2007 2011 Ericsson AB, All Rights Reserved The contents of this file are subject to the Erlang Public License, Version 1.1, (the "License"); you may not use this file except in compliance with the License. You should have received a copy of the Erlang Public License along with this software. If not, it can be retrieved online at http://www.erlang.org/. Software distributed under the License is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See the License for the specific language governing rights and limitations under the License. The Initial Developer of the Original Code is Ericsson AB. Constructing and matching binaries Bjorn Gustavsson 2007-10-12 binaryhandling.xml

In R12B, the most natural way to write binary construction and matching is now significantly faster than in earlier releases.

To construct at binary, you can simply write

DO (in R12B) / REALLY DO NOT (in earlier releases)

my_list_to_binary(List, <<>>). my_list_to_binary([H|T], Acc) -> my_list_to_binary(T, <>); my_list_to_binary([], Acc) -> Acc.]]>

In releases before R12B, Acc would be copied in every iteration. In R12B, Acc will be copied only in the first iteration and extra space will be allocated at the end of the copied binary. In the next iteration, H will be written in to the extra space. When the extra space runs out, the binary will be reallocated with more extra space.

The extra space allocated (or reallocated) will be twice the size of the existing binary data, or 256, whichever is larger.

The most natural way to match binaries is now the fastest:

DO (in R12B)

>) -> [H|my_binary_to_list(T)]; my_binary_to_list(<<>>) -> [].]]>
How binaries are implemented

Internally, binaries and bitstrings are implemented in the same way. In this section, we will call them binaries since that is what they are called in the emulator source code.

There are four types of binary objects internally. Two of them are containers for binary data and two of them are merely references to a part of a binary.

The binary containers are called refc binaries (short for reference-counted binaries) and heap binaries.

Refc binaries consist of two parts: an object stored on the process heap, called a ProcBin, and the binary object itself stored outside all process heaps.

The binary object can be referenced by any number of ProcBins from any number of processes; the object contains a reference counter to keep track of the number of references, so that it can be removed when the last reference disappears.

All ProcBin objects in a process are part of a linked list, so that the garbage collector can keep track of them and decrement the reference counters in the binary when a ProcBin disappears.

Heap binaries are small binaries, up to 64 bytes, that are stored directly on the process heap. They will be copied when the process is garbage collected and when they are sent as a message. They don't require any special handling by the garbage collector.

There are two types of reference objects that can reference part of a refc binary or heap binary. They are called sub binaries and match contexts.

A sub binary is created by split_binary/2 and when a binary is matched out in a binary pattern. A sub binary is a reference into a part of another binary (refc or heap binary, never into a another sub binary). Therefore, matching out a binary is relatively cheap because the actual binary data is never copied.

A match context is similar to a sub binary, but is optimized for binary matching; for instance, it contains a direct pointer to the binary data. For each field that is matched out of a binary, the position in the match context will be incremented.

In R11B, a match context was only using during a binary matching operation.

In R12B, the compiler tries to avoid generating code that creates a sub binary, only to shortly afterwards create a new match context and discard the sub binary. Instead of creating a sub binary, the match context is kept.

The compiler can only do this optimization if it can know for sure that the match context will not be shared. If it would be shared, the functional properties (also called referential transparency) of Erlang would break.

Constructing binaries

In R12B, appending to a binary or bitstring

> <>]]>

is specially optimized by the run-time system. Because the run-time system handles the optimization (instead of the compiler), there are very few circumstances in which the optimization will not work.

To explain how it works, we will go through this code

>, %% 1 Bin1 = <>, %% 2 Bin2 = <>, %% 3 Bin3 = <>, %% 4 Bin4 = <>, %% 5 !!! {Bin4,Bin3} %% 6]]>

line by line.

The first line (marked with the %% 1 comment), assigns a heap binary to the variable Bin0.

The second line is an append operation. Since Bin0 has not been involved in an append operation, a new refc binary will be created and the contents of Bin0 will be copied into it. The ProcBin part of the refc binary will have its size set to the size of the data stored in the binary, while the binary object will have extra space allocated. The size of the binary object will be either twice the size of Bin0 or 256, whichever is larger. In this case it will be 256.

It gets more interesting in the third line. Bin1 has been used in an append operation, and it has 255 bytes of unused storage at the end, so the three new bytes will be stored there.

Same thing in the fourth line. There are 252 bytes left, so there is no problem storing another three bytes.

But in the fifth line something interesting happens. Note that we don't append to the previous result in Bin3, but to Bin1. We expect that Bin4 will be assigned the value <<0,1,2,3,17>>. We also expect that Bin3 will retain its value (<<0,1,2,3,4,5,6,7,8,9>>). Clearly, the run-time system cannot write the byte 17 into the binary, because that would change the value of Bin3 to <<0,1,2,3,4,17,6,7,8,9>>.

What will happen?

The run-time system will see that Bin1 is the result from a previous append operation (not from the latest append operation), so it will copy the contents of Bin1 to a new binary and reserve extra storage and so on. (We will not explain here how the run-time system can know that it is not allowed to write into Bin1; it is left as an exercise to the curious reader to figure out how it is done by reading the emulator sources, primarily erl_bits.c.)

Circumstances that force copying

The optimization of the binary append operation requires that there is a single ProcBin and a single reference to the ProcBin for the binary. The reason is that the binary object can be moved (reallocated) during an append operation, and when that happens the pointer in the ProcBin must be updated. If there would be more than on ProcBin pointing to the binary object, it would not be possible to find and update all of them.

Therefore, certain operations on a binary will mark it so that any future append operation will be forced to copy the binary. In most cases, the binary object will be shrunk at the same time to reclaim the extra space allocated for growing.

When appending to a binary

>]]>

only the binary returned from the latest append operation will support further cheap append operations. In the code fragment above, appending to Bin will be cheap, while appending to Bin0 will force the creation of a new binary and copying of the contents of Bin0.

If a binary is sent as a message to a process or port, the binary will be shrunk and any further append operation will copy the binary data into a new binary. For instance, in the following code fragment

>, PortOrPid ! Bin1, Bin = <> %% Bin1 will be COPIED ]]>

Bin1 will be copied in the third line.

The same thing happens if you insert a binary into an ets table or send it to a port using erlang:port_command/2.

Matching a binary will also cause it to shrink and the next append operation will copy the binary data:

>, <> = Bin1, Bin = <> %% Bin1 will be COPIED ]]>

The reason is that a match context contains a direct pointer to the binary data.

If a process simply keeps binaries (either in "loop data" or in the process dictionary), the garbage collector may eventually shrink the binaries. If only one such binary is kept, it will not be shrunk. If the process later appends to a binary that has been shrunk, the binary object will be reallocated to make place for the data to be appended.

Matching binaries

We will revisit the example shown earlier

DO (in R12B)

>) -> [H|my_binary_to_list(T)]; my_binary_to_list(<<>>) -> [].]]>

too see what is happening under the hood.

The very first time my_binary_to_list/1 is called, a match context will be created. The match context will point to the first byte of the binary. One byte will be matched out and the match context will be updated to point to the second byte in the binary.

In R11B, at this point a sub binary would be created. In R12B, the compiler sees that there is no point in creating a sub binary, because there will soon be a call to a function (in this case, to my_binary_to_list/1 itself) that will immediately create a new match context and discard the sub binary.

Therefore, in R12B, my_binary_to_list/1 will call itself with the match context instead of with a sub binary. The instruction that initializes the matching operation will basically do nothing when it sees that it was passed a match context instead of a binary.

When the end of the binary is reached and second clause matches, the match context will simply be discarded (removed in the next garbage collection, since there is no longer any reference to it).

To summarize, my_binary_to_list/1 in R12B only needs to create one match context and no sub binaries. In R11B, if the binary contains N bytes, N+1 match contexts and N sub binaries will be created.

In R11B, the fastest way to match binaries is:

DO NOT (in R12B)

my_complicated_binary_to_list(Bin, 0). my_complicated_binary_to_list(Bin, Skip) -> case Bin of <<_:Skip/binary,Byte,_/binary>> -> [Byte|my_complicated_binary_to_list(Bin, Skip+1)]; <<_:Skip/binary>> -> [] end.]]>

This function cleverly avoids building sub binaries, but it cannot avoid building a match context in each recursion step. Therefore, in both R11B and R12B, my_complicated_binary_to_list/1 builds N+1 match contexts. (In a future release, the compiler might be able to generate code that reuses the match context, but don't hold your breath.)

Returning to my_binary_to_list/1, note that the match context was discarded when the entire binary had been traversed. What happens if the iteration stops before it has reached the end of the binary? Will the optimization still work?

>) -> T; after_zero(<<_,T/binary>>) -> after_zero(T); after_zero(<<>>) -> <<>>. ]]>

Yes, it will. The compiler will remove the building of the sub binary in the second clause

>) -> after_zero(T); . . .]]>

but will generate code that builds a sub binary in the first clause

>) -> T; . . .]]>

Therefore, after_zero/1 will build one match context and one sub binary (assuming it is passed a binary that contains a zero byte).

Code like the following will also be optimized:

{lists:reverse(Acc),Buffer}; all_but_zeroes_to_list(<<0,T/binary>>, Acc, Remaining) -> all_but_zeroes_to_list(T, Acc, Remaining-1); all_but_zeroes_to_list(<>, Acc, Remaining) -> all_but_zeroes_to_list(T, [Byte|Acc], Remaining-1).]]>

The compiler will remove building of sub binaries in the second and third clauses, and it will add an instruction to the first clause that will convert Buffer from a match context to a sub binary (or do nothing if Buffer already is a binary).

Before you begin to think that the compiler can optimize any binary patterns, here is a function that the compiler (currently, at least) is not able to optimize:

>) -> non_opt_eq(T1, T2); non_opt_eq([_|_], <<_,_/binary>>) -> false; non_opt_eq([], <<>>) -> true.]]>

It was briefly mentioned earlier that the compiler can only delay creation of sub binaries if it can be sure that the binary will not be shared. In this case, the compiler cannot be sure.

We will soon show how to rewrite non_opt_eq/2 so that the delayed sub binary optimization can be applied, and more importantly, we will show how you can find out whether your code can be optimized.

The bin_opt_info option

Use the bin_opt_info option to have the compiler print a lot of information about binary optimizations. It can be given either to the compiler or erlc

or passed via an environment variable

Note that the bin_opt_info is not meant to be a permanent option added to your Makefiles, because it is not possible to eliminate all messages that it generates. Therefore, passing the option through the environment is in most cases the most practical approach.

The warnings will look like this:

To make it clearer exactly what code the warnings refer to, in the examples that follow, the warnings are inserted as comments after the clause they refer to:

>) -> %% NOT OPTIMIZED: sub binary is used or returned T; after_zero(<<_,T/binary>>) -> %% OPTIMIZED: creation of sub binary delayed after_zero(T); after_zero(<<>>) -> <<>>.]]>

The warning for the first clause tells us that it is not possible to delay the creation of a sub binary, because it will be returned. The warning for the second clause tells us that a sub binary will not be created (yet).

It is time to revisit the earlier example of the code that could not be optimized and find out why:

>) -> %% INFO: matching anything else but a plain variable to %% the left of binary pattern will prevent delayed %% sub binary optimization; %% SUGGEST changing argument order %% NOT OPTIMIZED: called function non_opt_eq/2 does not %% begin with a suitable binary matching instruction non_opt_eq(T1, T2); non_opt_eq([_|_], <<_,_/binary>>) -> false; non_opt_eq([], <<>>) -> true.]]>

The compiler emitted two warnings. The INFO warning refers to the function non_opt_eq/2 as a callee, indicating that any functions that call non_opt_eq/2 will not be able to make delayed sub binary optimization. There is also a suggestion to change argument order. The second warning (that happens to refer to the same line) refers to the construction of the sub binary itself.

We will soon show another example that should make the distinction between INFO and NOT OPTIMIZED warnings somewhat clearer, but first we will heed the suggestion to change argument order:

>, [H|T2]) -> %% OPTIMIZED: creation of sub binary delayed opt_eq(T1, T2); opt_eq(<<_,_/binary>>, [_|_]) -> false; opt_eq(<<>>, []) -> true.]]>

The compiler gives a warning for the following code fragment:

>) -> %% INFO: matching anything else but a plain variable to %% the left of binary pattern will prevent delayed %% sub binary optimization; %% SUGGEST changing argument order done; . . .]]>

The warning means that if there is a call to match_body/2 (from another clause in match_body/2 or another function), the delayed sub binary optimization will not be possible. There will be additional warnings for any place where a sub binary is matched out at the end of and passed as the second argument to match_body/2. For instance:

>) -> %% NOT OPTIMIZED: called function match_body/2 does not %% begin with a suitable binary matching instruction match_body(List, Data).]]>
Unused variables

The compiler itself figures out if a variable is unused. The same code is generated for each of the following functions

>, Count) -> count1(T, Count+1); count1(<<>>, Count) -> Count. count2(<>, Count) -> count2(T, Count+1); count2(<<>>, Count) -> Count. count3(<<_H,T/binary>>, Count) -> count3(T, Count+1); count3(<<>>, Count) -> Count.]]>

In each iteration, the first 8 bits in the binary will be skipped, not matched out.