On Fri, Nov 02, 2018 at 01:48:11PM +0000, Ferruh Yigit wrote:
> On 11/2/2018 12:42 PM, Ferruh Yigit wrote:
> > On 10/23/2018 10:23 PM, Yongseok Koh wrote:
> >> This is a workaround to prevent a crash, which might be caused by
> >> optimization of newer gcc (7.3.0) on Intel Skylake.
> >>
> >> Bugzilla ID: 97
> >
> > After checking the defect description again, this is the issue observed in
> > rte_memcpy() implementation for AVX2, compiler uses AVX512F instructions
> > while
> > compiling it which causes the failure, so this may be a compiler defect but
> > we
> > don't know the root cause yet.
>
> Is the issue only with gcc, and only with specific version of gcc?
> If so can we reduce the disabling avx512 only to that gcc version?
>
> >
> > I think best solution is to find the root cause and fix either avx2
> > implementation or compiler, but this seems won't be soon, at least for rc2.
> >
> > What this patch does is to prevent compiler to use avx512f instruction when
> > "CONFIG_RTE_ENABLE_AVX512=n".
> >
> > Concern is this will affect all DPDK generated code for x86, but since
> > rte_memcpy() in header file there is no way to disable using avx512f
> > instructions locally for rte_memcpy().
> > I can't think of any other solution for now, so OK to go with this patch for
> > now. Please find below comment.
> >
> >>
> >> Cc: sta...@dpdk.org
> >>
> >> Signed-off-by: Yongseok Koh <ys...@mellanox.com>
> >> ---
> >> config/x86/meson.build | 5 +++++
> >> mk/rte.cpuflags.mk | 5 +++++
> >> 2 files changed, 10 insertions(+)
> >>
> >> diff --git a/config/x86/meson.build b/config/x86/meson.build
> >> index 33efb5e547..e10ba872ac 100644
> >> --- a/config/x86/meson.build
> >> +++ b/config/x86/meson.build
> >> @@ -47,6 +47,11 @@ endif
> >> if cc.get_define('__AVX512F__', args: march_opt) != ''
> >> dpdk_conf.set('RTE_MACHINE_CPUFLAG_AVX512F', 1)
> >> compile_time_cpuflags += ['RTE_CPUFLAG_AVX512F']
> >> +else
> >> +# disable compiler's AVX512F support as a workaround for Bug 97
> >> + if cc.has_argument('-mavx512f')
> >> + machine_args += '-mno-avx512f'
> >> + endif
> >> endif
> >>
> >> dpdk_conf.set('RTE_CACHE_LINE_SIZE', 64)
> >> diff --git a/mk/rte.cpuflags.mk b/mk/rte.cpuflags.mk
> >> index 43ed84155b..8fdb0cc2c3 100644
> >> --- a/mk/rte.cpuflags.mk
> >> +++ b/mk/rte.cpuflags.mk
> >> @@ -68,6 +68,11 @@ endif
> >> ifneq ($(filter $(AUTO_CPUFLAGS),__AVX512F__),)
> >> ifeq ($(CONFIG_RTE_ENABLE_AVX512),y)
> >> CPUFLAGS += AVX512F
> >> +else
> >> +# disable compiler's AVX512F support as a workaround for Bug 97
> >> +ifeq ($(shell $(CC) --target-help | grep -q mavx512f && echo 1), 1)
> >
> > This will not work for ICC, and do we need this? AUTO_CPUFLAGS already
> > should
> > have what you are looking for, so I think this check can be removed.
This is different from AUTO_CPUFLAGS as it tries to check compiler flag support.
And per your question, I have only tested it with gcc, so I agree on applying it
only for gcc. Will submit v2. But, I don't think we need to check gcc version as
there's no fix reported yet in a newer gcc version and this patch would have
very limited impact. avx512f support is quite new and kinda experimental so
far. Dropping a bit of performance would be better than crash. :-)
Thanks for your review,
Yongseok
> >> +MACHINE_CFLAGS += -mno-avx512f
> >> +endif
> >> endif
> >> endif
> >>
> >>
> >
>
