Define cast<i32>(u32) overflow behavior

[rootjalex]

The simplifier normalizes reinterpret<i32>(u32) (which has defined behavior) to cast<i32>(u32), which has undefined behavior. Currently, the following code produces inconsistent output:

{
    Expr x = UIntImm::make(UInt(32), 0xFFFFFFF0u);
    Expr e = i32(x + 1);
    e = simplify(e);
    std::cout << e << std::endl;

    e = reinterpret(Int(32), x + 1);
    e = simplify(e);
    std::cout << e << std::endl;
}
{
    Expr x = Variable::make(UInt(32), "x");
    Expr e = i32(x + 1);
    e = simplify(e);
    e = substitute("x", UIntImm::make(UInt(32), 0xFFFFFFF0), e);
    e = simplify(e);
    std::cout << e << std::endl;

    e = reinterpret(Int(32), x + 1);
    e = simplify(e);
    e = substitute("x", UIntImm::make(UInt(32), 0xFFFFFFF0), e);
    e = simplify(e);
    std::cout << e << std::endl;
}

Output:

signed_integer_overflow(0)
-15
signed_integer_overflow(1)
signed_integer_overflow(2)

@abadams and I agreed to simply define cast<i32>(u32) to have the same 2's complement wrapping behavior as reinterpret<i32>(u32). I think the only place this behavior is encoded is in the simplifier (LLVM codegen produces a no-op for either case because LLVM types don't have signedness).

Now the output of the above code is:

-15
-15
-15
-15

[steven-johnson]

SGTM

[steven-johnson]

No idea if these are related to this change or not, but:

#7810
#7811

(Just reported by Google-internal fuzz testing after our most recent merge from GitHub)