This reverts commit 5dda6c427a.
This breaks rust-llvm builds with:
ERROR: QA Issue: rust-llvm-native: LIC_FILES_CHKSUM points to an invalid file: TOPDIR/BUILD/work/x86_64-linux/rust-llvm-native/1.47.0-r0/rustc-1.47.0-src/src/llvm-project/llvm/COPYRIGHT [license-checksum]
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
Because rust-llvm-ncsa.inc requires rust-llvm.inc we do not want that
the latter overrides the license checksum set in the former.
Signed-off-by: Florin Sarbu <florin@balena.io>
Place generated artifacts in the directory suggested by bitbake (${B})
instead of directly in the source directory. This has multiple
advantages such as the ability to share source directory between
multiple machine types without risking cross contamination.
arm-unknown-linux-gnueabihf was incorrectly used as llvm-target instead
of armv7-unknown-linux-gnueabihf when building for some Cortex A SoCs.
This may cause segfaults in non trivial rust applications on ARMv7
when e.g. +a7 is passed to LLVM. It seems to differ between different
versions of the compiler and LLVM versions.
* BB_NUMBER_THREADS is number of bitbake tasks running in
parallel, not parallelization inside individual tasks
* use oe.utils.parallel_make_argument to make sure it
works even when people add e.g. "-l 10" in PARALLEL_MAKE
* with the recent improvements for rust-native build time, I wanted
to rerun some build time tests from
https://github.com/shr-project/test-oe-build-time
here are the results on AMD Threadripper 3970x with 128GB ram:
BB is BB_NUMBER_THREADS
PM is PARALLEL_MAKE
TIME BB PM Description
20:50 * * zeus based build with 1.37 rust (BB/PM has no impact)
14:50 8 32,48,64 dunfell, 1.43 as in d2ff87ca55 (PM has no impact, because bootstrap uses BB)
9:50 8 32,48,64 dunfell, 1.43 with this patch applied, it doesn't get faster after some threashold of PM
13:32 8 64 dunfell, 1.43 with this patch applied and "rust.inc: cut build time in half" (afcb58e5b9) reverted
13:30 8 64 dunfell, 1.43 with "rust.inc: cut build time in half" (afcb58e5b9) as well as "rust.inc: run bootstrap.py in parallel" (40a6bd8a8d) reverted to see if 1.43 builds faster without any meta-rust improvements compared to 1.37
94:47 1 64 dunfell, 1.43 as in d2ff87ca55 (either something went wrong or explicit "-j 1" disables some parallelism done by default already
13:40 * * zeus based build with 1.39 rust (BB/PM has no impact)
10:50 8 64 zeus based build with 1.39 rust as proposed in jansa/new-zeus-branch
80:51 8 1 zeus based build with 1.39 rust as proposed in jansa/new-zeus-branch with "-j 1" PARALLEL_MAKE
Signed-off-by: Martin Jansa <Martin.Jansa@gmail.com>
d55cce6b8b6b510bf4905f19b949f7995af57a4d added a use of
BB_NUMBER_THREADS which is not whitelisted in Poky. This caused machines
with a different number of CPUs to have different sstate for
rust-native.
As a string this was actually being ignored. It mostly didn't matter
because max-atomic-width falls back to target-pointer-size, and they are
usually the same. However, at least on i586, 64-bit atomics are
supported with a 32-bit pointer size.
The whole point of rust-native is that it should be common for all
targets. If we reference TARGET variables during the build of
rust-native, then bitbake will build a different version for different
TARGETS.
Since ac83d22eb5
The armvX arch definition is not present in TUNE_FEATURES but is now in
MACHINEOVERRIDES. Improve parsing so that MACHINEOVERRIDES is also checked.
Fixes#240
Tested on raspberrypi3 with balenaOS which uses rust.
Signed-off-by: Zubair Lutfullah Kakakhel <zubair@balena.io>
Since 1.31.0 cargo requires curl to support http2. Enabling http2 in
curl results in a dependency loop. This commit disables the use of
multiplexing and http2.
