On Thu, 8 Apr 2021 21:12:29 +0100, Ximin Luo <[email protected]> wrote: > Stephen Kitt: > > On Thu, 8 Apr 2021 13:11:46 +0100, Ximin Luo <[email protected]> > > wrote: > >>> On Thu, 08 Apr 2021 01:35:49 +0100, Ximin Luo <[email protected]> > >>> wrote: > >>>> Please mark gcc-mingw-w64-base "Multi-Arch: same" so that different > >>>> architectures can be co-installed. This is necessary to support > >>>> co-installing the i686 and x86-64 toolchain packages, both of which > >>>> depend on this package but for their respective architectures. > >>>> > >>>> As far as I can tell this package has the same content on all > >>>> architectures, so this should be safe: > >>>> https://wiki.debian.org/MultiArch/Hints#set_Multi-Arch:_same > >>> > >>> It should be, yes, but I’m curious about why you’d want to install the > >>> packages for both architectures simultaneously. Could you explain the > >>> scenario? > >> > >> Hi, yes my scenario is to co-install both libstd-rust-dev-windows:i386 > >> and libstd-rust-dev-windows:amd64 so that my rustc can cross-compile to > >> both 32-bit and 64-bit windows at the same time. > > > > OK, I understand. I didn’t follow that route with the GCC MinGW-w64 > > package because my initial goal (still relevant, albeit not something > > I’ll have time to work on in the foreseeable future) was to add real > > arches for the targets, so the library packages would ultimately have > > been :win32 and :win64 (or something like that). > > > > By "that route" do you mean the fact that I have a single package name > libstd-rust-dev-windows:$arch ? > > To avoid a potential misunderstanding - this works for LLVM-based compilers > only, not GCC-based compilers like MinGW. In other words, I think what you > are doing currently with gcc-mingw-w64 is fine (except for the missing > Multi-Arch labels, as per this bug) and it matches how the regular gcc > cross compilers work, i.e. they are named > > gcc-$target:$host > > e.g. > gcc-powerpc-linux-gnu:{amd64, i386, .. etc} > gcc-mingw-w64-i686-posix:{amd64, i386, .. etc} > > With LLVM-based compilers such as rustc, there is no need to build a > separate "cross compiler", rustc itself via LLVM already has the ability to > cross-compile to any target. What you need to install however is the > standard library for each target architecture that you want to > cross-compile to, and this is the same regardless of whatever host > architecture you are compiling on, so our packages are named like: > > libstd-rust-dev:$target > libstd-rust-dev-windows:$target > > (The latter convention for the windows packages is a slight hack because > "windows" isn't a Debian architecture, but it works for now.)
That’s what I’m referring to, yes; it would be possible on GCC too — we need different compilers, but the libraries could be packaged using foreign architectures (although support for that isn’t great, hence all the -cross packages for the non-MinGW-w64 GCC cross-compiler packages in Debian). > Regardless of what host architecture you are compiling from, you just need > to install rustc:$host and libstd-rust-dev:$target. > > However with a GCC-based compiler you really do need separate > gcc-$target:$host packages. For example if you are on ARM64 and you wanted > to cross-compile to windows i686 and x86-64, you would need to install > gcc-mingw-w64-i686-posix:arm64 and gcc-mingw-w64-x86-64-posix:arm64, so in > this case it is essential that you do actually have separate packages > gcc-mingw-w64-i686-posix:$host and gcc-mingw-w64-x86-64-posix:$host, and > don't merge them into one package name. I’m fairly familiar with all this ;-). > >> To do that, I need to co-install both gcc-mingw-w64-i686-posix:i386 and > >> gcc-mingw-w64-x86-64-posix, but this is not yet possible due to this bug. > >> > >> Alternatively (and thinking about it, you should probably both), you can > >> mark the compiler packages gcc-mingw-w64-*-* as "Multi-Arch: foreign" - > >> this would e.g. allow gcc-mingw-w64-i686-posix:amd64 to satisfy the > >> dependency from libstd-rust-dev-windows:i386 -> gcc-mingw-w64-i686-posix > >> [no explicit architecture annotation] - without Multi-Arch: foreign, this > >> dependency can only be satisfied by "gcc-mingw-w64-i686-posix:i386". It > >> seems that gcc-mingw-w64-i686-posix:amd64 > >> gcc-mingw-w64-x86-64-posix:amd64 are already co-installable, so this > >> would work for me too. > >> > >> Multi-Arch: foreign is what the gcc cross-compilers do, e.g. > >> gcc-powerpc-linux-gnu. So in summary, amending my original instructions: > >> > >> - Multi-Arch: foreign - gcc-mingw-w64-*-* > >> - Multi-Arch: same - gcc-mingw-w64-base - in fact the Debian QA > >> Multi-Arch hinter[1] goes a step further, it suggests you convert this > >> package to Architecture: all and mark it Multi-Arch: foreign; I'm not > >> sure why this package is arch-dependent in the first place. > >> > >> [1] > >> https://qa.debian.org/[email protected]&piuparts=1#gcc-mingw-w64 > >> > > > > The reason -base is arch-dependent is that it carries the changelog for > > all the arch-dependent packages, so it has to remain arch-dependent to > > provide binNMU changelogs. I’m somewhat reliant on binNMUs for GCC > > updates so I’d rather keep it that way. > > > >> (Note, libstd-rust-dev-windows:i386 doesn't exist in the Debian FTP > >> archive yet, but I'm testing it locally - I'm only now just getting > >> around to updating the rustc package to support cross-compiling to > >> 32-bit windows after you fixed #540782, so thanks for that.) > > > > So that means all this is not for Bullseye, right? That’s somewhat > > reassuring ;-). > > > > Yes this is not for bullseye. It is just adding Multi-Arch: to > debian/control though so I hope it can be done quickly. :) OK; it might not fly for Bullseye but I’ll get it into experimental soon. Regards, Stephen
pgp7DsTSQeszP.pgp
Description: OpenPGP digital signature
