5682 – Silently wrong CTFE result possibly related to operator overloading and expression order

The test function below will produce different results depending on whether it is executed normally or via CTFE with latest DMD (git master at a47637b0):
---
import std.stdio;
struct A {
 int n;
 auto opBinary(string op : "*")(A rhs) {
 return A(n * rhs.n);
 }
}
A foo(A[] lhs, A[] rhs) {
 A current;
 for (size_t k = 0; k < rhs.length; ++k) {
 current = lhs[k] * rhs[k]; // This is the crucial line.
 }
 return current;
}
auto test() {
 return foo([A(1), A(2)], [A(3), A(4)]);
}
void main() {
 enum tc = test();
 writefln("compile-time: %s; run-time: %s", tc.n, test().n);
}
---
compile-time: 4; run-time: 8
---
A few observations:
 - At compile-time, the first element from the first factor (lhs[0]) seems to be read twice, replacing A(2) with another value doesn't change the CTFE result.
 - Swapping lhs[k] and rhs[k] will change the CTFE result to 6 (supporting the above assumption).
 - If foo() is modified to work on e.g. int instead of A values, the problem doesn't occur.
 - If the assignment in question is replaced with »A a1 = lhs[k]; A a2 = rhs[k]; current = a1 * a2;«, the bug does not longer occur.

D2 fix:
https://github.com/D-Programming-Language/dmd/commit/0c8750880ad31e0715546daa68be4d69f4af88c0
D1:
https://github.com/D-Programming-Language/dmd/commit/669cafeb9fecc98d5f4689f379f638a9661f0b35