Gentle ping for this patch. v2 addresses the non-zero-offset base case Jason pointed out and adds coverage for both zero- and non-zero-offset bases.
Bootstrapped and regression-tested on x86_64-pc-linux-gnu. Is this OK for trunk? Thanks, Odysseas ________________________________ From: Odysseas Georgoudis <[email protected]> Sent: 26 July 2026 22:15 To: [email protected] <[email protected]> Cc: Jason Merrill <[email protected]> Subject: [PATCH v2] c++: Adjust constexpr covariant call results [PR126324] Hi Jason, Thanks for catching the non-zero-offset case. In v2, the result-thunk path now preserves null pointers, evaluates the call only once, and adjusts the result to the thunk’s static return type. The test covers both zero- and non-zero-offset bases. Thanks, Odysseas ________________________________ From: Jason Merrill <[email protected]> Sent: 26 July 2026 16:41 To: Odysseas Georgoudis <[email protected]>; [email protected] <[email protected]> Subject: Re: [PATCH] c++: Adjust constexpr covariant call results [PR126324] On 7/24/26 6:48 PM, Odysseas Georgoudis wrote: > This patch adjusts successful scalar virtual-call results to the static > call type after caching the underlying function result. The cache > therefore retains the final overrider's declared result type for direct > calls. > + /* A virtual call with a covariant return type can evaluate the final > + overrider directly, whose result has its declared type rather than > + the static type of the call. Adjust after caching so a direct call > + can reuse the result with its original type. */ > + if (!*non_constant_p > + && !*overflow_p > + && !*jump_target > + && DECL_VIRTUAL_P (fun) > + && result != void_node > + && scalarish_type_p (TREE_TYPE (t)) > + && !same_type_ignoring_top_level_qualifiers_p > + (TREE_TYPE (result), TREE_TYPE (t))) > + result = adjust_temp_type (TREE_TYPE (t), result); What if the base is at a non-zero offset from the derived type? If I change your testcase to add another base, i.e. struct B { virtual constexpr B *clone(bool null) { return null ? nullptr : this; } }; struct C { virtual void dummy () { } }; struct D : C, B { constexpr D *clone(bool null) override { return null ? nullptr : this; } }; then we still hit the same ICE. Jason
