From what I can read the author got really unlucky with some kind of radical API changes. Maybe at that time the LLVM team was a bit less serious with deprecations ?
I use LLVM since v9, nowadays I'm stuck on v15 (that's not because of LLVM btw).
Between the two versions there's been a radical change too, i.e "opaque pointers", but the transition was rather smooth as we were provided, for a long time, the two versions of the functions affected by the change. Maybe the LLVM team got more serious since the author experienced the said difficulties ?
Other thing I note is that the author uses the CPP API. I use the C one which exposes only a high-level subset of the CPP one. This encourages a saner use of LLVM, a more concrete separation between the front-end and the mid-end, although sometimes there are limitations.
A simple example of what encourages the C API, especially since opaque ptrs are added, is not to rely on LLVM to retrieve the IR type of an IR value. That should always be done using the AST, eg with an `.ir` field in your nodes.
Another one I remark, after a brief overview of LLVM-CRAP, is that the author had to change the internal data structure used, depending on the LLVM version [0]. Using the C API that would never had happened. The C API essentially allows to create block refs, instructions refs, value refs, type refs, contexts. Then you choose the containers you want to use to hold them. No need to switch to another stdcpp one, even if internally LLVM does so.
[0]: https://github.com/c3d/xl/blob/master/src/llvm-crap.cpp#L265