IDA Colors

Parser for color tags in IDA.

Usage

$ uv run ida-colors <path-to-idb>
...
    "LDR             X8, [X19,#0x10]"
        Insn("LDR")
        Opnd1(Reg("X8"))
        Opnd2(Symbol("["), Reg("X19"), Symbol(",#"), Number("0x10"), Symbol("]"))
...

Numbers

numbers() returns the integers of a parsed line, with the - printed in front of each one applied:

>>> from ida_colors import numbers, parse_full
>>> numbers(parse_full(ida_lines.generate_disasm_line(ea, 0)))  # STP XZR, XZR, [X1,#-0xD8]
[Number(value=-216, operand=3, standalone=False)]

standalone is true when the number is the entire operand, give or take a #, = or -: #0x10, -1 and =0xCBF29CE4 are, while [X1,#0x10] and [rsp+28h+var_20] are not.

u8() to u64() and i8() to i64() reinterpret the value at a width, so -1 and 0xFFFFFFFF compare equal as u32(), and return None when it doesn't fit. Spellings other than plain decimal, 0x10 and 10h (floats, chars, other radixes) are skipped rather than guessed.

Strings

strings() returns the strings and chars that IDA shows in a line's automatic comments, as printed: quotes removed, escapes kept.

>>> from ida_colors import parse_full, strings
>>> strings(parse_full(ida_lines.generate_disasm_line(ea, 0)))  # lea rsi, aSslSession ; "SSL-Session:\n"
[String(text='SSL-Session:\\n', char=False, addr=844343, truncated=False)]

addr is where the string starts, also when IDA shows it through a pointer ("/dev/urandom" ...). truncated is true when IDA cut it short ("abi.NewName: tag too long: hw.optional."...). Chars (sub esi, 35h ; '5') have char set and no addr. Names shown the same way (; __gmon_start__) and anything else that doesn't look exactly like a string or a char are skipped.

The CLI needs IDA 9+ and the ida extra (idapro). The parser itself has no dependencies.

Development

$ uv run ruff check && uv run ruff format --check
$ uv run ty check
$ uv run pytest

tests/test_ida_constants.py checks the constants copied from IDA against a local IDA install, and is skipped when idapro can't load one. tests/test_ida_numbers.py checks parsed immediates against IDA's operand values on every .i64 in $IDA_COLORS_IDBS (opened as temporary copies), and tests/test_ida_strings.py checks parsed strings against the string literals there. Both are skipped when it's unset.

uv run python bench/bench.py times parse_full(), numbers() and strings() per line on the sample of real disassembly in tests/data/lines.json, by processor. It's not part of CI, as timings on shared runners are too noisy.

tests/test_oracle.py checks that trees, results and errors match tests/ida_colors_ref, a frozen copy of the package, on those lines and on mangled versions of them. Speedups must keep it passing. A deliberate behavior change replaces the copy with the new code.

Why?

IDA's disassembly isn't just the decoded instruction. Between decode_insn and the line you see, the processor module's printer applies a lot of heuristics and user annotations: offsets, stack variables, struct offsets, enums, chars, @PAGE/@PAGEOFF pairs, hidden implicit operands. The printed line is the only place where all of that comes together, and IDA marks it up with color tags. Parsing the tags gets you what IDA decided, without reimplementing its printer for every processor.

The textbook alternative, decode_insn and op_t.type (o_imm, o_mem, o_displ, ...), tells you what the bytes encode, not what they mean: