If you want the decompiled code to produce a 1:1 match with the original binary (even if it takes some finessing by hand at the end) you need something rigorous approaching arithmetic. A fuzzy decompiler that just approximates the intent of the original code won't get you there (and this is mostly what you get out of GPT for many tasks), but it could still be useful for something.
There's been research into the latter but its in early stages. https://github.com/nforest/awesome-decompilation?tab=readme-...
decomp.me gives us a large database of C(++) <-> target asm to train a model on ;)