[OE-core] [BUG REPORT] failed to build valgrind on qemuarm

[Andre McCurdy]

On Tue, Jan 19, 2016 at 3:23 PM, Khem Raj <raj.khem at gmail.com> wrote:
>
>> On Jan 19, 2016, at 1:38 PM, Andre McCurdy <armccurdy at gmail.com> wrote:
>>
>> On Tue, Jan 19, 2016 at 1:00 PM, Khem Raj <raj.khem at gmail.com> wrote:
>>>
>>>> On Jan 18, 2016, at 11:27 PM, Rongqing Li <rongqing.li at windriver.com> wrote:
>>>>
>>>> 2: on qemuarma9
>>>>
>>>> | arm-wrs-linux-gnueabi-gcc -march=armv7-a -mfloat-abi=softfp  -marm -mthumb-interwork --sysroot=/work/wr/buildarea/arm/bitbake_build/tmp/sysroots/qemuarma9 -DHAVE_CONFIG_H -I. -I/work/wr/buildarea/arm/bitbake_build/tmp/work/armv7a-vfp-wrs-linux-gnueabi/valgrind/3.11.0-r0/valgrind-3.11.0/memcheck/tests -I../.. -I/work/wr/buildarea/arm/bitbake_build/tmp/work/armv7a-vfp-wrs-linux-gnueabi/valgrind/3.11.0-r0/valgrind-3.11.0 -I/work/wr/buildarea/arm/bitbake_build/tmp/work/armv7a-vfp-wrs-linux-gnueabi/valgrind/3.11.0-r0/valgrind-3.11.0/include -I/work/wr/buildarea/arm/bitbake_build/tmp/work/armv7a-vfp-wrs-linux-gnueabi/valgrind/3.11.0-r0/valgrind-3.11.0/coregrind -I../../include -I/work/wr/buildarea/arm/bitbake_build/tmp/work/armv7a-vfp-wrs-linux-gnueabi/valgrind/3.11.0-r0/valgrind-3.11.0/VEX/pub -I../../VEX/pub -DVGA_arm=1 -DVGO_linux=1 -DVGP_arm_linux=1 -DVGPV_arm_linux_vanilla=1   -Winline -Wall -Wshadow -Wno-long-long -g -fno-stack-protector    -O2 -pipe -g  -fno-omit-frame-pointer -fvisibility=default -O0 -c -o sh-mem-random.o /work/wr/buildarea/arm/bitbake_build/tmp/work/armv7a-vfp-wrs-linux-gnueabi/valgrind/3.11.0-r0/valgrind-3.11.0/memcheck/tests/sh-mem-random.c
>>>> | {standard input}: Assembler messages:
>>>> | {standard input}:1107: Error: selected processor does not support ARM mode `vld1.64 {d7},[r3]'
>>>> | {standard input}:1107: Error: selected processor does not support ARM mode `vst1.64 {d7},[r2]'
>>>> | Makefile:2467: recipe for target 'sh-mem-random.o’ failed
>>>
>>>
>>> So you disable neon but do not choose a particular implementation which allows so. e.g. you should have -mtune=cortex-a9 which allows neon to be optional
>>> but thats probably just the semantic thing. Below code should check for __ARM_NEON__ internal define being set as well before resorting to use neon instructions.
>>
>> It's a test app which is failing, so I think we have 3 solutions:
>>
>> 1) Disable tests which don't build for all supported targets (the
>> traditional OE approach). See:
>>
>>  http://git.openembedded.org/openembedded-core/tree/meta/recipes-devtools/valgrind/valgrind/remove-arm-variant-specific.patch
>>
>> 2) Tweak CFLAGS so that the test always compiles but fails at run-time
>> on non-neon targets.
>>
>> 3) Tweak test code so that it avoids NEON instructions on non-neon
>> targets (risk then is that the run-time test will appear to pass on
>> non-neon targets when it should actually fail).
>>
>> Any preference?
>
> option 3
>

Yes, I think that's right in this case. The sh-mem-random.c test isn't
really about testing neon, so falling back to C for non-neon targets
seems correct.

Patch sent.