[TypeChecker] ICE in memoryMemberTypes() about solidity

[r0qs]

Thanks @bshastry. I reduced the example to this:

pragma solidity >=0.0;

contract C {
  struct S {
    mapping(uint224 => int64) b;
  }
  function f() external returns(bool)  {}
  fallback() external virtual
  {
    (, C.S[2] memory ret) = abi.decode(abi.encodeCall(this.f, ()), (bool, C.S[2]));
  }
}

I'm wondering whether this issue might be related to the fact that if a struct includes a map, it's restricted to storage use only. So, this might not being detected by the compiler when doing low-level calls, and if that is the case, we may need to remove that assert and properly handle this case. For instance, if you do a direct call:

C.S[2] memory ret = this.f();

It will throw the error:

Error: Type struct C.S[2] memory is only valid in storage because it contains a (nested) mapping.

Maybe @ekpyron or @cameel has a better idea of the issue?

[ekpyron]

It's just a fluke of type-checking. The order in which we check things is a mess in our current analysis, so it happens easily like in this case that we assert against something assuming that it ought to already have triggered an error, while the error is only actually emitted later. So in the end this boils down to "this should be a better, more graceful error", but it's not particularly high priority to fix.