diff --git a/crates/core/src/util/mod.rs b/crates/core/src/util/mod.rs
index 9d12c9626..ef87f17f6 100644
--- a/crates/core/src/util/mod.rs
+++ b/crates/core/src/util/mod.rs
@@ -1,8 +1,12 @@
 mod fast_stack;
 mod gurantee;
+mod segment;
 
 #[allow(unused_imports)]
 pub use crate::util::fast_stack::*;
 
 #[allow(unused_imports)]
 pub use crate::util::gurantee::*;
+
+#[allow(unused_imports)]
+pub use crate::util::segment::*;
diff --git a/crates/core/src/util/segment.rs b/crates/core/src/util/segment.rs
new file mode 100644
index 000000000..b48b952e7
--- /dev/null
+++ b/crates/core/src/util/segment.rs
@@ -0,0 +1,241 @@
+use crate::util::gurantee::gurantee;
+
+use super::fast_stack::FastStack;
+
+// The order of these cases is important for performance
+// please do not change it without significant profiling
+#[derive(Copy, Clone)]
+enum Case {
+  Other,
+
+  // This makes things faster
+  // It has to be the 2nd case
+  Dummy1,
+
+  Close, // b')'
+  ParenL, // b'('
+  BracketL, // b'['
+  CurlyL, // b'{'
+
+  Escape, // b'\'
+  QuoteDouble, // b'"'
+  QuoteSingle, // b'\''
+  Separator, // other
+}
+
+const fn generate_cases(separator: u8) -> [Case; 256] {
+  let mut table = [Case::Other; 256];
+
+  table[b'\\' as usize] = Case::Escape;
+  table[b'"' as usize] = Case::QuoteDouble;
+  table[b'\'' as usize] = Case::QuoteSingle;
+  table[b'(' as usize] = Case::ParenL;
+  table[b'[' as usize] = Case::BracketL;
+  table[b'{' as usize] = Case::CurlyL;
+  table[b')' as usize] = Case::Close;
+  table[b']' as usize] = Case::Close;
+  table[b'}' as usize] = Case::Close;
+  table[separator as usize] = Case::Separator;
+
+  return table;
+}
+
+/**
+ * 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.
+ *
+ * Expected behavior using commas:
+ * var(--a, 0 0 1px rgb(0, 0, 0)), 0 0 1px rgb(0, 0, 0)
+ *        ┬              ┬  ┬    ┬
+ *        x              x  x    ╰──────── Split because top-level
+ *        ╰──────────────┴──┴───────────── Ignored b/c inside >= 1 levels of parens
+ */
+#[inline(never)]
+#[no_mangle]
+pub fn segment(input: &[u8], separator: u8) -> Vec<&[u8]> {
+  return segment_table(input, generate_cases(separator))
+}
+
+#[inline(always)]
+fn segment_table(input: &[u8], cases: [Case; 256]) -> Vec<&[u8]> {
+  let mut closing_bracket_stack = FastStack::new();
+
+  let mut parts: Vec<&[u8]> = vec![];
+  let mut last_pos = 0;
+
+  let mut idx = 0;
+  while idx < input.len() {
+    // SAFETY: last_pos can never be greater than idx at the start of the loop
+    // body. It's set to one greater than idx on a separator but idx is always
+    // incremented before the next iteration.
+    gurantee(last_pos <= idx);
+
+    match cases[input[idx] as usize] {
+      Case::Separator => {
+        if closing_bracket_stack.is_empty() {
+          parts.push(&input[last_pos..idx]);
+          last_pos = idx + 1;
+        }
+      }
+
+      // The next character is escaped, so we skip it.
+      Case::Escape => idx += 1,
+
+      // Strings should be handled as-is until the end of the string. No need to
+      // worry about balancing parens, brackets, or curlies inside a string.
+      Case::QuoteDouble => {
+        loop {
+          idx += 1;
+
+          // Ensure we don't go out of bounds.
+          if idx >= input.len() {
+            break
+          }
+
+          match cases[input[idx] as usize] {
+            Case::Escape => idx += 1,
+            Case::QuoteDouble => break,
+            _ => {},
+          }
+        }
+      }
+
+      // Strings should be handled as-is until the end of the string. No need to
+      // worry about balancing parens, brackets, or curlies inside a string.
+      Case::QuoteSingle => {
+        loop {
+          idx += 1;
+
+          // Ensure we don't go out of bounds.
+          if idx >= input.len() {
+            break
+          }
+
+          match cases[input[idx] as usize] {
+            Case::Escape => idx += 1,
+            Case::QuoteSingle => break,
+            _ => {},
+          }
+        }
+      }
+
+      Case::ParenL => {
+        closing_bracket_stack.push(b')');
+
+        if closing_bracket_stack.overgrown() {
+          break
+        }
+      }
+
+      Case::BracketL => {
+        closing_bracket_stack.push(b']');
+
+        if closing_bracket_stack.overgrown() {
+          break
+        }
+      }
+
+      Case::CurlyL => {
+        closing_bracket_stack.push(b'}');
+
+        if closing_bracket_stack.overgrown() {
+          break
+        }
+      }
+
+      Case::Close => {
+        if !closing_bracket_stack.is_empty() && closing_bracket_stack.peek() == input[idx] {
+          closing_bracket_stack.pop();
+        }
+      }
+
+      Case::Other => {},
+      Case::Dummy1 => {},
+    }
+
+    idx += 1;
+  }
+
+  // SAFETY: last_pos will be at most `input.len() - 1` ensuring that the slice
+  // is always within bounds.
+  gurantee(last_pos < input.len());
+
+  parts.push(&input[last_pos..]);
+
+  return parts
+}
+
+#[cfg(test)]
+mod tests {
+  use super::*;
+
+  #[test]
+  fn should_result_in_a_single_segment_when_the_separator_is_not_present() {
+    assert_eq!(segment(b"foo", b':'), vec![b"foo"])
+  }
+
+  #[test]
+  fn should_split_by_the_separator() {
+    assert_eq!(segment(b"foo:bar:baz", b':'), vec![b"foo" as &[u8], b"bar", b"baz"])
+  }
+
+  #[test]
+  fn should_not_split_inside_of_parens() {
+    assert_eq!(segment(b"a:(b:c):d", b':'), vec![b"a" as &[u8], b"(b:c)", b"d"])
+  }
+
+  #[test]
+  fn should_not_split_inside_of_brackets() {
+    assert_eq!(segment(b"a:[b:c]:d", b':'), vec![b"a" as &[u8], b"[b:c]", b"d"])
+  }
+
+  #[test]
+  fn should_not_split_inside_of_curlies() {
+    assert_eq!(segment(b"a:{b:c}:d", b':'), vec![b"a" as &[u8], b"{b:c}", b"d"])
+  }
+
+  #[test]
+  fn should_not_split_inside_of_double_quotes() {
+    assert_eq!(segment(b"a:\"b:c\":d", b':'), vec![b"a" as &[u8], b"\"b:c\"", b"d"])
+  }
+
+  #[test]
+  fn should_not_split_inside_of_single_quotes() {
+    assert_eq!(segment(b"a:'b:c':d", b':'), vec![b"a" as &[u8], b"'b:c'", b"d"])
+  }
+
+  #[test]
+  fn should_not_crash_when_double_quotes_are_unbalanced() {
+    assert_eq!(segment(b"a:\"b:c:d", b':'), vec![b"a" as &[u8], b"\"b:c:d"])
+  }
+
+  #[test]
+  fn should_not_crash_when_single_quotes_are_unbalanced() {
+    assert_eq!(segment(b"a:'b:c:d", b':'), vec![b"a" as &[u8], b"'b:c:d"])
+  }
+
+  #[test]
+  fn should_skip_escaped_double_quotes() {
+    assert_eq!(segment(b"a:\"b:c\\\":d\":e", b':'), vec![b"a" as &[u8], b"\"b:c\\\":d\"", b"e"])
+  }
+
+  #[test]
+  fn should_skip_escaped_single_quotes() {
+    assert_eq!(segment(b"a:'b:c\\':d':e", b':'), vec![b"a" as &[u8], b"'b:c\\':d'", b"e"])
+  }
+
+  #[test]
+  fn should_skip_escaped_separators() {
+    assert_eq!(segment(b"a:b\\:c:d", b':'), vec![b"a" as &[u8], b"b\\:c", b"d"])
+  }
+
+  #[test]
+  fn should_split_by_the_escape_sequence_which_is_escape_as_well() {
+    assert_eq!(segment(b"a\\b\\c\\d", b'\\'), vec![b"a" as &[u8], b"b", b"c", b"d"]);
+    assert_eq!(segment(b"a\\(b\\c)\\d", b'\\'), vec![b"a" as &[u8], b"(b\\c)", b"d"]);
+    assert_eq!(segment(b"a\\[b\\c]\\d", b'\\'), vec![b"a" as &[u8], b"[b\\c]", b"d"]);
+    assert_eq!(segment(b"a\\{b\\c}\\d", b'\\'), vec![b"a" as &[u8], b"{b\\c}", b"d"]);
+  }
+}