mysterymath wrote: I do think that the way LTO behavior depends on both the ambient context (e.g. flags) at time of LTO codegen *and* the ambient context at the time of the creation of each TU's worth of bitcode is *less than ideal*, particularly given how poorly documented and surprising the interactions are.
But, the original context of my RFC was https://github.com/llvm/llvm-project/issues/67698, where the assembler instantiated by LLVM for symbol extraction from module level assembly hard crashes because it has no earthly idea how to parse the assembly. Narrowly scoped to that, I've been happy to watch this patch going along from afar, as I think it basically does resolve that issue. Accordingly, it's kind of independent to the "ambient context problem": having the frontend embed the symbols is a solution that doesn't depend on knowledge of the target features at all. It doesn't really improve anything, but it also doesn't seem to make it any worse. https://github.com/llvm/llvm-project/pull/174995 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
