ritter-x2a wrote: > If we are adding a special case for handling overloads, perhaps a better > approach would be to consider inferring the caller context from the > enveloping function declaration attributes, and allow overload resolution to > pick a device function instead. It would avoid the errors you're trying to > suppress, and it will arguably make things more consistent -- the function > declaration will have the same signature in both host and device compilations.
One problem I encountered when I looked into such a solution is that the host/device attribute(s) are not necessarily parsed at the time when the overload resolution in template arguments of the return type happens, because the `__attribute__((device))` can come after the return type specifier (cf. the `DeviceUserOverloadFunHostDepr2` function in the test case). So far I haven't found an existing mechanism in clang to solve this without a bigger change that introduces some sort of backtracking. I'd be happy about pointers in that direction. https://github.com/llvm/llvm-project/pull/93546 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
