Ok the MUL workaround got me. MUL does (ab)/64 so you have to DIV first to get a64, then MUL finally gives you a*b. And recursive FDEFs because there's no WHILE? All in 6.5KB? What kind of frame rate do you actually get out of this?

Frame rate depends on the browser - Chrome gives around 30-60fps on my machine, but the bottleneck is actually Chrome deciding whether to re-run hinting at all (had to add axis jitter to force it). The TT bytecode itself executes fast, it's maybe a few thousand instructions per frame

  The recursive FDEF thing is the worst part honestly. Every while loop is a function that calls itself, and FreeType kills you at ~64 deep. So you're constantly juggling how many columns vs how many ray steps you can afford