Memory use of dart2wasm

[mkustermann]

Currently it seems that dart2wasm (a relatively simple compiler), uses 80% more memory than binaryen

When compiling flute specifically:

• dart2wasm uses 600 MB
• binaryen uses 330 MB

There's several reasons for this

• the AST is large
• the AST uses growable arrays instead of fixed-length arrays in nodes (in particular FunctionNode, FunctionType, Arguments)
• the AST seems to include the source code as bytes
• the dart2wasm code generator keeps the IL instruction objects around for all functions until the very end

We could consider investigating if there's some low hanging fruits.

[mkustermann]

/cc @jensjoha

[osa1]

the dart2wasm code generator keeps the IL instruction objects around for all functions until the very end

dart2wasm used to directly serialize generated instructions, but that meant generating the code after generating imports, globals etc. as instructions refer to these by index.

In most cases that's not a problem, when you need an import you can generate it (give it an index) and use it in the code that needed it, but IIRC there were some cases where we needed the flexibility to re-order things.

I don't remember what the those cases were exactly, but I remember discussing it quite a bit at meetings as the change to introduce instruction classes etc. were quite large and came with memory costs. (runtime costs didn't matter that much as wasm-opt uses most of the runtime)

Do you have an idea on how to improve this?

We could generate 0s for all indices, keep track of which index used where, and patch the code at the end when copying the function code buffers to the module. That would store one Uint8List per function and reduce the GC pressure quite a bit potentially.

[osa1]

Something else that we could improve: dart2wasm currently doesn't know about --strip-wasm and generates names always. Most of the names should be the strings from the kernel AST (TODO: confirm this) so the names themselves shouldn't cost extra memory, but we still allocate the vectors, maps etc. to map function, type, field indices to names. We could make dart2wasm aware of --strip-wasm and not pupolate these data structures to save some memory.