Author:         Patrik Nyblom <pan(at)erlang(dot)org>,
            Fredrik Svahn <Fredrik(dot)Svahn(at)gmail>
Status:         Final/R14A Proposal is implemented in OTP release R14A
Type:           Standards Track
Created:        28-Nov-2009
Erlang-Version: R13B03
Post-History:

EEP 31: Binary manipulation and searching module

Abstract

This EEP contains developed suggestions regarding the module binary first suggested in EEP 9.

EEP 9 suggests several modules and is partially superseded by later EEP's (i.e. EEP 11), while still containing valuable suggestions not yet implemented. The remaining modules from EEP 9 will therefore appear in separate EEP's. This construction is made in agreement with the original author of EEP 9.

The module binary is suggested to contain fast searching algorithms together with some common operations on binaries already present for lists (in the lists module).

Motivation

While efficient searching is already present in the re library, dedicated search functions can further speed up searching in binaries, given an efficient implementation (i.e. Boyer-More and Aho-Corassick algorithm). Another important advantage of separate searching algorithms are ease of use to the programmer, as the suggested interfaces do not require knowledge about regular expression syntax and special characters in the binaries need not be escaped. It's interesting to note how often regular expressions are used for simple sub-string searching or replacement, which can with this suggested module be done easily.

Decomposition of binaries are usually done by using bit-syntax. However some common operations are useful to have as ordinary functions, both for performance and to support a more traditiona functional programming style.

Some operations for converting lists to binaries and v.v. are today in the erlang module. BIFs concerning binaries now present have varied view of zero vs. one-based positioning in binaries. I.e. binary_to_list/3 uses one-based while split_binary/2 uses zero-based. As the convention is to use zero-based, new functions for converting binaries to lists and v.v. are needed.

Binaries are in fact a shared data-type, with small binaries often referencing parts of larger binaries in a way not controllable by the programmer in a simple way. The bitstring data-type further complicate things to the programmer in a way hard to easily manage. I therefore also suggest some low level functions to inspect binary representation and to clone binaries to ensure a minimal representation.

As matching is not allowed in guard expressions, I furthermore suggest that a function for extracting parts of binaries is added to the set of guard BIFs. This would be consistent with the function element/2 being allowed in guards.

Rationale

For the lists data type there is a help library providing functions for common operations such as searching and splitting lists. This EEP suggests that a similar set of library functions should be created for binaries. Many of the proposed functions are based on answers to questions regarding binaries on the erlang-questions mailing list, e.g. "how do I convert a number to a binary?". This EEP therefore suggests the addition of one module in stdlib, namely a module binary which will implement the requested functionality in an efficient way. Most of this module will need to be implemented in native code (residing in the virtual machine) why the proposed implementation will be delivered as "beta" functionality in a forthcoming Erlang release.

The functionality suggested is the following:

  • Functionality for searching, splitting and replacing in binaries. The functionality in some ways will overlap that of the regular expression library already present in Erlang, but will be even more efficient and will have a simpler interface.

  • Common operations on binaries that have their counterparts for lists already in the stdlib module lists. While not all interfaces in the lists module are applicable to binaries, many are. This module also provides a good place for future operations on binaries, operations that are not applicable to lists or that we still don't know the need for.

  • Functions for converting lists to binaries and v.v. These functions should have a consistent view of zero-based indexing in binaries.

  • Operations on binaries concerning their internal representation. This functionality is sometimes necessary to avoid extensive use of memory due to the shared nature of the binaries. As operations on binaries do not involve copying when binaries are taken apart, programs can unknowingly (or at least unintentionally) keep references to large binaries by holding seemingly small amounts of data in the process. The O(1) nature of many operations on binaries makes the data sharing necessary, but the effects can sometimes be surprising. On the other hand, O(n) complexity and instant memory explosions when splitting a binary would be even more surprising, why the current behavior need to be retained. It is suggested that functions for both inspecting the nature of sharing of a binary and to clone a copy of a binary to avoid sharing effects is present in this suggested module.

All functionality is to be applied to byte oriented binaries, never bitstrings that do not have a bitlength that is a multiple of eight. All binaries supplied to and returned by these functions should pass the is_binary/1 test, otherwise an error will be raised.

Suggested module reference

I suggest the following functionality (presented as an excerpt of an Erlang manual pages). A discussion about the interface can be found below.

DATA TYPES

cp()

Opaque data-type representing a compiled search-pattern. guaranteed to be a tuple() to allow programs to distinguish it from non precompiled search patterns.

part() = {Pos,Length}

Start = int()
Length = int()

A representaion of a part (or range) in a binary. Start is a zero-based offset into a binary() and Length is the length of that part. As input to functions in this module, a reverse part specification is allowed, constructed with a negative Length, so that the part of the binary begins at Start + Length and is -Length long. This is useful for referencing the last N bytes of a binary as {size(Binary), -N}. The functions in this module always return part()'s with positive Length.

EXPORTS

compile_pattern(Pattern) -> cp()

Types:

Pattern = binary() | [ binary() ]

Builds an internal structure representing a compilation of a search-pattern, later to be used in the find, split or replace functions. The cp() returned is guaranteed to be a tuple() to allow programs to distinguish it from non precompiled search patterns

When a list of binaries is given, it denotes a set of alternative binaries to search for. I.e if [<<"functional">>, <<"programming">>] is given as Pattern, this means ''either <<"functional">> or <<"programming">>''. The pattern is a set of alternatives; when only a single binary is given, the set has only one element.

If pattern is not a binary or a flat proper list of binaries, a badarg exception will be raised.

match(Subject, Pattern) -> Found | no

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Found = part()

The same as match(Subject, Pattern, []).

match(Subject,Pattern,Options) -> Found | no

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Found = part()
Options = [ Option ]
Option = {scope, part()}

Searches for the first occurrence of Pattern in Subject and returns the position and length.

The function will return {Pos,Length} for the binary in Pattern starting at the lowest position in Subject.

Example::

1> binary:find(<<"abcde">>, [<<"bcde">>,<<"cd">>],[]).
{1,4}

Even though <<"cd">> ends before <<"bcde">>, <<"bcde">> begins first and is therefore the first match. If two overlapping matches begins at the same position, the longest is returned.

Summary of the options:

  • {scope, {Start, Length}}
    Only the given part is searched. Return values still have offsets from the beginning of Subject. A negative Length is allowed as described in the TYPES section of this manual.

    The found part() is returned, if none of the strings in Pattern is found, the atom no is returned.

    For a descrition of Pattern, see compile_pattern/1.

    If {scope, {Start,Length}} is given in the options such that Start is larger than the size of Subject, Start + Length is less than zero or Start + Length is larger than the size of Subject, a badarg exception is raised.

matches(Subject, Pattern) -> Found

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Found = [ part() ] | []

The same as matches(Subject, Pattern, []).

matches(Subject,Pattern,Options) -> Found

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Found = [ part() ] | []
Options = [ Option ]
Option = {scope, part()}

Works like match, but the Subject is search until exhausted and a list of all non-overlapping parts present in Pattern are returned (in order).

The first and longest match is preferred to a shorter, which is illustrated by the following example::

1> binary:matches(<<"abcde">>, [<<"bcde">>,<<"bc">>>,<<"de">>],[]).
[{1,4}]

The result shows that <<"bcde">>> is selected instead of the shorter match <<"bc">> (which would have given raise to one more match,<<"de">>). This corresponds to the behavior of posix regular expressions (and programs like awk), but is not consistent with alternative matches in re (and Perl), where instead lexical ordering in the search pattern selects which string matches.

If none of the strings in pattern is found, an empty list is returned.

For a descrition of Pattern, see compile_pattern/1 and for a desctioption of available options, see match/3.

If {scope, {Start,Length}} is given in the options such that Start is larger than the size of Subject, Start + Length is less than zero or Start + Length is larger than the size of Subject, a badarg exception is raised.

split(Subject,Pattern) -> Parts

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Parts = [ binary() ]

The same as split(Subject, Pattern, []).

split(Subject,Pattern,Options) -> Parts

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Parts = [ binary() ]
Options = [ Option ]
Option = {scope, part()} | trim | global

Splits Binary into a list of binaries based on Pattern. If the option global is not given, only the first occurrence of Pattern in Subject will give rise to a split.

The parts of Pattern actually found in Subject are not included in the result.

Example::

1> binary:split(<<1,255,4,0,0,0,2,3>>, [<<0,0,0>>,<<2>>],[]).
[<<1,255,4>>, <<2,3>>] 
2> binary:split(<<0,1,0,0,4,255,255,9>>, [<<0,0>>, <<255,255>>],[global]).
[<<0,1>>,<<4>>,<<9>>]

Summary of options:

  • {scope, part()}
    Works as in binary:match/3 and binary:matches/3. Note that this only defines the scope of the search for matching strings, it does not cut the binary before splitting. The bytes before and after the scope will be kept in the result. See example below.

  • trim
    Removes trailing empty parts of the result (as does trim in re:split/3)

  • global
    Repeats the split until the Subject is exhausted. Conceptually the global option makes split work on the positions returned by binary:matches/3, while it normally works on the position returned by binary:match/3.

Example of the difference between a scope and taking the binary apart before splitting::

1> binary:split(<<"banana">>,[<<"a">>],[{scope,{2,3}}]).
[<<"ban">>,<<"na">>]
2> binary:split(binary:part(<<"banana">>,{2,3}),[<<"a">>],[]).
[<<"n">>,<<"n">>]

The return type is always a list of binaries which are all referencing Subject. This means that the data in Subject is not actually copied to new binaries and that Subject cannot be garbage collected until the results of the split are no longer referenced.

For a descrition of Pattern, see compile_pattern/1.

replace(Subject,Pattern,Replacement) -> Result

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Replacement = binary()
Result = binary()

The same as replace(Subject,Pattern,Replacement,[]).

replace(Subject,Pattern,Replacement,Options) -> Result

Types:

Subject = binary()
Pattern = binary() | [ binary() ] | cp()
Replacement = binary()
Result = binary()
Options = [ Option ]
Option = global | {scope, part()} | {insert_replaced, InsPos}
InsPos = OnePos | [ OnePos ]
OnePos = int() =< byte_size(Replacement)

Constructs a new binary by replacing the parts in Subject matching Pattern with the content of Replacement.

If the matching sub-part of Subject giving raise to the replacement is to be inserted in the result, the option {insert_replaced, InsPos} will insert the matching part into Replacement at the given position (or positions) before actually inserting Replacement into the Subject. Example::

1> binary:replace(<<"abcde">>,<<"b">>,<<"[]">>,[{insert_replaced,1}]).
<<"a[b]cde">>
2> binary:replace(<<"abcde">>,[<<"b">>,<<"d">>],<<"[]">>,
                 [global,{insert_replaced,1}]).
<<"a[b]c[d]e">>
3> binary:replace(<<"abcde">>,[<<"b">>,<<"d">>],<<"[]">>,
                 [global,{insert_replaced,[1,1]}]).
<<"a[bb]c[dd]e">>
4> binary:replace(<<"abcde">>,[<<"b">>,<<"d">>],<<"[-]">>,
                 [global,{insert_replaced,[1,2]}]).
<<"a[b-b]c[d-d]e">>

If any position given in InsPos is greater than the size of the replacement binary, a badarg exception is raised.

The options global and {scope, part()} works as for binary:split/3. The return type is always a binary.

For a descrition of Pattern, see compile_pattern/1.

longest_common_prefix(Binaries) -> int()

Types:

Binaries = [ binary() ]

Returns the length of the longest common prefix of the binaries in the list Binaries. Example::

1> binary:longest_common_prefix([<<"erlang">>,<<"ergonomy">>]).
2
2> binary:longest_common_prefix([<<"erlang">>,<<"perl">>]).
0

If Binaries is not a flat list of binaries, a badarg exception is raised.

longest_common_suffix(Binaries) -> int()

Types:

Binaries = [ binary() ]

Returns the length of the longest common suffix of the binaries in the list Binaries. Example::

1> binary:longest_common_suffix([<<"erlang">>,<<"fang">>]).
3
2> binary:longest_common_suffix([<<"erlang">>,<<"perl">>]).
0

If Binaries is not a flat list of binaries, a badarg exception is raised.

first(Subject) -> int()

Types:

Subject = binary()

Returns the first byte of the binary as an integer. If the binary length is zero, a badarg exception is raised.

last(Subject) -> int()

Types:

Subject = binary()

Returns the last byte of the binary as an integer. If the binary length is zero, a badarg exception is raised.

at(Subject, Pos) -> int()

Types:

Subject = binary()
Pos = int() >= 0

Returns the byte at position Pos (zero-based) in the binary Subject as an integer. If Pos >= byte_size(Subject), a badarg exception is raised.

part(Subject, PosLen) -> binary()

Types:

Subject = binary()
PosLen = part()

Extracts the part of the binary described by PosLen.

Negative length can be used to extract bytes at the end of a binary::

1> Bin = <<1,2,3,4,5,6,7,8,9,10>>.
2> binary:part(Bin,{byte_size(Bin), -5)).
<<6,7,8,9,10>>

If PosLen in any way references outside the binary, a badarg exception is raised.

part(Subject, Pos, Len) -> binary()

Types:

Subject = binary()
Pos = int()
Len = int()

The same as part(Subject, {Pos, Len}).

bin_to_list(Subject) -> list()

Types:

Subject = binary()

The same as bin_to_list(Subject,{0,byte_size(Subject)}).

bin_to_list(Subject, PosLen) -> list()

Subject = binary()
PosLen = part()

Converts Subject to a list of int(), each int representing the value of one byte. The part() denotes which part of the binary() to convert. Example::

1> binary:bin_to_list(<<"erlang">>,{1,3}).
"rla"
%% or [114,108,97] in list notation.

If PosLen in any way references outside the binary, a badarg exception is raised.

bin_to_list(Subject, Pos, Len) -> list()

Types:

Subject = binary()
Pos = int()
Len = int()

The same as bin_to_list(Subject,{Pos,Len}).

list_to_bin(ByteList) -> binary()

Types:

ByteList = iodata() (see module erlang)

Works exactly like erlang:list_to_binary/1, added for completeness.

copy(Subject) -> binary()

Types:

Subject = binary()

The same as copy(Subject, 1).

copy(Subject,N) -> binary()

Types:

Subject = binary()
N = int() >= 0

Creates a binary with the content of Subject duplicated N times.

This function will always create a new binary, even if N = 1. By using copy/1 on a binary referencing a larger binary, one might free up the larger binary for garbage collection.

NOTE! By deliberately copying a single binary to avoid referencing a larger binary, one might, instead of freeing up the larger binary for later garbage collection, create much more binary data than needed. Sharing binary data is usually good. Only in special cases, when small parts reference large binaries and the large binaries are no longer used in any process, deliberate copying might be a good idea.

If N < 0, a badarg exception is raised.

referenced_byte_size(binary()) -> int()

If a binary references a larger binary (often described as being a sub-binary), it can be useful to get the size of the actual referenced binary. This function can be used in a program to trigger the use of copy/1. By copying a binary, one might dereference the original, possibly large, binary which a smaller binary is a reference to.

Example::

store(Binary, GBSet) ->
  NewBin = 
      case binary:referenced_byte_size(Binary) of
          Large when Large > 2 * byte_size(Binary) ->
             binary:copy(Binary);
          _ ->
         Binary
      end,
  gb_sets:insert(NewBin,GBSet).

In this example, we chose to copy the binary content before inserting it in the gb_set() if it references a binary more than twice the size of the data we're going to keep. Of course different rules for when copying will apply to different programs.

Binary sharing will occur whenever binaries are taken apart, this is the fundamental reason why binaries are fast, decomposition can always be done with O(1) complexity. In rare circumstances this data sharing is however undesirable, why this function together with copy/1 might be useful when optimizing for memory use.

Example of binary sharing::

1> A = binary:copy(<<1>>,100).
<<1,1,1,1,1 ...
2> byte_size(A).
100
3> binary:referenced_byte_size(A)
100
4> <<_:10/binary,B:10/binary,_/binary>> = A.
<<1,1,1,1,1 ...
5> byte_size(B).
10
6> binary:referenced_byte_size(B)
100

NOTE! Binary data is shared among processes. If another process still references the larger binary, copying the part this process uses only consumes more memory and will not free up the larger binary for garbage collection. Use this kind of intrusive functions with extreme care, and only if a real problem is detected.

encode_unsigned(Unsigned) -> binary()

Types:

Unsigned = int() >= 0

The same as encode_unsigned(Unsigned,big).

encode_unsigned(Unsigned,Endianess) -> binary()

Types:

Unsigned = int() >= 0
Endianess = big | little

Converts a positive integer to the smallest possible representation in in a binary digit representation, either big or little endian.

Example:

1> binary:encode_unsigned(11111111,big). 
<<169,138,199>>

decode_unsigned(Subject) -> Unsigned

Types:

Subject = binary()
Unsigned = int() >= 0

The same as encode_unsigned(Subject,big).

decode_unsigned(Subject, Endianess) -> Unsigned

Types:

Subject = binary()
Endianess = big | little
Unsigned = int() >= 0

Converts the binary digit representation, in big or little endian, of a positive integer in Subject to an Erlang int().

Example::

1> binary:decode_unsigned(<<169,138,199>>,big). 
11111111

Guard BIF

I suggest adding the functions binary:part/2 and binary:part/3 to the set of BIFs allowed in guard tests. As guard BIFs are traditionally put in the erlang module, the following names for the guard BIFs are suggested::

erlang:binary_part/2
erlang:binary_part/3

They should both work exactly as their counterparts in the binary module.

Interface design discussion

As with all modules, there are a lot of arguments about the actual interface, sometimes more than about the functionality. In this case a number of parameters has to be considered.

  • Effectiveness - The interface should be constructed so that fast implementation is possible and so that code using the interface can be written in an effective way. To not create unnecessary garbage is one parameter, to allow for general code is another.

  • Parameter ordering - I've chosen to make the binary subject the first parameter in all applicable calls. Putting the subject first corresponds to the re interface. The lists module, however, usually has the subject as last parameter. We could go for that instead, but unfortunately the lists:sublist/{2,3} interface, which corresponds to the part function, has the subject first, why following the conventions of lists would not only break conformance with re, it would also give us a generally non-stringent interface. The effect of not conforming to the lists interface is that using function names from that module would lead to confusion and therefore is avoided.

  • Function naming - We have two related modules to take into account when naming functions here. The module re is related to the searching function (match, replace etc), while the lists module is related to the decomposition functions (first, last etc).

    I've basically retained the names from re when I find the functionality, both in concept and interface to be similar enough. The nature of regular expressions as small executable programs, which is to much to say for a collection of binaries as the patterns are in this module, prohibits the use of the function name run for actually doing the searching. We use match and matches instead of run.

    As this module is more general than re, a function name like compile is not really good. re:compile means "compile a regular expression", but what would binary:compile mean? Therefore the pre-processing function is instead called compile_pattern.

    When it comes to the lists module, the parameter ordering has prevented me from reusing any function names but last, which only takes one parameter in lists and there is no real alternative there.

  • Options or multiple functions - I believe a good rule of thumb is to not have options that change the return type of the function, which would have been the case if we i.e. had a global option to match/3 instead of a separate matches/3 function.

    The fact that there are a manageable set of possible return types for the searching and decomposition functions allows us to follow that rule of thumb.

    (Unfortunately that rule could not be easilly followed in re, as the rich assortment of options would have given rise to a non-manageable amount of function names).

Performance

Although the decomposition functions are not really faster than using bit-syntax for decomposition, they create slightly less garbage than the bit syntax. As they are not slower than bit-syntax, they also have a purpose in allowing for a different programming style.

The match/replace/split functionality should be compared to similar functionality in the re module. Implementation methods has to be chosen so that this modules search functions are faster, or possibly even significantly faster, than re.

Reference implementation

A reference implementation was available on GitHub development branch before the final inclusion in R14A.

Copyright

This document is licensed under the Creative Commons license.

Powered by Erlang Web