perf: NonZeroU64 to optimize encoded_len_varint (#1192)

The leading zeros count may perform better on many architectures
when the zero case is excluded.
Also use ilog2 as shorthand for the leading zeros trick because
it makes more clearly what we mean to get, and should be ideally
optimized by the compiler.

prost/src/encoding/varint.rs

@@ -1,4 +1,5 @@
 use core::cmp::min;
+use core::num::NonZeroU64;
 
 use ::bytes::{Buf, BufMut};
 
@@ -25,8 +26,10 @@ pub fn encode_varint(mut value: u64, buf: &mut impl BufMut) {
 #[inline]
 pub fn encoded_len_varint(value: u64) -> usize {
     // Based on [VarintSize64][1].
-    // [1]: https://github.com/google/protobuf/blob/3.3.x/src/google/protobuf/io/coded_stream.h#L1301-L1309
-    ((((value | 1).leading_zeros() ^ 63) * 9 + 73) / 64) as usize
+    // [1]: https://github.com/protocolbuffers/protobuf/blob/v28.3/src/google/protobuf/io/coded_stream.h#L1744-L1756
+    // Safety: value | 1 is non-zero.
+    let log2value = unsafe { NonZeroU64::new_unchecked(value | 1) }.ilog2();
+    ((log2value * 9 + (64 + 9)) / 64) as usize
 }
 
 /// Decodes a LEB128-encoded variable length integer from the buffer.