In this proposal, offsets in memargs were updated to always be u64. For i32 memories, this means it is possible to express an offset that is larger than 4GiB, and this is not rejected in validation. This means that 32-bit runtimes (or runtimes in 32-bit mode) actually now have to store offsets differently and make sure that they handle 64-bit offsets in their bounds checks. This is very strange and wholly unnecessary, since an offset 4GiB or greater will always trap.

I think we should restrict the maximum offset in validation so that this is no longer a runtime concern. The syntax definition and binary format can probably continue to use u64, but in validation, we can enforce that the maximum offset for an i32 memory is 2^32 and the maximum offset for an i64 memory is 2^64.

Thoughts? I can make this change easily if others agree.