Improve optimization of generated code
The Delphi compiler could generate more efficient assembly code if it would do some simple optimizations. Thing to look for : Make better use of the fact that lots of operations already set flags (so a comparison can be avoided). Improve in the register-allocation (don't do "MOV eax, ebx; SHL eax, cx" when eax will be re-used anyway - SHL ebx itself then). Don't emit opcodes that have already been emitted and will give the same result - try to re-use the previous result instead. Etc, etc.
I will recomend to use the SIMD/AVX opcodes when it's possible.
Support for old processors is what left the Delphi compiler very badly optimized for todays computers, you should focus on new processors not some old Pentium processor or so that nobody uses anymore.
And use MMX and SSE code (but provide fallbacks for older processors).