| | - Athlon-XP version was faster on Athlon64 than version compiled with -march=athlon64 and -m32 |
| | - Using -mtune=i686 is even faster, not to mention it works on all older CPUs as well |
| | - -O2 is faster than -O3. Even more, -O2 with -fomit-frame-pointer gains more than -O3 with -fomit-frame-pointer |
| | - All that SSE and SSE2 optimizations which should make FPU operations faster don't do a thing |
| | - Some other options, like loop unrolling and fast math do not produce any measurable boost |
| | - Compiling with -fprofile-generate/-fprofile-use produces a small boost (~1%) and also makes the code somewhat smaller |
| | - And finally - using the Improved flag, EA, LEA, and XCHG handling in dynamic core from DOSBox sourceforge patch page gains another 5-7% speed |
Using these findings, today's version should be faster on all machines (the previous i686 version was rather unoptimized since I didn't pay much attention to it :)) and works on all CPUs. I also fixed some stupid bugs in Direct3D, fullscreen mode should be much better now. ALT-Tabbing from fullscreen is still a bit strange...it works on my laptop but not my desktop. I'd still recommend not to use it :). The cvs also fixed mouse in Win3.1 Enhanced mode, so I guess w3.1 are finally runable in DOSBox!