Re: [PATCH] net/mlx5: restrict workaround of gcc AVX512F bug

[Thomas Monjalon <thomas@monjalon.net>]

13/11/2018 01:14, Stephen Hemminger:
> On Mon, Nov 12, 2018, 4:01 PM Thomas Monjalon <thomas@monjalon.net wrote:
> 
> > A bug was found when the inline function mlx5_tx_complete()
> > is optimized with AVX512F instructions. It corrupts an offset
> > in the instructions vmovdqu8 of the AVX2 version of rte_mov128(),
> > used in rte_memcpy(), which is called in rte_mempool_put_bulk().
> >
> > All the above functions are inline. So the workaround is
> > to disable AVX512F optimization for the functions calling the
> > top-level function of this call stack, i.e. mlx5_tx_complete().
> > All GCC versions supporting AVX512 are supposed to be affected.
> >
> > The root cause is not identified yet. It may be thought that
> > more related bugs may happen in other functions.
> > That's why the initial workaround was to disable AVX512F globally.
> > This patch takes the risk of applying the workaround only for the
> > functions known to be affected, in order to preserve the optimization
> > everywhere else.
> >
> > Bugzilla ID: 97
> > Fixes: 8d07c82b239f ("mk: disable gcc AVX512F support")
> >
> > Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
> 
> The additional annotations clutter the code.
> How big a performance hit is it to disable for whole driver? Or just use
> memcpy instead of rte_memcpy?

rte_memcpy() is used via rte_mempool_put_bulk().
I am not going to change it to memcpy...

About disabling AVX512F for the whole driver, the goal
of this patch is to reduce the scope of the workaround.
If a per-function scope is not chosen, then we can stay with
a global safe scope.

If you are interested to know more, the bugzilla has tons of infos:
	https://bugs.dpdk.org/show_bug.cgi?id=97

Given that we don't get much help on this major GCC bug,
we are probably going to stay on the safe side.
Anyway I must stop working (alone) on this bug, and instead,
focus on making 18.11 out.