[sahvx655-wq]

Description

Quadratic node removal in the --line-ranges conversion

Profiling black --line-ranges over a synthetic module with a few thousand elif branches, convert_unchanged_lines dominated the run and blib2to3's Base.remove alone accounted for roughly half of it. When a range leaves most of a file unchanged, every unchanged block is turned into a single STANDALONE_COMMENT by _convert_node_to_standalone_comment / _convert_nodes_to_standalone_comment, and those call remove and insert_child on the block's parent. Each call walks and shifts the parent's whole children list, so a parent that owns many sibling blocks (a long if/elif chain, a match with many cases, or a module full of top-level definitions) is O(n^2). A 4000-branch chain sat at about 3.6s and grew quadratically; left unfixed it is a cheap way to make --line-ranges on a large file crawl.

All the blocks under a parent are known up front, so there is no need to mutate the tree one block at a time. I collect the conversions in a small _NodeReplacements helper and rewrite each affected parent's children list in a single pass, splicing the comment leaves in place of the removed runs; nodes already scheduled for conversion are skipped so the deferred pass keeps the old "first conversion wins" behaviour (an async block keeps its ASYNC leaf on a different parent, and stacked decorators collapse to one node, both of which the single-pass rewrite handles). Output is byte-identical: I diffed the formatter over the tests/data/cases corpus and a batch of generated if/elif, match, decorator, try, with and mixed modules against many random line ranges, and the existing test_ranges suite still passes. The same 4000-branch chain now formats in about 0.6s and scales linearly.

Checklist - did you ...

• Implement any code style changes under the --preview style, following the stability policy? (no style change; output is unchanged)
• Add an entry in CHANGES.md if necessary?
• Add / update tests if necessary? (behaviour is unchanged, so this is covered by the existing test_ranges suite)
• Add new / update outdated documentation? (not applicable)

[github-actions]

diff-shades results comparing this PR (a4d4f79) to main (451316b):

--preview style: no changes

--stable style: no changes

---

What is this? | Workflow run | diff-shades documentation

[cobaltt7]

Tests are failing, please take a look

[sahvx655-wq]

Sorted. The failure was line_ranges_decorator_edge_case: the batched conversion mangled single-line decorated blocks, emitting @overload\n\n\ndef foo(): ... instead of keeping them together.

Root cause was an ordering the old per-node path quietly relied on. It converted the decorator in place first, so by the time the enclosing decorated node was converted the decorator was already a single leaf with its trailing newline stripped and its prefix migrated onto the standalone. Batching the conversions dropped that in-place step, so re-inserting a newline doubled the blank line and the outer conversion then read an empty prefix off the leaf the decorator run had already cleared. I removed the now-redundant newline insert, kept each first leaf's taken prefix so the inner and outer conversions reuse it, and zeroed that leaf's lineno to mirror the line-0 standalone the in-place path used to leave behind. test_ranges and test_format both pass locally now.