Speed up number formatting by caching locale and pattern parsing - #1437

A typical format_decimal(x, locale='de_DE') or format_currency(x, 'EUR', format='¤#,##0.00', locale=...) call parses the same locale identifier and often the same number pattern every time. This PR removes that repeated work in four small commits. Each one can be reviewed (or dropped) on its own:

1. core: cache parse_locale() results. A 1024-entry lru_cache keyed by the exact (identifier, sep) strings. The result is a tuple of strings. Locale objects are still created on every Locale.parse() call, so nothing mutable is shared. str subclasses and non-string inputs take the uncached path, so their behaviour and error messages don't change. Exceptions are never cached.
2. core: try the exact locale before defining the Locale.parse fallbacks. Locale.parse() created two closures on every call, but they're only needed when the locale doesn't exist. It now tries cls(*parts) first. The likely-subtag fallback code is unchanged.
3. numbers: cache parsed number patterns. Uses the same approach as babel.dates.parse_pattern, which already keeps a 1024-entry lru_cache. Here the cache holds the parsed constructor arguments, which are all strings, tuples, ints and None. NumberPattern is mutable, so a new instance is still built on every call. This helps any call that passes a custom format= string. Locale default patterns are already stored as NumberPattern objects and skip the cache.
4. numbers: skip quote cleanup when the result has no quotes. The re.sub that strips pattern quotes can't match without an apostrophe. Skipping it saves a regex call and a lambda on almost every formatted number.

Both caches are bounded, and their keys are the function's only inputs. Neither result depends on locale data or global state.

Benchmarks

The workload, for each of 12 locales (en_US, de_DE, fr_FR, ja_JP, ar_EG, hi_IN, pt_BR, ru_RU, es_ES, zh_Hans_CN, fi_FI, de_CH), is: format_decimal with the default pattern and with '#,##0.00', format_currency with the default pattern and with '¤#,##0.00', format_percent, format_compact_decimal, and parse_decimal. That's 84 calls per run. Each sample was a separate process pinned to one core on Linux. I ran master and this branch in alternating order and give the median per-pair ratio:

Eight pairs on 3.12 and 3.15, six on 3.9. Per commit, on the string-locale workload (Python 3.12, 8 pairs each): parse_locale cache 1.056x (1.023–1.090x), Locale.parse reorder 1.030x (1.009–1.036x), number-pattern cache 1.080x (1.063–1.139x), quote check 1.055x (1.020–1.084x). When callers pass Locale objects, Locale.parse returns early and only the two numbers commits apply.

Checks

• I compared master and this branch in separate processes and checked that the output was byte-identical. The comparison covered parse_locale and Locale.parse (with and without likely-subtag resolution, both separators) for every sixth CLDR locale plus aliases, POSIX/C, modifiers, encodings and invalid identifiers. It also covered format_decimal, format_currency, format_currency(format_type='name'), format_percent, format_scientific, format_compact_decimal, format_compact_currency and parse_decimal across those locales, 19 numbers (including NaN, ±Infinity, -0.0 and 40-digit Decimals) and 24 patterns (quotes, '', currency signs, per-mille, exponents, negative subpatterns and invalid patterns). Each case ran twice so that cache hits were compared too, and I mutated a returned NumberPattern between parses. That's 452,106 results, 8,776 of them exceptions, all identical on Python 3.9, 3.12 and 3.15.
• New tests check that cached results aren't shared: a mutated Locale or NumberPattern doesn't affect the next call, and pickling still works. They also cover str subclasses, unhashable and invalid inputs (same exceptions as before), the cache size limits, Locale subclasses whose constructors return other objects or raise, and quote handling after pattern mutation and currency-symbol replacement.
• The full test suite passes on Python 3.9, 3.12 and 3.15, and pre-commit run --all-files passes.

Overlap: #1380 rewrites the quote cleanup that commit 4 touches. If #1380 is merged first, commit 4 can be dropped or redone on top of it. #1382 and #1370 change parse_pattern internals; they apply cleanly on this branch and work with the cache.

🤖 Generated with Claude Code