On Mon, 3 Oct 2022, Martin Storsjö wrote: > On Mon, 3 Oct 2022, Mark Harmstone wrote: > >> Hi Martin, >> >>> As I assume you're aware, lld's mingw port also supports PDB generation - >>> and the description of this option also sounds like it's chosen to match >>> lld's option for outputting PDB files - that's good! >> >> Yes, that's right. One notable difference is that the parameter here is >> optional, unlike with lld, making it a lot easier to fit this into e.g. >> CMake toolchain files or LDFLAGS. > > LLD also has got that behaviour, since > https://github.com/llvm/llvm-project/commit/2c52ddf31f5421c5373923535b958b84c79772e3 > in 2019. That's in particular why I wanted to make sure that this case works > the same in binutils too. > >> It looks like the equals sign is mandatory when providing optional >> parameters, otherwise it interprets the filename as another parameter. > > Yep, that's the case in LLD too. > > Unfortunately I didn't think of this behaviour initially when I first added > this option - otherwise we could have had e.g. --pdb as a boolean option to > just output to the default name, and e.g. --output-pdb= if you wanted > to specify the name. But oh well, "-pdb=" works, and I guess it isn't the > worst thing in the world. > >> But it does mean that the form "-pdb=out.pdb" will work on both ld and lld, >> which I think is the most important thing. > > TBH, I consider the "-pdb=" case equally important too - that's what most > people would use in the end. FWIW, I'm actually a bit concerned about the interop between binutils and lld here. I don't want interop between binutils and lld to work only for some subset of the used parameter forms, I'd like it to work for all commonly used forms. First off, the (slightly awkward) syntax that lld uses for an optional empty output name, "-pdb=" really should be handled by binutils too - handling that doesn't conflict with anything else and should be simple to support. This is the format of the option that I've been recommending people to use, and this has been in use in third party projects for years already - e.g. this: https://code.videolan.org/videolan/vlc/-/blob/master/configure.ac#L429 This should be trivial to support in your patch: diff --git a/ld/emultempl/pep.em b/ld/emultempl/pep.em index 11216830dd3..538fdf5054b 100644 --- a/ld/emultempl/pep.em +++ b/ld/emultempl/pep.em @@ -926,7 +926,7 @@ gld${EMULATION_NAME}_handle_option (int optc) if (emit_build_id == NULL) emit_build_id = xstrdup (DEFAULT_BUILD_ID_STYLE); pdb = 1; - if (optarg) + if (optarg && optarg[0]) pdb_name = xstrdup (optarg); break; } (And the same for pe.em.) Secondly, for explicitly naming an output file, I've documented to end users that they can use either -Wl,-pdb= or -Wl,-pdb, - see https://github.com/mstorsjo/llvm-mingw/blob/master/README.md?plain=1#L175. In the original implementation in the mingw frontend in lld in 2018, the "-pdb " format was the only format for the option: https://github.com/llvm/llvm-project/commit/b7d50115ba4900da6db7afb6460ad42ff19ba6a2 Only one year later with the implicit output name, the "-pdb=" and "-pdb=" form was added: https://github.com/llvm/llvm-project/commit/2c52ddf31f5421c5373923535b958b84c79772e3 In one of my test scripts, I use the initial form of the option, -Wl,-pdb,: https://github.com/mstorsjo/llvm-mingw/blob/master/run-tests.sh#L234 It seems like Wine has picked up on the -Wl,-pdb, form: https://gitlab.winehq.org/wine/wine/-/blob/wine-7.18/tools/winegcc/winegcc.c#L467 Also here are a couple of other cases I found that all seem to use that form: https://youtrack.jetbrains.com/issue/KT-47175/How-to-generate-kotlin-native-debug-info-filesPDB-on-windows-platform https://git.kernel.dk/?p=fio.git;a=commitdiff;h=76bc30ca118fda404f19c17d97bafdba9779c4c2 So with all these users, I'd be kinda hesitant to change lld's interpretation of this option form, and to have binutils ld in parallel interpreting that form differently. What do you think? // Martin