Banning eval's visibility of lexicals is the number 1 way to optimize perl to C speed. You have to compile out the "english". Also banning magic would allow common subexpression elimination and convert SVs into more C like data/machine words. Reducing usage of the perl stack might be a worth idea to pursue. Considered taking a profiler to all the perl opcodes and see which are the slowest? I've modified NYTProf for myself to time per opcode times (part of NYTProf's CORE:: profiling system). There is a list in some header of NYTProf which says which opcodes to profile.

A random idea I thought of today is, stop pp_sassign for target aware opcodes. sv_setsv does bail out almost instantly if source and dest SV*s are the same, but pp_sassign goes through alot of code until it calls sv_setsv.

Sadly, your best bet is to write a Perl 5 to .NET CIL or Java bytecode compiler since pretty much all CPUs nowadays are tuned to running one of those 2 VMs.