Reimplement idna on top of ICU4X

[hsivonen]

Opening as draft PR to enable early feedback while the dependency remains unlanded in ICU4X.

The motivation of reformulating the idna crate on top of ICU4X is to be able to move Firefox's IDNA handling to use ICU4X (instead of the current combination of ICU4C and very old code). The ICU4X normalizer is faster than unicode-normalization and the ICU4X normalizer represents UTS 46 data as a normalization as opposed to representing it separately like the idna crate currently does.

The benchmarks in the idna crate itself show this PR to result in faster performance. This is also more correct than the old code: I removed skipping of the ContextJ tests from the harness that runs the UTS 46 test suite.

See the added README for removed capabilities. I searched for GitHub for public code using the idna crate, and I believe the removals to be mostly not need action from the ecosystem and to be tolerable when they do.

For projects that use ICU4X for normalization (or collation), this change has the benefit of deduplicating data across normalization and IDNA handling. There is the ecosystem risk of causing projects that use unicode-normalization for normalization in ways other than as a dependency of idna to end up with more data. One way to mitigate that (already preliminarily discussed with the maintainer) would be to introduce a cargo feature to unicode-normalization that would delegate the unicode-normalization internals to ICU4X (better performance, more crates in the dependency tree).

Not properly investigated yet: Binary size impact.

[valenting]

I think you need to explicitly add a dependency for the icu crate, instead of using a relative path

[valenting]

We need to add a direct dependency on the icu crates

[hsivonen]

Yeah, the dependency declaration will change when this becomes a non-draft.

[hsivonen]

Using the demo https://github.com/hsivonen/urldemo (strip=true, lto=true, opt_level="z") and binaryen wasm-opt -Oz on the result, I get 215085 bytes with this patch and 310986 without, so this should not only improve performance but should also make (Wasm at least) binary size smaller.

[codecov]

Codecov Report

Attention: Patch coverage is 78.66795% with 221 lines in your changes are missing coverage. Please review.

Please upload report for BASE (main@de947ab). Learn more about missing BASE report.

Files	Patch %	Lines
idna/src/uts46.rs	76.89%	165 Missing ⚠️
idna/src/punycode.rs	73.13%	18 Missing ⚠️
idna/tests/deprecated.rs	82.35%	18 Missing ⚠️
idna/src/lib.rs	44.44%	10 Missing ⚠️
idna/tests/uts46.rs	85.41%	7 Missing ⚠️
idna/src/deprecated.rs	96.29%	3 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #923   +/-   ##
=======================================
  Coverage        ?   79.94%           
=======================================
  Files           ?       22           
  Lines           ?     4208           
  Branches        ?        0           
=======================================
  Hits            ?     3364           
  Misses          ?      844           
  Partials        ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[hsivonen]

I looked at the red lines for non-test code in the coverage report, and it's not as awful as the percentages suggest. In particular, it looks like const-evaluated code shows up as uncovered. Could be better, of course.

[hsivonen]

I did the minimum possible version bump for url before realizing that url now rejects URLs whose domain violates the ContextJ rule. (Additionally, the error classification for URLs that violate the forbidden domain code point rule changed, but that change doesn't change whether or not these is an error.)

[hsivonen]

From the point of view of url users, the changes are:

• Change in dependency tree. I believe not a semver break per community standards.
• Change in MSRV. Not a semver break per community standards.
• Change in which error kind is reported for forbidden domain code point. Rather silly to treat as a semver break.
• URLs whose domain violates the ContextJ rule are now rejected (as they should have been all along). Semver break or not?

[hsivonen]

The dependency version increment is about dealing with an issue in the transitive dependency declarations which caused a problem in the case where a dependent already had the _data crates downloaded at 1.4.0.

Note that 1.5 versions have also already been released.

GitHub shows this changeset as "Changes requested". I believe I've addressed the request already, but I don't see what GitHub UI action I need to take to dismiss the "Changes requested" state.

[djc]

From the point of view of url users, the changes are:

* Change in dependency tree. I believe not a semver break per community standards.

* Change in MSRV. Not a semver break per community standards.

* Change in which error kind is reported for forbidden domain code point. Rather silly to treat as a semver break.

* URLs whose domain violates the ContextJ rule are now rejected (as they should have been all along). Semver break or not?

Doesn't sound like a semver break is needed.

GitHub shows this changeset as "Changes requested". I believe I've addressed the request already, but I don't see what GitHub UI action I need to take to dismiss the "Changes requested" state.

I think only the maintainers themselves can dismiss this.