tejohnson added a comment. In https://reviews.llvm.org/D35081#805761, @yunlian wrote:
> I've sent a reproduce test case to tejohnson. FYI I tracked down what is going on here and reproduced with a small test case. Essentially, two copies of a linkonce odr function were optimized somewhat differently by the compile step and have different instruction counts, with the prevailing one being larger. Depending on the order they are encountered during the thin link's callgraph traversal, we may initially decide to import the smaller non-prevailing copy, and later decide to import the larger prevailing copy. This should be fixed in the function importer, which needs to make a decision about which copy should be imported (the fix is simple, I just need to decide which fix is "right"). This is different than the situation Florian encountered, which is due to passing the full combined index to clang when invoking a distributed backend process, which we don't want users to do. Florian, I'd still like to understand better what the third party you are fixing this for is trying to do, so I can advise. https://reviews.llvm.org/D35081 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
