On 6/20/19 10:46 AM, Marek Polacek wrote:
The first test is rejected because build_noexcept_spec first converts a TARGET_EXPR using a user-defined conversion, creatingR::operator bool (&TARGET_EXPR <D.2323, {}>) which then fails when instantiating ("taking address of rvalue"). The second test ICEs in verify_ctor_sanity because there was no ctx.ctor when evaluating the TARGET_EXPR. That is expected, because we create constructors for class and vector types (and PMFs I guess) only, not for scalars. We need finish_compound_literal to turn TARGET_EXPR <D.2306, {}> into TARGET_EXPR <D.2307, 0> Fixed both by handling the noexcept expr like we handle the explicit expr in build_explicit_specifier. Bootstrapped/regtested on x86_64-linux, ok for trunk and 9?
OK, though this pattern still seems fragile. I'll try to poke at it soon. Jason
