Speed up segmentation (#13415)

* Speed up segmentation

* Add segment benchmark

* Add and move comments

* Update

* Tweak comments

* Tweak variable name

* Tweak

---------

Co-authored-by: Jordan Pittman <jordan@cryptica.me>

packages/tailwindcss/src/utils/segment.bench.ts

@@ -0,0 +1,14 @@
+import { bench } from 'vitest'
+import { segment } from './segment'
+
+const values = [
+  ['hover:focus:underline', ':'],
+  ['var(--a, 0 0 1px rgb(0, 0, 0)), 0 0 1px rgb(0, 0, 0)', ','],
+  ['var(--some-value,env(safe-area-inset-top,var(--some-other-value,env(safe-area-inset))))', ','],
+]
+
+bench('segment', () => {
+  for (let [value, sep] of values) {
+    segment(value, sep)
+  }
+})

packages/tailwindcss/src/utils/segment.ts

@@ -1,8 +1,23 @@
+// This is a shared buffer that is used to keep track of the current nesting level
+// of parens, brackets, and braces. It is used to determine if a character is at
+// the top-level of a string. This is a performance optimization to avoid memory
+// allocations on every call to `segment`.
+const closingBracketStack = new Uint8Array(256)
+
+// All numbers are equivalent to the value returned by `String#charCodeAt(0)`
+const BACKSLASH = 0x5c
+const OPEN_PAREN = 0x28
+const OPEN_BRACKET = 0x5b
+const OPEN_CURLY = 0x7b
+const CLOSE_PAREN = 0x29
+const CLOSE_BRACKET = 0x5d
+const CLOSE_CURLY = 0x7d
+
 /**
  * This splits a string on a top-level character.
  *
- * Regex doesn't support recursion (at least not the JS-flavored version).
- * So we have to use a tiny state machine to keep track of paren placement.
+ * Regex doesn't support recursion (at least not the JS-flavored version),
+ * so we have to use a tiny state machine to keep track of paren placement.
  *
  * Expected behavior using commas:
  * var(--a, 0 0 1px rgb(0, 0, 0)), 0 0 1px rgb(0, 0, 0)
@@ -11,43 +26,50 @@
  *        ╰──────────────┴──┴───────────── Ignored b/c inside >= 1 levels of parens
  */
 export function segment(input: string, separator: string) {
-  // Stack of characters to close open brackets. Appending to a string because
-  // it's faster than an array of strings.
-  let closingBracketStack = ''
+  // SAFETY: We can use an index into a shared buffer because this function is
+  // synchronous, non-recursive, and runs in a single-threaded envionment.
+  let stackPos = 0
   let parts: string[] = []
   let lastPos = 0
 
-  for (let idx = 0; idx < input.length; idx++) {
-    let char = input[idx]
+  let separatorCode = separator.charCodeAt(0)
 
-    if (closingBracketStack.length === 0 && char === separator) {
+  for (let idx = 0; idx < input.length; idx++) {
+    let char = input.charCodeAt(idx)
+
+    if (stackPos === 0 && char === separatorCode) {
       parts.push(input.slice(lastPos, idx))
       lastPos = idx + 1
       continue
     }
 
     switch (char) {
-      case '\\':
+      case BACKSLASH:
         // The next character is escaped, so we skip it.
         idx += 1
         break
-      case '(':
-        closingBracketStack += ')'
+      case OPEN_PAREN:
+        closingBracketStack[stackPos] = CLOSE_PAREN
+        stackPos++
         break
-      case '[':
-        closingBracketStack += ']'
+      case OPEN_BRACKET:
+        closingBracketStack[stackPos] = CLOSE_BRACKET
+        stackPos++
         break
-      case '{':
-        closingBracketStack += '}'
+      case OPEN_CURLY:
+        closingBracketStack[stackPos] = CLOSE_CURLY
+        stackPos++
         break
-      case ')':
-      case ']':
-      case '}':
-        if (
-          closingBracketStack.length > 0 &&
-          char === closingBracketStack[closingBracketStack.length - 1]
-        ) {
-          closingBracketStack = closingBracketStack.slice(0, closingBracketStack.length - 1)
+      case CLOSE_BRACKET:
+      case CLOSE_CURLY:
+      case CLOSE_PAREN:
+        if (stackPos > 0 && char === closingBracketStack[stackPos - 1]) {
+          // SAFETY: The buffer does not need to be mutated because the stack is
+          // only ever read from or written to its current position. Its current
+          // position is only ever incremented after writing to it. Meaning that
+          // the buffer can be dirty for the next use and still be correct since
+          // reading/writing always starts at position `0`.
+          stackPos--
         }
         break
     }