On Thu, Jan 16, 2020 at 12:10:28PM +0100, Ard Biesheuvel wrote:

On Thu, 16 Jan 2020 at 12:02, Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jan 15, 2020 at 02:23:39PM -1000, Richard Henderson wrote:

quoted

On 1/15/20 4:26 AM, Catalin Marinas wrote:

quoted

Until the ifunc resolver can work with CPUID, I think we should keep
adding HWCAPn bits. We can revisit this with the toolchain people before
introducing HWCAP3.

Why would the ifunc resolver not be able to use HWCAP_CPUID?

It can indeed check the HWCAP_CPUID but I haven't seen any plans to
implement the next part, actual use of an MRS instruction to read the
corresponding ID_AA64* regs. This MRS emulation was requested by (some
of) the toolchain people, even the architecture gained a feature to
simplify the emulation, but followed by complete silence from the
toolchain folk.

But what infrastructure would the toolchain folks need to provide
here? An ifunc resolver would simply do

void generic_func(...);
void foo_func(...);

void *resolve_foo(long hwcap)
{
   if (hwcap & HWCAP_CPUID) {
       long l;
       asm ("mrs %1, ID_AA64_...") : "=r"(l));
       if (l has 'foo')
         return foo_func;
   }
   return generic_func;
}

so all that is needed for using ID registers to do ifunc resolution is
already there.

If you write the resolver yourself, it should work. I was thinking of
function multiversioning (which I thought using ifunc behind the scenes)
but I'm not sure what the aarch64 support level is (in gcc or clang).

Anyway, I'm not aware of anyone using the MRS emulation (maybe they do
and I haven't heard). I guess it doesn't help that we keep adding HWCAP
bits ;).

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel