[VishnuSrivatsava]

Problem

Hir::alternation calls lift_common_prefix, which rewrites XA|XB into X(?:A|B). When the common prefix X can match in more than one way, this rewrite is not leftmost-first equivalent. In XA|XB, every way of matching X followed by A is preferred over every way of matching X followed by B. But in X(?:A|B), X's most preferred way followed by B completes before X's less preferred way followed by A, since the lifted X never backtracks into its less preferred ways once the continuation matches.

Minimal example: a+a|a+b against aab.

Regex::new("a+a|a+b").unwrap().find("aab")
// expected: Some(0..2), matching "aa" (first branch, Perl style)
// actual:   Some(0..3), matching "aab"

This is #1400 (diagnosis credit to @ranflarion, who connected the lift_common_prefix rewrite to this behavior by differential testing against java.util.regex). It also manifests through ripgrep: BurntSushi/ripgrep#3568. The behavior has been present since regex-syntax 0.7.0 (regex 1.8.0), where the lift was introduced.

Solution

Skip the lift when the common prefix can match in more than one way. A new conservative analysis is_multi_way reports whether an expression has more than one distinct way to match; the lift now only applies to single way prefixes (literals, classes, exact count repetitions of single way expressions and concatenations of those). Some safe lifts may now be skipped, but no unsafe lift is performed.

Test

A unit test in regex-syntax asserts that a+a|a+b remains an alternation after translation while a single way prefix case still lifts. An integration test in tests/regression.rs asserts the correct match spans.

Verified: full workspace test suites pass (about 1100 tests), and the match span is correct across engine configurations (default and minimal size limit).

Fixes #1400

[VishnuSrivatsava]

Closing this as a duplicate of #1401, which implements the same fix with a better shape: rather than skipping the lift entirely, the prefix walk stops at the first element that can match in more than one way, so partial single way prefixes are still lifted. Sorry for the noise — I checked this issue for existing discussion but missed the linked pull request in the timeline.

Before closing I verified #1401 independently: with its branch, Regex::new("a+a|a+b").find("aab") returns Some(0..2) across engine configurations (default and minimal size limit), and the full regex and regex-syntax test suites pass on the branch. The verification table in this PR's description (all working output modes unchanged, both failure-shaped inputs corrected) applies to #1401 as well. @ranflarion the HIR equivalence test style in your PR is nice; an end-to-end span assertion like the one in this PR's tests/regression.rs commit could be a useful addition if you want it, but it is not required.