speedups for more architectures
speedups for more microarchitectures
merge tables across implementations/compilers
consider automatically eliminating rarely-used implementations
merge subroutines in source to the extent possible
compile each asm subroutine with only one compiler
after compilation, merge object files that are essentially repeated
sort object files (for, e.g., improved cache utilization)
automatically build+use table of known exceptional cpuids that want other implementations
improve automatic order of implementations to reduce number of exceptional cpuids
optionally allow post-installation patching of current cpu as another exceptional cpuid
  (based on benchmarks and, with more CPU time, full functionality tests)
dispatch: eliminate, e.g., avx2 if avx is higher priority
speed up dispatch cpuid tests (lazy evaluation, merging cpuid calls)
randombytes: support getrandom, getentropy
verify constbranch, constindex
full functional verification