SYS calls for accelerating interpreted programs

The virtual CPU or interpreter is now complete. At least, its code page is full, even overflowing a bit, and nothing can be added without removing something else. It has 25 opcodes, the last added is SYS. This is an escape: with SYS the application can jump to arbitrary native code. Here I test this capability in the “clear screen” loop, right at the beginning of the recording.

Without acceleration, clearing the screen pixel by pixel with interpreted code takes more than 2 seconds. Here the application uses a SYS call that clears 8 pixels at a time. The result is that the screen now clears in 200 ms. Nice proof of concept.

Restrictions apply if you don’t want to break the video timing: the callee must be timing-aware and cooperate with the interpreter loop. But it can run for longer than a regular vCPU instruction (these are 28 cycles max) and the code can do whatever it wants, as long as it completes in the available time slice on the scan line. In reality this means that SYS calls can last for up to 146 cycles. If they need more cycles, they could still use some trickery and restart themselves by messing with the virtual program counter… They are “parasitic” after all.

Another novelty is that this recording is not from an actual Gigatron! Martin @ the talkchess forum has made an incredible visualisizer for the project. I didn’t get it running with the Mac binary, but it runs ok with Wine. Emulation in emulation in emulation, but it can still keep up with real time if you use “Run sketch (JIT)”. Amazing.

There is some latency with the sound, but that is fine. This tool already saved me a tons of EPROM burning cycles. For the real experience, just use a real Gigatron…

P.S: The Christmas card is there to give my phone something to focus on. Without it the recording is blurry.

– Marcel