error recovery improvement for stmtexprs

[lattner]

Bugzilla Link	4070
Version	unspecified
OS	All

Extended Description

In this example:

struct fooo {} ;
struct fooo x;

void test(void) {

  *({
    0;
    x/x;
    });
}

We emit two diagnostics: one about x/x (good) and one about "indirection requires a pointer operand". The problem is that we correctly kill the expression tree for x/x, but that means that we build and return a compound statement with just { 0; } and the stmt expr gets built with type int. We should somehow propagate invalidity of the last statement in a stmt expr up into the stmt expr itself.

This came from #4437 .

[philnik777]

This has been fixed in Clang 11: https://godbolt.org/z/o8axnbrd6