Why does wasm-opt consume all cores?

jjpe · June 13, 2020, 1:45am

When compiling a Rust crate to WASM in release mode, there is a utility being used called wasm-opt that, when called, ramps up my CPU like almost no software I've ever seen*: it has work for all cores, which is quite a feat since I have 16 of them. In addition, it can keep all those cores fed for roughly 3-4 minutes straight.

So I'm wondering: what kinds of optimizations does wasm-opt perform that consume so much computing power?

*exceptions are compiling a crate that has a lot of (transitive) Rust dependencies, and (if I were so inclined) video rendering.

daboross · June 13, 2020, 2:16am

I don't know of any documentation on describing all the optimization passes wasm-opt does, but the source is public?

In particular, each file in binaryen/src/passes at main · WebAssembly/binaryen · GitHub implements a different optimization pass. If you're interested, the names of each file are fairly descriptive (things like "Inlining" and "DeadCodeElimination", "MergeBlocks", etc.).

binaryen is essentially an entire optimizing compiler, and the wasm-opt frontend just happens to compile from WASM to WASM.

bjorn3 · June 13, 2020, 4:15pm

Binaryen uses a thread pool to optimize all functions in parallel:

github.com

WebAssembly/binaryen/blob/65d495f50a8e804c2d38505201ef5afc448dab86/src/passes/pass.cpp#L600


      
                abort();
              }
            }
          } else {
            // non-debug normal mode, run them in an optimal manner - for locality it is
            // better to run as many passes as possible on a single function before
            // moving to the next
            std::vector<Pass*> stack;
            auto flush = [&]() {
              if (stack.size() > 0) {
                // run the stack of passes on all the functions, in parallel
                size_t num = ThreadPool::get()->size();
                std::vector<std::function<ThreadWorkState()>> doWorkers;
                std::atomic<size_t> nextFunction;
                nextFunction.store(0);
                size_t numFunctions = wasm->functions.size();
                for (size_t i = 0; i < num; i++) {
                  doWorkers.push_back([&]() {
                    auto index = nextFunction.fetch_add(1);
                    // get the next task, if there is one
                    if (index >= numFunctions) {

system · September 11, 2020, 4:15pm

This topic was automatically closed 90 days after the last reply. We invite you to open a new topic if you have further questions or comments.

Topic		Replies	Views
State of wasm with rust help	9	1087	March 22, 2022
Wasm-bindgen performance	7	1059	May 19, 2021
Is calling a Rust function from JS (through WASM) every few ms inefficient? help	6	409	November 23, 2022
Is there a transpiler or something that converts Rust code to JavaScript?	5	1613	March 25, 2022
Treeshaking wasm?	32	991	January 27, 2024

Why does wasm-opt consume all cores?

Related Topics