On Mon, Jun 21, 2021 at 11:46 AM Michael Ellerman [off-list ref] wrote:

Nathan Chancellor [off-list ref] writes:

quoted

On 6/20/2021 4:59 PM, Nicholas Piggin wrote:

quoted

Excerpts from kernel test robot's message of April 3, 2021 8:47 pm:

quoted

vim +/kvmhv_enter_nested_guest +264 arch/powerpc/kvm/book3s_hv_nested.c

Not much changed here recently. It's not that big a concern because it's
only called in the KVM ioctl path, not in any deep IO paths or anything,
and doesn't recurse. Might be a bit of inlining or stack spilling put it
over the edge.

It appears to be the fact that LLVM's PowerPC backend does not emit
efficient byteswap assembly:

https://github.com/ClangBuiltLinux/linux/issues/1292

https://bugs.llvm.org/show_bug.cgi?id=49610

quoted

powerpc does make it an error though, would be good to avoid that so the
robot doesn't keep tripping over.

Marking byteswap_pt_regs as 'noinline_for_stack' drastically reduces the
stack usage. If that is an acceptable solution, I can send it along
tomorrow.

Yeah that should be OK.

That's fine with me as well.

Can you post the before/after disassembly when
you post the patch?

It should just be two extra function calls, which shouldn't be enough
overhead to be measurable.

The thing I remember is that the 'before' code here is some seriously bad
output from llvm, and it would be helpful to have someone get the compiler
to emit the correct powerpc byteswap instructions and avoid the excessive
stack spilling.

The warning here is just a symptom of a missed optimization and the
same thing probably happens elsewhere on powerpc, even if it doesn't
exceed the stack warning limit.

        Arnd