Data from Binaryen implementation

[tlively]

Support for opportunistically emitting compact imports has landed in Binaryen: WebAssembly/binaryen#8926. Binaryen is just opportunistically and greedily emitting compact imports whenever it sees two imports in a row that have the same module and type or just the same module. It's not reordering imports to try to take advantage of compact imports at all.

I took a look at how compact imports affect the sizes of example programs from 4 different toolchains. The compact-imports and no-compact-imports modules were both run through the same Binaryen pipeline because Binaryen can change module sizes slightly even without running any optimizations. This makes it a more apples-to-apples comparison.

Here's the command I used to produce the files:

for w in java clang dart kotlin; do wasm-opt -O1 -all --disable-compact-imports ${w}.wasm --generate-stack-ir --optimize-stack-ir -o ${w}-no-ci.wasm && wasm-opt -O1 -all ${w}.wasm --generate-stack-ir --optimize-stack-ir -o ${w}-ci.wasm && brotli -9 < ${w}-no-ci.wasm > ${w}-no-ci.wasm.br && brotli -9 < ${w}-ci.wasm > ${w}-ci1.wasm.br; done

I ran the same command again with the logic for emitting encoding 2 (where both module name and externtype are shared) commented out to produce the *-ci1 results.

Here's the raw data:

 3964129 java-no-ci.wasm
 1045621 java-no-ci.wasm.br
 3935537 java-ci1.wasm
 1044310 java-ci1.wasm.br
 3881021 java-ci.wasm
 1037808 java-ci.wasm.br 

32123330 clang-no-ci.wasm
 8706737 clang-no-ci.wasm.br
32122969 clang-ci1.wasm
 8703684 clang-ci1.wasm.br
32122969 clang-ci.wasm
 8703684 clang-ci.wasm.br

 5335989 dart-no-ci.wasm
 1584270 dart-no-ci.wasm.br
 5327196 dart-ci1.wasm
 1584533 dart-ci1.wasm.br
 5327194 dart-ci.wasm
 1584406 dart-ci.wasm.br

 7179987 kotlin-no-ci.wasm
 1403186 kotlin-no-ci.wasm.br
 7174215 kotlin-ci1.wasm
 1403456 kotlin-ci1.wasm.br
 7174232 kotlin-ci.wasm
 1403343 kotlin-ci.wasm.br

And here are the takeaways:

Language	Encoding 1 savings	Additional encoding 2 savings	Encoding 1 savings (compressed)	Additional encoding 2 savings (compressed)
Java	28592 (0.72%)	54516 (1.39%)	1311 (0.13%)	6502 (0.62%)
Clang	361 (0.001%)	0 (0.0%)	3053 (0.04%)	0 (0.0%)
Dart	8793 (0.16%)	2 (0.0%)	-263 (-0.02%)	127 (0.01%)
Kotlin	5772 (0.08%)	-17 (-0.0%)	-270 (-0.02%)	113 (0.01%)

[bvisness]

Thanks for sharing those numbers. I'm also running Binaryen against our original corpus and will share the results soon (probably just during the CG meeting at this point).

Anything to report from the Binaryen implementation process? Any difficulties or surprises?

[tlively]

Our text parser runs in several phases. It needs to store the positions of module declarations found in the initial phase to parse them again later (after it has parsed types). This logic needed some updating to support encoding 2, but it turned out to be a nice refactoring anyway.

Apart from that I'd say it was pretty straightforward. WAT is already such a pain to parse that this doesn't move the needle much, and the binary format is fairly straightforward. Opportunistically emitting compact imports was easy for both encodings.