Re: [PATCH 2/5] x86_64, -march=native: POPCNT support

[Alexey Dobriyan]

On Tue, Jul 23, 2019 at 12:15:39AM +0300, Alexey Dobriyan wrote:
> On Mon, Jul 22, 2019 at 11:12:10PM +0200, Peter Zijlstra wrote:
> > On Mon, Jul 22, 2019 at 11:27:20PM +0300, Alexey Dobriyan wrote:
> > > Detect POPCNT instruction support and inline hweigth*() functions
> > > if it is supported by CPU.
> > > 
> > > Detect POPCNT at boot time and conditionally refuse to boot.
> > > 
> > > Signed-off-by: Alexey Dobriyan <adobriyan@xxxxxxxxx>
> > > ---
> > >  arch/x86/include/asm/arch_hweight.h           | 24 +++++++++++++++++++
> > >  arch/x86/include/asm/segment.h                |  1 +
> > >  arch/x86/kernel/verify_cpu.S                  |  8 +++++++
> > >  arch/x86/lib/Makefile                         |  5 +++-
> > >  .../drm/i915/display/intel_display_power.c    |  2 +-
> > >  drivers/misc/sgi-gru/grumain.c                |  2 +-
> > >  fs/btrfs/tree-checker.c                       |  4 ++--
> > >  include/linux/bitops.h                        |  2 ++
> > >  lib/Makefile                                  |  2 ++
> > >  scripts/kconfig/cpuid.c                       |  7 ++++++
> > >  scripts/march-native.sh                       |  2 ++
> > >  11 files changed, 54 insertions(+), 5 deletions(-)
> > 
> > *WHY* ?
> > 
> > AFAICT this just adds lines and complexity and wins aboslutely nothing.
> 
> If CPU is know to have POPCNT, it doesn't make sense to go through RDI.
> Additionally some CPUs (still?) have fake dependency on the destination,
> so "popcnt rax, rdi" is suboptimal.

More general argument is that if -march=native is accepted, compiler will
generate new instructions which will throw #UD on CPUs which aren't
capable, so it doesn't make sense to _not_ go deeper and use all the
knowledge about current CPU.