On Wed, 20 Sep 2023 10:27:58 +0200,
Richard Fitzgerald wrote:
>
> On 20/9/23 07:51, Takashi Iwai wrote:
> > On Tue, 19 Sep 2023 22:44:28 +0200,
> > Nick Desaulniers wrote:
> >>
> >> On Mon, Sep 18, 2023 at 10:51:29AM +0100, Richard Fitzgerald wrote:
> > (snip)
> >>> +static void cirrus_scodec_test_set_gpio_ref_arg(struct software_node_ref_args *arg,
> >>> + int gpio_num)
> >>> +{
> >>> + struct software_node_ref_args template =
> >>> + SOFTWARE_NODE_REFERENCE(&cirrus_scodec_test_gpio_swnode, gpio_num, 0);
> >>
> >> I'm observing the following error when building with:
> >>
> >> $ make LLVM=1 -j128 allmodconfig sound/pci/hda/cirrus_scodec_test.o
> >>
> >> sound/pci/hda/cirrus_scodec_test.c:151:60: error: initializer element is not a compile-time constant
> >> 151 | SOFTWARE_NODE_REFERENCE(&cirrus_scodec_test_gpio_swnode, gpio_num, 0);
> >> | ^~~~~~~~
> >> /builds/linux/include/linux/property.h:291:37: note: expanded from macro 'SOFTWARE_NODE_REFERENCE'
> >> 291 | .nargs = ARRAY_SIZE(((u64[]){ 0, ##__VA_ARGS__ })) - 1, \
> >> | ^~~~~~~~~~~
> >> /builds/linux/include/linux/kernel.h:57:75: note: expanded from macro 'ARRAY_SIZE'
> >> 57 | #define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]) + __must_be_array(arr))
> >> | ^~~
> >> /builds/linux/include/linux/compiler.h:228:59: note: expanded from macro '__must_be_array'
> >> 228 | #define __must_be_array(a) BUILD_BUG_ON_ZERO(__same_type((a), &(a)[0]))
> >> | ^
> >> /builds/linux/include/linux/compiler_types.h:366:63: note: expanded from macro '__same_type'
> >> 366 | #define __same_type(a, b) __builtin_types_compatible_p(typeof(a), typeof(b))
> >> | ^
> >> /builds/linux/include/linux/build_bug.h:16:62: note: expanded from macro 'BUILD_BUG_ON_ZERO'
> >> 16 | #define BUILD_BUG_ON_ZERO(e) ((int)(sizeof(struct { int:(-!!(e)); })))
> >> | ^
> >
> > Hm, this looks like some inconsistent handling of the temporary array
> > passed to ARRAY_SIZE() in the SOFTWARE_NODE_REFERENCE macro. LLVM
> > can't treat it if it contains a variable in the given array, while GCC
> > doesn't care.
> >
> > A hackish workaround would be the patch like below, but it's really
> > ugly. Ideally speaking, it should be fixed in linux/properties.h, but
> > I have no idea how to fix there for LLVM.
> >
> > Adding more relevant people to Cc.
> >
> >
> > thanks,
> >
> > Takashi
> >
> > --- a/sound/pci/hda/cirrus_scodec_test.c
> > +++ b/sound/pci/hda/cirrus_scodec_test.c
> > @@ -148,8 +148,9 @@ static void cirrus_scodec_test_set_gpio_ref_arg(struct software_node_ref_args *a
> > int gpio_num)
> > {
> > struct software_node_ref_args template =
> > - SOFTWARE_NODE_REFERENCE(&cirrus_scodec_test_gpio_swnode, gpio_num, 0);
> > + SOFTWARE_NODE_REFERENCE(&cirrus_scodec_test_gpio_swnode, 0, 0);
> > + template.args[0] = gpio_num;
> > *arg = template;
> > }
> >
>
> The tests must generate sw nodes dynamically, not a fixed const struct.
> I wanted to avoid directly accessing the internals of the SW node
> structures. Use only the macros.
>
> I can rewrite this code to open-code the construction of the
> software_node_ref_args. Or I can wait a while in case anyone has a
> suggested fix for the macros.
>
> Its seems reasonable that you should be able to create software nodes
> dynamically. A "real" use of these might need to construct the data from
> some other data that is not known at runtime (for example, where the
> ACPI provides some information but not the necessary info).
Yeah, fixing the macro would be ideal.
An easy and compromise solution would be to factor out the
ARRAY_SIZE() call and allow giving nargs explicitly.
e.g.
--- a/include/linux/property.h
+++ b/include/linux/property.h
@@ -285,13 +285,18 @@ struct software_node_ref_args {
u64 args[NR_FWNODE_REFERENCE_ARGS];
};
-#define SOFTWARE_NODE_REFERENCE(_ref_, ...) \
+#define __SOFTWARE_NODE_REFERENCE(_ref_, _nargs_, ...) \
(const struct software_node_ref_args) { \
.node = _ref_, \
- .nargs = ARRAY_SIZE(((u64[]){ 0, ##__VA_ARGS__ })) - 1, \
+ .nargs = _nargs_, \
.args = { __VA_ARGS__ }, \
}
+#define SOFTWARE_NODE_REFERENCE(_ref_, ...) \
+ __SOFTWARE_NODE_REFERENCE(_ref_,\
+ ARRAY_SIZE(((u64[]){ 0, ##__VA_ARGS__ })) - 1,\
+ ##__VA_ARGS__)
+
/**
* struct property_entry - "Built-in" device property representation.
* @name: Name of the property.
--- a/sound/pci/hda/cirrus_scodec_test.c
+++ b/sound/pci/hda/cirrus_scodec_test.c
@@ -148,7 +148,7 @@ static void cirrus_scodec_test_set_gpio_ref_arg(struct software_node_ref_args *a
int gpio_num)
{
struct software_node_ref_args template =
- SOFTWARE_NODE_REFERENCE(&cirrus_scodec_test_gpio_swnode, gpio_num, 0);
+ __SOFTWARE_NODE_REFERENCE(&cirrus_scodec_test_gpio_swnode, 2, gpio_num, 0);
*arg = template;
}
... though I'm not convinced by this change, either.
Takashi
