feat/perf: abi benchmarks

[DaniPopes]

Results

sol-types: good
dyn-abi: not so good, optimizations needed, probably in DynSolType::encode_single

ethabi/encode/single    time:   [86.689 ns 87.177 ns 87.735 ns]
ethabi/encode/struct    time:   [172.32 ns 172.65 ns 173.02 ns]
ethabi/decode/word      time:   [33.044 ns 33.168 ns 33.310 ns]
ethabi/decode/dynamic   time:   [72.122 ns 72.273 ns 72.452 ns]

dyn-abi/encode/single   time:   [91.784 ns 92.412 ns 93.107 ns]
dyn-abi/encode/struct   time:   [724.03 ns 729.16 ns 734.79 ns]
dyn-abi/decode/word     time:   [72.234 ns 72.568 ns 72.967 ns]
dyn-abi/decode/dynamic  time:   [99.585 ns 99.764 ns 99.945 ns]

sol-types/encode/single time:   [43.869 ns 44.193 ns 44.581 ns]
sol-types/encode/struct time:   [55.131 ns 55.286 ns 55.423 ns]
sol-types/decode/word   time:   [1.3890 ns 1.4000 ns 1.4141 ns]
sol-types/decode/dynamic
                        time:   [36.848 ns 36.983 ns 37.135 ns]

[prestwich]

can you also bench these against the previous implementation?

[DaniPopes]

can you also bench these against the previous implementation?

which previous implementation? If you mean the one before my changes in #52, there is no difference in ABI logic/performance

[prestwich]

I mean the abi impl relied on by ethers@2.0.0

[DaniPopes]

that is ethabi

[prestwich]

oh i'm just blind. got it. thank you

[DaniPopes]

After 281cec3:

dyn-abi/encode/single   time:   [69.210 ns 69.292 ns 69.382 ns]                                  
                        change: [-26.237% -25.649% -25.227%] (p = 0.00 < 0.05)
                        Performance has improved.
dyn-abi/encode/struct   time:   [626.60 ns 627.97 ns 629.56 ns]                                   
                        change: [-12.694% -12.201% -11.654%] (p = 0.00 < 0.05)
                        Performance has improved.

dyn-abi/decode/word     time:   [49.180 ns 49.331 ns 49.572 ns]                                 
                        change: [-33.199% -32.358% -31.439%] (p = 0.00 < 0.05)
                        Performance has improved.
dyn-abi/decode/dynamic  time:   [84.247 ns 84.431 ns 84.620 ns]                                   
                        change: [-18.215% -17.543% -17.087%] (p = 0.00 < 0.05)
                        Performance has improved.

[prestwich]

well, let's keep investigating :)

[DaniPopes]

Maybe we can skip tokenizing by encoding directly the DynSolValue? Looking at the "struct" bench, ethabi is much more concise because of ethabi::encode(tokens: &[ethabi::Token]), while dyn-abi has to declare all the types, and then also clone the input on each call.

[prestwich]

approving this. can you rebase, and open issues for future optimization work?