PR #1064

PR #1068

PR #1069

Best guess is that the parser used to use something other than a `&str`,
but I can't remember.

This fixes a bug in how prefilters were applied for multi-regexes
compiled with "all" semantics. It turns out that this corresponds to the
regex crate's RegexSet API, but only its `is_match` routine.

See the comment on the regression test added in this PR for an
explanation of what happened. Basically, it came down to incorrectly
using Aho-Corasick's "standard" semantics, which doesn't necessarily
report leftmost matches. Since the regex crate is really all about
leftmost matching, this can lead to skipping over parts of the haystack
and thus lead to missing matches.

Fixes #1070

The main reason we used mips before was to get test coverage on a big
endian target. Now that mips no longer seems to work[1], I wanted to
add at least one other big endian target. From the tier 2 supported
platforms[2], the only big endian targets I could find were powerpc and
s390x. So we just add both here.

[1]: rust-lang/compiler-team#648
[2]: https://doc.rust-lang.org/nightly/rustc/platform-support.html#tier-2-with-host-tools