[PATCH 1/2] gcc-target: Ensure buildtools-extended-tarball doesn't use arch=native

[PATCH 1/2] gcc-target: Ensure buildtools-extended-tarball doesn't use arch=native

[Richard Purdie]

A nativesdk BBCLASSEXTEND was added to gcc-target without realising this
would pass arch=native through to it for x86-64. This heavily optimises
gcc output for the host its running on meaning it can't be reused via
sstate on other machines.

Add class-target overrides here to get the desired behaviour. All
targets have been covered for completeness.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
---
 meta/recipes-devtools/gcc/gcc-target.inc | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/meta/recipes-devtools/gcc/gcc-target.inc b/meta/recipes-devtools/gcc/gcc-target.inc
index 8cb432780f5..6263e814971 100644
--- a/meta/recipes-devtools/gcc/gcc-target.inc
+++ b/meta/recipes-devtools/gcc/gcc-target.inc
@@ -15,11 +15,11 @@ EXTRA_OECONF_append_linuxstdbase = " --enable-clocale=gnu"
 
 ARMFPARCHEXT ?= ""
 
-EXTRA_OECONF_append_armv6 = " --with-arch=armv6${ARMFPARCHEXT}"
-EXTRA_OECONF_append_armv7a = " --with-arch=armv7-a${ARMFPARCHEXT}"
-EXTRA_OECONF_append_armv7ve = " --with-arch=armv7ve${ARMFPARCHEXT}"
-EXTRA_OECONF_append_arc = " --with-cpu=${TUNE_PKGARCH}"
-EXTRA_OECONF_append_x86-64 = " --with-arch=native"
+EXTRA_OECONF_append_armv6_class-target = " --with-arch=armv6${ARMFPARCHEXT}"
+EXTRA_OECONF_append_armv7a_class-target = " --with-arch=armv7-a${ARMFPARCHEXT}"
+EXTRA_OECONF_append_armv7ve_class-target = " --with-arch=armv7ve${ARMFPARCHEXT}"
+EXTRA_OECONF_append_arc_class-target = " --with-cpu=${TUNE_PKGARCH}"
+EXTRA_OECONF_append_x86-64_class-target = " --with-arch=native"
 
 # libcc1 requres gcc_cv_objdump when cross build, but gcc_cv_objdump is
 # set in subdir gcc, so subdir libcc1 can't use it, export it here to
-- 
2.25.1

[PATCH 2/2] abi_version/staging: Bump versions to force rebuild after sstate corruption

[Richard Purdie]

A "broken" buildtools-extended-tarball has been released into the wild
where it is optimising binaries for the host processsor. This is fine in
local usage but in a non-homogeneous cluster like our autobuilder, this
results in SIGILL on other machines when the sstate is shared amongst them
and is painful to debug.

The buildtools tarball has been fixed but we need to invalidate the hash
equivalence and sstate data. This change does that. Adding to OE-Core
rather than autobuilder local changes as its good to illustrate how to
do this and the issue is potentially wider than just Yocto Project
infrastructure.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
---
 meta/classes/staging.bbclass | 1 +
 meta/conf/abi_version.conf   | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/meta/classes/staging.bbclass b/meta/classes/staging.bbclass
index 5b04f88b2d8..de3a19815a3 100644
--- a/meta/classes/staging.bbclass
+++ b/meta/classes/staging.bbclass
@@ -93,6 +93,7 @@ SYSROOT_PREPROCESS_FUNCS ?= ""
 SYSROOT_DESTDIR = "${WORKDIR}/sysroot-destdir"
 
 python do_populate_sysroot () {
+    # SYSROOT 'version' 2
     bb.build.exec_func("sysroot_stage_all", d)
     bb.build.exec_func("sysroot_strip", d)
     for f in (d.getVar('SYSROOT_PREPROCESS_FUNCS') or '').split():
diff --git a/meta/conf/abi_version.conf b/meta/conf/abi_version.conf
index 2bdc55695bd..e04343b47c1 100644
--- a/meta/conf/abi_version.conf
+++ b/meta/conf/abi_version.conf
@@ -12,4 +12,4 @@ OELAYOUT_ABI = "12"
 # a reset of the equivalence, for example when reproducibility issues break the
 # existing match data. Distros can also append to this value for the same effect.
 #
-HASHEQUIV_HASH_VERSION  = "1"
+HASHEQUIV_HASH_VERSION  = "3"
-- 
2.25.1

Re: [OE-core] [PATCH 1/2] gcc-target: Ensure buildtools-extended-tarball doesn't use arch=native

[Khem Raj]

On 5/1/20 9:46 AM, Richard Purdie wrote:
> A nativesdk BBCLASSEXTEND was added to gcc-target without realising this
> would pass arch=native through to it for x86-64. This heavily optimises
> gcc output for the host its running on meaning it can't be reused via
> sstate on other machines.
> 
> Add class-target overrides here to get the desired behaviour. All
> targets have been covered for completeness.
> 

yeah perhaps it should have been --with-arch=generic for non target option,

this looks ok

> Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
> ---
>   meta/recipes-devtools/gcc/gcc-target.inc | 10 +++++-----
>   1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/meta/recipes-devtools/gcc/gcc-target.inc b/meta/recipes-devtools/gcc/gcc-target.inc
> index 8cb432780f5..6263e814971 100644
> --- a/meta/recipes-devtools/gcc/gcc-target.inc
> +++ b/meta/recipes-devtools/gcc/gcc-target.inc
> @@ -15,11 +15,11 @@ EXTRA_OECONF_append_linuxstdbase = " --enable-clocale=gnu"
>   
>   ARMFPARCHEXT ?= ""
>   
> -EXTRA_OECONF_append_armv6 = " --with-arch=armv6${ARMFPARCHEXT}"
> -EXTRA_OECONF_append_armv7a = " --with-arch=armv7-a${ARMFPARCHEXT}"
> -EXTRA_OECONF_append_armv7ve = " --with-arch=armv7ve${ARMFPARCHEXT}"
> -EXTRA_OECONF_append_arc = " --with-cpu=${TUNE_PKGARCH}"
> -EXTRA_OECONF_append_x86-64 = " --with-arch=native"
> +EXTRA_OECONF_append_armv6_class-target = " --with-arch=armv6${ARMFPARCHEXT}"
> +EXTRA_OECONF_append_armv7a_class-target = " --with-arch=armv7-a${ARMFPARCHEXT}"
> +EXTRA_OECONF_append_armv7ve_class-target = " --with-arch=armv7ve${ARMFPARCHEXT}"
> +EXTRA_OECONF_append_arc_class-target = " --with-cpu=${TUNE_PKGARCH}"
> +EXTRA_OECONF_append_x86-64_class-target = " --with-arch=native"
>   
>   # libcc1 requres gcc_cv_objdump when cross build, but gcc_cv_objdump is
>   # set in subdir gcc, so subdir libcc1 can't use it, export it here to
> 
> 
> 
> 

Re: [OE-core] [PATCH 1/2] gcc-target: Ensure buildtools-extended-tarball doesn't use arch=native

[Adrian Bunk]

On Fri, May 01, 2020 at 05:46:01PM +0100, Richard Purdie wrote:
> A nativesdk BBCLASSEXTEND was added to gcc-target without realising this
> would pass arch=native through to it for x86-64. This heavily optimises
> gcc output for the host its running on meaning it can't be reused via
> sstate on other machines.
> 
> Add class-target overrides here to get the desired behaviour. All
> targets have been covered for completeness.
>...
> +EXTRA_OECONF_append_x86-64_class-target = " --with-arch=native"
>...

What do you expect --with-arch=native to do for the target
of a cross compiler?

cu
Adrian

Re: [OE-core] [PATCH 1/2] gcc-target: Ensure buildtools-extended-tarball doesn't use arch=native

[Richard Purdie]

On Tue, 2020-05-05 at 13:14 +0300, Adrian Bunk wrote:
> On Fri, May 01, 2020 at 05:46:01PM +0100, Richard Purdie wrote:
> > A nativesdk BBCLASSEXTEND was added to gcc-target without realising
> > this
> > would pass arch=native through to it for x86-64. This heavily
> > optimises
> > gcc output for the host its running on meaning it can't be reused
> > via
> > sstate on other machines.
> > 
> > Add class-target overrides here to get the desired behaviour. All
> > targets have been covered for completeness.
> > ...
> > +EXTRA_OECONF_append_x86-64_class-target = " --with-arch=native"
> > ...
> 
> What do you expect --with-arch=native to do for the target
> of a cross compiler?

gcc-target isn't a cross compiler.

Cheers,

Richard