From: Florian Fainelli Date: Fri, 16 Feb 2018 15:00:31 -0800 Subject: [Buildroot] [RFC 0/6] Support 32-bit binaries on a 64-bit architecture In-Reply-To: <20180216215928.4ada222c@windsurf.lan> References: <20180216005612.69593-1-mmayer@broadcom.com> <20180216064557.d6zfemlr7c2zj4x3@tarshish> <20180216215928.4ada222c@windsurf.lan> Message-ID: <71d93c8c-3a75-aa9b-2dd9-283ab9c6178a@gmail.com> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: buildroot@busybox.net On 02/16/2018 12:59 PM, Thomas Petazzoni wrote: > Hello, > > On Fri, 16 Feb 2018 08:45:57 +0200, Baruch Siach wrote: > >> On Thu, Feb 15, 2018 at 04:56:06PM -0800, Markus Mayer wrote: >>> This series contains a proposal for supporting 32-bit libraries and >>> binaries on a 64-bit platform. >>> >>> There are some limitations and prerequisites as to the scope this has >>> been tested. >>> >>> - It has only been tested on ARM/ARM64. >>> - It has only been tested with an external toolchain >>> (https://github.com/Broadcom/stbgcc-6.3/releases). >>> - It requires that the Aarch64 compiler be compiled with "multi-lib" >>> enabled. This ensures that the sysroot doesn't contain a lib64 -> >>> lib symlink. Instead, lib and lib64 are separate directories. >>> - It has only been tested with a fairly limited number of packages. >>> There might be other packages that don't correctly pass --libdir, >>> similar to bzip2 as mentioned below. >> >> There is one thing I am missing here. AFAICS, this series does not add support >> for building 32-bit libraries in addition to their 64-bit versions. So the >> 32-bit binaries that are to run on target must have their dependencies >> satisfied from the toolchain provided libraries. Is that correct? > > Yes, indeed, I'm interested in understanding that part as well. It > seems like the assumption is that the external toolchain has a 32-bit > sysroot, and only the libraries from this sysroot are necessary. Correct, that is what the stbgcc-6.3 toolchain contains. > > This feels very limited/restricted to very specific use cases and > therefore not very usable in a broader context. I suppose it could be reasonably easy to extend on that and either require a second toolchain to be built or just point to a different location? > > And extending that to building packages is really difficult with > Buildroot: > > (1) Kconfig has no way to easily allow to select for each and every > package whether it should be built 32-bit, 64-bit or both. > > (2) Our package infrastructure is entirely based on the fact that a > given package is built only once. So if you assume that you > solved problem (1) by adding gazillions of Config.in options to > each and every package, there is still no way the zlib package > will be built two times, once 32-bit and 64-bit. No without major > changes to the package infrastructure. > > So supporting multilib has already been discussed several times, but > nobody ever came up with a proposal that could reasonably solve > problems (1) and (2). Unfortunately, unless there are some proposals to > solve those problems, I don't think we want to have some very limited > and highly specific multilib support, that only works with external > toolchains and that only works for 32-bit libraries provided by the > external toolchain. Would you recommend that this logic be sort of done using a two pass build, by that I mean something like an external top-level Makefile would be controlling the buildroot build process and do something like: - build all relevant packages for the 32-bit architecture (that could include the toolchain as well) - build all relevant packages for the 64-bit architecture, and copy the 32-bit binaries and libraries of interest from the previously build 32-bit environment? I suppose such a thing could be done, it has the advantage of not tweaking buildroot too much, although we still need to teach the 64-bit variant to accept seeing "foreign" binaries. -- Florian