Robust ABC (ActionScript Bytecode) [Dis-]Assembler

RABCDAsm is a collection of utilities including an ActionScript 3 assembler/disassembler, and a few tools to manipulate SWF files. These are:

rabcdasm - ABC disassembler
rabcasm - ABC assembler
abcexport - extracts ABC from SWF files
abcreplace - replaces ABC in SWF files
swfdecompress - decompresses zlib-compressed SWF files
swf7zcompress - (re-)compress the contents of a SWF using 7-Zip

abcexport and abcreplace are reimplementations of similar utilities from my swfutilsex Java package, however these work faster as they do not parse the SWF files as deeply.
swfdecompress is ancilliary and is only useful for debugging and studying of the SWF file format, and not required for ABC manipulation. It is functionally equivalent to flasm's -x option.
swf7zcompress is an utility to further reduce the size of SWF files. It uses 7-Zip to compress the data better than the standard zlib library would. It requires that the 7z command-line program be installed and in PATH.

Motivation and goals

This package was created due to lack of similar software out there. Particularly, I needed an utility which would allow me to edit ActionScript 3 bytecode with the following properties:

Speed. Less waiting means more productivity. rabcasm can assemble large projects (>200000 LOC) in under a second on modern machines.
Comfortably-editable output. Each class is decompiled to its own file, with files arranged in subdirectories representing the package hierarchy. Class files are #included from the main file.
Most importantly - robustness! If the Adobe AVM can load and run the file, then it must be editable - no matter if the file is obfuscated or otherwise mutilated to prevent reverse-engineering. RABCDAsm achieves this by using a textual representation closer to the ABC file format, rather than to what an ActionScript compiler would generate.

Compiling from source

RABCDAsm is written in the D programming language, version 1. It uses one third-party library, std2 (for some fancy template stuff).

Assuming you have dmd, git and svn installed, compiling should be as straight-forward as:

git clone git://github.com/CyberShadow/RABCDAsm.git
cd RABCDAsm
svn co http://svn.dsource.org/projects/std2/trunk/std2/std2
dmd rabcdasm abcfile asprogram disassembler autodata murmurhash2a
dmd rabcasm  abcfile asprogram    assembler autodata murmurhash2a
dmd abcexport     swffile zlibx
dmd abcreplace    swffile zlibx
dmd swfdecompress swffile zlibx
dmd swf7zcompress swffile zlibx

Pre-compiled binaries

You can find pre-compiled Windows binaries in the Downloads section on GitHub. However, please don't expect them to be up-to-date with the latest source versions.

Usage

To begin hacking on a SWF file:

abcexport file.swf

This will create file0.abc ... fileN.abc (often just file0.abc). Each file corresponds to an ABC block inside the SWF file.

To disassemble one of the .abc files:

rabcdasm file0.abc

This will create a file0 directory, which will contain file0.main.asasm (the main program file), file0.privatens.asasm (private namespace alias definitions) and a file per ActionScript class.

To assemble the .asasm files back, and update the SWF file:

rabcasm file0/file0.main.asasm
abcreplace file0.swf 0 file0/file0.main.abc

The second abcreplace argument represents the index of the ABC block in the SWF file, and corresponds to the number in the filename created by abcexport.

Syntax

The syntax of the disassembly was designed to be very simple and allow fast and easy parsing. It is a close representation of the .abc file format, and thus it is somewhat verbose. All constant pool elements (signed/unsigned integers, doubles, strings, namespaces, namespace sets, multinames) are always expanded inline, for ease of editing. Similarly, classes, instances, methods and method bodies are also defined inline, in the context of their "parent" object. By-index references of classes and methods (used in the newclass, newfunction and callstatic instructions) are represented via automatically-generated unique "reference strings", declared as refid fields.

If you haven't yet, I strongly recommend that you look through Adobe's ActionScript Virtual Machine 2 (AVM2) Overview. You will most likely need to consult it for the instruction reference anyway (although you can also use this handy list as well). You will find it difficult to understand the disassembly without good understanding of concepts such as namespaces and multinames.

Overview

In order to guarantee unambiguity and data preservation, all strings read from the input file - including identifiers (variable/function/class names) - are represented as string literals. Thus, the syntax does not have any "reserved words" or such - an unrecognized word is treated as an error, not as an identifier.

Whitespace (outside string literals, of course) is completely ignored, except where required to separate words. Comments are Intel-assembler-style: a single ; demarks a comment until the next end-of-line. Control directives (such as #include) are allowed anywhere where whitespace is allowed.

The syntax is comprised of hierarchical blocks. Each block contains a number of fields - starting with a keyword specifying the field type. A block is terminated with the end keyword. Some fields contain a limited number of parameters, and others are, or contain blocks.

Hierarchy

The topmost block in the hierarchy is the program block. This must be the first block in the file (thus, program must be the first word in the file as well). The program block contains script fields, and class / method fields for "orphan" classes and methods (not owned by other objects in the hierarchy). Orphan methods are usually anonymous functions. The file version is also specified in the program block, using the minorversion and majorversion fields (both unsigned integers).

script blocks have one mandatory sinit field (the script initialization method) and trait fields.

A "trait" can be one of several kinds. The kind is specified right after the trait keyword, followed by the trait name (a multiname). Following the name are the trait fields, varying by trait kind:

slot / const : slotid (unsigned integer), type (multiname), value
class : slotid, class (the actual class block)
function : slotid, method (the actual method block)
method / getter / setter : dispid (unsigned integer), method

Additionally, all traits may have flag fields, describing the trait's attributes (FINAL / OVERRIDE / METADATA).

class blocks have mandatory instance and cinit fields, defining the class instance and the class initializer method respectively. They may also have trait fields and a refid field (the refid field is not part of the file format - it's an unique string to allow referencing the class, see above).

instance blocks - always declared inline of their class block - must contain one iinit field (the instance initializer method), and may contain one extends field (multiname), implements fields (multinames), flag fields (SEALED / FINAL / INTERFACE / PROTECTEDNS), one protectedns field (namespace), and trait fields.

method blocks may contain one name field (multiname), a refid field, param fields (multinames - this represents the parameter types), one returns field (multiname), flag fields (NEED_ARGUMENTS / NEED_ACTIVATION / NEED_REST / HAS_OPTIONAL / SET_DXNS / HAS_PARAM_NAMES), optional fields (values), paramname fields (strings), and a body field (method body).

body blocks - always declared inline of their method block - must contain the maxstack, localcount, initscopedepth and maxscopedepth fields (unsigned integers), and a code field. It may also contain try and trait fields.

code blocks - always declared inline of their body block - are somewhat different in syntax from other blocks - mostly in that they may contain labels. Labels follow the most common syntax - a word followed by a : character, optionally followed by a relative byte offset (in case of pointers inside instructions). Multiple instruction arguments are comma-separated. Instruction arguments' types depend on the instruction - see the OpcodeInfo array in abcfile.d for a reference.

try blocks - always declared inline of their body block - represent an "exception" (try/catch) block. They contain five mandatory fields: from, to and target (names of labels representing start and end of the "try" block, and start of the "catch" block respectively), and type and name (multinames), representing the type and name of the exception variable.

Values have the syntax type ( value ) . type can be one of Integer, UInteger, Double, Utf8, Namespace, PackageNamespace, PackageInternalNs, ProtectedNamespace, ExplicitNamespace, StaticProtectedNs, PrivateNamespace, True, False, Null or Undefined. The type of the value depends on type. Types True, False, Null and Undefined have no value.

Constants

Multinames have the syntax type ( parameters ) . type can be one of QName / QNameA, RTQName / RTQNameA, RTQNameL / RTQNameLA, Multiname / MultinameA, MultinameL / MultinameLA, or TypeName. parameters depends on type:

QName / QNameA ( namespace , string )
RTQName / RTQNameA ( string )
RTQNameL / RTQNameLA ( )
Multiname / MultinameA ( string , namespace-set )
MultinameL / MultinameLA ( namespace-set )
TypeName ( multiname < multiname [ , multiname ... ] > )

Namespace sets have the syntax [ [ namespace [ , namespace ... ] ] ] (that is, a comma-separated list of namespaces in square brackets). Empty namespace sets can be specified using [].

Namespaces have the syntax type ( parameters ) . For types other than PrivateNamespace there is only one parameter - a string. PrivateNamespace namespaces have a second parameter, a named alias for a particular private namespace. Internally (the ABC file format), private namespaces are distinguished by a numerical index - rabcdasm will attempt to give them descriptive names based on their context. Aliases can be defined using the #privatens directive. rabcdasm will create a separate file containing the aliases (file0.privatens.asasm).

Strings have a syntax similar to C string literals. Strings start and end with a ". Supported escape sequences (a backslash followed by a letter) are \n (generates ASCII 0x0A), \r (ASCII 0x0D), and \x followed by two hexadecimal digits, which inserts the ASCII character with that code. Any other characters following a backslash generate that character - thus, you can escape backslashes using \\ and double quotes using \". When decompiling, high-ASCII characters (usually UTF-8) are not escaped - if you see gibberish instead of international text, configure your editor to open the files in UTF-8 encoding.

Additionally, constant pool types (signed/unsigned integers, doubles, strings, namespaces, namespace sets and multinames) may also have the value null (which represents the index 0 in the ABC file). Note that null is conceptually different from zero, an empty string or empty namespace set.

Macros

RABCDAsm has some basic macro-like capabilities, controlled by directives and variables. These bear some similarity to the C preprocessor, however these are processed in-loop rather than as a separate pre-processing step.

Directives

Directives start with a #, followed by a word identifying the directive:

#include string - inserts the contents of the file by the specified filename inline. Functionally equivalent to #mixin #get string , but faster.
#mixin string - inserts the contents of the specified string inline. Not very useful on its own.
#call string ( [ string [ , string ... ] ] ) - same as #mixin, however it additionally sets the special variables $1, $2 etc. to the contents of the specified arguments. When the end of the inserted string is reached, the old values of $1, $2 etc. are restored.
#get string - inserts a string containing the contents of the file by the specified filename inline. Similar to #include, but it inserts a string (surrounded by " etc.) instead.
#set word string - assigns the contents of the string to the variable word.
#unset word - deletes the variable word.
#privatens defines a private namespace alias, as described above.

Variables

Variables are manipulated with the #set and #unset directives, and can be instantiated in two ways:

$name - this inserts the contents of the variable inline. Note that although variables are defined using a string syntax, they are not inserted as a string using this syntax. Thus, the code:

#set str "Hello, world!" ... pushstring $str

will expand to pushstring Hello, world!, which will result in an error. To correct the problem, add escaped quotes around the variable contents ( #set str "\"Hello, world!\"" ), or use the second syntax:
$"name" - this inserts a string containing the contents of the variable inline. This syntax also works for #call arguments (e.g. $"1").

Example

Here's an example of how to use the above features to create a macro which logs a string literal and the contents of a register:

#set log "
    findpropstrict      QName(PackageNamespace(\"\"), \"log\")
    pushstring          $\"1\"
    getlocal            $2
    callpropvoid        QName(PackageNamespace(\"\"), \"log\"), 2
"

; ...

pushbyte 2
pushbyte 2
add_i
setlocal1
#call $"log"("two plus two equals", "1")

Highlighting

Included with the project is the file asasm.hrc, a simple syntax definition for the Colorer take5 syntax highlighting library. It should be straight-forward to adapt it to other syntax highlighting systems.

Hacking

ABC is internally represented in two forms. The ABCFile class stores the raw data structures, as they appear in the binary file. ASProgram uses pointers instead of indexes, allowing easy manipulation without having to worry about record order or constant pools. Conversion between various states is done as follows:

                             file.abc
                               |  ^
              ------ ABCReader |  | ABCWriter ----
             /                 v  |               \
            /                ABCFile               \
           /                   |  ^                 \
    rabcdasm---------- ABCtoAS |  | AStoABC --------rabcasm
           \                   v  |                 /
            \               ASProgram              /
             \                 |  ^               /
              --- Disassembler |  | Assembler ----
                               v  |
                            file.asasm

AStoABC will rebuild the constant pools, in a manner similar to Adobe's compilers (reverse-sorted by reference count). The exact order will almost surely be different, however. Also, some records (classes and methods) are currently not sorted by reference count, which may cause rabcasm to generate slightly larger files than the originals (due to variable-length encoding of integers).

Should you need to write an utility to manipulate ABC, you can use the existing code to load the file to either an ABCFile or ASProgram instance, and perform the necessary manipulations using those classes.

Tips

The following tips come from the author's experience and may be useful for RABCDAsm users.

Once you have disassembled a SWF file you intend to modify, you should immediately add the directory to a distributed source control system, such as Git or Mercurial. This will allow you to easily track and undo your changes, and easily merge your changes with new versions of SWF files.

If you plan on making non-trivial changes to SWF files, you should install the debug Flash Player. This will allow you to see validation and run-time error messages, instead of simply getting a blank rectangle.

The Fiddler Web Debugging Proxy can be very useful for analysing websites with SWF content. The following script fragment (which is to be placed in the OnBeforeResponse function) will automatically save all SWF files while preserving the directory structure.

if (oSession.oResponse.headers.ExistsAndContains("Content-Type", 
        "application/x-shockwave-flash")) {
    // Set desired path here
    var path:String = "C:\\Temp\\FiddlerCapture\\" + 
        oSession.host + oSession.PathAndQuery;
    if (path.Contains('?'))
        path = path.Substring(0, path.IndexOf('?'));
    var dir:String = Path.GetDirectoryName(path);
    if (!Directory.Exists(dir))
        Directory.CreateDirectory(dir);
    oSession.utilDecodeResponse();
    oSession.SaveResponseBody(path);
}

Once you have edited a SWF file, you can use Fiddler's AutoResponder to replace the original file with your modified version.

Limitations

Metadata is currently ignored. I haven't noticed any metadata blocks in any SWF files I've disassembled.
Floating point numbers may not be disassembled with adequate precision.
rabcasm may create a broken file due to not ordering classes by ancestry.

The problem originates from the fact that a class's ancestors (extended class and implemented interfaces) are stored as multinames, and not as class indices. (This makes sense, since classes may extend objects outside the current ABC file.) Since rabcasm currently doesn't decode multinames, it is unaware of the class dependencies, and may thus write the classes out of order. This results in a file that, when opened, will fail to load with an error message similar to:

VerifyError: Error #1014: Class AncestorClassName could not be found.

The simple work-around is to re-order the classes as they are declared in the .main.asasm file, and place ancestors before descendants.

License

RABCDAsm is distributed under the terms of the GPL v3 or later, with the exception of murmurhash2a.d and zlibx.d, which are in the public domain, and asasm.hrc, which is tri-licensed under the MPL 1.1/GPL 2.0/LGPL 2.1. The full text of the GNU General Public License can be found in the file COPYING.

BruceJawn/RABCDAsm