[Custom Descriptors] Optimize RefCast to RefCastDesc in GlobalStructInference

[kripken]

No description provided.

[tlively]

It's not enough to check that we have a single global for the descriptor type; we must also check that there are no allocations of that descriptor type besides the initializer for that one global.

[kripken]

That is something the pass verifies,

binaryen/src/passes/GlobalStructInference.cpp

Lines 113 to 115 in 46b37d3

	// We cannot optimize types that appear in a struct.new in a function, which
	// we just collected and merge now.
	HeapTypes unoptimizable;

Unless I misunderstood you?

[tlively]

Great, I had forgotten we already did that.

[kripken]

There are some TODOs here, but the larger question is whether this seems worthwhile. I see a 1% code size regression on j2cl for example - do we expect the benefit to be larger than that? @tlively @jakobkummerow

[tlively]

This is primarily a performance improvement. ref.cast_desc is faster than other casts.

[kripken]

Makes sense, but do we think it is fast enough to justify a 1% size regression? Given that iiuc we think the VM can do much of this too.