git clone 'https://github.com/cosmos72/hyperluminal-mem.git'

(ql:quickload :hyperluminal-mem)



Hyperluminal-mem is a high-performance serialization/deserialization library for Common Lisp.


Hyperluminal-mem is designed and optimized for the following objectives: - speed: serializing and deserializing data can have sustained rates exceeding 1GB/s on a single CPU core. - safety: it can be used on untrusted and possibly malicious data, as for example serialized packets or files received from the internet. - portability: the serialization format is fairly portable. It is independent from the Lisp implementation, and only depends on endianity (small or big endian) and on user's choice between 32 and 64 bit formats. Conversion between small and big endian format is trivial. - ease of use: adding support for user-defined types is usually straightforward.

Latest news, 2nd March 2015

Hyperluminal-mem 0.6.1 is included in the newest Quicklisp distribution. You can now load it with: (ql:quickload "hyperluminal-mem")

News, 24th January 2015

Released version 0.5.2. License change from GPLv3 to LLGPL!

Older versions were bundled together with Hyperluminal-DB in a single GPLv3 package. Hyperluminal-DB is now a separate project, still under GPLv3.

News, 9th February 2014

Released version 0.5.0.

The serialization library is tested, documented and ready to use. It may still contain some rough edges and small bugs.

News, 1st February 2014

The serialization library works and is in BETA status.

The memory-mapped database (built on top of the serialization library) is in the early-implementation stage, not yet ready for general use.

Supported systems

Hyperluminal-mem is currently tested on the following Common Lisp implementations:

Note: on ABCL, memory buffers are implemented using java.nio.ByteBuffer instead of CFFI-SYS raw memory pointers due to currently limited compatibility between ABCL and CFFI/OSICAT libraries. Memory-mapped files are supported, and internally use java.nio.channels.FileChannel.map() instead of OSICAT-POSIX (mmap)

Partially supported systems

Unsupported systems

Other systems

Hyperluminal-mem requires CFFI, OSICAT and STMX libraries to work. While reasonably portable, they exploit features well beyond ANSI Common Lisp and their support for the various Common Lisp implementations varies widely.

For this reason no general guarantees can be given: Hyperluminal-mem may or may not work on other, untested Common Lisp implementations.

Installation and loading

From Quicklisp

Since 2nd March 2015, hyperluminal-mem is available from Quicklisp. The simplest way to obtain it is to first install Quicklisp then run these commands from REPL:

CL-USER> (ql:quickload "hyperluminal-mem")
;; lots of output...
CL-USER> (use-package :hlmem)

If all goes well, this will load Hyperluminal-mem and its dependencies, CFFI, OSICAT and STMX.

Since hyperluminal-mem was added to QuickLisp quite recently (2 March 2015), it may happen that your Quicklisp installation can't find it. In such case, you need to first update your QuickLisp installation as described here - search for “To get updated software” in the page.

Latest version - from GitHub

In case you want to use the “latest and greatest” version directly from the author, in order to get the newest features, improvements, bug fixes, and occasionally new bugs, you need to download it into your Quicklisp local-projects folder. Open a shell and run the commands:

$ cd ~/quicklisp/local-projects
$ git clone git://github.com/cosmos72/hyperluminal-mem.git

then open a REPL and run:

CL-USER> (ql:quickload "hyperluminal-mem")
;; lots of output...
CL-USER> (use-package :hlmem)

If all goes well, this will load Hyperluminal-mem and its dependencies, CFFI, OSICAT and STMX.


In case you get errors:

should print something like

    #<SYSTEM hyperluminal-mem / hyperluminal-mem-20150302-git / quicklisp 2015-03-02>
    #<SYSTEM hyperluminal-mem-test / hyperluminal-mem-20150302-git / quicklisp 2015-03-02>

If it doesn't, you need to update Quicklisp as described here - search for “To get updated software” in the page.

Testing that it works

After loading Hyperluminal-mem for the first time, it is recommended to run the test suite to check that everything works as expected. From the REPL, run:

CL-USER> (ql:quickload "hyperluminal-mem-test")
;; lots of output...

CL-USER> (fiveam:run! 'hyperluminal-mem-test:suite)
;; even more output...
 Did 3364 checks.
    Pass: 3364 (100%)
    Skip: 0 ( 0%)
    Fail: 0 ( 0%)

Note: (ql:quickload "hyperluminal-mem-test") intentionally works only after (ql:quickload "hyperluminal-mem") has completed successfuly.

The test suite should report zero Skip and zero Fail; the number of Pass may vary. You are welcome to report any failure you get while running the test suites, please include in the report: - operating system name and version (example: Debian GNU/Linux x86_64 version 7.0) - Common Lisp implementation and version (example: SBCL, x86_64) - exact output produced by the test suite - any other relevant information

See “Contacts, help, discussion” below for the preferred method to send the report.


Hyperluminal-mem reads and writes serialized data to raw memory, using CFFI foreign pointers - equivalent to C/C++ pointers.

The most direct way to save serialized data to disk, and to load it back, is to open a file then map it to memory with the POSIX mmap() system call provided by OSICAT library.

An alternative, suitable both for files and network sockets, is to allocate a raw memory buffer with (hlmem:malloc-words) then use the POSIX read() and write() calls provided by OSICAT library.

Basic usage

Hyperluminal-mem offers the following Lisp types, constants, macros and functions, also documented in the sources - remember (describe 'some-symbol) at REPL.

For the curious, in practice it is (unsigned-byte 30) on 32-bit systems, (unsigned-byte 61) on 64-bit systems, and so on…

It is actually a simple alias for the function (cffi-sys:%foreign-alloc n-bytes) and it is equivalent to the function void * malloc(size_t n_bytes) found in C/C++ languages.


    (defun malloc (n-bytes)
      (declare (type unsigned-byte n-bytes))

      (the maddress (cffi-sys:%foreign-alloc n-bytes)))

Remember that, as in C/C++, the memory returned by malloc must be deallocated manually: call mfree on it when no longer needed.


    (defun malloc-words (n-words)
      (declare (type mem-size n-words))

      (the maddress #| ...implementation... |# ))


    (defun mfree (ptr)
      (declare (type maddress ptr))

      (cffi-sys:foreign-free ptr))

with-mem-words is an alternative to malloc and malloc-words, useful if you know in advance that the raw memory can be deallocated after BODY finishes. It is a wrapper around the CFFI macro (cffi-sys:with-foreign-pointer (var size &optional size-var) &body body) which performs the same task but counts memory size in bytes, not in words.

Other alternatives to obtain raw memory include at least: * using the functions make-static-vector and static-vector-pointer from STATIC-VECTORS library (remember to call free-static-vector when done) * using memory-mapped files, for example with the function mmap from OSICAT library (remember to call munmap when done)

It is useful to know how large a raw memory block must be in order to write a serialized value into it. It is defined as:

    (defun msize (index value)
      (declare (type t value)
               (type mem-size index))
      (the mem-size (+ index
                       #| ...implementation... |#)))

The argument INDEX is useful to compute the total size of composite values, as for example lists, arrays, hash-tables and objects: the value returned by msize is increased by the value of index, so the following three code snippets are equivalent

    (+ (msize 0 "foo") (msize 0 'bar))

    (let ((index (msize 0 "foo")))
      (msize index 'bar))

    (msize (msize 0 "foo") 'bar)

with the advantage that the second and third versions automatically check for length overflows and can exploit tail-call optimizations.

msize supports the same types as MWRITE below, and can be extended similarly to support arbitrary types, see MSIZE-OBJECT and MWRITE-OBJECT for details.

To use it, you need three things beyond the value to serialize: * a pointer to raw memory, obtained for example with one of malloc-words, malloc or with-mem-words described above. * the offset (in words) where you want to write the serialized value. It must be passed as the index argument * the available length (in words) of the raw memory. It must be passed as the end-index argument

mwrite returns an offset pointing immediately after the serialized value. This allows to easily write consecutive serialized values into the raw memory.

Any kind of raw memory is supported, thus it is also possible to call mwrite on memory-mapped files. This is actually the mechanism that allows Hyperluminal-mem to implement an object store backed by memory-mapped files.

mwrite supports the following standard Lisp types: * integers - both fixnums and bignums * ratios, as for example 2/3 * single-floats, double-floats and complexes * characters - full Unicode range is supported * symbols and keywords * cons cells and their aggregates: lists, alists, plists, trees * vectors and arrays of any rank (i.e. any number of dimensions) * strings * pathnames * hash-tables

It also supports the following types implemented by STMX (requires latest STMX from GitHub): * tcell - simple transactional variable * tstack - transactional first-in last-out stack * tmap and rbmap - sorted maps, both transactional (tmap) or non-transactional (rbmap) * thash-table and ghash-table - hash tables, both transactional (thash-table) and non-transactional (ghash-table)

Finally, it can be easily extended to support arbitrary types, see MWRITE-OBJECT for details.

It returns two values: the value itself, and an offset pointing immediately after the serialized value inside raw memory. This allows to easily read consecutive serialized values from the raw memory.

mread supports the same types as mwrite and it can be extended similarly, see MREAD-OBJECT and MWRITE-OBJECT for details.

Programmers can extend Hyperluminal-mem by defining specialized methods for it, see MWRITE-OBJECT for details.

Programmers can extend Hyperluminal-mem by defining specialized methods for it, see MWRITE-OBJECT for details.

Programmers can extend Hyperluminal-mem by defining specialized methods for msize-object, mwrite-object and mread-object. Such methods are invoked automatically by msize, mwrite and mread when they encounter a user-defined object, i.e. an instance of structure-object or standard-object or their subclasses.

The task of msize-object, mwrite-object and mread-object is relatively straightforward: they are supposed to cycle through the relevant instance slots (or accessors) and recursively call msize, mwrite or mread on each slot.

For example, if a POINT3D class is defined as

    (defclass point3d ()
      ((x :initarg :x :initform 0.0 :accessor point3d-x)
       (y :initarg :y :initform 0.0 :accessor point3d-y)
       (z :initarg :z :initform 0.0 :accessor point3d-z)))

then a reasonable specialization of msize-object is:

    (defmethod msize-object ((p point3d) index)
      (let* ((index-x (msize index   (point3d-x p)))
             (index-y (msize index-x (point3d-y p)))
             (index-z (msize index-y (point3d-z p))))

note how the result of each msize call is passed to the next call - this ensures that the sum of the sizes is computed automatically, and also takes care of signalling an error in case of overflow.

A shorter, slightly automagic alternative is to use the macro msize* which expands to multiple calls of msize and correctly passes around the intermediate index values:

    (defmethod msize-object ((p point3d) index)
      (msize* index (point3d-x p) (point3d-y p) (point3d-z p)))

Similarly, mwrite-object can be specialized as:

    (defmethod mwrite-object ((p point3d) ptr index end-index)
      (let* ((index-x (mwrite (point3d-x p) ptr index   end-index))
             (index-y (mwrite (point3d-y p) ptr index-x end-index))
             (index-z (mwrite (point3d-z p) ptr index-y end-index)))

which uses the same index-passing mechanism to compute the serialized value total size. Again a shorter, slightly automagic alternative is to use the macro mwrite* which expands to multiple calls of mwrite and correctly passes around the intermediate index values:

    (defmethod mwrite-object ((p point3d) ptr index end-index)
      (mwrite* ptr index end-index (point3d-x p) (point3d-y p) (point3d-z p)))

Defining mread-object specialization is slightly more complicated, for two reasons: first, it also needs to instantiate an appropriate object and fill its slots and second, mread and mread-object return multiple values.

The result is a painstaking nest of multiple-value-bind:

    (defmethod mread-object ((type (eql 'point3d)) ptr index end-index &key)
      (multiple-value-bind (x index-x) (mread ptr index end-index)
        (multiple-value-bind (y index-y) (mread ptr index-x end-index)
          (multiple-value-bind (z index-z) (mread ptr index-y end-index)
              (make-instance 'point3d :x x :y y :z z)

The with-mread* macro comes to the rescue, removing most boilerplate code:

    (defmethod mread-object ((type (eql 'point3d)) ptr index end-index &key)
      (with-mread* (x y z new-index) (ptr index end-index)
          (make-instance 'point3d :x x :y y :z z)

An alternative approach to implement msize-object, mread-object and mwrite-object for standard-objects is to take advantage of the generic functions mlist-objects-slots, msize-object-slots, mread-object-slots and mwrite-object-slots. See MWRITE-OBJECT-SLOTS for details.

The mechanism described in MWRITE-OBJECT above is very powerful and general, but sometimes all you need is to serialize/deserialize the slots of a standard-object: in this case it surely feels overcomplicated.

For such purpose, Hyperluminal-mem provides the functions msize-object-slots, mread-object-slots and mwrite-object-slots which automatically obtain the slots of an object (more details below) and call the appropriate function among msize, mread, mwrite on each slot.

This allows programmers to implement msize-object, mread-object and mwrite-object with the following six lines of code:

    (defmethod msize-object ((object point3d) index)
      (msize-object-slots object index))

    (defmethod mwrite-object ((object point3d) ptr index end-index)
      (mwrite-object-slots object ptr index end-index))

    (defmethod mread-object ((type (eql 'point3d) ptr index end-index &key)
      (mread-object-slots (make-instance 'point3d) ptr index end-index))

This simplified approach has some limitations:

  1. it only works on standard-objects, i.e. on classes defined with (defclass …)

  2. it can only serialize/deserialize (some or all) the object slots with plain msize, mread and mwrite, i.e. it is not possible to specify a customized logic to serialize/deserialize the slots.

  3. it must be possible to construct the object with some initial, dummy slot values in order to pass it to mread-object-slots. This function will then set the actual slot values.

  4. by default, all slots are serialized/deserialized. To override this behaviour, programmers can specialize the generic function mlist-object-slots, which must return the list of slots (either slot names or closer-mop:slot-definition) to be serialized/deserialized. The methods on mlist-object-slots have the form:

       (defmethod mlist-object-slots ((object point3d))
         '(x y z))

    i.e. they must be specialized on objects, not on their class names.

  5. MWRITE-MAGIC to be documented…

  6. MREAD-MAGIC to be documented…

  7. (HLMEM-VERSION) is a function that returns the current version of Hyperluminal-mem. The returned value is a list having the form '(major minor patch) as for example '(0 5 2)

Serialization format and ABI

By default, Hyperluminal-mem serialization format and ABI is autodetected to match Lisp idea of CFFI-SYS pointers: * 32 bit when CFFI-SYS pointers are 32 bit, * 64 bit when CFFI-SYS pointers are 64 bit, * and so on…

In other words, mem-word is normally autodetected to match the width of underlying CPU registers (exposed through CFFI-SYS foreign-type :pointer) and +msizeof-word+ is set accordingly.

It is possible to override such autodetection by adding an appropriate entry in the global variable *FEATURES* before compiling and loading Hyperluminal-mem. Doing so disables autodetection and either tells Hyperluminal-mem the desired size of mem-word or, in alternative, the CFFI-SYS type it should use for mem-word.

For example, to force 64 bit (= 8 bytes) file format and ABI, execute the following form before compiling and loading Hyperluminal-mem:

(pushnew :hyperluminal-mem/word-size/8 *features*)

on the other hand, to force 32 bit (= 4 bytes) file format and ABI, execute the form

(pushnew :hyperluminal-mem/word-size/4 *features*)

in both cases, the Hyperluminal-mem internal function (choose-word-type) will recognize the override and define mem-word and +msizeof-word+ to match a CFFI-SYS unsigned integer type having the specified size among the following candidates: * :unsigned-char * :unsigned-short * :unsigned-int * :unsigned-long * :unsigned-long-long

In case it does not find a type with the requested size, it will signal an error.

Forcing the same value that would be autodetected is fine and harmless. Also, the chosen type must be at least 32 bits wide, but there is no upper limit: Hyperluminal-mem is designed to automatically support 64 bits systems, 128 bit systems, and anything else that will exist in the future. It even supports ‘unusual’ configurations where the size of mem-word is not a power of two (ever heard of 36-bit CPUs?).

For the far future (which arrives surprisingly quickly in software) where CFFI-SYS will know about further unsigned integer types, it is also possible to explicitly specify the type to use by executing a form like

(pushnew :hyperluminal-mem/word-type/<SOME-CFFI-SYS-TYPE> *features*)

as for example:

(pushnew :hyperluminal-mem/word-type/unsigned-long-long *features*)

Hyperluminal-mem will honour such override, intern the type name to convert it to a keyword, use it as the definition of mem-word, and derive +msizeof-word+ from it.


As of February 2015, Hyperluminal-mem is being written by Massimiliano Ghilardi and it is considered by the author to be fairly stable, tested and documented. It may still contain some rough edges and minor bugs.

Contacts, help, discussion

As long as the traffic is low enough, GitHub Issues can be used to report test suite failures, bugs, suggestions, general discussion etc.

If the traffic becomes high, more appropriate discussion channels will be set-up.

The author will also try to answer support requests, but gives no guarantees.


Hyperluminal-mem is released under the terms of the Lisp Lesser General Public License, known as the LLGPL.