[Bug ipa/123459] [16 Regression] speculative devirtualization and always_inlining doesn't play nice together

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123459

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
And implementation-wise we should disregard always-inline on indirect calls
(which would mean putting always-inline on the actual call stmt and preserve
this when devirtualizing).  So IMO it's a QOI issue with the way we document
this as "invalid".