We often generate IR like: Check(Int32:@x) ... Foo(KnownInt32:@x) It would be valid for any optimization that somehow proves the type of @x to remove the Check node entirely. But then, AI might fail on an assertion at Foo() because of the KnownInt32 use kind, if AI isn't smart enough to construct the same proof that the former optimization used for removing the Check. The correct solution is probably to remove the compile-time assertions about Known use kinds having already been checked. It's OK for those to be debug-only JIT assertions.
I might end up fixing this before I do property type inference. It looks like that code might reveal this bug.
Created attachment 260369 [details] the patch
Comment on attachment 260369 [details] the patch r=me
Landed in http://trac.webkit.org/changeset/189219