On Mon, Oct 1, 2018 at 2:37 PM Christophe Lyon christophe.lyon@linaro.org wrote:
Hi Arnd,
After our discussion during Connect, I've uploaded to tarballs of binary toolchains for arm and aarch64: http://people.linaro.org/~christophe.lyon/toolchain-arm-none-linux-gnueabihf... http://people.linaro.org/~christophe.lyon/toolchain-aarch64-none-linux-gnu.t...
I've managed to build the Linux kernel with them except a problem with popcount on aarch64. I know Kugan worked on this, but I'm not sure about the current status.
I can update my scripts to upload these toolchains daily (after "Daily bump"), let me know if that suits you. They only support the C language, that's the stage1 compiler, I think it's OK for you.
These toolchains are built in the ST compute farm, on RHEL6 hosts, but it ran correctly on my Ubuntu-18 workstation.
Is that OK with you?
I gave the 32-bit toolchain a quick spin on my own box, it worked great here. I've discussed this with Kevin during Connect, he can probably knows more about how close we are to having multi-toolchain support ready to integrate this in kernelci.
I'm also adding the kernel-build-reports list to Cc here, so others can comment. For background: There was a request about including limited testing of daily gcc snapshots in kernelci, which would catch both regressions introduced in the compiler and newly added warnings about possible kernel bugs early. Christophe is already building arm and aarch64 compilers for his own toolchain testing every day and said he could upload those every day to a public location.
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
Christophe: I noticed that your arm compiler is over 50% slower than the gcc-8.1 release build I have. I assume this is mostly the result of your binary being a debug build rather than a release build that skips a lot of the checks. Does that sound right? Is this what you intended? I suppose using the debug build of the toolchain catches more bugs in the compiler itself, but it also adds some cost to the build servers.
Arnd
On Mon, Oct 01, 2018 at 04:34:28PM +0200, Arnd Bergmann wrote:
On Mon, Oct 1, 2018 at 2:37 PM Christophe Lyon
These toolchains are built in the ST compute farm, on RHEL6 hosts, but it ran correctly on my Ubuntu-18 workstation.
I gave the 32-bit toolchain a quick spin on my own box, it worked great here. I've discussed this with Kevin during Connect, he can probably knows more about how close we are to having multi-toolchain support ready to integrate this in kernelci.
You'd be better off talking to Matt Hart (CCed) about that. It'll be a little while yet.
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
I think we should be covering x86 as well, it's probably mainstream enough :)
On Mon, Oct 1, 2018 at 4:40 PM Mark Brown broonie@kernel.org wrote:
On Mon, Oct 01, 2018 at 04:34:28PM +0200, Arnd Bergmann wrote:
On Mon, Oct 1, 2018 at 2:37 PM Christophe Lyon
These toolchains are built in the ST compute farm, on RHEL6 hosts, but it ran correctly on my Ubuntu-18 workstation.
I gave the 32-bit toolchain a quick spin on my own box, it worked great here. I've discussed this with Kevin during Connect, he can probably knows more about how close we are to having multi-toolchain support ready to integrate this in kernelci.
You'd be better off talking to Matt Hart (CCed) about that. It'll be a little while yet.
Right, of course. I also meant to talk to Matt during Connect, but didn't get get around to catching up with him.
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
I think we should be covering x86 as well, it's probably mainstream enough :)
Agreed. The thing here is that the binaries that Christophe produces are a byproduct of his more elaborate toolchain tests, and he's not doing those for any architectures other than arm/aarch64 today. Before we ask him to do any additional work that is outside of his normal scope, I think we should try to integrate the binaries we already have. Once that works, we can go looking for volunteers that can build x86 (and possibly ppc, mips, microblaze and rv32 to cover all the kernelci targets) compilers, either in an automated way, or manually but less frequently.
Arnd
On Mon, Oct 01, 2018 at 04:55:08PM +0200, Arnd Bergmann wrote:
On Mon, Oct 1, 2018 at 4:40 PM Mark Brown broonie@kernel.org wrote:
On Mon, Oct 01, 2018 at 04:34:28PM +0200, Arnd Bergmann wrote:
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
I think we should be covering x86 as well, it's probably mainstream enough :)
Agreed. The thing here is that the binaries that Christophe produces are a byproduct of his more elaborate toolchain tests, and he's not doing those for any architectures other than arm/aarch64 today. Before we ask him to do any additional work that is outside of his normal scope, I think we should try to integrate the binaries we already have. Once that works, we can go looking for volunteers that can build x86 (and possibly ppc, mips, microblaze and rv32 to cover all the kernelci targets) compilers, either in an automated way, or manually but less frequently.
Or for something like this perhaps we should just be keying off the source anyway rather than trying to integrate random binaries from different places? Question for Matt anyway.
On Mon, 1 Oct 2018 at 16:34, Arnd Bergmann arnd@arndb.de wrote:
On Mon, Oct 1, 2018 at 2:37 PM Christophe Lyon christophe.lyon@linaro.org wrote:
Hi Arnd,
After our discussion during Connect, I've uploaded to tarballs of binary toolchains for arm and aarch64: http://people.linaro.org/~christophe.lyon/toolchain-arm-none-linux-gnueabihf... http://people.linaro.org/~christophe.lyon/toolchain-aarch64-none-linux-gnu.t...
I've managed to build the Linux kernel with them except a problem with popcount on aarch64. I know Kugan worked on this, but I'm not sure about the current status.
I can update my scripts to upload these toolchains daily (after "Daily bump"), let me know if that suits you. They only support the C language, that's the stage1 compiler, I think it's OK for you.
These toolchains are built in the ST compute farm, on RHEL6 hosts, but it ran correctly on my Ubuntu-18 workstation.
Is that OK with you?
I gave the 32-bit toolchain a quick spin on my own box, it worked great here. I've discussed this with Kevin during Connect, he can probably knows more about how close we are to having multi-toolchain support ready to integrate this in kernelci.
I'm also adding the kernel-build-reports list to Cc here, so others can comment. For background: There was a request about including limited testing of daily gcc snapshots in kernelci, which would catch both regressions introduced in the compiler and newly added warnings about possible kernel bugs early. Christophe is already building arm and aarch64 compilers for his own toolchain testing every day and said he could upload those every day to a public location.
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
Christophe: I noticed that your arm compiler is over 50% slower than the gcc-8.1 release build I have. I assume this is mostly the result of your binary being a debug build rather than a release build that skips a lot of the checks. Does that sound right? Is this what you intended? I suppose using the debug build of the toolchain catches more bugs in the compiler itself, but it also adds some cost to the build servers.
Indeed, I configure GCC with --enable-checking, so as to catch more problems. In the past I missed regressions for lack of using this flag.
Arnd
On Mon, 1 Oct 2018 at 15:34, Arnd Bergmann arnd@arndb.de wrote:
On Mon, Oct 1, 2018 at 2:37 PM Christophe Lyon christophe.lyon@linaro.org wrote:
Hi Arnd,
After our discussion during Connect, I've uploaded to tarballs of binary toolchains for arm and aarch64: http://people.linaro.org/~christophe.lyon/toolchain-arm-none-linux-gnueabihf... http://people.linaro.org/~christophe.lyon/toolchain-aarch64-none-linux-gnu.t...
I've managed to build the Linux kernel with them except a problem with popcount on aarch64. I know Kugan worked on this, but I'm not sure about the current status.
I can update my scripts to upload these toolchains daily (after "Daily bump"), let me know if that suits you. They only support the C language, that's the stage1 compiler, I think it's OK for you.
These toolchains are built in the ST compute farm, on RHEL6 hosts, but it ran correctly on my Ubuntu-18 workstation.
Is that OK with you?
I gave the 32-bit toolchain a quick spin on my own box, it worked great here. I've discussed this with Kevin during Connect, he can probably knows more about how close we are to having multi-toolchain support ready to integrate this in kernelci.
I'm also adding the kernel-build-reports list to Cc here, so others can comment. For background: There was a request about including limited testing of daily gcc snapshots in kernelci, which would catch both regressions introduced in the compiler and newly added warnings about possible kernel bugs early. Christophe is already building arm and aarch64 compilers for his own toolchain testing every day and said he could upload those every day to a public location.
I want to stress here that the current work regarding toolchains in KernelCI is aimed solely at testing a limited static matrix of toolchains. Reliable regression testing of the Kernel relies on the rest of the stack (userspace, toolchains) changing infrequently. Before I started this work, I don't believe the toolchains used in KernelCI have been changed ever.
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
I don't speak for all of KernelCI, but I feel like this is... "ToolchainCI" and is possibly out of scope.
Christophe: I noticed that your arm compiler is over 50% slower than the gcc-8.1 release build I have. I assume this is mostly the result of your binary being a debug build rather than a release build that skips a lot of the checks. Does that sound right? Is this what you intended? I suppose using the debug build of the toolchain catches more bugs in the compiler itself, but it also adds some cost to the build servers.
Arnd
Kernel-build-reports mailing list Kernel-build-reports@lists.linaro.org https://lists.linaro.org/mailman/listinfo/kernel-build-reports
On Mon, Oct 01, 2018 at 05:38:18PM +0100, Matt Hart wrote:
On Mon, 1 Oct 2018 at 15:34, Arnd Bergmann arnd@arndb.de wrote:
Before I started this work, I don't believe the toolchains used in KernelCI have been changed ever.
We have bumped the versions several times as the kernel
We probably don't want to add all combinations to the build farm, but doing arm and aarch64 builds on linux-next with the compiler of the day seems like a nice addition to me.
I don't speak for all of KernelCI, but I feel like this is... "ToolchainCI" and is possibly out of scope.
We have talked about it in the past (years ago, when the feature was first floated) - from a kernel point of view the benefit is that when new GCC releases come out there's less breakage to triage on both sides, less likelyhood that we have to carry workarounds or whatever.
kernel-build-reports@lists.linaro.org
-
Arnd Bergmann -
Christophe Lyon -
Mark Brown -
Matt Hart