* Alfie
On Wed, Jul 23, 2025 at 09:51:50AM +0100, Alfie Richards wrote:
> On 23/07/2025 09:45, Yury Khrustalev wrote:
> > gcc/testsuite/ChangeLog:
> >
> > * g++.target/aarch64/mv-cpu-features.C: new test.
> > ...
> >
> > +extern "C" void
> > +__init_cpu_features_resolver (unsigned long hwcap, const ifunc_arg_t *arg);
> > +
> > +extern "C" void *
> > +fun_resolver (uint64_t a0, const ifunc_arg_t *a1)
> > +{
> > + ifunc_arg_t arg = {};
> > + arg.size = sizeof (ifunc_arg_t);
> > + arg.hwcap = HWCAP_ATOMICS;
> > + arg.hwcap2 = HWCAP2_RNG;
> > + __init_cpu_features_resolver (arg.hwcap | _IFUNC_ARG_HWCAP, &arg);
> > + return (void *)(uintptr_t)impl;
> > +}
> > +
> > +extern "C" int fun (void) __attribute__((ifunc ("fun_resolver")));
> > +
> > +int main ()
> > +{
> > + int res = fun ();
> > + if (res == 0 && foo () == 2)
> > + return 0;
>
>
> The FMV part of this test doesn't seem quite right.
This tests triggers FMV automatic resovler and checks that it selects the
right implementation when both features are present. Sorry, I should have
added a comment!
> For FMV it is using the auto-generated resolver, which will use
> __init_cpu_features_resolver and get a real read on the host's arch.
To ensure we have stable test results, I use manual ifunc resolver that
calls __init_cpu_features_resolver() with a custom set of hwcap values.
We set one bit in hwcap and one bit in hwcap2, the rest are zero. Because
__init_cpu_features_resolver() will only call __init_cpu_features_constructor()
once during runtime, the subsequent invocation of the automatic FMV resolver
will not read the real hwcap bits from OS.
> This
> test then seems to check that the lse version is selected. So it seems like
> this would only work on systems with lse?
>
> The ifunc part looks good to my eye.
Thanks,
Yury
