http://www.winuae.net/files/b/winuae_2400b21.zip

Beta 21: (Fastest possible vsync should be fixed finally, only some testing left to do)

- Direct3D masks and scanlines fixed, forgotten testing code.. (b20)
- 2.3.2 "stop the cpu and wait until blitter has finished if any blitter register is accessed while blitter is busy and cpu mode is fastest possible" was broken and could cause side-effects in some situations.
- Direct3D to DirectDraw fallback if DirectX is not recent enough crashed. (b5)
- Keyboard lost sync state emulated, needed with some programs that don´t handshake all buffered key codes but still clear the CIA keyboard interrupt flag, this caused dead keyboard after b1 update. Now keyboard should be fully emulated.
- Quickstart panel disk insert always reset track position to zero. (forgot to restore old track after checking disk type and bootable state, bug since Quickstart was introduced!) Fixes Wrath of Demon disk swaps.
- Z2 RTG changes broke chip RAM "memory barrier" causing crash if JIT executed code at the very end of chip RAM. (b16)

I finally really examined and even understood (I think..) how fastest possible and JIT modes handle CPU emulation, (in reality fastest possible without JIT didn´t really handle it all, other UAE ports most likely have exact same issue):
Some unnecessary and big event code inlining also removed (shorter code, should be faster on modern CPUs due to less cache trashing)

- Huge improvement in fastest possible CPU (with or without JIT) + low latency vsync performance.
- Fastest possible without JIT performance improved in non-vsync modes. It now executes small chunks of extra code after each scanline instead (as long as there is time, the faster the host CPU, the more extra time there is) of single huge chunk just before vsync. JIT basically does the same but because it has "unknown" timing, it may execute "too much" code first and then skip multiple scanlines until there is enough time again.
- Immediate blitter is now 100% immediate. Fastest possible (with or without JIT) performance greatly improved (10x+ possible!) if program does lots of small blits. Previously, even in immediate blitter mode, blitter wait caused CPU emulator to waste its extra "fastest possible time slot" by doing absolutely nothing else than waiting the blitter that never happened. It can´t happen until next scanline, during extra fastest CPU "slots" chipset emulation has to be paused.

Fastest possible CPU throttling option will be also possible, this will be done later..

NOTE: Above changes WILL break some games/demos that (accidentally) worked previously. It is 100% guaranteed!