You can find the source at https://hg.ulukai.org/ecm/inicomp/file/1229f30865d2/x.asm

As noted in the initial commit it is based on git 71afd5b05eb4.

To save on the memory usage, I limited the ctFreq field to 16 bits. (Carry out of the top bit is detected as an error.) Even so, the context tables require 256 KiB. I also had to add some layer threshold memory (the word "layout" here is in error) so there's enough space to decompress subsequent layers. (The 96 kB kernel compresses to 62 kB and uses two layers.)

I did not yet inline function calls only used once. However, I do not expect much speed-up from that. I also did not yet add support for overlapping source and destination, which all four of the other decompressors do support.

Its speed seems to place it as the slowest of all five decompressors yet. I think the reason for that is that every byte has to call into bio_read_gr, increment_frequency, and update_model. I was able to modify/create all other (mostly LZ-like) decompressors to use repeated string instructions to write several literal or match bytes in one step.

It is also the first decompressor to need additional memory (other than stack frame variables). And the compression ratio is not especially good either. However, the port was interesting to me. Thanks for your work!