Yet another CL impementation of rfc1951 deflate decompression (optionally with rfc1950 zlib or rfc1952 gzip wrappers), with support for reading from foreign pointers (for use with mmap and similar, etc), and from CL octet vectors and streams.
Still somewhat WIP, but approaching usability.
Performance for vectors/pointers is somewhere between FFI to libz and chipz, still needs some low-level optimization of copy routines and checksums. Stream API is very slow, and may be replaced at some point.
API isn't completely stable yet, needs some actual use to figure out the details.
decompress-vector (compressed &key (format :zlib) output
;; pass a (simple-array (unsigned-byte 8) (*)) (3bz:decompress-vector (alexandria:read-file-into-byte-vector "foo.gz") :format :gzip) ;; accepts :deflate, :zlib, :gzip ;; -> #(....) 1234 ;; get back decompressed data and size as 2 values
If decompressed size is known, you can save some copies by passing in
(simple-array (unsigned-byte 8) (*)) vector with
Allows input and output in multiple pieces, as well as input from vectors, FFI pointers, or streams (streams are currently very slow though).
step 1: make a decompression state:
:output-buffer octet-vector to provide initial
step 2: make an input context
pass source of coresponding type, + optional
to specify valid region within source. For FFI pointers, use
(with-octet-pointer (octet-pointer ffi-pointer size) ...) to wrap a
raw pointer + size into
octet-pointer to pass to
make-octet-pointer-context. (If you need indefinite scope pointers,
file an issue so support that can be added to API.)
step 3: decompress
(decompress context state)returns current offset into output buffer.
step 4: check decompression state
(finished state), you are done.
(input-underrun state), you need to supply more input by creating a new input context, and call
decompressagain with new context.
(output-overflow state), you need to provide a new output buffer with
(replace-output-buffer state new-buffer)and call
Streams API is currently very slow, and will probably be rewritten at some point.
Output in small pieces is slow:
deflateneeds to maintain a 32kb history window, so every time
output buffer overflows, it copies 32kb of history. If output buffer is smaller than 32kb, that means more history is copied than total output, so avoid small output buffers.
Ideally, if output size is known in advance and output buffer never
overflows, there is no extra copies for history window.
If output size isn't known, if you start with something like `(min
input-size 32768)` and double (or at least scale by 1.x) on each overflow, there will only be O(N) bytes copied to combine output buffers, and O(lgN) byytes copied for history.