smith: Fix types recorded in ref.func
#1965
Conversation
This commit fixes a minor issue in `wasm-smith` when generating the `ref.func` instruction when the GC proposal is disabled but shared-everything-threads is enabled. In this situation all `ref.func` instructions were recorded as `RefType::FUNCREF` but this wasn't correct for shared function types. The logic here is applicable both with and without the GC proposal is enabled so the precise type is always pushed now.
It's no longer present on `main`
But shared-everything extends the GC proposal, so I don't think we should even allow this set up. It's like disabling bulk memory but enabling reference types. |
| // If GC is disabled then favor the "top" type of the input type. | ||
| if !self.config.gc_enabled { | ||
| return Ok(ty); | ||
| return Ok(match ty { | ||
| HeapType::Concrete(i) => { | ||
| let CompositeType { inner, shared } = &self.ty(i).composite_type; | ||
| let ty = match inner { | ||
| CompositeInnerType::Array(_) => AbstractHeapType::Array, | ||
| CompositeInnerType::Struct(_) => AbstractHeapType::Struct, | ||
| CompositeInnerType::Func(_) => AbstractHeapType::Func, |
There was a problem hiding this comment.
The only case that this would be valid for is (ref null? func). If GC is disabled, then struct and array references will still be invalid.
Which again points me towards this trying to do something that doesn't make sense...
Alternative fix from bytecodealliance#1965
|
Seems reasonable to apply that fix too yeah. Personally the two proposals are only loosely coupled in my head because we'll probably want pretty separate gates/paths for both, but it doesn't really matter in the interim.
Technically we do allow this, but we don't allow disabling reference types and enabling gc. Or at least not in
True, but if GC is disabled, then the struct/array stuff won't show up. I gambled that an insertion of |
* smith: Disable shared-everything-threads without GC Alternative fix from #1965 * Review comments
This commit fixes a minor issue in
wasm-smithwhen generating theref.funcinstruction when the GC proposal is disabled but shared-everything-threads is enabled. In this situation allref.funcinstructions were recorded asRefType::FUNCREFbut this wasn't correct for shared function types. The logic here is applicable both with and without the GC proposal is enabled so the precise type is always pushed now.