Re: Dynamic instance properties in TYPE_ARM_CPU

[Eduardo Habkost]

On Thu, Oct 29, 2020 at 11:16:27AM +0100, Andrew Jones wrote:
> On Mon, Oct 26, 2020 at 01:22:19PM -0400, Eduardo Habkost wrote:
> > I've been trying to clean up the qdev property code (to bridge
> > the gaps between qdev and QOM, and between QOM and QAPI), and
> > I've noticed that TYPE_ARM_CPU is the only remaining user of
> > qdev_property_add_static().
> > 
> > qdev_property_add_static() has a misleading name: it won't
> > register a static property.  It is actually a hack to use a
> > static Property variable (defined using DEFINE_PROP*), but
> > register it as a dynamic instance property.
> > 
> > Dynamic instance properties make introspection hard.  What can we
> > do to get rid of them in TYPE_ARM_CPU?
> > 
> > Can we just register all the properties unconditionally, and
> > error out on realize if the requested CPU configuration is
> > incompatible with the available CPU features?
> > 
> > The following properties are registered as dynamic instance
> > properties at arm_cpu_post_init():
> > "cntfrq", "reset-cbar", "reset-hivecs", "rvbar", "has_el2",
> > "has_el3", "cfgend", "vfp", "neon", "dsp", "has-mpu",
> > "pmsav7-dregion", "secure-memory", "pmu", "idau", "init-svtor",
> > "tag-memory", "secure-tag-memory".
> 
> Hi Eduardo,
> 
> Are properties added at runtime with
> 
>   if (this_config_wants_this_property)
>     object_property_add(...)
> 
> also considered dynamic instance properties? If so, then
> arm as many more than the ones listed above. And, if not,
> then it should be pretty easy to convert the ones listed
> above to this if-want-add pattern.

Yeah, all object_property_add() calls are also dynamic instance
properties.

Being conditional makes them more difficult to convert to
introspectable class properties (unless the condition is a
function of class information only).

-- 
Eduardo