The LONG_BITS make variable, if set, will force the size of a long
    as well as forcing the USB8, SB8, USB16, SB16, USB32, SB32,
    HAVE_B64, USB64, SB64, U(x) and L(x) types.  If the longbits
    program is given an arg (of 32 or 64), then it will output
    based on a generic 32 or 64 bit machine where the long is
    the same size as the wordsize.

    Fixed how the SVAL and HVAL macros were formed for BASEB==16 machines.

    Dropped explicit Makefile support for MIPS r8k since these processors
    no longer need special compiler flags.

    SGI 6.2 and later uses -xansi.


Following is the change from calc version 2.10.2t33 to 2.10.2t34:

    Fixed a bug related to fact().

    Thanks to Ernest Bowen <ernie@neumann.une.edu.au>, for two or three
    arguments,

	    search(x, val, start);
	    rsearch(x, val, start);

    and for matrix, list or association x:

	    search(f, str, start);
	    rsearch(f, str, start);

    for a file stream f open for reading, behave as before except for a few
    differences:

	(1) there are no limits on the integer-valued start.

	(2) negative values of start are interpreted as offsets from the size of
	    x and f.  For example,

		    search(x, val, -100)

	    searches the last 100 elements of x for the first i for which
	    x[[i]] = val.

	(3) for a file f, when start + strlen(str) >= size(f) and
	    search(f, str, start) returns null, i.e. str is
	    not found, the file position after the search will be

		    size(f) - strlen(str) + 1

	    rather than size(f).

    For four arguments:

	    search(a, b, c, d)
	    rsearch(a, b, c, d),

    a has the role of x or f, and b the role of val or str as described
    above for the three-argument case, and for search(), c is
    essentially "start" as before, but for rsearch() is better for c
    and d to be the same as for search().  For a non-file case, if:

             0 <= c < d <= size(a),

    the index-interval over which the search is to take place is:

             c <= i < d.

    If the user has defined a function accept(v,b), this is used rather
    than the test v == b to decide for matrix, list, or association
    searches when a "match" of v = a[[i]] with b occurs. E.g.  after:

             define accept(v,b) = (v >= b);

    then calling:

             search(a, 5, 100, 200)

    will return, if it exists, the smallest index i for which
    100 <= i < 200 and a[[i]] >= 5.  To restore the effect of
    the original "match" function, one would then have to:

             define accept(v,b) == (v == b).

    Renamed the calc symbol BYTE_ORDER to CALC_BYTE_ORDER in order
    to avoid conflict.

    Added beer.cal and hello.cal lib progs in support of:   :-)

	http://www.ionet.net/~timtroyr/funhouse/beer.html
	http://www.latech.edu/~acm/HelloWorld.shtml


Following is the change from calc version 2.10.2t25 to 2.10.2t32:

    Eliminated use of VARARG and <varargs.h>.  Calc supports only
    <stdarg.h>.  The VARARGS Makefile variable has been eliminated.

    Source is converted to ANSI C.  In particular, functions
    will now have ANSI C style args.  Any comments from old K&R
    style args have been moved to function comment section.

    Removed prototype.h.  The PROTO() macro is no longer needed
    or supported.

    Added mfactor.cal to find the smallest factor of a Mersenne number.

    The built .h file: have_times.h, determines if the system has
    <time.h>, <times.h>, <sys/time.h> and <sys/times.h>.

    Because shs.c depends on HASHFUNC, which in turn depends on
    VALUE, shs.o has been moved out of libcalc.a.  For the same
    reasons, hash.h and shs.h are not being installed into
    the ${LIBDIR} for now.

    A number of the regression tests that need random numbers now
    use different seeds.

    Fixes for compiling under BSDI's BSD/OS 2.0.  Added a Makefile
    section for BSD/OS.

    Added a Makefile compile section for Dec Alpha without gcc ...
    provides a hack-a-round for Dec Alpha cc bug.

    Minor comment changes to lucas.cal.

    Added pix.cal, a slow painful but interesting way to compute pix(x).

    Confusion over the scope of static and global values has been reduced
    by a patch from Ernest Bowen <ernie@neumann.une.edu.au>.

	The change introduced by the following patch terminates the
	scope of a static variable at any static declaration with the
	same name at the same level, or at any global declaration with
	the same name at any level.  With the example above, the scope
	of the static "a" introduced in the third line ends when the
	"global a" is read in the last line.  Thus one may now use the
	same name in several "static" areas as in:

			> static a = 10;
			> define f(x) = a + x;
			> static a = 20;
			> define g(x) = a + x;
			> global a;

	The first "a" exists only for the definition of f(); the second
	"a" only for the definition of g().  At the end one has only
	the global "a".

	Ending the scope of a static variable in this way is consistent
	with the normal use of static variables as in:

			> static a = 10;
			> define f(x) {static a = 20; return a++ + x;}
			> define g(x) = a + x;
			> global a;

	The scope of the first "a" is temporarily interrupted by the
	"static a" in the second line; the second "a" remains active
	until its scope ends with the ending of the definition of f().
	Thus one ends with g(x) = 10 + x and on successive calls to
	f(), f(x) returns 20 + x, 21 + x, etc.  With successive "static
	a" declarations at the same level, the active one at any stage
	is the most recent; if the instructions are being read from a
	file, the scope of the last "static a" ends at the end-of-file.

	Here I have assumed that no "global a" is encountered.  As
	there can be only one global variable with name "a", it seems
	to me that its use must end the scope of any static "a".  Thus
	the changes I introduce are such that after:

			> global a = 10;
			> define f(x) = a + x;
			> static a = 20;
			> define g(x) = a + x;
			> define h(x) {global a = 30; return a + x;}
			> define i(x) = a + x;

	g(x) will always return 20 + x, and until h(x) has been called,
	f(x) and i(x) will return 10 + x; when h(x) is called, it
	returns 30 + x and any later call to f(x) or i(x) will return
	30 + x.  It is the reading of "global a" in the definition of
	h() that terminates the scope of the static a = 20, so that the
	"a" for the last line is the global variable defined in the
	first line.  The "a = 30" is executed only when h() is called.

	Users who find this confusing might be well advised to use
	different names for different variables at the same scope level.

	The other changes produced by the patch are more straightforward,
	but some tricky programming was needed to get the possibility of
	multiple assignments and what seems to be the appropriate order
	of executions and assignments.  For example, the order for the
	declaration:

		global a, b = expr1, c, d = expr2, e, f

	will be:

		evaluation of expr1;
		assignment to b;
		evaluation of expr2;
		assignment to d;

	Thus the effect is the same as for:

		a = 0; b = expr1; c = 0; d = expr2; e = 0; f = 0;

	The order is important when the same name is used for different
	variables in the same context.  E.g. one may have:

		define f(x) {
			global a = 10;
			static a = a;
			local a = a--;

			while (--a > 0)
				x++;
			return x;
		}

	Every time this is called, the global "a" is assigned the value
	10.  The first time it is called, the value 10 is passed on to
	the static "a" and then to the local "a".  In each later call
	the "static a = a" is ignored and the static "a" is one less than
	it was in the preceding call.  I'm not recommending this style of
	programming but it is good that calc will be able to handle it.

	I've also changed dumpop to do something recent versions do not do:
	distinguish between static and global variables with the same name.

	Other changes: commas may be replaced by spaces in a sequence of
	identifiers in a declaration. so one may now write:

		global a b c = 10, d e = 20

	The comma after the 10 is still required.  Multiple occurrences
	of an identifier in a local declaration are now acceptable as
	they are for global or static declarations:

		local a b c = 10, a = 20;

	does the same as:

		local a b c;
		a = b = c = 10;
		a = 20;

	The static case is different in that:

		static a b c = 10, a = 20;

	creates four static variables, the first "a" having a very short and
	useless life.

    Added new tests to verify the new assugnments above.

    Added the builtin test(x) which returns 1 or 0 according as x tests
    as true or false for conditions.

    Added have_posscl.c which attempts to determine if FILEPOS is
    a scalar and defines HAVE_FILEPOS_SCALAR in have_posscl.h
    accordingly.  The Makefile variable HAVE_POSSCL determines
    if have_posscl.c will test this condition or assume non-scalar.

    Added have_offscl.c which attempts to determine if off_t is
    a scalar and defines HAVE_OFF_T_SCALAR in have_posscl.h
    accordingly.  The Makefile variable HAVE_OFFSCL determines
    if have_offscl.c will test this condition or assume non-scalar.

    Reading to EOF leaves you positioned one character beyond
    the last character in the file, just like Un*x read behavior.

    Calc supports files and offsets up to 2^64 bytes, if the OS
    and file system permits.


Following is the change from calc version 2.10.2t4 to 2.10.2t24:

    Added makefile debugging rules:

	make chk	like a 'make check' (run the regression tests)
			except that only a few lines around interesting
			(and presumable error messages) are printed.
			No output if no errors are found.

	make env	print important makefile values

	make mkdebug	'make env' + version information and a
			make with verbose output and printing of
			constructed files

	make debug	'make mkdebug' with a 'make clobber'
			so that the entire make is verbose and
			a constructed files are printed

     Improved instuctions in 'BUGS' section on reporting problems.
     In particular we made it easy for people to send in a full
     diagnostic output by sending 'debug.out' which is made as follows:

	make debug > debug.out

     Added -v to calc command line to print the version and exit.

     Fixed declarations of memcpy(), strcpy() and memset() in the
     case of them HAVE_NEWSTR is false.

     Fixed some compile time warnings.

     Attempting to rewind a file this is not open generates an error.

     Noted conversion problems in file.c in tripple X comments.

     Some extremely braindead shells cannot correctly deal with if
     cluases that do not have a non-empty else statement.  Their
     exit bogosity results in make problems.  As a work-a-round,
     Makefile if clauses have 'else true;' clauses for if statements
     that previously did not have an else cluause.

     Fixed problems where the input stack depth reached the 10 levels.

     The show keyword is now a statement instead of a command:

	> define demo() {local f = open("foo", "w"); show files; fclose(f);}
	> demo()

     Added a new trace option for display of links to real and complex
     numbers.  This is activated by config("trace", 4).  The printing
     of a real number is immediately followed by "#" and the number of
     links to that number; complex numbers are printed in the same
     except for having "##" instead of "#".  <ernie@neumann.une.edu.au>

     The number of links for a number value is essentially the number of value
     locations at which it is either stored or deemed to be stored.  Here a
     number value is the result of a reading or evaluation; when the result
     is assigned to lvalues, "linking" rather than copying occurs.  Different
     sets of mutually linked values may contain the same number.  For example:

	a = b = 2 + 3; x, y = 2 + 3;

     a and b are linked, and x and y are linked, but a and x are not linked.

     Revised the credits help file and man page.  Added archive help
     file to indicate where recent versions of calc are available.

     The regression test suite output has been changed so that it will
     output the same information regardless of CPU performance.  In
     particular, cpu times of certain tests are not printed.  This allows
     one to compare the regression output of two different systems easier.

     A matrix or object declaration is now considered an expression
     and returns a matrix or object of the specified type.  Thus one may
     use assignments like:

	A = mat[2];		/* same as: mat A[2]; */
	P = obj point;		/* same as: obj point P; */

     The obj and mat keywords may be with "local", "global", "static" as in:

	local mat A[2];

     Several matrices or objects may be assigned or declared in the one
     statement, as in:

	mat A, B[2], C[3];	/* same as: mat A[2], B[2], C[3] */

     except that only one matrix creation occurs and is copied as in:

	A = B = mat[2];

     Initialization of matrices and objects now occur before assignments:

	mat A, B [2] = {1,2};	/* same as: A = B = (mat[2] = {1,2}); */

     Missing arguments are considered as "no change" rather than
     "assign null values".  As in recent versions of calc, the default
     value assigned to matrix elements is zero and the default for object
     elements is a null value).  Thus:

	mat A[2] = {1,2};
	A = { , 3};

     will change the value of A to {1,3}.

     If the relevant operation exists for matrices or has been defined for
     the type of object A is, the assignment = may be combined with +, -, *,
     etc. as in:

	A += {3, 4};		/* same as: A[0] += 3; A[1] += 4; */
	A += { };		/* same as: A += A; */

     In (non-local) declarations, the earlier value of a variable may be
     used in the initialization list:

	mat A[3]={1,2,3}; mat A[3]={A[2],A[1],A[0]}; /* same as: A={3,2,1} */

     Also:

	mat A[3] = {1,2,3};
	mat A[3] = {A, A, A};

     produces a 3-element matrix, each of whose elements is a 3-element matrix.

     The notation A[i][j] requires A[i] to be a matrix, whereas B[i,j]
     accesses an element in a 2-dimensional matrix.  Thus:

	B == A[i]	implies		A[i][j] = B[j]

     There is requirement in the use of A[i][j] that the matrices A[i]
     for i = 0, 1, ... all be of the same size.  Thus:

	mat A[3] = {(mat[2]), (mat[3]), (mat[2])};

     produces a matrix with a 7-element structure:

	A[0][0], A[0][1], A[1][0], A[1][1], A[1][2], A[2][0], A[2][1]

     One can initialize matrices and objects whose elements are matrices
     and/or objects:

	obj point {x,y}
	obj point P;
	obj point A = {P,P};

     or:

	obj point {x,y};
	obj point P;
	mat A[2] = {P,P};
	A = {{1,2}, {3,4}};

     The config("trace", 8) causes opcodes of newly defined functions
     are displayed.  Also show can now show the opcides for a function.
     For example:

	config("trace", 8);
	define f(x) = x^2;
	show opcodes f;
	define g(x,y) {static mat A[2]; A += {x,y}; return A;}
	show opcodes g
	g(2,3);
	show opcodes g;
	g(3,4);

     The two sequences displayed for f should show the different ways
     the parameter is displayed.  The third sequence for g should also
     show the effects of the static declaration of A.

     Fixed a number of compiler warning and type cast problems.

     Added a number of new error codes.

     Misc bug fixes for gcc2 based Sparc systems.

     Fixed a bug in the SVAL() macro on systems with 'long long'
     type and on systems with 16 bit HALFs.

     Reduced the Makefile CC set:

	 CCOPT are flags given to ${CC} for optimization
	 CCWARN are flags given to ${CC} for warning message control
	 CCMISC are misc flags given to ${CC}

	 CFLAGS are all flags given to ${CC}
		[[often includes CCOPT, CCWARN, CCMISC]]
	 ICFLAGS are given to ${CC} for intermediate progs

	 CCMAIN are flags for ${CC} when files with main() instead of CFLAGS
	 CCSHS are flags given to ${CC} for compiling shs.c instead of CFLAGS

	 LCFLAGS are CC-style flags for ${LINT}
	 LDFLAGS are flags given to ${CC} for linking .o files
	 ILDFLAGS are flags given to ${CC} for linking .o files
		  for intermediate progs

	 CC is how the the C compiler is invoked

    Added more tests to regress.cal.

    Port to HP-UX.

    Moved config_print() from config.c to value.c so prevent printvalue()
    and freevalue() from being unresolved symbols for libcalc.a users.

    Calc will generate "maximum depth reached" messages or errors when
    reading or eval() is attempted at maximum input depth.

    Now each invocation of make is done via ${MAKE} and includes:

	MAKE_FILE=${MAKE_FILE}
	TOPDIR=${TOPDIR}
	LIBDIR=${LIBDIR}
	HELPDIR=${HELPDIR}

    Setting MAKE_FILE= will cause make to not re-make if the Makefile
    is edited.

    Added libinit.c which contains the function libcalc_call_me_first().
    Users of libcalc.a MUST CALL libcalc_call_me_first BEFORE THEY USE
    ANY OTHER libcalc.a functions!

    Added support for the SGI IRIX6.2 (or later) Mongoose 7.0 (or later)
    C Compiler for the r4k, r8k and r10k.  Added LD_NO_SHARED for
    non-shared linker support.

    Re-ordered and expanded options for the DEBUG make variable.

    Make a few minor cosmetic comment changes/fixes in the main Makefile.

    Statements such as:

                mat A[2][3];

    now to the same as:

                mat M[3];
                mat A[2] = {M, M};

    To initialize such an A one can use a statement like

                A = {{1,2,3}, {4,5,6}};

    or combine initialization with creation by:

                mat A[2][3] = {{1,2,3}, {4,5,6}};

    One would then have, for example, A[1][0] = 4.  Also, the inner braces
    cannot be removed from the initialization for A:

                mat A[2][3] = {1,2};

    results in exactly the same as:

                mat A[2] = {1,2};

    Added rm("file") builtin to remove a file.

    The regress test sections that create files also use rm() to remove
    them before and afterwards.

    Added 4400-4500 set to test new mat and obj initializaion rules.

    Added 4600 to test version file operations.

    Added CCZPRIME Makefile variable to the set for the short term
    to work around a CC -O2 bug on some SGI machines.

    Added regression test of _ variables and function names.

    Added read of read and write, including read and write test for
    long strings.

    Fixed bug associated with read of a long string variable.

    Renumbered some of the early regress.cal test numbers to make room
    for more tests.  Fixed all out of sequence test numbers.  Fixed some
    malformatted regression reports.

    Renamed STSIZE_BITS to OFF_T_BITS.  Renamed SWAP_HALF_IN_STSIZE to
    SWAP_HALF_IN_OFF_T.


Following is the change from calc version 2.10.2t1 to 2.10.2t3:

    Fixed bug in the regression suite that made test3400 and test4100
    fail on correct computations.

    The randbit() builtin, when given to argument, returns 1 random bit.

    Fixed a bug in longlong.c which made is generate a syntax error
    on systems such as the PowerPC where the make variable LONGLONG
    was left empty.

    By default, the Makefile leaves LONGLONG_BITS empty to allow for
    testing of 64 bit data types.  A few hosts may have problems with
    this, but hopefully not.  Such hosts can revert back to LONGLONG_BITS=0.

    Improved SGI support.  Understands SGI IRIX6.2 performance issues
    for multiple architectures.

    Fixed a number of implicit conversion from unsigned long to long to avoid
    unexpected rounding, sign extension, or loss of accuracy side effects.

    Added SHSCC because shs.c contains a large expression that some
    systems need help in optimizing.

    Added "show files" to display information about all currently open files.

    Calc now prevents user-defined function having the same name as a
    builtin function.

    A number of new error codes (more than 100) have been added.

    Added ctime() builtin for date and time as string value.
    Added time() builtin for seconds since 00:00:00 1 Jan 1970 UTC.
    Added strerror() builtin for string describing error type.
    Added freopen() builtin to reopen a file.
    Added frewind() builtin to rewind a file.
    Added fputstr() builtin to write a null-terminated string to a file.
    Added fgetstr() builtin to read a null-terminated string from a file.
    Added fgetfield() builtin to read next field from file.
    Added strscan() builtin to scan a string.
    Added scan() builtin to scan of a file.
    Added fscan() builtin to scan of a file.
    Added fscanf() builtin to do a formatted scan of a file.
    Added scanf() builtin to do a formatted scan of stdin.
    Added strscanf() builtin to do a formatted scan of a string.
    Added ungetc() builtin to unget character read from a file.

    As before, files opened with fopen() will have an id different from
    earlier files.  But instead of returning the id to the FILEIO slot
    used to store information about it, calc simply uses consecutive
    numbers starting with 3.  A calc file retains its id, even when the
    file has been closed.

    The builtin files(i) now returns the file opened with id == i
    rather than the file with slot number i.  For any i <= lastid,
    files(i) has at some time been opened.  Whether open or closed, it
    may be "reopened" with the freopen() command.  This write to a file
    and then read it, use:

	f = fopen("junk", "w")
	freopen(f, "r")

	To use the same stream f for a new file, one may use:

	    freopen(f, mode, newfilename)

	which closes f (assuming it is open) and then opens newfilename on f.

	And as before:

	    f = fopen("curds", "r")
	    g = fopen("curds", "r")

	results in two file ids (f and g) that refer to the same file
	name but with different pointers.

    Calc now understands "w+", "a+" and "r+" file modes.

    If calc opens a file without a mode there is a "guess" that mode
    "r+" will work for any files with small descriptors found to be
    open.  In case it doesn't (as apparently happens if the file had
    not been opened for both reading and reading) the function now also
    tries "w" and "r", and if none work, gives up.  This avoids having
    "open" files with null fp.

    The buildin rewind() calls the C rewind() function, but one may
    now rewind several files at once by a call like rewind(f1, f2).
    With no argument, rewind() rewinds all open files with id >= 3.

    The functions fputstr(), fgetstr() have been defined to include the
    terminating '\0' when writing a string to a file.  This can be done
    at present with a sequence of instructions like:

	fputs(f, "Landon"); fputc(f, 0);
	fputs(f, "Curt"); fputc(f, 0);
	fputs(f, "Noll"); fputc(f, 0);

	One may now do:

	    fputstr(f, "Landon", "Curt", "Noll");

	and read them back by:

	    rewind(f);
	    x = fgetstr(f);	/* returns "Landon" */
	    y = fgetstr(f);	/* returns "Curt" */
	    z = fgetstr(f);	/* returns "Noll" */

    The buildin fgetfield() returns the next field of non-whitepsace
    characters.

    The builtins scan(), fscan(), strscan() read tokens (fields of
    non-whitepsace characters) and evaluates them.  Thus:

	global a,b,c;
	strscan("2+3  4^2\n c=a+b", a, b, 0);

	results in a = 5, b = 16, c = 21

    The functions scanf, fscanf, strscanf behave like the C functions
    scanf, fscanf, sscanf.   The conversion specifiers recognized are "%c",
    "%s", "%[...]" as in C, with the options of *, width-specification,
    and complementation (as in [^abc]), and "%n" for file-position, and
    "%f", "%r", "%e", "%i" for numbers or simple number-expressions - any
    width-specification is ignored; the expressions are not to include any
    white space or characters other than decimal digits, +, -, *, /, e, and i.
    E.g. expressions like 2e4i+7/8 are acceptable.

    The builtin size(x) now returns the size of x if x is an open file
    or -1 if x is a file but not open.  If s is a string, size(s) returns
    characters in s.

    Added buildin access("foo", "w") returns the null value if a file
    "foo" exists and is writeable.

    Some systems has a libc symbolc qadd() that conflicted with calc's
    qadd function.  To avoid this, qadd() has been renamed to qqadd().

    The calc error codes are produced from the the calcerr.tbl file.
    Instead of changing #defines in value.h, one can not edit calcerr.tbl.
    The Makefile builds calcerr.h from this file.

    Calc error codes are now as follows:

	<0 			invalid
	0 .. sys_nerr-1		system error ala C's errno values
	sys_nerr .. E__BASE-1	reserved for future system errors
	E__BASE .. E__HIGHEST	calc internal errors
	E__HIGHEST+1 .. E_USERDEF-1	invalid
	E_USERDEF ..		user defined errors

    Currently, E__BASE == 10000 and E_USERDEF == 20000.  Of course,
    sys_nerr is system defined however is likely to be < E__BASE.

    Renamed CONST_TYPE (as defined in have_const.h) to just CONST.
    This symbol will either be 'const' or an empty string depending
    on if your compiler understands const.

    CONST is beginning to be used with read-only tables and some
    function arguments.  This allows certain compilers to better
    optimize the code as well as alerts one to when some value
    is being changed inappropriately.  Use of CONST as in:

	int foo(CONST int curds, char *CONST whey)

    while legal C is not as useful because the caller is protected
    by the fact that args are passed by value.  However, the
    in the following:

	int bar(CONST char *fizbin, CONST HALF *data)

    is useful because it calls the compiler that the string pointed
    at by 'fizbin' and the HALF array pointer at by 'data' should be
    treated as read-only.


Following is the change from calc version 2.10.1t21 to 2.10.2t0:

    Bumped patch level 2.10.2t0 in honor of having help files for
    all builtin functions.  Beta release will happen at the end of
    the 2.10.2 cycle!!!

    Fewer items listed in BUGS due to a number of bug fixes.

    Less todo in the help/todo file because more has already been done.  :-)

    All builtin functions have help files!  While a number need cleanup
    and some of the LIMITS, LIBRARY and SEE ALSO sections need fixing
    (or are missing), most of it is there.  A Big round of thanks goes
    to <ernie@neumann.une.edu.au> for his efforts in initial write-ups
    for many of these files!

    The recognition of '\' as an escape character in the format argument
    of printf() has been dropped.  Thus:

	printf("\\n");

    will print the two-character string "\n" rather than the a
    one-character carriage return.  <ernie@neumann.une.edu.au>

    Missing args to printf-like functions will be treated as null values.

    The scope of of config("fullzero") has been extended to integers,
    so that for example, after config("mode","real"), config("display", 5),
    config("fullzero", 1), both:

	print 0, 1, 2;
	printf("%d %d %d\n", 0, 1, 2);

    print:

	.00000 1.00000, 2.00000

    The bug which caused calc to exit on:

	b = "print 27+"
	eval(b)

    has been fixed.  <ernie@neumann.une.edu.au>

    Fixed bugs in zio.c which caused eval(str(x)) == x to fail
    in non-real modes such as "oct".  <ernie@neumann.une.edu.au>

    The following:

	for (i = 1; i < 10; i++) print i^2,;

    now prints the same as:

	for (i = 1; i < 10; i++) print i^2,;

    The show globals will print '...' in the middle of large values.
    <ernie@neumann.une.edu.au>

    The param(n) builtin, then n > 0, returns the address rather than
    the value of the n-th argument to save time and memory usage.  This
    is useful when a matrix with big number entries is passed as an arg.
    <ernie@neumann.une.edu.au>

    The param(n) builtin, then n > 0, may be used as an lvalue:

	> define g() = (param(2) = param(1));
	> define h() = (param(1)++, param(2)--);
	> u = 5
	> v = 10
	> print g(u, &v), u, v;
	5 5 5
	> print h(&u, &v), u, v;
	5 6 4

    Missing args now evaluate to null as in:

	A = list(1,,3)
	B = list(,,)
	mat C[] = {,,}
	mat D[] = { }


Following is the change from calc version 2.10.1t20 to 2.10.1t20:

    Changes made in preparation for Blum Blum Shub random number generator.

    REDC bug fixes: <ernie@neumann.une.edu.au>

	Fixed yet another bug in zdiv which occasionally caused the "top digit"
	of a nonzero quotient to be zero.

	Fixed a bug in zredcmul() where a rarely required "topdigit" is
	sometimes lost rather than added to the appropriate carry.

    A new function zredcmodinv(ZVALUE z, ZVALUE *res) has been defined for
    evaluating rp->inv in zredcalloc().  <ernie@neumann.une.edu.au>

    New functions zmod5(ZVALUE *zp) and zmod6(ZVALUE z, ZVALUE *res) have
    been defined to give O(N^1.585)-runtime evaluation of z % m for
    large N-word m.  These require m and BASE^(2*N) // m to have been
    stored at named locations lastmod, lastmodinv.  zmod5() is essentially
    for internal use by zmod6() and zpowermod().  <ernie@neumann.une.edu.au>

    Changes to rcmul(x,y,m) so that the result is always in [0, m-1].
    <ernie@neumann.une.edu.au>

    Changes to some of the detail of zredcmul() so that it should run slightly
    faster.  Also changes to zredcsq() in the hope that it might achieve
    something like the improvement in speed of x^2 compared with x * x.
    <ernie@neumann.une.edu.au>

    A new "bignum" algorithm for evaluating pmod(x,k,m) when
    N >= config("pow2").  For the multiplications and squarings
    modulo m, or their equivalent, when N >= config("redc2"),
    calc has used evaluations correponding to rcout(x * y, m),
    for which the runtime is essentially that of three multiplications.
    <ernie@neumann.une.edu.au>

    Yet more additions to the regress.cal test suite.

    Fixed some ANSI-C compile nits in shs.c and quickhash.c.

    Plugs some potential memory leaks in definitions in func.c.
    Expressions such as qlink(vals[2]) in some circumstances are
    neither qfreed nor returned as function values.
    <ernie@neumann.une.edu.au>

    The nextcand() and prevcand() functions handle modval, modulus
    and skip by using ZVALUE rather than ZVALUE * and dropping
    the long modulus, etc.  <ernie@neumann.une.edu.au>

    Changed a couple of occurrences of itoq(1) or itoq(0) to &_qone_
    and &_qzero_.  <ernie@neumann.une.edu.au>

    In definition of f_primetest, changed ztolong(q2->num) to ztoi(q2->num)
    so that the sign of count in ptest(n, count, skip) is not lost; and
    ztolong(q3->num) to q3->num so that skip can be any integer.
    <ernie@neumann.une.edu.au>

    In zprime.c, in definition of small_factor(), adds "&& *tp != 1" to
    the exit condition in the for loop so that searching for a factor
    will continue beyond the table of primes, as required for e.g.
    factor(2^59 - 1).  <ernie@neumann.une.edu.au>

    Changed zprimetest() so that skip in ptest(n, count, skip)
    determines the way bases for the tests are selected.  Neg values of
    n are treated differently.   When considering factorization,
    primeness, etc. one is concerned with equivalence classes which for
    the rational integers are {0}, {-1, 1}, {-2, 2}, etc.  To refer to
    an equivalence class users may use any of its elements but when
    returning a value for a factor the computer normally gives the
    non-negative member.  The same sort of thing happens with integers
    modulo an integer, with fractions, etc., etc.  E.g. users may refer
    to 3/4 as 6/8 or 9/12, etc.  A simple summary of the way negative n
    is treated is "the sign is ignored". E.g. isprime(-97) and
    nextprime(-97) now return the same as isprime(97) and nextprime(97).
    <ernie@neumann.une.edu.au>


Following is the change from calc version 2.10.1t11 to 2.10.1t19:

    Added many more regression tests to lib/regress.cal.  Some
    due to <ernie@neumann.une.edu.au>.

    Added many help files, most due to <ernie@neumann.une.edu.au>.

    Fixed exp() and ln() so that when they return a complex value with a
    zero imaginary component, isreal() is true.  <ernie@neumann.une.edu.au>

    Fixed cast problem in byteswap.c.  <ernie@neumann.une.edu.au>

    Fixed memory leak problem where repeated assignments did not
    free the previous value.  <ernie@neumann.une.edu.au>

    Complex number ordering/comparison has been changed such that:

	a < b implies a + c < b + c
	a < b and c > 0 implies a * c < b * c
	a < b implies -a > -b

    To achieve a "natural" partial ordering of the complex numbers
    with the above properties, cmp(a,b) for real or complex numbers
    may be considered as follows:

	cmp(a,b) = sgn(re(a) - re(b)) + sgn(im(a) - im(b)) * 1i

    The cmp help file has been uptdated.

    Change HASH type to QCKHASH.  The HASH type is a name better suited
    for the upcoming one-way hash interface.

    Added the CONFIG type; a structure containing all of the configuration
    values under the control of config().  Added V_CONFIG data type.
    The call config("all") returns a V_CONFIG.  One may now save/restore
    the configuration state as follows:

	x = config("all")
	...
	config("all",x)

    Added two configuration aliases, "oldstd" (for old backward compatible
    standard configuration) and "newstd" (for new style configuration).
    One may set the historic configuration state by:

	config("all", "oldstd")

    One may use what some people consider to be a better but not backward
    compatible configuration state by:

	config("all", "newstd")

    Renamed config.h (configuration file built during the make) to conf.h.
    Added a new config.h to contain info on thw V_CONFIG type.

    Fixed some ANSI C compile warnings.

    The show config output is not indented by only one tab, unless
    config("tab",0) in which case it is not indented.

    The order of show config has been changed to reflect the config
    type values.

    Changed declaration of sys_errlst in func.c to be char *.

    Added quo(x,y,rnd) and mod(x,y,rnd) to give function interfaces
    to // and % with rounding mode arguments.  Extended these functions
    to work for list-values, complex numbers and matrices.
    <ernie@neumann.une.edu.au>

    For integer x, cfsim(x,8) returns 0.  <ernie@neumann.une.edu.au>

    Fixed config("leadzero").  <ernie@neumann.une.edu.au>

    Set config("cfsim",8) by default (in "oldstd").  Setup initial idea for
    config("all", "newstd") to be the default with the following changes:

	display		10
	epsilon		1e-10
	quo		0
	outround        24
	leadzero	1
	fullzero	1
	prompt		"; "		(allows full line cut/paste)
	more		";; "		(allows full line cut/paste)

    The "newstd" is a (hopefully) more perferred configuration than the
    historic default.

    The fposval.h file defines DEV_BITS and INODE_BITS giving the
    bit size of the st_dev and st_ino stat elements.  Also added
    SWAP_HALF_IN_DEV and SWAP_HALF_IN_STSIZE.

    Added sec(), csc(), cot(), sech(), csch(), coth(), asec(), acsc(),
    acot(), asech(), acsch() and acoth() builtins. <ernie@neumann.une.edu.au>

    The initmasks() call is no longer needed.  The bitmask[] array
    is a compiled into zmath.c directly.

    Added isconfig(), ishash(), isrand() and israndom() builtins to
    test is something is a configuration state, hash state, RAND
    state or RANDOM state.

    The lib/cryrand.cal library now no longer keeps the Blum prime
    factors used to formt he Blum modulus.  The default modulus has
    been expanded to 1062 bits product of two Blum primes.

    The function hash_init() is called to initialize the hash function
    interface.

    Misc calc man page fixes and new command line updates.

    Fixed bug related to srand(1).

    Cleaned up some warning messages.

    All calls to math_error() now have a /*NOTREACHED*/ comment after
    them.  This allows lint and compiler flow progs to note the jumpjmp
    nature of math_error().  Unfortunately some due to some systems
    not dealing with /*NOTREACHED*/ comments correctly, calls of the form:

	if (foo)
		math_error("bar");

    must be turned into:

	if (foo) {
		math_error("bar");
		/*NOTREACHED*/
	}

    The ploy() function can take a list of coefficients.  See
    the help/poly file.  Added poly.c.  <ernie@neumann.une.edu.au>

    Fixes and performance improvemtns to det().  <ernie@neumann.une.edu.au>

    Renamed atoq() and atoz() to str2q() and str2z() to avoid conflicts
    with libc function names.

    Fixed use of ${NROFF_ARG} when ${CATDIR} and ${NROFF} are set.

    Fixed SWAP_HALF_IN_B64 macro use for Big Endian machines without
    long long or with LONGLONG_BITS=0.

    Added error() and iserror() to generate a value of a given error type.
    See help/error for details.  <ernie@neumann.une.edu.au>

    Added singular forms of help files.  For example one can now get
    help for binding, bug, change, errorcode and type.

    The builtin mmin(x, md) has been changed to return the same as
    mod(x, md, 16).  The old mmin(x, md) required md to be a positive
    integer and x to be an integer.  Now md can be any real number; x
    can be real, complex, or a matrix or list with real elements, etc.
    <ernie@neumann.une.edu.au>

    The builtin avg(x_1, x_2, ...) has been changed to accept list-valued
    arguments:  a list x_i contributes its elements to the list of items to
    be averaged.  E.g. avg(list(1,2,list(3,4)),5) is treated as if it were
    avg(1,2,3,4,5).  If an error value is encountered in the items to be
    averaged, the first such value is returned.  If the number of items to be
    averaged is zero, the null value is returned.  <ernie@neumann.une.edu.au>

    The builtin hmean(x_1, x_2, ...) has been changed to admit types
    other than real for x_1, x_2, ...; list arguments are treated in
    the same way as in avg().  <ernie@neumann.une.edu.au>

    The builtin eval(str) has been changed so that when str has a
    syntax error, instead of call to math_error(), an error value is
    returned.  <ernie@neumann.une.edu.au>

    The old frem(x,y) builtin returned the wrong value when y was a power of
    2 greater than 2, e.g. f(8,4) is returned as 4 when its value should be 2.
    This has been fixed by a small change to the definition of zfacrem().
    Calc used to accept with no warning or error message, gcdrem(0,2) or
    generally gcdrem(0,y) for any y with abs(y) > 1, but then went into an
    infinite loop.  This has been fixed by never calling zfacrem() with zero x.
    Both frem(x,y) and gcdrem(x,y) now reject y = -1, 0 or 1 as errors.  For
    nonzero x, and y == -1 or 1, defining frem(x,y) and gcdrem(x,y) to equal
    abs(x) is almost as natural as defining x^0 to be 1.  Similarly, if x is
    not zero then gcdrem(x,0) == 1.
    <ernie@neumann.une.edu.au>

    Plugged some more memory leaks.

    Fixed bug related randbit(x) skip (where x < 0).

    Added seedrandom.cal to help users use the raw random() interface well.

    Made extensive additions and changes to the rand() and random() generator
    comments in zrand.c.

    Fixed a bug in fposval.c that prevented calc from compiling on systems
    with 16 bit device and/or inodes.  Fixed error messages in fposval.c.

    Fixed bug that would put calc into an infinite loop if it is ran
    with errors in startup files (calc/startup, .calcrc).
    Ha Lam <hl@kuhep5.phsx.ukans.edu>


Following is the change from calc version 2.10.0t13 to 2.10.1t10:

    Added SB8, USB8, SB16, USB16, SB32, USB32 typedefs, determined by
    longbits and declared in longbits.h, to deal with 8, 16 and 32 bit
    signed and unsigned values.

    The longbits.h will define HAVE_B64 with a 64 bit type (long or
    longlong) is available.   If one is, then SB64 abd US64 typedefs
    are declared.

    The U(x) and L(x) macros only used to define 33 to 64 bit signed
    and unsigned constants.  Without HAVE_B64, these macros cannot
    be used.

    Changed the way zmath.h declares types such as HALF and FULL.

    Changed the PRINT typedef.

    The only place where the long long type might be used is in longlong.c
    and if HAVE_LONGLONG, in longbits.h if it is needed.  The only place
    were a long long constant might be used is in longlong.c.  Any
    long long constants, if HAVE_LONGLONG, are hidden under the U(x) and
    L(x) macros on longbits.h.  And of course, if you don't have long long,
    then HAVE_LONGLONG will NOT be defined and long long's will not be used.

    The longlong.h file is no longer directly used by the main calc source.
    It only comes into play when compiling the longbits tool.

    Added config("prompt") to change the default interactive prompt ("> ")
    and config("more") to change the default continuation prompt (">> ").

    Makefile builds align32.h with determines if 32 bit values must always
    be aligned on 32 bit boundaries.

    The CALCBINDINGS file is searched for along the CALCPATH.  The Makefile
    defines the default CALCBINDINGS is "bindings" (or "altbind") which
    is now usualy found in ./lib or ${LIBDIR}.

    Per Ernest Bowen <ernie@neumann.une.edu.au>, an optional third argument
    was added  sqrt() so that in sqrt(x,y,z), y and z have essentially the
    same role as in appr(x,y,z) except that of course what is being
    approximated is the sqrt of x.  Another difference is that two more
    bits of z are used in sqrt: bit 5 gives the option of exact results
    when they exist (the value of y is then ignored) and bit 6 returns
    the nonprincipal root rather than the principal value.

    If commands are given on the command line, leading tabs are not
    printed in output.  Giving a command on the command line implies
    that config("tab",0) was given.

    Pipe processing is enabled by use of -p.  For example:

	echo "print 2^21701-1, 2^23209-1" | calc -p | fizzbin

    In pipe mode, calc does not prompt, does not print leading tabs
    and does not print the initial version header.

    Calc will now form FILE objects for any open file descriptor > 2
    and < MAXFILES.  Calc assumes they are available for reading
    and writing.  For example:

	$ echo "A line of text in the file on descriptor 5" > datafile
	$ calc 5<datafile
	C-style arbitrary precision calculator (version 2.10.1t3)
	[Type "exit" to exit, or "help" for help.]

	> files(5)
		FILE 5 "descriptor[5]" (unknown_mode, pos 0)
	> fgetline(files(5))
		"A line of text in the file on descriptor 5"

    The -m mode flag now controls calc's ability to open files
    and execute programs.  This mode flag is a single digit that
    is processed in a similar way as the octal chmod values:

	0   do not open any file, do not execute progs
	1   do not open any file
	2   do not open files for reading, do not execute progs
	3   do not open files for reading
	4   do not open files for writing, do not execute progs
	5   do not open files for writing
	6   do not execute any program
	7   allow everything (default mode)

    Thus if one wished to run calc from a privledged user, one might
    want to use -m 0 in an effort to make calc more secure.

    The -m flags for reading and writing apply on open.
    Files already open are not effected.  Thus if one wanted to use
    the -m 0 in an effort to make calc more secure, but still be
    able to read and write a specific file, one might do:

	calc -m 0 3<a.file 4>b.file

	NOTE: Files presented to calc in this way are opened in an unknown
	      mode.  Calc will try to read or write them if directed.

    The maximum command line size it MAXCMD (16384) bytes.  Calc objects to
    command lines that are longer.

    The -u flag cause calc to unbuffer stdin and stdout.

    Added more help files.  Improved other help files.

    Removed trailing blanks from files.

    Removed or rewrite the formally gross and disgusting hacks for
    dealing with various sizes and byte sex FILEPOS and off_t types.

    Defined ilog2(x), ilog10(x), ilog(x,y) so that sign of x is ignored,
    e.g. ilog2(x) = ilog2(abs(x)).

    The sixth bit of rnd in config("round", rnd) and config("bround", rnd)
    is used to specify rounding to the given number of significant
    digits or bits rather than places, e.g. round(.00238, 2, 32)
    returns .0023, round(.00238, 2, 56) returns .0024.

Following is the change from calc version 2.9.3t11 to 2.10.0t12:

    The default ${LIBDIR}/bindings CALCBINDINGS uses ^D for editing.
    The alternate CALCBINDINGS ${LIBDIR}/altbind uses ^D for EOF.

    The Makefile CC flag system has been changed.  The new CC flag system
    includes:

	CCMAIN are flags for ${CC} when compiling only files with main()
	CCOPT are flags given to ${CC} for optimization
	CCWARN are flags given to ${CC} for warning message control
	CCMISC are misc flags given to ${CC}

	CNOWARN are all flags given to ${CC} except ${CCWARN} flags
	CFLAGS are all flags given to ${CC}
	ICFLAGS are given to ${CC} for intermediate progs

	LCFLAGS are CC-style flags for ${LINT}
	LDFLAGS are flags given to ${CC} for linking .o files
	ILDFLAGS are given to ${CC} for linking .o's for intermediate progs

	CC is how the the C compiler is invoked

    The syntax error:

	print a[3][[4]]

    used to send calc into a loop printing 'missing expression'.  This
    has been fixed.

    Added config("maxerr") and config("maxerr",val) to control the
    maximum number of errors before a computation is aborted.

    Removed regress.cal test #952 and #953 in case calc's stdout or
    stderr is re-directed to a non-file by some test suite.

    Changed how <stdarg.h>, <varags.h> or simulate stdarg is determined.
    Changed how vsprintf() vs sprintf() is determined.  The args.h file
    is created by Makefile to test which combination works.  Setting
    VARARG and/or HAVE_VSPRINTF in the Makefile will alter these tests
    and direct a specific combination to be used.  Removed have_vs.c,
    std_arg.h and try_stdarg.c.  Added have_stdvs.c and have_varvs.c.

    Added 3rd optional arg to round(), bround(), appr() to specify the type of
    rounding to be used.

    Moved fnvhash.c to quickhash.c.

    Fixed a bug in appr rounding mode when >= 16.

    Added test2600.cal and test2700.cal. They are used by the regress.cal
    to provide a more extensive test suite for some builtin numeric
    functions.

Following is the change from calc version 2.9.3t9.2+ to 2.9.3t10:

    Added many help files for builtin functions and some symbols.
    More help files are needed, see help/todo.

    Removed the calc malloc code.  Calc now uses malloc and free to
    manage storage since these implementations are often written to
    work best for the local system.  Removed CALC_MALLOC code and
    Makefile symbol.  Removed alloc.c.

    Added getenv("name"), putenv("name=val") and putenv("name, "val")
    builts for environment variable support thanks to "Dr." "D.J." Picton
    <dave@aps2.ph.bham.ac.uk>.

    Added system("shell command") builtin to execute shell commands,
    thanks to "Dr." "D.J." Picton <dave@aps2.ph.bham.ac.uk>.

    Added isatty(fd) builtin to determine if fd is attached to a tty
    thanks to "Dr." "D.J." Picton <dave@aps2.ph.bham.ac.uk>.

    Added cmdbuf() builtin to return the command line executed by calc's
    command line args thanks to "Dr." "D.J." Picton <dave@aps2.ph.bham.ac.uk>.

    Added strpos(str1,str2) builtin to determine the first position where
    str2 is found in str1 thanks to "Dr." "D.J." Picton
    <dave@aps2.ph.bham.ac.uk>.

    Fixed bug that caused:

	global a,b,c 		(newline with no semicolon)
        read test.cal

    the read command to not be recognized.

    The show command looks at only the first 4 chars of the argument so
    that:

	show globals
	show global
	show glob

    do the same thing.

    Added show config to print the config values and parameters thanks
    to Ernest Bowen <ernie@neumann.une.edu.au>.

    Added show objtypes to print the defined objects thanks to Ernest Bowen
    <ernie@neumann.une.edu.au>.

    Added more builtin function help files.

    Fixed the 3rd arg usage of the root builtin.

    Expanded the regress.cal regression test suite.

    Fixed -- and ++ with respect to objects and asignment (see the 2300
    series in regress.cal).

    Added isident(m) to determine if m is an identity matrix.

    The append(), insert() and push() builtins can now append between
    1 to 100 values to a list.

    Added reverse() and join() builtins to reverse and join lists
    thanks to Ernest Bowen <ernie@neumann.une.edu.au>.

    Added sort() builtin to sort lists thanks to Ernest Bowen
    <ernie@neumann.une.edu.au>.

    Added head(), segment() and tail() builtins to return the head, middle or
    tail of lists thanks to Ernest Bowen <ernie@neumann.une.edu.au>.

    Added more and fixed some help files.

    The builtin help file is generated by the help makefile.  Thus it will
    reflect the actual calc builtin list instead of the last time someone
    tried to update it correctly.  :-)

    Fixed non-standard void pointer usage.

    Fixed base() bug with regards to the default base.

    Renamed MATH_PROTO() and HIST_PROTO() to PROTO().  Moved PROTO()
    into prototype.h.

    Fixed many function prototypes.  Calc does not declare functions
    as static in one place and extern in another.  Where reasonable
    function prototypes were added.  Several arg mismatch problems
    were fixed.

    Added support for SGI MIPSpro C compiler.

    Changes the order that args are declared to match the order
    of the function.  Some source tools got confused when:
    arg order did not match as in:

	void
	funct(foo,bar)
		int bar;	/* this caused a problem */
		char *foo;	/* even though it should not! */
	{
	}

Following is the change from calc version 2.9.3t8 to 2.9.3t9.2:

    Use of the macro zisleone(z) has been clarified.  The zisleone(z) macro
    tests if z <= 1.  The macro zisabsleone(z) tests of z is 1, 0 or -1.
    Added zislezero(z) macro.  Bugs are related to this confusion have
    been fixed.

    Added zge64b(z) macro to zmath.h.

    Added the macro zgtmaxufull(z) to determine if z will fit into a FULL.
    Added the macro zgtmaxlong(z) to determine if z will fit into a long.
    Added the macro zgtmaxulong(z) to determine if z will fit into a unsigned
    long.

    Added the macro ztoulong(z) to convert an absolute value of a ZVALUE to
    an unsigned long, or to convert the low order bits of a ZVALUE.
    Added the macro ztolong(z) to convert an absolute value of a ZVALUE to
    an long, or to convert the low order bits of a ZVALUE.

    Some non-ANSI C compilers define __STDC__ to be 0, whereas all ANSI
    C compiles define it as non-zero.  Code that depends on ANSI C now
    uses #if defined(__STDC__) && __STDC__ != 0.

    Fixed ptest(a,b) bug where (a mod 2^32) < b.  Previously ptest()
    incorrectly returned 1 in certain cases.

    The second ptest() argument, which is now optional, defaults to 1.
    This ptest(x) is the same as ptest(x,1).

    Added an optional 3rd argument to ptest().  The 3rd arg tells how many
    tests to skip.  Thus ptest(a,10) performs the same probabilistic
    tests as ptest(a,3) and ptest(a,7,3).

    The ptest() builtin by default will determine if a value is divisible
    by a trivial prime.  Thus, ptest(a,0) will only perform a quick trivial
    factor check.  If the test count is < 0, then this trivial factor check
    is omitted.  Thus ptest(a,10) performs the same amount of work as
    ptest(a,3) and ptest(a,-7,3) and the same amount of work as
    ptest(a,-3) and ptest(a,7,3).

    Added nextcand(a[,b[,c]]) and prevcand(a[,b[,c]]) to search for the
    next/previous value v > a (or v < a) that passes ptest(v[,b[,c]]).
    The nextcand() and prevcand() builtins take the same arguments
    as ptest().

    Added nextprime(x) and and prevprime(x) return the next and
    previous primes with respect to x respectively.  As of this
    release, x must be < 2^32.  With one argument, they will return
    an error if x is out of range.  With two arguments, they will
    not generate an error but instead will return y.

    Fixed some memory leaks, particularly those related with pmod().

    Fixed some of the array bounds reference problems in domult().

    Added a hack-a-round fix for the uninitialized memory reference
    problems in zsquare/dosquare.

    The LIBRARY file has been updated to include a note about calling
    zio_init() first.  Also some additional useful macros have been noted.

    The lfactor() function returns -1 when given a negative value.
    It will not search for factors beyond 2^32 or 203280221 primes.
    Performance of lfactor() has been improved.

    Added factor(x,y) to look for the smallest factor < min(sqrt(x),y).

    Added libcalcerr.a for a math_error() routine for the convince of
    progs that make use of libcalc.a.  This routine by default will
    print an message on stderr and exit.  It can also be made to
    longjump instead.  See the file LIBRARY under ERROR HANDING.

    Added isprime() to test if a value is prime.  As of this release,
    isprime() is limited to values < 2^32.  With one argument,
    isprime(x) will return an error if x is out of range.  With
    two arguments, isprime(x,y) will not generate an error but
    instead will return y.

    Added pix(x) to return the number of primes <= x.  As of this
    release, x must be < 2^32.  With one argument, pix(x) will
    return an error if x is out of range.  With two arguments,
    pix(x,y) will not generate an error but instead will return y.

    Fixed the way *.h files are formed.  Each file guards against
    multiple inclusion.

    Fixed numeric I/O on 64 bit systems.  Previously the print and
    constant conversion routines assumed a base of 2^16.

    Added support for 'long long' type.  If the Makefile is setup
    with 'LONGLONG_BITS=', then it will attempt to detect support
    for the 'long long' type.  If the Makefile is setup with
    'LONGLONG_BITS=64', then a 64 bit 'long long' is assumed.
    Currently, only 64 bit 'long long' type is supported.
    Use of 'long long' allows one to double the size of the
    internal base, making a number of computations much faster.
    If the Makefile is setup with 'LONGLONG_BITS=0', then the
    'long long' type will not be used, even if the compiler
    supports it.

    Fixed avg() so that it will correctly handle matrix arguments.

    Fixed btrunc() limit.

    The ord("string") function can now take a string of multiple
    characters.  However it still will only operate on the first
    character.

    Renamed stdarg.h to std_arg.h and endian.h endian_calc.h to
    avoid name conflicts with /usr/include on some systems that
    have make utilities that are too smart for their own good.

    Added additive 55 shuffle generator functions rand(), randbits()
    and its seed function srand().  Calling rand(a,b) produces a
    random value over the open half interval [a,b).  With one arg,
    rand(a) is equivalent to rand(0,a).  Calling rand() produces
    64 random bits and is equivalent to rand(0,2^64).

    Calling randbit(x>0) produces x random bits.  Calling randbit(skip<0)
    skips -skip bits and returns -skip.

    The srand() function will return the current state.  The call
    srand(0) returns the initial state.  Calling srand(x), where
    x > 0 will seed the generator to a different state.  Calling
    srand(mat55) (mat55 is a matrix of integers at least 55 elements long)
    will seed the internal table with the matrix elements mod 2^64.
    Finally calling srand(state) where state is a generator state
    also sets/seeds the generator.

    The cryrand.cal library has been modified to use the builtin
    rand() number generator.  The output of this generator is
    different from pervious versions of this generator because
    the rand() builtin does not match the additive 55 / shuffle
    generators from the old cryrand.cal file.

    Added Makfile support for building BSD/386 releases.

    The cmp() builtin can now compare complex values.

    Added the errno() builtin to return the meaning of errno numbers.

    Added fputc(), fputs(), fgets(), ftell(), fseek() builtins.

    Added fsize() builtin to determine the size of an open file.

    Supports systems where file positions and offsets are longer than 2^32
    byte, longer than long and/or are not a simple type.

    When a file file is printed, the file number is also printed:

	FILE 3 "/etc/motd" (reading, pos 127)

    Added matsum() to sum all numeric values in a matrix.

    The following code now works, thanks to a fix by ernie@neumann.une.edu.au
    (Ernest Bowen):

		mat A[3] = {1, 2, 3};
		A[0] = A;
		print A[0];

    Also thanks to ernie, calc can process compound expressions
    such as 1 ? 2 ? 3 : 4 : 5.

    Also^2 thanks to ernie, the = operator is more general:

		(a = 3) = 4		(same as a = 3; a = 4)
		(a += 3) *= 4		(same as a += 3; a *= 4)
		matfill(B = A, 4)	(same as B = A; matfill(B, 4);)

    Also^3 thanks to ernie, the ++ and -- operators are more general.

		a = 3
		++(b = a)		(a == 3, b == 4)
		++++a			(a == 5)
        	(++a)++ == 6		(a == 7)
		(++a) *= b		(a == 32, b == 4)

    Fixed a bug related to calling epsilon(variable) thanks to ernie.

    Removed trailing whitespace from source and help files.

    Some compilers do not support the const type.  The file have_const.h,
    which is built from have_const.c will determine if we can or should
    use const.  See the Makefile for details.

    Some systems do not have uid_t.  The file have_uid_t.h, which is
    built from have_uid_t.c will determine if we can or should depend
    on uid_t being typefed by the system include files.  See the Makefile
    for details.

    Some systems do not have memcpy(), memset() and strchr().  The
    file have_newstr.h, which is built from have_newstr.c will
    determine if we can or should depend libc providing these
    functions.  See the Makefile for details.

    The Makefile symbol DONT_HAVE_VSPRINTF is now called HAVE_VSPRINTF.
    The file have_vs.h, which is built from have_vs.c will determine if
    we can or should depend libc providing vsprintf().  See the Makefile
    for details.

    Removed UID_T and OLD_BSD symbols from the Makefile.

    A make all of the upper level Makefile will cause the all rule
    of the lib and help subdirs to be made as well.

    Fixed bug where reserved keyword used as symbol name caused a core dump.

Following is the change from calc version 2.9.3t7 to 2.9.3t7:

    WARNING: This patch is an beta test patch by chongo@toad.com
	     (Landon Curt Noll).

    The 'show' command by itself will issue an error message
    that will remind one of the possible show arguments.
    (thanks to Ha S. Lam <hl@kuhep4.phsx.ukans.edu>)

    Fixed an ANSI-C related problem with the use of stringindex()
    by the show command.  ANSI-C interprets "bar\0foo..." as if
    it were "bar\017oo...".

    Added a cd command to change the current directory.
    (thanks to Ha S. Lam <hl@kuhep4.phsx.ukans.edu>)

    Calc will not output the initial version string, startup
    message and command prompt if stdin is not a tty.  Thus
    the shell command:

	echo "fact(100)" | calc

    only prints the result.  (thanks to Ha S. Lam <hl@kuhep4.phsx.ukans.edu>)

    The zmath.h macro zisbig() macro was replaced with zlt16b(),
    zge24b(), zge31b(), zge32b() and zgtmaxfull() which are
    independent of word size.

    The 'too large' limit for factorial operations (e.g., fact, pfact,
    lcmfact, perm and comb) is now 2^24.  Previously it depended on the
    word size which in the case of 64 bit systems was way too large.

    The 'too large' limit for exponentiation, bit position (isset,
    digit, ), matrix operations (size, index, creation), scaling,
    shifting, rounding and computing a Fibonacci number is 2^31.
    For example, one cannot raise a number by a power >= 2^31.
    One cannot test for a bit position >= 2^31.  One cannot round
    a value to 2^31 decimal digit places.  One cannot compute
    the Fibonacci number F(2^31).

    Andy Fingerhut <jaf@dworkin.wustl.edu> (thanks!) supplied a fix to
    a subtle bug in the code generation routines.  The basic problem was
    that addop() is sometimes used to add a label to the opcode table
    of a function.  The addop() function did some optimization tricks,
    and if one of these labels happens to be an opcode that triggers
    optimization, incorrect opcodes were generated.

    Added utoz(), ztou() to zmath.c, and utoq(), qtou() to qmath.c
    in preparation for 2.9.3t9 mods.

Following is the change from calc version 2.9.2 to 2.9.3t7:

    WARNING: This patch is an beta test patch by chongo@toad.com
	     (Landon Curt Noll).

    Calc can now compile on OSF/1, SGI and IBM RS6000 systems.

    A number of systems that have both <varargs.h> and <stdarg.h> do
    not correctly implement both types.  On some System V, MIPS and DEC
    systems, vsprintf() and <stdarg.h> do not mix.  While calc will
    pass the regression test, use of undefined variables will cause
    problems.  The Makefile has been modified to look for this problem
    and work around it.

    Added randmprime.cal which find a prime of the form h*2^n-1 >= 2^x
    for some given x.  The initial search points for 'h' and 'n'
    are selected by a cryptographic pseudo-random generator.

    The library script nextprim.cal is now a link to nextprime.cal.
    The lib/Makefile will take care of this link and install.

    The show command now takes singular forms.  For example, the
    command 'show builtin' does the same as 'show builtins'.  This
    allows show to match the historic singular names used in
    the help system.

    Synced 'show builtin' output with 'help builtin' output.

    Fixed the ilog2() builtin.  Previously ilog2(2^-20) returned
    -21 instead of -20.

    The internal function qprecision() has been fixed.  The changes
    ensure that for any e for which 0 < e <= 1:

	1/4 < sup(abs(appr(x,e) - x))/e  <= 1/2.

    Here 'sup' denotes the supremum or least upper bound over values of x.
    Previousld calc did: 1/4 <= sup(abs(appr(x,e) - x))/e  < 1.

    Certain 64 bit processors such as the Alpha are now supported.

    Added -once to the READ command.  The command:

	read -once filename

    like the regular READ expect that it will ignore filename if
    is has been previously read.

    Improved the makefile.  One now can select the compiler type.  The
    make dependency lines are now simple foo.o: bar.h lines.  While
    this makes for a longer list, it is easier to maintain and will
    make future Makefile patches smaller.  Added special options for
    gcc version 1 & 2, and for cc on RS6000 systems.

    Calc compiles cleanly under the watchful eye of gcc version 2.4.5
    with the exception of warnings about 'aggregate has a partly
    bracketed initializer'.  (gcc v2 should allow you to disable
    this type of warning with using -Wall)

    Fixed a longjmp bug that clobbered a local variable in main().

    Fixed a number of cases where local variables or malloced storage was
    being used before being set.

    Fixed a number of fence post errors resulting in reads or writes
    just outside of malloced storage.

    A certain parallel processor optimizer would give up on
    code in cases where math_error() was called.  The obscure
    work-a-rounds involved initializing or making static, certain
    local variables.

    The cryrand.cal library has been improved.  Due to the way
    the initial quadratic residues are selected, the random numbers
    produced differ from previous versions.

    The printing of a leading '~' on rounded values is now a config
    option.  By default, tilde is still printed.  See help/config for
    details.

    The builtin function base() may be used to set the output mode or
    base.  Calling base(16) is a convenient shorthand for typing
    config("mode","hex").  See help/builtin.

    The printing of a leading tab is now a config option.  This does not
    alter the format of functions such as print or printf.  By default,
    a tab is printed.  See help/config for details.

    The value atan2(0,0) now returns 0 value in conformance with
    the 4.3BSD ANSI/IEEE 754-1985 math library.

    For all values of x, x^0 yields 1.  The major change here is
    that 0^0 yields 1 instead of an error.

    Fixed gcd() bug that caused gcd(2,3,1/2) to ignore the 1/2 arg.

    Fixed ltol() rounding so that exact results are returned, similar
    to the way sqrt() and hypot() round, when they exist.

    Fixed a bug involving ilog2().

    Fixed quomod(a,b,c,d) to give correct value for d when a is between
    0 and -b.

    Fixed hmean() to perform the necessary multiplication by the number of
    arguments.

    The file help/full is now being built.

    The man page is not installed by default.  One may install either
    the man page source or the cat (formatted man) page.  See the
    Makefile for details.

    Added a quit binding.  The file lib/bindings2 shows how this new
    binding may be used.

    One can now do a 'make check' to run the calc regression test
    within in the source tree.

    The regression test code is now more extensive.

    Updated the help/todo list.  A BUGS file was added.  Volunteers are
    welcome to send in patches!

Following is the change from calc version 2.9.1 to 2.9.1:

    Fixed floor() for values -1 < x < 0.

    Fixed ceil() for values -1 < x < 0.

    Fixed frac() for values < 0 so that int(x) + frac(x) == x.

    Fixed wild fetch bug in zdiv, zquo and zmod code.

    Fixed bug which caused regression test #719 to fail on some machines.

    Added more regression test code.

Following is the change from calc version 2.9.0 to 2.9.0:

    A major bug was fixed in subtracting two numbers when the first
    number was zero.  The problem caused wrong answers and core dumps.

Following is a list of visible changes to calc from version 1.27.0 to 2.8.0:

    Full prototypes have been provided for all C functions, and are used
    if calc is compiled with an ANSI compiler.

    Newly defined variables are now initialized to the value of zero instead
    of to the null value.  The elements of new objects are also initialized
    to the value of zero instead of null.

    The gcd, lcm, and ismult functions now work for fractional values.

    A major bug in the // division for fractions with a negative divisor
    was fixed.

    A major bug in the calculation of ln for small values was fixed.

    A major bug in the calculation of the ln and power functions for complex
    numbers was fixed.

    A major lack of precision for sin and tan for small values was fixed.

    A major lack of precision for complex square roots was fixed.

    The "static" keyword has been implemented for variables.  So permanent
    variables can be defined to have either file scope or function scope.

    Initialization of variables during their declaration are now allowed.
    This is most convenient for the initialization of static variables.

    The matrix definition statement can now be used within a declaration
    statement, to immediately define a variable as a matrix.

    Initializations of the elements of matrices are now allowed.  One-
    dimensional matrices may have implicit bounds when initialization is
    used.

    The obj definition statement can now be used within a declaration
    statement, to immediately define a variable as an object.

    Object definitions can be repeated as long as they are exactly the same
    as the previous definition.  This allows the rereading of files which
    happen to define objects.

    The integer, rational, and complex routines have been made into a
    'libcalc.a' library so that they can be used in other programs besides
    the calculator.  The "math.h" include file has been split into three
    include files: "zmath.h", "qmath.h", and "cmath.h".

Following is a list of visible changes to calc from version 1.26.4 to 1.26.4:

    Added an assoc function to return a new type of value called an
    association.  Such values are indexed by one or more arbitrary values.
    They are stored in a hash table for quick access.

    Added a hash() function which accepts one or more values and returns
    a quickly calculated small non-negative hash value for those values.

Following is a list of visible changes to calc from version 1.26.2 to 1.26.4:

    Misc fixes to Makefiles.

    Misc lint fixes.

    Misc portability fixes.

    Misc typo and working fixes to comments, help files and the man page.

Following is a list of visible changes to calc from version 1.24.7 to 1.26.1:

    There is a new emacs-like command line editing and edit history
    feature.  The old history mechanism has been removed.  The key
    bindings for the new editing commands are slightly configurable
    since they are read in from an initialization file.  This file is
    usually called /usr/lib/calc/bindings, but can be changed by the
    CALCBINDINGS environment variable.  All editing code is
    self-contained in the new files hist.c and hist.h, which can be
    easily extracted and used in other programs.

    Two new library files have been added: chrem.cal and cryrand.cal.
    The first of these solves the chinese remainder problem for a set
    of modulos and remainders.  The second of these implements several
    very good random number generators for large numbers.

    A small bug which allowed division by zero was fixed.

    A major bug in the mattrans function was fixed.

    A major bug in the acos function for negative arguments was fixed.

    A major bug in the strprintf function when objects were being printed
    was fixed.

    A small bug in the library file regress.cal was fixed.