On Thu, 12 Mar 2026 13:16:26 +0000
Amit Machhiwal <[email protected]> wrote:
> Hi Christhophe,
>
> Thanks for looking at the patch. Please find my comments inline:j
>
> On 2026/03/10 11:54 AM, Christophe Leroy (CS GROUP) wrote:
> >
> >
> > Le 10/03/2026 à 11:15, Amit Machhiwal a écrit :
> > > GCC 15 reports the below false positive '-Wmaybe-uninitialized' warning
> > > in vphn_unpack_associativity() when building the powerpc selftests.
> > >
> > > # make -C tools/testing/selftests TARGETS="powerpc"
> > > [...]
> > > CC test-vphn
> > > In file included from test-vphn.c:3:
> > > In function ‘vphn_unpack_associativity’,
> > > inlined from ‘test_one’ at test-vphn.c:371:2,
> > > inlined from ‘test_vphn’ at test-vphn.c:399:9:
> > > test-vphn.c:10:33: error: ‘be_packed’ may be used uninitialized
> > > [-Werror=maybe-uninitialized]
> > > 10 | #define be16_to_cpup(x) bswap_16(*x)
> > > | ^~~~~~~~
> > > vphn.c:42:27: note: in expansion of macro ‘be16_to_cpup’
> > > 42 | u16 new = be16_to_cpup(field++);
> > > | ^~~~~~~~~~~~
> > > In file included from test-vphn.c:19:
> > > vphn.c: In function ‘test_vphn’:
> > > vphn.c:27:16: note: ‘be_packed’ declared here
> > > 27 | __be64 be_packed[VPHN_REGISTER_COUNT];
> > > | ^~~~~~~~~
> > > cc1: all warnings being treated as errors
> > >
> > > When vphn_unpack_associativity() is called from hcall_vphn(), this error
> > > is not seen during compilation because GCC 15 seems to consider 'retbuf'
> > > always populated from the hypervisor which is eventually referred by
> > > 'be_packed'. However, GCC 15's dataflow analysis can’t prove the same
> > > before the first dereference when vphn_unpack_associativity() is called
> > > from test_one() with pre-initialized array of 'struct test'. This
> > > results in a false positive warning which is promoted to an error under
> > > '-Werror'. This problem is not seen when the compilation is performed
> > > with GCC 13 and 14.
> > >
> > > Suppress the warning locally around the offending statement when
> > > building with GCC 15 using a diagnostic pragma. This keeps the build
> > > working while limiting the scope of the suppression to the specific
> > > statement that triggers the false positive. An issue [1] has also been
> > > created on GCC bugzilla.
> >
> > Usually when we get this kind of warning this is because the code is too
> > complex. We should try to make it more obvious instead of just hiding the
> > warning.
>
> The real issue here is that GCC 15 emits '-Wmaybe-uninitialized' due to
> type punning between __be64[] and __b16* when accessing the buffer via
> be16_to_cpup(). The underlying object is fully initialized but GCC 15
> fails to track the aliasing due to the strict aliasing violation here.
Nope, I think it is tracking it correctly.
The writes to be_packed[] are of 64bit values.
The only reads of that memory are 16bit ones through field[].
With 'strict aliasing' the compiler doesn't have to order those accesses.
Indeed, it is allowed to completely optimise away the first loop.
If you cast to 'unsigned char *' then the accesses do have to be ordered.
gcc will also treat accesses to different members of a union as being ordered
(the C stand doesn't require this, IIRC s/union/struct/ is valid).
David
> Please refer [1] and [2].
>
> The selftest compiles fine with '-fno-strict-aliasing'. I see that when
> we build vphn.c while compiling the kernel, the top level Makefile
> includes '-fno-strict-aliasing' flag always.
>
> So, I believe the same flag should be used to build vphn tests when
> compiling vphn.c via the selftests. I'll send the v2 to achieve this
> thus avoiding the compilation failure.
>
> Please let me know you have different thoughts.
>
> [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124427
> [2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99768
>
> ~Amit
>
> >
> > Here the for loop is a bit misleading.
> >
> > >
> > > [1]
> > > https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgcc.gnu.org%2Fbugzilla%2Fshow_bug.cgi%3Fid%3D124427&data=05%7C02%7Cchristophe.leroy%40csgroup.eu%7C06a4d55b55f24c5cf00208de7e8e3676%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C639087346428583316%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=xEfO94N6IfGYhmmapNFduv3OrMarxpjTpZR6B38uR1s%3D&reserved=0
> > >
> > > Fixes: 58dae82843f5 ("selftests/powerpc: Add test for VPHN")
> > > Reviewed-by: Vaibhav Jain <[email protected]>
> > > Signed-off-by: Amit Machhiwal <[email protected]>
> > > ---
> > > arch/powerpc/platforms/pseries/vphn.c | 15 +++++++++++++++
> > > 1 file changed, 15 insertions(+)
> > >
> > > diff --git a/arch/powerpc/platforms/pseries/vphn.c
> > > b/arch/powerpc/platforms/pseries/vphn.c
> > > index 3f85ece3c872..9bc891143fec 100644
> > > --- a/arch/powerpc/platforms/pseries/vphn.c
> > > +++ b/arch/powerpc/platforms/pseries/vphn.c
> > > @@ -39,7 +39,22 @@ static int vphn_unpack_associativity(const long
> > > *packed, __be32 *unpacked)
> > > be_packed[i] = cpu_to_be64(packed[i]);
> > > for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
> > > +/*
> > > + * When this function is called from hcall_vphn(), GCC 15 seems to
> > > consider
> > > + * 'retbuf' always populated from the hypervisor which is eventually
> > > referred by
> > > + * 'be_packed'. However, GCC 15's dataflow analysis can’t prove the same
> > > before
> > > + * the first dereference when this function is called from test_one()
> > > with
> > > + * pre-initialized array of 'struct test'. This results in a false
> > > positive
> > > + * '-Wmaybe-uninitialized' warning which is promoted to an error under
> > > + * '-Werror'. This problem is not seen when the compilation is performed
> > > with
> > > + * older GCC versions.
> > > + */
> > > +#pragma GCC diagnostic push
> > > +#if defined(__GNUC__) && __GNUC__ >= 15
> > > +#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
> > > +#endif
> > > u16 new = be16_to_cpup(field++);
> > > +#pragma GCC diagnostic pop
> > > if (is_32bit) {
> > > /*
> > >
> > > base-commit: 1f318b96cc84d7c2ab792fcc0bfd42a7ca890681
> >
>
