fingerprint-toolchain
Use at the start of a byte-matching decompilation of a DOS executable (or, with the armcc5 profile, an Arm Cortex-M firmware image; with the ti-c2000 profile, a TI C2000 DSP image), before any matching, to identify the exact compiler, assembler, linker, memory model and switches that built it, and to prove the identification by compiling a function to the same bytes.
abedegno · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add abedegno/Exhume --skill fingerprint-toolchain技能说明
name: fingerprint-toolchain description: Use at the start of a byte-matching decompilation of a DOS executable (or, with the armcc5 profile, an Arm Cortex-M firmware image; with the ti-c2000 profile, a TI C2000 DSP image), before any matching, to identify the exact compiler, assembler, linker, memory model and switches that built it, and to prove the identification by compiling a function to the same bytes.
Fingerprint the toolchain
Paths are relative to the Exhume checkout. Nothing else in the method works until one function compiles to the original's bytes, so this step ends with that proof, not with a guess.
Steps
- Run
python3 tools/fingerprint.py GAME.EXE. Note the MZ facts, whether there is an overlay block (FBOV is Borland VROOMM), the counts of standard frames (55 8B EC),enter, far calls, INT 3Fh thunks and far returns, and which profile's runtime strings it finds. - Read the runtime strings for a version, not just a vendor. "Turbo C++ - Copyright 1990 Borland Intl." is Turbo C++ 1.0x, not Borland C++ 2.0 or 3.x. Get that exact version: switches, code generation and the runtime library differ between point releases.
- Find the memory model from the code: far calls (
9A) between segments with globals addressed near (mov bx,[828Ah]) mean medium, as in UW2; data reached throughles bx,[...]is an explicit far pointer there. UW2 also showed DS equal to SS: a stack array is indexed with noss:override. - Set DGROUP: fingerprint.py suggests a paragraph from C0's copyright string. Confirm it with a string the code references by DS offset (UW2: "trap" at DS:1AEB is at file 0x6A57B, so DS:0 is file 0x68A90, paragraph 0x65E9). Put it in exhume.toml as
dgroup_para. - Get the toolchain from the user's own disk images and unpack it with the profile's setup scripts (
profiles/borland-tc101/setup-tc.sh,setup-tasm.sh). They check MD5s. Never add compiler binaries or game files to any repository. Then choose the DOS it runs in (docs/method.md, "Choosing the DOS"): build emu2 withsh tools/setup-emu2.sh(a compile then takes about half a second rather than several in js-dos, which matters over the hundreds of compiles fingerprinting takes), and check withnode tools/dosbackend.mjs. - Pick one small, self-contained function (no calls, or calls only to other files). Write C for it, put
/* target: SEG */and/* opts: ... */lines at the top, make a target table row for it, and runpython3 tools/match.py src/FILE.C --dis NAME. Iterate on the source and the switches until it says MATCH. Then do two more functions from different segments. - Find the switches by the differences they make, one at a time:
-G(add sp,Ncleanup, noenter),-O,-Z(register reuse across stores),-Y(overlay code: a same-file function address is a relocated segment, notmov ax,cs),-d(duplicate strings stored once in the data segment),-k-(no frame in argument-less functions). Expect them to vary per file; each file names its own in its opts line. - For assembly, assemble a routine rebuilt from its disassembly. Padding is the tell: TASM 2.0 single-pass writes an unshortened forward jump as the short jump plus
nops. If the padding comes out the same, you have the assembler.
Arm Compiler 5 (the armcc5 profile)
The same proof, other tells; profiles/armcc5/compiler.md has the evidence.
- Switches from the code: a short conditional as compare and branch is
-O1, as an IT block-O2or-O3; a 16-bit divide by a constant as multiply and shift needs-Otime; a literal pool after every function means--split_sections. MicroLIB shows in its compact__mainand__scatterload. - The release, not just the version: armcc 5's updates compile nearly all code alike, so one
matching function proves little. Install every MDK release you can get (each is one toolchain
home), match every merged file under each with
profiles/armcc5/tools/armbuilds.py, and read the files that only some releases match. On one firmware every file matched under 5.05u1 to 5.06u2, and a single library function, 2 bytes longer from 5.06u3 on, settled the range. Repeat this when the tree is nearly done: the last stubborn functions are the best test, and a function nothing matches may be the compiler's, not the source's. - The libraries are evidence too: MicroLIB's members come from the MDK release (link the image's
_printf_coreagainst each release'smc_w.l), and a vendor peripheral library's functions can match one release of its source and not the next (compile every release). armcc --vsn, and the.commentsection of any object, name the build ("ARM Compiler 5.06 update 2 (build 183)"); several MDK releases ship the same build.
TI C2000 (the ti-c2000 profile)
The same proof; profiles/ti-c2000/compiler.md, "Fingerprinting the version", has the evidence.
- The runtime library first:
profiles/ti-c2000/tools/c28libfp.py --libs 'GLOB/lib/rts2800_fpu32.lib'places each release's members in the image with relocated words masked and prints how many each places. The library can be older than the compiler, so it bounds the build, not the compiler. - Then the compiler: install every release you can get from TI's public server into one Wine
prefix, and run
match.py SRC --all-cgton matched files. Defect fixes and changed code shapes pin versions (an interrupt epilogue'sSETC INTM|DBGMbeforePOP RBfrom 15.12.2; a struct member past 255 words throughMOVL XAR0,#kfrom 15.12.7 and 16.9.5; the -Ooff frame layout from 16.9.0). Neighbouring releases break different functions; the right one breaks none. - A file that "needs" another compiler is evidence about the whole tree: try the whole tree under
that compiler before giving one file a
/* cgt: */line. On one firmware every file matched under the release a few files seemed to need, once functions bent to fit the first guess were written plainly again. -O1, not -O2: o