bit-ops

2017-02-27

Bit-Ops - Tools for Writing Optimized Bit-Vector Operations

Build Status

In the modern Common Lisp implementations, bit-vector operation functions such as bit-and are compiled into word-size iterations where 32 or 64 bits are processed at once. However, it requires a careful handling when multiple operations are combined, e.g. for avoiding the generation of intermediate vectors by destructively modifying the existing vectors. For example,

(bit-and a (bit-and b c))

causes consing since the intermediate value (bit-and b c) is generated on the heap. This library addresses this problem as well as the scarsity of useful bit-vector functions in ANSI CL.

This section provides a review of related libraries, which I believe is important for alleviating choise paralysis.

BIT-SMASHER provides functions for converting bit-vector to/from integers, octets and hex strings. BIT-OPS does not have such conversion functions. BIT-SMASHER also provides functions for arithmetic, such as addition, subtraction, shifting. However, note that those operations are not always optimized and runs bitvec->integer->bitvec conversion each time, with possibly consing.

BITFIELD-SCHEMA provides several functions analogous to DPB and LPB for integers (GET/PUT-INTEGER). It also provides a DSL for writing accessors to bit-vectors (ala union type in C).

Example from the doc:

(defbitfield-schema tree-node (:offset offt)
  (disabled-p   :width 1)
  (values       :width 16 :count 10)
  (left-child   :width 24)
  (right-child  :width 7))

BINARY-TYPES provides DEFINE-BITFIELD and DEFINE-BINARY-CLASS whose role is similar to BITFIELD-SCHEMA, but is for parsing machine integers, not bit-vectors. TRIVIAL-BIT-STREAMS provides a buffered stream of bits. NIBBLES provides optimized access to octet vectors, especially on SBCL by defining several SSE VOP operations.

Usage

Macro AS-BITWISE-OPERATIONS (&key result) &body body

Compute bitwise operations using bit vector arithmetic. BODY accepts a single form. Within BODY, one can use variables holding bit-vectors as arguments to bitwise operations, as well as constants 0 and 1, which maps to the bit vector filled with 0 or 1, respectively. All bit-vectors that are involved in bitwise operations should be of the same length.

Primitive operators corresponds to ANSI CL functions: For example, (not subform) is compiled into (bit-not subform <temporary storage>) . Following primitive operators are available:

not and andc1 andc2 eqv ior nand nor orc1 orc2 xor

Additionally, (SUBSEQ FORM OFFSET) operator evaluates FORM and extracts its window starting from OFFSET and of the length equal to the other variables. FORM is a regular lisp expression, not a bitwise operation, and the result may be different from the other bit-vectors.

The final computation result is stored in a newly allocated vector, or in RESULT if specified, in spirit similar to the optional argument of Common Lisp bit-vector functions. The entire form returns the bit-vector which contains the result.

The compiler does various optimizations:

  • Nested expressions store the results into dynamic-extent temporary vectors.
  • Common subexpressions are eliminated.
  • The number of temporary vectors are minimized/shared in spirit similar to register allocation.
  • Macros for bitwise operations can be defined with DEFINE-BITWISE-OPERATION.

(as-bitwise-operations ()
  (and a b c))

->

(LET* ((#:LEN835 (LENGTH C)) (#:G833 (MAKE-BIT-VECTOR #:LEN835)))
  (LET* ((+ZERO+ (MAKE-ZERO #:LEN835))
         (+ONE+ (MAKE-ONE #:LEN835)))
    (DECLARE (DYNAMIC-EXTENT +ZERO+ +ONE+))
    (DECLARE (IGNORABLE +ZERO+ +ONE+))
    (BIT-AND B C #:G833)
    (BIT-AND A #:G833 #:G833)
    #:G833))



(define-bitwise-operation if (condition then else)
  `(ior (and ,condition ,then)
        (andc1 ,condition ,else)))

(as-bitwise-operations (:result r)
  (if a b c))

->

(LET* ((#:LEN839 (LENGTH C)) (#:G836 R))
  (LET* ((+ZERO+ (MAKE-ZERO #:LEN839))
         (+ONE+ (MAKE-ONE #:LEN839))
         (#:G837 (MAKE-BIT-VECTOR #:LEN839)))
    (DECLARE (DYNAMIC-EXTENT +ZERO+ +ONE+ #:G837))
    (DECLARE (IGNORABLE +ZERO+ +ONE+))
    (BIT-AND A B #:G836)
    (BIT-ANDC1 A C #:G837)
    (BIT-IOR #:G836 #:G837 #:G836)
    #:G836))

Dependencies

This library is at least tested on implementation listed below:

  • SBCL 1.3.13 on X86-64 Linux 4.4.0-59-generic (author's environment)
  • CCL 1.11 X86-64 Linux 4.4.0-59-generic (author's environment)

Also, it depends on the following libraries:

  • iterate by ** : Jonathan Amsterdam's iterator/gatherer/accumulator facility
  • alexandria by Nikodemus Siivola , and others. : Alexandria is a collection of portable public domain utilities.
  • trivia :

Installation

Author

  • Masataro Asai (guicho2.71828@gmail.com)

Copyright (c) 2017 Masataro Asai (guicho2.71828@gmail.com)

License

Licensed under the LLGPL License.

Author
Masataro Asai
License
LLGPL