Rollup merge of rust-lang#154077 - jhpratt:optimize-128-bit-formatting, r=dtolnay

Optimize 128-bit integer formatting

The compiler is unaware of the restricted range of the input, so it is unable to optimize out the final division and modulus. By doing this manually, we get a nontrivial performance gain.

r? @dtolnay

(copied from dtolnay/itoa#68)

Author: Zalathar

library/core/src/fmt/num.rs

@@ -823,7 +823,7 @@ fn enc_16lsd<const OFFSET: usize>(buf: &mut [MaybeUninit<u8>], n: u64) {
     let mut remain = n;
 
     // Format per four digits from the lookup table.
-    for quad_index in (0..4).rev() {
+    for quad_index in (1..4).rev() {
         // pull two pairs
         let quad = remain % 1_00_00;
         remain /= 1_00_00;
@@ -834,6 +834,14 @@ fn enc_16lsd<const OFFSET: usize>(buf: &mut [MaybeUninit<u8>], n: u64) {
         buf[quad_index * 4 + OFFSET + 2].write(DECIMAL_PAIRS[pair2 * 2 + 0]);
         buf[quad_index * 4 + OFFSET + 3].write(DECIMAL_PAIRS[pair2 * 2 + 1]);
     }
+
+    // final two pairs
+    let pair1 = (remain / 100) as usize;
+    let pair2 = (remain % 100) as usize;
+    buf[OFFSET + 0].write(DECIMAL_PAIRS[pair1 * 2 + 0]);
+    buf[OFFSET + 1].write(DECIMAL_PAIRS[pair1 * 2 + 1]);
+    buf[OFFSET + 2].write(DECIMAL_PAIRS[pair2 * 2 + 0]);
+    buf[OFFSET + 3].write(DECIMAL_PAIRS[pair2 * 2 + 1]);
 }
 
 /// Euclidean division plus remainder with constant 1E16 basically consumes 16