perf(codegen): remove branch from print_space_before_identifier (#27842)

Perf optimization. Continuation of #27841.

That PR simplified `print_space_before_identifier`. Remove the branch on "was last byte ASCII?" by using a small lookup table.

This removes 2 or 3 instructions from this very hot path.

The previous implementation already used a lookup table for ASCII bytes. The table here is twice as large (128 bytes -> 256 bytes) but the 2nd half of it is only read when the source code contains an identifier ending in a non-ASCII character, which is extremely rare. So in 99.99% of cases, the cache usage of this larger table is no greater than the previous version.

crates/oxc_codegen/src/lib.rs

@@ -47,6 +47,39 @@ pub use options::{CodegenOptions, CommentOptions, LegalComment};
 // Re-export `IndentChar` from `oxc_data_structures`
 pub use oxc_data_structures::code_buffer::IndentChar;
 
+#[repr(C, align(128))]
+struct Align128<T>(pub(crate) T);
+
+/// Lookup table for whether a byte which is last byte of output when printing an identifier
+/// is one that needs a space printed after it.
+///
+/// Used by [`print_space_before_identifier`].
+///
+/// Non-ASCII characters can appear in various places (e.g. in strings),
+/// but in such cases an ASCII character (e.g. `"`) is always printed after them.
+/// Only case in which last byte of output can be non-ASCII when printing an identifier
+/// or keyword (when [`print_space_before_identifier`] is called) is when what was
+/// last printed was an identifier ending with a non-ASCII character.
+/// So non-ASCII must be an identifier, and a space is required.
+///
+/// Non-ASCII characters in identifiers are extremely rare in real-world code,
+/// so only the first 128 bytes of this table will ever be read.
+/// Aligned on 128, so occupies 2 x 128-byte cache lines on Apple Silicon,
+/// and 2 pairs of 64-byte cache lines on x86.
+///
+/// [`print_space_before_identifier`]: Codegen::print_space_before_identifier
+static IDENT_BYTES: Align128<[bool; 256]> = {
+    let mut slots = [true; 256];
+
+    let mut i = 0u8;
+    while i < 128 {
+        slots[i as usize] = is_identifier_part_ascii(i as char);
+        i += 1;
+    }
+
+    Align128(slots)
+};
+
 /// Output from [`Codegen::build`]
 #[non_exhaustive]
 pub struct CodegenReturn<'a> {
@@ -683,19 +716,11 @@ impl<'a> Codegen<'a> {
     fn print_space_before_identifier(&mut self) {
         let Some(byte) = self.last_byte() else { return };
 
-        if self.need_space_before_identifier != self.code.len() {
-            // Non-ASCII characters can appear in various places (e.g. in strings),
-            // but in such cases an ASCII character (e.g. `"`) is always printed after them.
-            // Only case in which last byte of output can be non-ASCII when printing an identifier
-            // or keyword (when this function is called) is when what was last printed was
-            // an identifier ending with a non-ASCII character.
-            // So non-ASCII must be an identifier, and a space is required.
-            if byte.is_ascii() && !is_identifier_part_ascii(byte as char) {
-                return;
-            }
+        let force_print_space = self.need_space_before_identifier == self.code.len();
+        let last_byte_is_ident = IDENT_BYTES.0[byte as usize];
+        if force_print_space || last_byte_is_ident {
+            self.print_hard_space();
         }
-
-        self.print_hard_space();
     }
 
     #[inline]

crates/oxc_syntax/src/identifier.rs

@@ -131,7 +131,7 @@ pub fn is_identifier_part(c: char) -> bool {
 }
 
 #[inline]
-pub fn is_identifier_part_ascii(c: char) -> bool {
+pub const fn is_identifier_part_ascii(c: char) -> bool {
     ASCII_ID_FLAGS.0[c as usize] & ID_CONTINUE != 0
 }