Auto merge of #82562 - llogiq:one-up-82248, r=oli-obk

Optimize counting digits in line numbers during error reporting further

This one-ups #82248 by switching the strategy: Instead of dividing the value by 10 repeatedly, we compare with a limit that we multiply by 10 repeatedly. In my benchmarks, this took between 50% and 25% of the original time. The reasons for being faster are:

1. While LLVM is able to replace a division by constant with a multiply + shift, a plain multiplication is still faster. However, this doesn't even factor, because
2. Multiplication, unlike division, is const. We also use a simple for-loop instead of a more complex loop + break, which allows
3. rustc to const-fold the whole loop, and indeed the assembly output simply shows a series of comparisons.

1 files changed, 26 insertions, 12 deletions

diff --git a/compiler/rustc_errors/src/emitter.rs b/compiler/rustc_errors/src/emitter.rs
index 42c3d5e48fe..9b6f67166bd 100644
--- a/compiler/rustc_errors/src/emitter.rs
+++ b/compiler/rustc_errors/src/emitter.rs
@@ -1713,18 +1713,8 @@ impl EmitterWriter {
         let max_line_num_len = if self.ui_testing {
             ANONYMIZED_LINE_NUM.len()
         } else {
-            // Instead of using .to_string().len(), we iteratively count the
-            // number of digits to avoid allocation. This strategy has sizable
-            // performance gains over the old string strategy.
-            let mut n = self.get_max_line_num(span, children);
-            let mut num_digits = 0;
-            loop {
-                num_digits += 1;
-                n /= 10;
-                if n == 0 {
-                    break num_digits;
-                }
-            }
+            let n = self.get_max_line_num(span, children);
+            num_decimal_digits(n)
         };
 
         match self.emit_message_default(span, message, code, level, max_line_num_len, false) {
@@ -1952,6 +1942,30 @@ impl FileWithAnnotatedLines {
     }
 }
 
+// instead of taking the String length or dividing by 10 while > 0, we multiply a limit by 10 until
+// we're higher. If the loop isn't exited by the `return`, the last multiplication will wrap, which
+// is OK, because while we cannot fit a higher power of 10 in a usize, the loop will end anyway.
+// This is also why we need the max number of decimal digits within a `usize`.
+fn num_decimal_digits(num: usize) -> usize {
+    #[cfg(target_pointer_width = "64")]
+    const MAX_DIGITS: usize = 20;
+
+    #[cfg(target_pointer_width = "32")]
+    const MAX_DIGITS: usize = 10;
+
+    #[cfg(target_pointer_width = "16")]
+    const MAX_DIGITS: usize = 5;
+
+    let mut lim = 10;
+    for num_digits in 1..MAX_DIGITS {
+        if num < lim {
+            return num_digits;
+        }
+        lim = lim.wrapping_mul(10);
+    }
+    MAX_DIGITS
+}
+
 fn replace_tabs(str: &str) -> String {
     str.replace('\t', "    ")
 }