Re: Lua/LuaJIT2 performance on Android

[Adam Strzelecki <ono@...>]

> And even when the ARM JIT compiler is finished, you won't see such
> a drastic speedup for SciMark on a soft-float device. The underlying
> FP operations are very expensive (62 cycles for an FP addition!) and
> put a hard limit on the minimum execution time.

Mike, just wanted to ask whether support for VFP/NEON instruction somewhere close on roadmap? I can see that VFPs are already marked NYI, so I hope they will find their way into the code. FYI both iPad and iPhone CPUs have VFP, while newer devices got also NEON (SIMD) extensions. Also many Android devices are VFP/NEON capable.

> For integer-heavy benchmarks, the ARM JIT compiler _will_ get into
> the 10x-100x region. Some characteristics of these low-end devices
> heavily penalize interpreters in comparison to native code.

Does it mean for soft-float devices integer internal representation needs/will be used instead of doubles (as on X86/X64) and NaN tagging to achieve this performance?

> [I want to let the ARM interpreter sit in the git repo for a week
> or two of testing and then release beta7. Work on the ARM JIT
> compiler will start after that.]

Heading to test 'scimark.lua' on my iPhone 3GS :)

Best regards,
-- 
Adam Strzelecki