Cranelift: emit_stackmaps should avoid walking the IR if the function doesn't use r32 or r64

[fitzgen]

Enabling stack maps and GC safepoints adds another pass over the IR (emit_stackmaps) regardless whether the function in question actually uses reference types at all. One could imagine (especially with wasm-bindgen's "switch reference types into table indices at the boundaries" approach, that seems likely to be copied by C/C++) that many functions in a module don't use reference types. For these functions, that extra pass over the IR is a waste of time.

We could have a flag on functions that gets set if it uses reference types in any way (and then never gets unset when a particular reference type-using instruction is removed, so we don't have to precisely count them) and in emit_stackmaps, only walk the IR if the flag is set (and probably debug_assert! that it doesn't have any uses of reference types if the flag is not set).

This came up in #1832 (comment) for Wasmtime.

[github-actions]

Subscribe to Label Action

cc @bnjbvr

This issue or pull request has been labeled: "cranelift"

Thus the following users have been cc'd because of the following labels:

• bnjbvr: cranelift

To subscribe or unsubscribe from this label, edit the .github/subscribe-to-label.json configuration file.

Learn more.

[bnjbvr]

Thanks for opening an issue! Do you have any numbers on how much relative time is spent going over the IR, for a large module? This would help quantifying how urgent an issue it is. In any case, agreed that not doing wasteful work is generally good, of course.

The idea of using a flag on the function sounds good. Not sure where to put it, since as far as i can tell it's the first usage of a flag in a Function; could be directly there, or DataFlowGraph, etc.

[fitzgen]

Not sure how much time is spent doing the pass when there aren't any reference types used, it just came up during review.