emscripten optimisations

[atomictag]

Hi there, it's me again :)

I was trying out a few emcc options and added the following:

EMCC_OPT+= -flto 

then ran some benchmarks (BLS12-381, emcc 3.1.66):

MCL invVecInPlace (Fr, 1000 elements)
before: 0.31ms
after: 0.31ms 
==> no improvement on Fr math it seems

MCL invVecInPlace (Fp, 1000 elements)
before: 0.72ms
after: 0.44ms 
==> 39% improvement

MCL mulVec (MSM G1, 256 elements):
before:  30.22ms
after: 16.45ms
==> 45% improvement (note the fastest BLS MSM WASM I've ever used scored 13.5ms)

MCL pairing (averaged over 1000 random pairings):
before: 3.58ms
after: 2.86ms
==> 25% improvement

In my tests it appears that -flto produces significant improvements on math ops in G1/G2 (and possibly GT).

I ran this modified library over 600 complex tests on various algorithms and real-life applications that I have accumulated and encountered no issue. However my focus is solely on BLS12.381 in "Eth mode" so I can't speak for other curves/settings.

But it looks promising!

Other options I have tried

EMCC_OPT+=-O3 -DNDEBUG -std=c++11 # compiles fine, no measurable effects
EMCC_OPT+=-ffast-math # no measurable effects, might be unsafe
EMCC_OPT+= -s FILESYSTEM=0 -s ASSERTIONS=0 # this would make sense to reduce code size (maybe?)

What do you think? 🙏

[herumi]

Thank you for your interesing report. I've tried those options.

The size of mcl_c.js is reduced from 660097 bytes to 622165 bytes with -flto.

function/usec	no -flto	-flto
pairing	2412.492	2086.236
millerLoop	1006.681	917.331
finalExp	1334.797	1250.731
precomputedMillerLoop	735.137	741.364
G1::add	3.084	2.855
G1::dbl	2.285	2.015
G1::mul	279.962	203.932
G2::add	7.657	6.774
G2::dbl	3.462	3.243
G2::mul	459.43	405.152

-ffast-math may improve floating point operations but mcl does not use them. -s FILESYSTEM=0 -s ASSERTIONS=0 did not change the binary. So I will append -flto option.

[atomictag]

awesome - and thank you for the new release