unique_ptr fails to build with GCC 9 (and 10)

[shenki]

When building hostboot with a newer GCC (see https://github.com/open-power/op-build/pull/4595):

    ../../../../src/include/util/impl/unique_ptr.H:408:30: error: cannot
    bind rvalue reference of type ‘std::enable_if<true, STDUniquePtrTest::practice_struct&&>::type’
    {aka ‘STDUniquePtrTest::practice_struct&&’} to lvalue of type ‘STDUniquePtrTest::practice_struct’
      408 |                 return iv_ptr[i];

The failure is in the test case (no in hostboot seems to trigger it) here https://github.com/open-power/hostboot/blob/master-p10/src/usr/testcore/lib/unique_ptr.H#L116

            i[4].x = 99;

            if (i[4].x != 99)
            {
                TS_FAIL("Array access through unique_ptr returned incorrect value (expected 99)");
            }

            if (i[3].x != 4)
            {
                TS_FAIL("Array access through unique_ptr returned incorrect value (expected 4)");
            }

Removing this code allows the build to complete. However, this code seems to be valid. If we build it using GCC's unique_ptr.h, it compiles and the test case runs.

I don't understand what the templates are doing here, but it seems they are the cause of the problem:

            /* @brief Index operator for unique_ptr<T[]>.
             *
             * operator[] is not callable on a unique_ptr to non-array types.
             *
             * @param[in] i   Index into the array
             * @return        Value of T::operator[](i)
             */
            template<typename Index, typename X = T>
            typename enable_if<is_array<X>::value, decltype(declval<X>()[declval<Index>()])>::type
            operator[](const Index i)
            {
                return iv_ptr[i];
            }

[dcrowell77]

We've got some time scheduled later this year to bump up our compiler. In the meantime I'll see if someone can take a quick peek.

[ibmzach]

Looks like the way that decltype works has changed perhaps. Anyhow, fixed the issue with a commit that should make its way out of gerrit soon.

[shenki]

Thanks @ibmzach