6502bench SourceGen: Intro

Back to index

Overview

SourceGen converts 6502/65C02/65816 machine-language programs to assembly-language source.

SourceGen has two purposes. The first is to be a really nice disassembler for the 6502 and related CPUs. Code tracing with status flag tracking makes it easier to separate the code from the data, automatic formatting of ASCII strings and filled-data areas helps get the data regions sorted out, and modern IDE-style features like cross-reference generation and color-highlighted bookmarks help navigate the code while trying to figure out what it does. A disassembler should help you understand the code, not just dump the instructions to a text file.

The computer I built in 2014 has a 4GHz CPU and 8GB of RAM. I figured we should put that kind of power to good use.

The second purpose is to facilitate sharing and collaboration. Most disassemblers generate output for a specific assembler, or in a way that's generic enough to match most any assembler; either way, you're left with a text file in somebody's idea of the "correct" format. SourceGen keeps everything in an assembler-neutral format, and provides numerous options for customizing the display, so that multiple people viewing the same project can each do so with the conventions they are accustomed to. Code and data operands can be formatted in various numeric formats or as symbols. The project file uses a text format that is fairly diff-friendly, so sharing projects through git works reasonably well. If you want source code you can assemble, SourceGen will generate code optimized for the assembler of your choice.

The sharing and collaboration ideas only work if the formatting capabilities within SourceGen are sufficiently flexible. If you need to generate assembly source and tweak it a bunch to express the intent of the original code, then passing a SourceGen project around won't work. This sort of thing is a bit outside the bounds of what a typical disassembler does, so it remains to be seen whether SourceGen succeeds at what it's trying to do, and also whether what it's trying to do is something that people actually want.

You can get started by watching the demo video and playing with the tutorials.

Fundamental Concepts

The next few sections present some general concepts and terminology. The rest of the documentation assumes you've read and understood this. It will be helpful if you already understand something about the 6502 instruction set and assembly-language programming, but disassembling other programs is actually a pretty good way to learn how to code in assembly.

About 6502 Code

For brevity's sake, "6502 code" should be taken to mean "code for the 6502 CPU or any of its derivatives, including but not limited to the 65C02 and 65816". So let's talk about 6502 code.

Code usually arrives in a big binary blob. Some of it will be instructions, some of it will be data, some will be empty space used for variable storage. Part of the challenge of disassembly is identifying which parts of the file contain which.

Much of the code you'll find for the 6502 was written by humans, rather than generated by a compiler, which means it won't conform to a standard set of conventions. However, most programmers will use subroutines, which can be identified and analyzed in isolation. Subroutines are often interspersed with variable storage, referred to as a "stash". Variables may be single-byte or multi-byte, the latter typically in little-endian byte order.

Much of the data in a typical program is read-only, often in the form of graphics or ASCII string data. Graphics can be difficult to recognize automatically, but strings can be identified with a reasonable degree of confidence. Address tables, which are a collection of addresses to other things, are also fairly common.

A simple disassembler would start at the top of the file and just start converting bytes to instructions. Unfortunately there's no reliable way to tell the difference between instructions, data, and variable stashes. When the converter hits data bytes it'll start generating instructions that won't make sense. You'll have another problem when the data ends and code resumes: 6502 instructions are variable-length, so if the last byte of the data area appears to be a three-byte instruction, the first two bytes of the next instruction area will be gobbled up.

Some programmers will use a trick where they "embed" an instruction inside another instruction. This allows code to branch to two different entry points, one of which will set a flag or load a register, and then continue on to common code.

Another trick is to embed "inline data" after a JSR or JSL instruction. The caller pulls the calling address off the stack, uses it to access the parameters, then pushes the address back on after modifying it to point to an address past the inline data. This can be very confusing for the disassembler, which will try to interpret the inline data as instructions.

Sometimes code is loaded at one location, then moved to another and executed there. If you're disassembling an executing program you don't have to worry about this, but if you're disassembling the binary from the loadable file on disk then you need to track the address changes. The address is communicated to the assembler with a "pseudo-opcode", usually something like "ORG". Other pseudo-op directives are used to define external symbols and (for 65816 code) register widths.

The 8-bit CPUs have a 16-bit (64KiB) address space, so addresses can range from $0000 to $ffff. (I'm going to write hex values with a preceding '$', like "$12ab", rather than "0x12ab" or "12abh", because that's what 6502 systems commonly used.) The 65816 has a 24-bit address space, but it's not contiguous -- a branch that extends past the end will wrap around to the start of the 64KiB "bank". For 16-bit instruction operands, the bank is identified for instruction and data addresses by the program bank register and the data bank register, respectively. The disassembler can't generally know the contents of the data bank register, which makes life a bit more interesting.

The 6502 has an 8-bit processor status register ("P") with a bunch of flags in it. Some of the flags determine whether a conditional branch is taken or not, which is important because some branches appear to be conditional but actually are always or never taken in practice. The disassembler needs to be able to figure this out so that it doesn't try to disassemble the bytes that follow an always-taken branch. A more significant concern is the M and X flags found on the 65802/65816, which determine the width of the registers and of immediate load instructions. If you don't know what state the flags are in, you can't know whether LDA #value is two bytes or three, and the disassembly of the instruction stream will come out wrong.

How SourceGen Works

SourceGen employs a partial emulation technique that traces the flow of execution. Most of what a given instruction does isn't important; only its effect on the flow of execution matters.

The code tracing has to start somewhere, so SourceGen uses "code entry point hints" to identify places where execution may begin. By default, a hint is placed at the start of the file. From there, the tracing process walks through the code, pursuing all branches. In many cases, if you mark all external entry points, SourceGen will automatically find all executable code and separate it from variable storage and data areas.

As noted earlier, tracking the processor status flags can make the analysis more accurate. Identifying situations where a branch instruction is always or never taken avoids mis-categorizing a data region as code. On the 65816, it's absolutely crucial to track the M/X flags, since those affect the width of instructions. SourceGen tracks the value of the processor flags at every instruction, blending sets of flags together when multiple paths of execution converge.

Once instructions and data have been separated, the instruction operands can be examined. Branches, loads, and stores that reference an address that falls inside the address space covered by the file can be replaced with a symbol. Operands that refer to addresses outside the file, such as ROM or operating system routines, can be replaced with a symbol defined by an equate directive.

(For more details on how this works, see the analysis appendix.)

Extension Scripts

Extension scripts are C# source files that are compiled and executed by SourceGen. They can be added to a project from SourceGen's runtime data directory, or can live in the directory next to the project file.

In the current implementation, scripts are only called to examine JSR/JSL instructions. They can format nearby bytes as inline data, or apply symbols to operands.

To reduce the chances of a script causing problems, all scripts are executed in a sandbox with severely restricted access. Notably, nothing in the sandbox can access files, except to read files from the PluginDll directory.

The PluginDll directory lives next to the SourceGen executable, and contains all of the compiled script DLLs, as well as two pre-built application DLLs that plugins are allowed access to. The contents are persistent, to avoid recompiling the scripts every time SourceGen is launched, but may be manually deleted without harm.

Analyzer Hints

Sometimes SourceGen can't automatically find the start or end of an instruction stream, or gets confused by inline data. These situations can be resolved by adding an appropriate hint.

Code entry point hints tell the analyzer to add the offset to the list of instruction start points. Suppose you've got a code library that begins with jump vectors, like this:

1000: 4c0910    JMP     $1009
1003: 4cef10    JMP     $10ef
1006: 4c3012    JMP     $1230
1009: 18        CLC

When opened with SourceGen, it will look like this:

         .ORG    $1000
         JMP     L1009

         .DD1    $4c
         .DD1    $ef
         .DD1    $10
         .DD1    $4c
         .DD1    $30
         .DD1    $12
L1009    CLC

SourceGen doesn't see any code that jumps to $1003 or $1006, so it assumes those are data. Further, the functions at those addresses may also be considered data unless some bit of code reachable from L1009 calls into them. If you add a code hint to $1003 and $1006, you'll get better results: 

         .ORG    $1000
         JMP     L1009
         JMP     L10ef
         JMP     L1230
L1009    CLC

Be careful that you only add hints to the instruction opcode. If you applied hints to the full range of bytes from $1003 to $1008, you would end up with this:

         .ORG    $1000
         JMP     L1009
         JMP ⏩  L10ef
         BPL ⏩  L1053
         JMP ⏩  L1230
         BMI     L101b
L1009    CLC

The exact set of instructions shown depends on your CPU configuration. The problem is that the bytes in the middle of the instruction have been marked as entry points, and SourceGen is treating them as embedded instructions. $EF and $12 aren't valid 6502 opcodes, so they're being ignored, but $10 is BPL and $30 is BMI. Because hinting multiple consecutive bytes is rarely useful, SourceGen only applies code hints to the first byte in a selected line.

Data hints tell the analyzer when it should stop. For example, suppose address $ff00 is known to always be nonzero, and the code uses that fact to get a branch-always on the 6502:

         .ORG    $1000
         LDA     $ff00
         BNE     L1010
         BRK     $11

By placing a data hint on the BRK, you're telling the analyzer that it should stop the current path of execution. (Note that this example would actually be better solved by setting a status flag override on the BNE that sets Z=0, so the code tracer will know it's a branch-always and do the right thing.) It's only necessary to place a hint on the very first (opcode) byte. Placing a data hint in the middle of what SourceGen believes to be instruction will have no effect.

As with code hints, only the first byte in each selected line will be hinted.

Inline data hints identify bytes as being part of the instruction stream, but not instructions. A simple example of this is the ProDOS 8 call interface on the Apple II, which looks like this:

         JSR     $bf00
         .DD1    $function
         .DD2    $address
         BCS     BAD

The three bytes following the JSR $bf00 should be hinted as inline data, so that the code analyzer skips them and continues the analysis at the BCS. Because you need to hint *every* byte of inline data, all bytes in a selected line will receive hints.

If code branches into a region that is marked as inline data, the branch will be ignored.

SourceGen Concepts

As you work on a disassembled file, formatting operands and adding comments, everything you do is saved in the project file as "meta data". None of the data from the file being disassembled is included. This should allow project files to be shared without violating the copyright of the work being disassembled. (This will vary by region. Also, note that the mere act of disassembling a piece of software may be illegal in some cases.)

To avoid mix-ups where the wrong data file is used, the file's length and CRC are stored in the project file. SourceGen will refuse to open a project if the data file's length and CRC don't match.

Most of the data in the project file is associated with a file offset. When you create a comment, you aren't associating it with line 53, you're associating it with the 127th byte in the file. This ensures that, as the project evolves, the comment you wrote is always connected to the same instruction or data item. This also means you can't have two comments on the same line -- each offset only has room for one. By convention, file offsets are always shown as a six-digit hexadecimal value with a leading '+', e.g. "+0012ab". This makes it easy to distinguish between an address and a offset.

Instruction and data operands can be formatted in various ways. The formatting choice is associated with the first offset of the item. For instructions the number of bytes in the operand is determined by the opcode (and, on the 65816, the M/X status flags). For data items the length can be a single byte or an entire file. Operand formats are not allowed to overlap.

When an instruction or data operand references an address, we call it a numeric reference. When the target address has a label, and the operand uses that symbol, we call that a symbolic reference. SourceGen tries to establish symbolic references whenever possible, so that the generated assembly source doesn't refer to hard-coded locations within the program. Labels are generated automatically for the targets of numeric references.

As your understanding of the disassembled code develops, you will want to add comments explaining it. SourceGen projects have three kinds of comments:

  1. End-of-line comments. As the name implies, these appear at the end of a line, to the right of the opcode or operand.
  2. Long comments, also known as multi-line comments. These get a line all to themselves, and may span multiple lines.
  3. Notes. Like long comments, these get a line to themselves. Unlike long comments, these do not appear in generated assembly code. They are a way for you to leave notes to yourself, perhaps "don't forget to figure this out" or "this is the cool part".

Every file offset can have one of each.

Labels and comments may disappear if you associate them with a file offset that is in the middle of a multi-byte instruction or data item. For example, suppose you put a long comment at offset +000010, and then mark a 50-byte region starting at offset +000008 as an ASCII string. The comment won't be deleted, but won't be displayed either. The same thing can happen to labels. SourceGen will try to prevent this from happening by splitting formatted data into sub-regions at label boundaries.

All About Symbols

A symbol has two parts, a label and a value. The label is a short ASCII string; the value may be an 8-to-24-bit address or a numeric constant. Symbols can be defined in different ways, and applied in different ways.

The label syntax is restricted to a format that should be compatible with most assemblers:

Label comparisons are case-sensitive, as is customary for programming languages.

Platform symbols are defined in platform symbol files. These are named with a ".sym65" extension, and have a fairly straightforward name/value syntax. Several files for popular platforms come with SourceGen and live in the RuntimeData directory. You can also create your own, but they have to live in the same directory as the project file.

Platform symbols can be addresses or constants. If an instruction or data operand references an address outside the scope of the data file, SourceGen looks for a symbol with a matching address value. If it finds one, it automatically uses that symbol. Symbolic constants can be used the same way, but are not matched automatically. This makes them useful for things like operating system function numbers.

If two platform symbols have the same value, the one whose label comes first alphabetically is used. If two platform symbols have the same label, the most recently read one is kept.

Project symbols behave like platform symbols, but they are defined in the project file itself. The editor will prevent you from creating two symbols with the same name. If two symbols have the same value, the one whose label comes first alphabetically is used.

Project symbols always have precedence over platform symbols, allowing you to redefine symbols within a project. (You can "hide" a platform symbol by creating a project symbol with the same name and an unused value, such as $ffffffff.)

User labels are labels added to instructions or data by the user. The editor won't allow you to add a label that conflicts, but if you manage to do so, the user label takes precedence over project and platform symbols. User labels may be tagged as local, global, or global and exported. Local vs. global is important for the label localizer, while exported symbols can be pulled directly into other projects.

Auto labels are automatically generated labels placed on instructions or data offsets that are the target of operands. They're formed by appending the hexadecimal address to the letter "L", with additional characters added if some other symbol has already defined that label. Auto labels are only added where they are needed. Because auto labels may be redefined or disappear, the editor will try to prevent you from referring to them when editing operands.

Operands may use parts of symbols. For example, if you have a label MYSTRING, you can write:

MYSTRING .STR    "hello"
         LDA     #<MYSTRING
         STA     $00
         lda     #>MYSTRING
         STA     $01

The format editor allows you to choose which part of the symbol's value to use. If the value doesn't match exactly, an adjustment will be applied.

Weak References

Symbolic references in operands are "weak references". If the named symbol exists, the reference is used. If the symbol can't be found, the operand is formatted in hex instead.

It's important to know this when editing a project. Consider the following trivial chunk of code: 

1000: 4c0310     JMP     $1003
1003: ea         NOP

When you load it into SourceGen, it will be formatted like this:

         .ORG    $1000
         JMP     L1003
L1003    NOP

The analyzer found the JMP operand, and created an auto label for address $1003. It then created a weak reference to "L1003" in the JMP operand.

If you edit the JMP instruction's operand to use the symbol "FOO", the results are probably not what you want:

         .ORG    $1000
         JMP     $1003
         NOP

This happened because you added a weak reference to "FOO" in the operand, but the label doesn't exist. The operand is formatted as hex. Because there's no longer a reference to L1003, SourceGen removed the auto-label as well.

If you set the label "FOO" on the NOP instruction, you'll see what you probably wanted:

         .ORG    $1000
         JMP     FOO
FOO      NOP

You don't actually need the explicit reference in the JMP instruction. If you edit the JMP operand and set it back to "Default", the code will still look the same. This is because SourceGen identified the numeric reference, and automatically added a symbolic reference to the label on the NOP instruction.

However, suppose you didn't actually want FOO as the operand label. You can create a project symbol, BAR with the value $1003, and then edit the operand to reference BAR instead. Your code would then look like:

BAR      .EQ     $1003
         .ORG    $1000
         JMP     BAR
FOO      NOP

If you change the value of BAR in the project symbol file, the operand will continue to refer to it, but with an adjustment. For example, if you changed BAR from $1003 to $1007, the code would become:

BAR      .EQ     $1007
         .ORG    $1000
         JMP     BAR-4
FOO      NOP

If you rename a label, all references to that label are updated. For numeric references that happens implicitly. For explicit operand references, the weak references are updated individually. (Modern IDEs call this "refactoring".)

If you remove a label, all of the numeric references to it will reference something else, probably a new auto label. Weak references to the symbol will break and be formatted as hex, but will not be removed. Similarly, removing symbols from a platform or project file will break the reference but won't modify the operands.

Parts and Adjustments

Sometimes you want to use part of a label, or adjust the value slightly. (I use "adjustment" rather than "offset" to avoid confusing it with file offsets.) Consider the following example:

1000: a910      LDA     #$10
1002: 48        PHA
1003: a906      LDA     #$06
1005: 48        PHA
1006: 60        RTS
1007: 4c3aff    JMP     $ff3a

This pushes the address of the JMP instruction ($1007) onto the stack, and jumps to it with the RTS instruction. However, RTS requires the address of the byte before the target instruction, so we actually push $1006.

The disassembler won't know that offset $1007 is code because nothing appears to reference it. After adding a code hint at $1007, the project looks like this:

         LDA     #$10
         PHA
         LDA     #$06
         PHA
         RTS

         JMP     $ff3a

We set a label called "NEXT" on the JMP instruction, and then edit the two LDA instructions to reference the high and low parts, yielding:

         .ORG    $1000
         LDA     #>NEXT
         PHA
         LDA     #<NEXT-1
         PHA
         RTS

NEXT     JMP     $ff3a

SourceGen will adjust label values by whatever amount is required to generate the original value. If the adjustment seems wrong, make sure you're selecting the right part of the symbol.

Different assemblers use different syntaxes to form expressions. This is particularly noticeable in 65816 code. You can adjust how it appears on-screen from the app settings.

Automatic Use of Nearby Targets

Sometimes you want to use a symbol that doesn't match up with the operand. SourceGen tries to anticipate situations where that might be the case, and apply adjustments for you.

Suppose you have the following:

         .ORG    $1000
         LDA     #$00
         STA     L1010
         LDA     #$20
         STA     L1011
         LDA     #$e1
         STA     L1012
         RTS

L1010    .DD1    $00
L1011    .DD1    $00
L1012    .DD1    $00

Showing stores to three different labeled addresses is fine, but the code is actually setting up a single 24-bit address. For clarity, you'd like the output to reflect the fact that it's a single, multi-byte variable. So, if you set a label at $1010, SourceGen removes the nearby auto labels, and sets the numeric references to use your label: 

         .ORG    $1000
         LDA     #$00
         STA     DATA
         LDA     #$20
         STA     DATA+1
         LDA     #$e1
         STA     DATA+2
         RTS

DATA     .DD1    $00
         .DD1    $00
         .DD1    $00

If you decide that you really wanted each store to have its own label, you can set labels on the other two addresses. SourceGen won't search for alternate labels if the numeric reference target has a user-defined label.

This is also used for self-modifying code. For example:

1000: a9ff      LDA     #$ff
1002: 8d0610    STA     $1006
1005: 4900      EOR     #$00

The above changes the EOR #$00 instruction to EOR #$ff. The operand target is $1006, but we can't put a label there because it's in the middle of the instruction. So SourceGen puts a label at $1005 and adjusts it:

         LDA     #$ff
         STA     L1005+1
L1005    EOR     #$00

If you really don't like the way this works, you can disable the search for alternate targets entirely from the project properties. Self-modifying code will always be adjusted because of the limitation on mid-instruction labels.

Width Disambiguation

It's possible to interpret certain instructions in multiple ways. For example, "LDA $0000" might be an absolute load from a 16-bit address, or it might be a direct page load from an 8-bit address. Humans can infer from the fact that it was written with a 4-digit address that it's meant to be absolute, but assemblers often treat operands purely as numbers, and would just see "LDA 0". Common practice is to use the shortest instruction possible.

Every assembler seems to address the problem in a slightly different way. Some use opcode suffixes, others use operand prefixes, some allow both. You can configure how they appear in the application settings.

SourcGen will only add width disambiguators to opcodes or operands when they are needed, with one exception: the opcode suffix for long (24-bit address) operations is always applied. This is done because some assemblers require it, insisting on "LDAL" rather than "LDA" for an absolute long load, and because it can make 65816 code easier to read.

Data and Directive Pseudo-Opcodes

There are only three assembler directives that appear in the code list:

Every data item is represented by a pseudo-op. Some of them may represent hundreds of bytes and span multiple lines.

In addition, several pseudo-ops are defined for string constants:

If the characters have their high bits set -- commonly referred to as "high ASCII" -- an upward arrow will be added to the pseudo-op. How these strings are generated into assembly source varies.

You can configure these to look more like what your favorite assembler uses in the application settings.