NFC: Propagate global effects via strongly-connected components

[stevenfontanella]

Helps prevent huge visited sets and work queues for edges when traversing large call graphs. After adding support for indirect call effects, the previous algorithm would OOM or timeout when computing global effects for large binaries like calcworker.

Yields a small runtime improvement when tested on calcworker: 0.13451015 -> 0.1328012 (1.3%, averaged from 20 tries with a release build).

Part of #8615.

[tlively]

Nice, I hadn't even considered that you could just compute a single Effects for each SCC, then flow them along the edges between SCCs, completely avoiding any unbounded fixed point analysis.

[tlively]

What's the performance difference between the status quo and this solution?

[stevenfontanella]

What's the performance difference between the status quo and this solution?

Included that in the description, it's 0.13451015 -> 0.1328012 (1.3% improvement) on calcworker. Also after adding indirect effects (the next PR), the runtime is 0.146778 (+9% slower from main without this PR) (EDIT: after fixing some code, the runtime with indirect effects is about the same). All averaged across 20 runs with a release build.

[kripken]

Are the speedup numbers for this particular pass, or the total workload? (if the latter, which?)

[stevenfontanella]

This is from running BINARYEN_PASS_DEBUG=2 bin/wasm-opt -tnh -all --generate-global-effects calcworker_wasm.wasm on the already-optimized calcworker binary from blaze. This is just the runtime for the pass. The number came from output like this:

...
[PassRunner]   running pass: generate-global-effects... [PassRunner] running nested passes
[PassRunner]   running pass: ... 0.132686 seconds.
[PassRunner] nested passes took 0.132686 seconds.

[kripken]

Sounds good about the numbers.