https://sourceware.org/bugzilla/show_bug.cgi?id=30568
--- Comment #2 from Nikita Popov <npopov at redhat dot com> --- Thanks for the patch! > I have a couple of concerns with the patch. The first is that it uses a > simple heuristic to detect LTO generated filenames. I have no idea how > robust this will be. Yes, I agree. The checked sub-strings are pretty short, so it seems likely that there will be conflicts with real linker inputs. For example, LLVM has a tool called llvm-lto, so it seems pretty conceivable that "llvm-lto-" appears as part of a file name. It would be preferable if we could exclude all files added by the add_input_file hook from the dependency tracking. In the original report I suggested using the existing lto_output flag to distinguish this (for ld.bfd) -- is there any reason why that wouldn't work? (Another possibility would be to check whether the file still exists when writing the dependency file -- but I'm not sure whether these files have already been deleted at that point, or whether that happens later.) > The second is that the patch strips the filenames from the command line that > forms the first part of the dependency file, as well as the list of files > that comes afterwards. I ma not sure if this is the correct thing to do. > Maybe the command line should remain unfiltered. This is the right thing to do: The part at the start isn't a command line, but rather the actual dependency list in Makefile format, so we do want to filter that. -- You are receiving this mail because: You are on the CC list for the bug.
