Improve Linux performance

[jpsim]

86% of the time spent linting the SwiftLint repo on Linux is in CFStringFindCharacterFromSet.

Interactive version (open in a web browser): flamegraph.svg.zip

[norio-nomura]

We need to disable caching in NSString on Linux.
https://github.com/jpsim/SourceKitten/blob/master/Source/SourceKittenFramework/String%2BSourceKitten.swift#L184

[jpsim]

Actually, enabling caching on Linux has a great impact:

flamegraph.svg.zip

Just a naive implementation:

diff --git a/Source/SourceKittenFramework/String+SourceKitten.swift b/Source/SourceKittenFramework/String+SourceKitten.swift
index c40cffe..14163a1 100644
--- a/Source/SourceKittenFramework/String+SourceKitten.swift
+++ b/Source/SourceKittenFramework/String+SourceKitten.swift
@@ -35,6 +35,10 @@ private let commentLinePrefixCharacterSet: CharacterSet = {
 private var keyCacheContainer = 0
 
 extension NSString {
+
+    static private var stringCache = [NSString: CacheContainer]()
+    static private var stringCacheLock = NSLock()
+
     /**
     CacheContainer caches:
 
@@ -181,7 +185,14 @@ extension NSString {
     */
     private var cacheContainer: CacheContainer {
         #if os(Linux)
-        return CacheContainer(self)
+        NSString.stringCacheLock.lock()
+        defer { NSString.stringCacheLock.unlock() }
+        if let cache = NSString.stringCache[self] {
+            return cache
+        }
+        let cache = CacheContainer(self)
+        NSString.stringCache[self] = cache
+        return cache
         #else
         if let cache = objc_getAssociatedObject(self, &keyCacheContainer) as? CacheContainer {
             return cache

[jpsim]

With the caching above, linting this repo takes 15.04s rather than the previous 3m23s.

[norio-nomura]

Yup,. The disadvantage of that implementation is that "CacheContainer and String are never freed".

[jpsim]

True, but we already make that trade off on macOS. Linting this repo in debug mode with a dictionary cache above takes the same amount of memory as with ObjC associated objects (46MB).

[norio-nomura]

Certainly File will persist until the end of the SwiftLint process, so neither String nor CacheContainer will be freed until the end. 🤔
I don't know about another consumer of SourceKitten. Are memory leaks due to intentional circular references an issue for other SourceKittenFramework users?

[jpsim]

It probably is an issue, but not at the expense of performance. I'd say we should implement the dictionary based caching approach for now and eventually move that to File.

[norio-nomura]

I'd say we should implement the dictionary based caching approach for now and eventually move that to File.

I fully agree. 👍