[gnutix]

Stacked on #279 (it uses Stream.Memo): please merge that one first. Until then this diff also shows #279's commit.

Problem: many rules look for a word (sprintf, switch, null…) or a docblock that most files don't have, but still walk every token of every file to find out. PHP-CS-Fixer avoids this with isCandidate().

Change:

• Stream.Index(): built once per file, cached with Stream.Memo. It holds the token kinds present, and the lowercased words in a Bloom filter. It can say "present" for an absent word, never the reverse. A non-ASCII word makes every word "present", because case folding maps some non-ASCII letters to ASCII ones.
• fixer.Candidate: an optional IsCandidate(*Stream) bool. The runner skips Fix and Errors when it's false. Fixers without it, including blink/extend ones, run as before.
• 39 rules implement it, each with a trigger taken from its own Fix:
  • one or more function names: NoUselessSprintf, ImplodeCall, NoAliasFunctions, EregToPreg, FunctionToConstant, ModernizeTypesCasting, CombineNestedDirname, DirConstant, SetTypeToCast, ArrayPush, PowToExponentiation, IsNull, GetClassToClassKeyword, FopenFlagOrder, StringLengthToEmpty, ModernizeStrpos, ModernSerializationMethods
  • keywords: MagicConstantCasing, MagicMethodCasing, NoAliasLanguageConstructCall, SwitchCaseSemicolonToColon, SwitchCaseSpace, NoBreakComment, SwitchContinueToBreak, ListSyntax, NoUnsetCast, NoMixedEchoPrint, LogicalOperators, LowercaseStaticReference, NoNullPropertyInitialization, NullableTypeDeclaration, NullableTypeDeclarationForDefaultNullValue
  • a non-ASCII byte: NonPrintableCharacter
  • a docblock or comment: PhpdocOrder, PhpdocTypesOrder, PhpdocLineSpan, NoSuperfluousPhpdocTags, TypeToVarTag, RemoveUselessDefaultComment
• Shared helpers (applyToDocblocks, streamHas*, phpUnitTestClasses) check the index first, which helps about 30 more rules.
• Not gated: triggers in nearly every file (declare, use, class…), punctuation, text inside docblocks, whitespace rules.

Result: on a private 8.8k-file codebase, the 72 rules this touches: 5.91 s → 4.38 s of CPU. FunctionToConstant 140 → 5 ms, EregToPreg 139 → 6 ms. Building the index costs about 1%.

Tests:

• New tests for the index, the runner skip, and each gated rule with its trigger in odd case (SprintF, __Line__…).
• TestMain checks that no rule rules out a source its own tests fix.
• Existing tests pass, and the private codebase still has 0 diffs. On 28k files with 18.5k diffs, Blink: add a stream version, to share work between rules #279 and this branch give the same output.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y1vvcRhpVMjWLC4U5j78Xp

[TomasVotruba]

Nice speed improvement!

I think we should be able to get under 2 secs on such a project