The stuff that I have written so far is more of a logic manual than a presentation of ideas.
The main idea is to define a landmark that is a property of a string of bits that is suitably rare. Each file is scanned for landmarks and a checksum of the data around each landmark is made and included in a new file along with the address of where it was found. The new file is then sorted by checksum and then landmark sites with the same checksum are compared by reading the original files.
The main difficulty is finding a suitable landmark definition. I invented a stream hash function that is fairly efficient to compute. It amounts to a convolution of the bit stream of the searched file with a string of about 500 bits. I look for strings of about 11 ones in the hash and call this a landmark. Bits surrounding this string of ones are used as the checksum. This has the convenient property that if two file portions are alike except for different offsets then the transclusion will be found even if the offset difference is not a multiple of one byte.
This scheme is likely to find transclusions of more than 256 characters as such a stream is likely to have a landmark. Shorter transclusions are less likely to be found.
One piece of hair is removal of redundant reports resulting from multiple landmarks in one transclusion.
Local Stationary Bit Streams Functions
This is the theory of a string hash function. It is a function on a long string of bits. Its value is a few hundred bits longer than the argument. It is local, stationary and fairly fast to compute. We use “string” here to refer to arrays of bits indexed by the integers. If b is such a string then bj is bit j of that string. The strings are infinitely long in both directions but the bits are zero outside some finite range. Note that we picture bits with a larger index on the left! e.g. b2 b1 b0 b–1 b–2 ... —this is the classic order for polynomial coefficients. Both the argument and value of string hash is a string. b ≪ n is the string b shifted n bits to the left. (b≪n)i= bi–n for any integers i and n. We also write “b(x)” to represent the polynomial Σbixi over the field GF(2).
A string function s is stationary iff s(b≪1) = s(b)≪1 for all b. In this case s(b≪n) = s(b)≪n for all b and n.
A function s is local when there are integers a and b such that a<b and s(b)i is determined by bj only when i+a < j < i+b. This means that such functions depend on only bits in the neighborhood. More precisely L is local if there are integers a and b such that a<b and for any strings s and t (for all i, if (for all j if (i+a < j < i+b) then sj = tj) then L(s)i = L(t)i).
We want a stationary function S so that properties of a string that are preserved under shifting will remain visible in the yield of S if they were visible before. We want S to be local so that equality of sub-strings of modest length will lead to equality of the yield of S. We will look for landmarks in the yield such as strings of ones. No particular pattern can be presumed to have a useful density in the original data—zeros, for instance, are excluded from ascii text. Such patterns should have known densities in the randomized data that comes from S.
Our particular string hash function is based on a polynomial h over GF(2) of degree P≈50. We identify a bit string b with the polynomial h(x) = Σbixi summed over non-negative i.
When polynomials are used to checksum transmission blocks a checksum is computed and transmitted that would have caused the block to be a multiple of the polynomial had it been exclusive-ored into the right end of the block. In this case the checksum is a function of all of the bits in the block and is thus not local.Our string hash function S is a convolution function. S(b)i = Σbi–jpj. (In this context Σ means the one bit exclusive-or: ⊕) If h(x) = x64+x61+1 then p may be recursively defined as pj = 0 for j<0 or j≥B and p0 = 1 and pj = pj–64 ⊕ pj–61 for 0 < j < B. (B is some constant ≈ 500). Alternatively p may be defined by the relations xB + r(x) = p(x)h(x) and deg(r)≤deg(h). r(x) is the remainder after dividing xB by h(x) and p(x) = (xB + r(x))/h(x). It may be seen that S is stationary and local for the range a = –B, b = 0.
The “polynomial” associated with p for infinite B is the reciprocal of h in the sense that the terms of the product of p and h are the coefficients of the polynomial 1 despite the fact there are infinitely many terms in p. This is like 1/(1–x) = 1+x–1+x–2+x–3... You may have seen the old fallacious proof that 1+2+4+8... = –1. That proof is valid in this funny world.The desired convolution is m(x)p(x) where mi are the bits of the argument to S. If there are N bits in the argument then the first bit is mN–1 and the last bit is m0. We will process the bits from first to last.
m(x)p(x) = m(x)[(xB + r(x))/h(x)] = [m(x)(xB + r(x))]/h(x)
We first describe a conceptual method that does not have the factor B in its cost. Then we show how to do it several bits at a time instead of one. The first conceptual method of computation proceeds as follows. The string argument is delivered one bit at a time, largest index first. The algorithm does not know the value of that index. S being stationary, we can compute the answer as soon as the data begin to arrive. By the first and last bits of the string argument we refer to some range of the infinite input outside of which is all zero. We buy three left infinite accumulators m, s and q. m accumulates the argument, s will accumulate m(x)(xB + r(x)) and q will accumulate the quotient [m(x)(xB + r(x))]/h(x). We prepare the initial answer as if the input string were empty (all zero) by setting m, s and q to all zeros.
When ever a bit of m arrives:
The loop invariant for the above is:
s(x) = (xB + r(x))m(x) – q(x)h(x)x(B–P) and deg(s)<B.
Since s is driven to zero we have q(x) = (xB + r(x))m(x)x(B–P)/(h(x)x(B–P)) upon exit from the loop.
The extra factor x(B–P) in m(x)x(B–P) above arose when we fed B–P more zero bits into the algorithm after the last bit of the argument arrived.
Steps 1 and 2 constitute the multiplication (xB + r(x))m(x) and steps 3 and 4 perform the division (xB + r(x))m(x)/h(x).
We now have:
q(x) = (xB + r(x))m(x)/h(x) = m(x)p(x)
which is the desired answer.
....We reformulate the above in notation without mutation such as shifting and “exclusive-or”ing.
To accelerate the above computations we consider the data flow. The only parts of the computation that are proportional to B are the shifts of s, m and q. m is entirely unneeded—it was computed only to make the the loop invariant testable and thus make the computation easier to debug. The bits of q are produced as the answer and are unused again in this computation. Only the first B bits of s are used—indeed the rest are all zero. After sP is produced it is not used again, aside for shifting until it becomes sB–P. This is thus a fixed length FIFO queue and may be implemented without shifting. Alternatively we can dispense with the queue and perform steps 3 and 4 sooner if we can get an early edition of the input bits so as to perform the xB element of step 2. The latter is the only thing delaying the computation in steps 3 and 4. In either case there remain no factors of B in the computation.
Substantial further savings are possible by processing several bits of input at a time. For simplicity we will describe the 8 bit strategy. As the algorithm processes 8 bits, its storage, s, undergoes a linear transformations. s0 thru sP–1 are transformed by being shifted 8 to the left, becoming s0 thru sP+7, and having a P+8 bit string depending on the next 8 input bits, exclusive-ored into them. The P+8 bits may be found in a table of pre-computed constants indexed by the next input byte.
Bits sB–P thru sB–1 undergo a similar linear transformation. Any time after the input bits have been contributed at position xB eight iterations of step 4 may be done in an 8 bit left shift followed by an exclusive-or of P bits into sB–P thru sB–1. These P bits are taken from another constant table indexed by the old bits sB–8 thru sB–1. Yet another table indexed the same yields the 8 answer bits.
There is a simple but unsatisfying proof of the above. The output is a linear function of the input in the sense that input and output are in a vector space over GF(2). The answer is correct when the message has a single one bit. (you can work this out in your head.) It is therefore correct for any argument with a single bit since S is stationary. The messages of a single bit form a basis for the vector space. The algorithm above computes a linear function. Linear functions that agree for each member of a basis agree everywhere. Q.E.D.
This yields S32 = 131730944.
“options” is a string of letters, each suppressing a different form of diagnostic output to “terminal”. The more characters the less output. Without the -s option, “-s CL” is assumed. “directory_name” is traversed recursively.
Classes of diagnostic and amusement output:
“F” names each file as it is first read, and directory as it is entered.
“R” reports how many characters in each file and prints “b” for each block.
“L” reports details about where matches were found.
“V” is for vernier faults, unfortunate and rare.
There are two phases to the program and it is currently coded to finish one phase before starting the next. There is some flexibility to merge the phases and achieve trade-offs for space and against execution time.
Phase one reads the accessible files, eliminating duplication arising from links. The large output of this phase is a list of marks. Each mark is represented by an instance of the sx structure that comprises a 32 bit signature and a 32 bit mark locator. This is accumulated into a large array LM. A small output of phase one is a coded map which can translate a mark locator into a file name and a block within file.
At the end of phase one we sort the marks by signature and carry along the locations. Then in phase two we pass over the sorted list. When we come to a pair of adjacent equal marks we call locate and read the files where the marks were found. We extend the equality span as far as possible. Then, so as not to process this span again as triggered by other marks in the same span, we remove the marks from the sorted list associated with one of the two files. We do this by again passing over that file portion and for each mark find we do a binary search in the sorted mark list and “remove” the mark from the list. We cannot merely zap the entry for that would break the binary search code. Instead we disable the entry by changing its location field to –5. We keep such invalidated mark entries at the end of the run of entries with equal signatures.
We imagine the accessible files to be concatenated in directory order and their blocks numbered in that concatenation. This means that we must provide an auxiliary memory of which file holds block numbers within a given range. This memory will be large but not nearly so large as the landmark memory. I propose that we form a chronological and linear record of each directory and file visited on the first pass. If we merely record the size of each file then the file may be efficiently located. Each directory entry will record the total size of the files within and point to the previous directory entry at the same level. This allows rapid location of the file given the block locator. It is also used in conjunction with the opendir call to reconstruct the file name that is necessary to open the file to confirm the transclusion. This form of file memory is suitable to writing into a file during the first pass and landmark sort. The pointers between sibling directories are relative to the file memory.
We build a large FILO structure FM to hold our file memory. We build it in one order and consult it repeatedly in the other order. Abstractly FM is a stack of integers—it grows like a stack but does not shrink as it is used backwards. As we process a file yield of readdir we append something to FM. If the yield indicated a file we append the file’s block count. If we cannot process the file, perhaps because access was denied, we append zero. If the yield indicates a directory we note the FM cursor and the current global block count. We then process the entries in the new directory. After we exhaust that directory we append the size of the data appended to FM for this directory. We also append the total block count. We then a append –1 as a distinctive marker indicating a directory. This allows to skip this part of FM as we read FM. Since FM is large and most of its integers are small we compress them by techniques best understood by reading the code.
There is an infelicity in this design. It takes special logic to discover a string that is repeated in the same file block. We omit that logic for now.
The routines w_log and r_log store and retrieve a list if 32 bit integers. They compact leading zeros in these integers. r_log retrieves successive integers in the reverse order that w_log stored them. log_crsr is a variable that can be read and written to provide random access to the integer store. w_log and r_log are used to record an excessively clever data base called name-memory and located by log_crsr that is used to convert a global block number recorded in that mark data base back into a file name. It is very compact but a bit delicate. The delicacy is that it may fail if a directory in the path for the file name is changed between forming the memory (first phase) and consulting it (second phase). I don’t know a way overcome this delicacy short of vast increase in the size of the main mark data base. It could be made more self diagnostic by including some sort of name hash which could be used to detect most of these changes. This would double the size but not ultimately improve the situation for there would be nothing safe to do to find the real old file. The logic of the name-memory assumes constant directories. For each item in a directory (file or sub-directory) it remembers the number of global blocks traversed while processing that entry in phase one. When phase one finishes a directory It assumes that the order and number of the items in a directory as presented in the SVR4 system call readdir is the same in phase one and two. When this fails transclusions are lost and copious diagnostic messages arise unless suppressed. No other ill effects should result.
It calls a deeper routine with the bit offset within file and 32 bits of relevant bits of stream hash computed from array grunge, a small circular array or stream hash bits. These hash bits are aligned with the landmark.
Which deeper routine it calls, Rlm, sm or forget_sam, is determined by the current value of the file global variable pointer rlm.
X8 is the sole interface to the stream hash function. Its output depends exactly on its current and previous 57 (=(B–P)/8) inputs. zreg() means the same as {int j; for(j=0; j<58; ++j) X8(0);}—
it merely clears the memory.
core(much, who) slightly improves malloc by testing for exhaustion. “who” is merely for identifying the failing request.
GBlc is a global symbol that counts all blocks of all files. It supports locating stuff in files for later location and perusal.
IGBl is a copy of GBlc's value upon opening of the current file.
At the bottom, fread at the zx label in routine gaw in file fl.c reads the files in both phases.
The preprocessor symbol “C” is for “const” and I have tried to put it just about anywhere it could go. This is in line with my suspicion that the default should be with instead of without.
bgary is an epiphenomenon, merely for debugging.
bi is the index thru the file block buffer, viewed as a word array.
X8 is the function that does the stationary hash function, one character at a time.
in is a file opening variable local to fl.c. It gets its value as routine fl is called. It is for the current subject file.
The main code transformation for a little endian machine is end-for-end swapping of the bits in fields that represent file data. This consists largely of interchanging C’s ≪ and ≫ operators. The two preprocessor symbols SA (towards Small Addresses) and LA (towards Large Addresses) map to ≪ and ≫ respectively for big endian machines and vice-versa for little endian ones. When the preprocessor symbol L is 1, the high bit of a byte is considered to be adjacent to the low bit of the following byte from the file. It is 1 for little endian machines and 0 for others. If preprocessor symbol Ebug is set then ‘L must be wrong for the machine’. This funny mode allows one to check out code pretty well on a machine of the wrong endianess. Ebug entails some run-time cost.
The function X8 is an 8 bit interface to the hash magic. It suffices to bit invert both the argument and value of X8. We go a bit farther for efficiency. The magic constant tables, head and tail, are special. Their entries are reversed just as they are computed. The internal variables a, b, c, d, y and z, are all reversed. The array index is also reversed. This seemed simpler than reworking the synthetic division logic. rBy reverses bytes and rW reverses words. The code X8pr tests the X8 code and the code in Fast.c. Its yield depends on L but not the endianess of the machine! It can be debugged on a machine with the wrong end.