Investigate possibility to reduce rollbacks in parser

[hurricup]

According to #2703 we have around 10-15% of token rollbacks. Most of those are 1 and 2 tokens length and probably we could get rid of those.