When yocto is searching for lib dependencies it uses SONAMES and RPATHS to find
the right one. When we compile a rust package which dynamically links to rust's
libstd, yocto will not be able to find the matching dependency in shlib and
will complain with a warning. Even worse, if you have compiled go-runtime, the
build will find go-runtime as provider of libstd and will complain that you
missed a RDEPENDS to go-runtime - this is completely wrong obviously.
This patch adds SONAME to rust's stdlib. In this way yocto will create the
required provider and packages dynamically linking to it will be able to find
libstd-rs as a provider of stdlib.
Signed-off-by: Andrei Gherzan <andrei@resin.io>
Support for mips32 big (mips) and little endian (mipsel) for mips32
and mips32r2 cpus. The big endian target can be verified with the
qemumips machine.
The "target_c_int_width field is added to librustc_back since rust
1.22.0 because not all rust targets (e.g. msp430-none-elf, avr)
have 32-bit int types any more.
Switch the build process to using Rust's new build system called
rustbuild. At the same time stop using jemalloc on all platforms and
switch to the system allocator since that was simplier to do.
Technically these dependencies change between versions. It just happens
that these two versions didn't change the dependencies but future
versions will so preemptively make the adjustment.
The path to the rust source changes from 1.15 so preemptively make
changes to prepare for that.
Signed-off-by: Derek Straka <derek@asterius.io>
Signed-off-by: Doug Goldstein <cardoe@cardoe.com>
The way LLVM builds its native utilities is not safe (it runs a
recursive cmake in the same work directory as the main instance). To
prevent those instances from stepping on each other, we'll build the two
native utilities separately before starting the main build.
Fixes issue #143
This patches handle changing the directories that things are installed
in but appear to be unused. They were likely used before rust-cross was
split out to not install files.
Signed-off-by: Doug Goldstein <cardoe@cardoe.com>
Not sure what version this patch appears to have become unnecessary but
I verified it by checking the output in
build/tmp/sstate-control/manifest-x86_64-rust-native.populate_sysroot
and the installed files and their paths didn't change with and without
this patch.
Signed-off-by: Doug Goldstein <cardoe@cardoe.com>
Fix building a toolchain installer with bitbake
-c populate_sdk in packages compiler-rt,
compiler-rt-dev and compiler-rt-staticdev.
Signed-off-by: Leon Anavi <leon.anavi@konsulko.com>
(cherry picked from commit fcbebdf58f)
Redid how compiler-rt is built by using the Rust makefiles to build it
instead of compiler-rt's build system. Rust itself deviated from using
compiler-rt's build system and provided their own rules requiring the
source tree to be configured to use them.
We need the Rust snapshot (stage0) for the build environment not the
host environment. This is not affecting us currently because we are only
building native versions of this today but should improve the behavior
of #81 or if any Rust components are built for the target machine.
Since we no longer use shared source we can modify the source directory
as we build. This drops another patch we need to build and simplifies
the addition of Rust 1.12.1 in #97.
I'm honestly not sure what this patch does since everything appears to
build the same with and without it. I did a `tree` of the sysroots and
images and I don't see any file differences. Something like this used to
be necessary in Gentoo but I dropped it a few releases back.
The host system where things are being built really should control what
toolchain is used for stage0. This in theory should improve #23 but I'm
not sure how to specify a hash per host sys.
Jonathan Creekmore
Adam Schwalm
Andrei Gherzan
Derek Straka
Tyler Hall
Kyle Russell
Johan Anderholm
Tyler Hall
Doug Goldstein
Steven Walter
Leon Anavi
Cody P Schafer