I get results like hippo's with perl 5.36 on a recent AMD Ryzen

                    Rate length_subst         subst        match trans
+literation
length_subst      8970/s           --          -89%         -91%      
+      -97%
subst            78884/s         779%            --         -19%      
+      -78%
match            97126/s         983%           23%           --      
+      -73%
transliteration 355019/s        3858%          350%         266%      
+        --
[download]

Though, depending how many matches there are, does perl have to assemble a stack of N elements (copying each character into its own scalar) before assigning the list to the scalar to get the count? With the subst, the right optimizations could allow it to update that one character without changing the length of the string or copying anything, so it could be fast, and then doesn't need to assemble a list of the matches.