[CD] Fix TypeSSA on exact parameters in unreachable calls #7628
Conversation
There was a problem hiding this comment.
Hmm, pretty annoying that we have to do this, but the only way I see around it would be to update validation so that expressions with an unreachable child do not impose any constraints on their other children. That would be an even more invasive change.
In addition to returning early in the cases where we don't know how the children should be constrained, can we call a user callback like noteUnknown so that specific users of ChildTyper other than TypeSSA can assert that this never happens?
I don't follow, why would other users need/want that? |
|
The IRBuilder user would want to be able to assert that all the necessary type information has been provided. It would be a bad bug for any of the children to be silently skipped. We are also about to add another user to compute principal types for Outlining, and it would be similarly bad if that use case skipped any children (cc @ashleynh). |
|
Makes sense, thanks. I added a hook so this is customizable. |
| @@ -823,7 +832,9 @@ template<typename Subtype> struct ChildTyper : OverriddenVisitor<Subtype> { | |||
| void visitTupleExtract(TupleExtract* curr, | |||
| std::optional<size_t> arity = std::nullopt) { | |||
| if (!arity) { | |||
| assert(curr->tuple->type.isTuple()); | |||
There was a problem hiding this comment.
Let's keep this assertion for the case when skipUnreachable() is false.
TypeSSA ignored unreachable instructions, but we do validate parts of them,
like calls - a later unreachable param does not disable validation or other
params (see new testcase). Remove the ignoring of such code in TypeSSA and
add proper skipping of unreachable code, where necessary, in ChildTyper.