General optimisations for libsnark

[HarryR]

Need to verify or reflect on the following opportunities for optimisation:

• are vtable lookups used for G1, G2 and G12 arithmetic operations? if so, can they be avoided
  • Create test executables for different cases, load in IDA, do analysis work to see if vtables are being used
• Does rvalue usage reduce number of copy constructors, realistically, in any code that's used?
• Can improvements for general BN256 operations be improved by using the same improvements from https://github.com/cloudflare/bn256 ?
• Run through valgrind, cachegrind etc.
• Use -flto and -fwhole-programwith GCC, see: https://stackoverflow.com/questions/13674894/when-to-use-certain-optimizations-such-as-fwhole-program-and-fprofile-generate - all oflibsnarkcan be built like this, butlibgmp` uses the system implementation

[HarryR]

Interestingly, GMP is very slow in Emscripten with WebAssembly, and 30% slower than an optimised version of a 256bit multiply operation.

Improving performance on WebAssembly would require investigating field operations performance in WebAssembly. Likewise, a 30% speedup could be made by using an alternative to GMP?