Refactor char/string and byte search

[jakobnissen]

This is a refactoring of base/string/search.jl. It is purely internal, and comes with no changes in behaviour. It's based on #54593 and #54579, so those needs to get merged first, then this PR will be rebased onto master.

Included changes are:

• The char/string search functions now use the last byte to memchr, not the first byte. Because the last bytes are more varied, this is much faster on small non-ASCII alphabets (like searching Greek or Cyrillic text) and somewhat faster on large non-ASCII ones (like Japanese). Speed on ASCII alphabets (like English) in unchanged.
• Several unused or redundant methods have been removed
• Moved boundschecks from the inner _search and _rsearch functions to the outer top-level functions that call them. This is because the former may be called in a loop where repeated boundschecking is needless. This should speed up search a bit.
• Char/string search functions are now implemented in terms of an internal lazy iterator. This allows findall and findnext to share implementation, and will also make it trivially easy to implement a lazy findall in the future (see Implement lazy findall (Iterators.findall, perhaps?) #43737)

IMO there is still more work to be done on this file, but this requires a decision to be made on #43737, #54581 or #54584

Benchmarks

using BenchmarkTools
using Random

rng = Xoshiro(55)

greek = join(rand(rng, 'Α':'ψ', 100000)) * 'ω'
@btime findfirst('ω', greek)

@btime findfirst(==('\xce'), greek)

english = join(rand(rng, 'A':'y', 100000)) * 'z'
@btime findfirst('z', english)

@btime findall('A', english)
@btime findall('\xff', english)
nothing

1.11.0-beta2:

  100.049 μs (1 allocation: 16 bytes)
  474.084 μs (0 allocations: 0 bytes)
  689.110 ns (1 allocation: 16 bytes)
  93.536 μs (9 allocations: 21.84 KiB)
  72.316 μs (1 allocation: 32 bytes)

This PR:

  1.319 μs (1 allocation: 16 bytes)
  398.011 μs (0 allocations: 0 bytes)
  681.550 ns (1 allocation: 16 bytes)
  8.867 μs (8 allocations: 21.81 KiB)
  683.962 ns (1 allocation: 32 bytes)

[jakobnissen]

This is good to go now. Test failures are unrelated.

[jakobnissen]

Bump. CI failures are unrelated.

[jakobnissen]

Bump

[oscardssmith]

@nanosoldier runbenchmarks(!scalar)

[oscardssmith]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[oscardssmith]

The perf results here look great! We really should try to get this reviewed and merged ASAP

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[jakobnissen]

Nanosoldier found a regression for short ASCII searches, where extra book-keeping in this PR, which speeds up the special cases, but costs around 5 nanoseconds, becomes significant. I've addressed this by manually adding a fast path in findnext and findprev when the char is ASCII.
This fast path would have been hit eventually, but now a bunch of setup code is skipped, saving this handful of nanoseconds.

[oscardssmith]

@nanosoldier runbenchmarks(!"scalar", vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[oscardssmith]

I think this is good to go, but I do want someone else to get another set of eyes on this.

[KristofferC]

Just PkgEval it and if it looks good, merge?

[jakobnissen]

@nanosoldier runtests()

[nanosoldier]

The package evaluation job you requested has completed - possible new issues were detected.
The full report is available.

Report summary

❗ Packages that crashed

2 packages crashed only on the current version.

• A segmentation fault happened: 2 packages

13 packages crashed on the previous version too.

✖ Packages that failed

16 packages failed only on the current version.

• Illegal method overwrites during precompilation: 1 packages
• Package has test failures: 5 packages
• Package tests unexpectedly errored: 1 packages
• Tests became inactive: 2 packages
• Test duration exceeded the time limit: 7 packages

1111 packages failed on the previous version too.

✔ Packages that passed tests

30 packages passed tests only on the current version.

• Other: 30 packages

5325 packages passed tests on the previous version too.

~ Packages that at least loaded

20 packages successfully loaded only on the current version.

• Other: 20 packages

2935 packages successfully loaded on the previous version too.

➖ Packages that were skipped altogether

908 packages were skipped on the previous version too.