deflate_quick is only called at compression level 1 on x86_64.
There is too much code to implement for aarch64/aarch32 before the release.
We may revisit this after the release.

ARM doesn't have anything similar to the pcmpestri SSE4.2 instruction used by the compare258 function of deflate_quick. This instruction is a huge part of the performance: it's used to do a miniature strstr-like operation to find matching patterns in the dictionary.

It seems like we could replace match_len = compare258(s->window + s->strstart, s->window + s->strstart - dist) with match_len = longest_match(s, hash_head) if SSE4.2 not supported. If that is the case, perhaps we could extend compare258 to deflate_slow, deflate_medium, etc if SSE4.2 is supposed. So of like a functable.longest_match. What do you think?

compare258 is essentially compare of 256 bytes followed by compare of trailing 2 bytes... Only difference is that for all iterations of the loop, it needs to find the first non-matching index in the vector and calculate length from it... If we assume that it is likely that all lanes are equal most of the time, it should not have performance penalty when the trailing bytes are not equal. Trailing bytes can be handled using uint8_t, uint16_t, uint32_t and uint64_t.

This is now complete.