Merge compiler flags in linear time

[OskarEichler]

Summary

Retain the last occurrence of mergeable -D/-W flags using a reverse traversal and a seen set; preserve all nonmergeable flags and their order.

Reproduction and verification

The previous Array#delete on each mergeable flag repeatedly scans the output. A 1,000-distinct-flag input takes quadratic work. The candidate visits each input once and reverses the result once. A fixed-seed corpus of 1,000 mixed arrays matches the original algorithm exactly, including -Wa/-Wl/-Wp exceptions and frozen input arrays.

Paired Ruby 4.0.6 benchmark: 12 rounds, two warmups. For 100 merges of 1,000 distinct flags: median 0.250664s -> 0.014598s; repeated flags: 0.136190s -> 0.010624s. For 20,000 six-flag merges: 0.020347s -> 0.020816s, with one extra hash allocation per call. No universal speedup or production-throughput claim.

Existing suite on Ruby 4.0.6 through rbenv: 27 tests, 31 assertions, zero failures/errors, two upstream macOS omissions, baseline and this isolated patch. Combined installed-release consumer also passes the suite, 1,098 focused/model checks, four process cases, five linker-option cases and four path cases. rbenv exec ruby test/run.rb runs the existing suite.

Compatibility and limitations

No output/API change intended. Tradeoff: a seen hash and tiny measured small-input overhead in exchange for linear scaling.

No tests were added or modified under the consuming repository's no-new-tests policy. Verification uses existing tests and external scratch reproductions. No production changes; older Ruby/Linux/Windows matrices await upstream CI. No unrelated version/dependency changes.