notes/svndiff

This file describes the svndiff format used by the Subversion code.
Its design borrows many ideas from the vdelta and vcdiff encoding
formats from AT&T Research Labs, but it is much simpler and thus a
little less compact.

From the point of view of svndiff, a delta is a sequence of windows,
each containing a list of instructions for reconstructing a contiguous
section of the target using a contiguous section of the source as a
reference.  The section of the target being reconstructed is called
the "target view"; the section of the source being referenced is
called the "source view."  Source views must not slide backwards from
one window to the next; this allows svndiffs to be applied using a
single pass through the source file.  Instructions in a window direct
copies to be made into the target view from one of three places: from
the source view, from the portion of the target view which has already
been reconstructed, or from a block of new data encoded inside the
window.

An svndiff document begins with four bytes, "SVN" followed by a zero
byte which represents a version number.  After the header come one or
more windows, until the document ends.  (So the decoder must have
external context indicating when there is no more svndiff data.)

A window is the concatenation of the following:

	The source view offset
	The source view length
	The target view length
	The length of the instructions in bytes
	The length of the new data in bytes
	The window's instructions
	The window's new data (as raw data)

Integers (including the first five items listed above) are encoded
using a variable-length format.  The high bit of each byte is used as
a continuation bit; 1 indicates that there is more data and 0
indicates the final byte.  The other seven bits of each byte are data.
Higher-order bits are encoded before lower-order bits.  As an example,
130 would be encoded as two bytes, 10000001 followed by 00000010.

Instructions are encoded as follows: the two high bits of the first
byte compose an instruction selector, as follows:

	00	Copy from source view
	01	Copy from target view
	10	Copy from new data
	11	invalid

The remaining six bits of the first byte indicate the length of the
copy.  If those six bytes are all zero, then the length is encoded as
an integer immediately following the first byte of the instruction.
If the instruction selector is 00 or 01, then the instruction encoding
continues with an offset encoded as an integer.  If the instruction
selector is 10, then the offset into the new data is implicit; each
copy from the new data is always for "the next <length> bytes" after
the last copy.

A copy from the target view must begin at a location before than the
current position in the target view, but its length may extend past
the current position.  In this case, the target data copied is
repeated, as happens naturally if the copy is performed byte by byte
starting at the beginning.

Following are some example instruction encodings.

	Copy 11 bytes from offset 0 in source view:
	00001011 00000000

	Copy 64 bytes from offset 128 in target view:
	01000000 00100000 10000001 00000000

	Copy the next 63 bytes of new data:
	10111111

Following is a complete example of an svndiff between the source
document "aaaabbbbcccc" and the target document "aaaaccccdddddddd":

	01010011 01010110 01001110 00000000	Header ("SVN\0")

	00000000				Source view offset 0
	00001100				Source view length 12
	00010000				Target view length 16
	00000111				Instruction length 7
	00000001				New data length 1

	00000100 00000000			Source, len 4, offset 0
	00000100 00001000			Source, len 4, offset 8
	10000001				New, len 1
	01000111 00001000			Target, len 7, offset 8

	01100100				The new data: 'd'