On 3/11/26 11:03 AM, Hari Bathini wrote:
On 11/03/26 9:32 pm, Alexei Starovoitov wrote:
On Wed, Mar 11, 2026 at 8:10 AM Hari Bathini <[email protected]>
wrote:
+
+ /* Check zero-extension */
+ if (val != (unsigned long)a)
+ return 1;
+ /* Check no sign-extension */
+ if (val < 0)
+ return 2;
+
+ val = b;
+ if (val != (unsigned long)b)
+ return 3;
+ if (val < 0)
+ return 4;
+
+ val = c;
+ if (val != (unsigned long)c)
+ return 5;
+ if (val < 0)
+ return 6;
+
+ return 0;
+}
Overall this looks very useful.
I would expand with another test where a,b,c are s8,s16,s32.
Slightly different approach but kfunc_call_test4/bpf_kfunc_call_test4
cover signed arguments already?
Ahh. Then may be tweak it to adopt similar fine grained
error reporting as your bpf_kfunc_call_test5()
I Prefer this.
Does the below change to bpf_kfunc_call_test4 look fine:
diff --git a/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
b/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
index 48dcaf93bb9f..6237c2222633 100644
--- a/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
+++ b/tools/testing/selftests/bpf/test_kmods/bpf_testmod.c
@@ -760,8 +760,30 @@ __bpf_kfunc struct sock
*bpf_kfunc_call_test3(struct sock *sk)
__bpf_kfunc long noinline bpf_kfunc_call_test4(signed char a, short
b, int c, long d)
I might regret for bringing this up as it could be yet another ABI
fiasco between gcc and llvm.
As per C standard, sign of unadorned char (i.e. w/o explicit signed or
unsigned prefix) is ABI defined.
For gcc-bpf char is specified to be signed.
So test4 has s8, while new test5 has u8. Would it make sense to have an
additional test without signed/unsigned annotation for char ?
This will flag any discrepancy between the two compilers.
Thx,
-Vineet
{
+ /*
+ * Make val as volatile to avoid compiler optimizations.
+ * Verify that negative signed values remain negative after
+ * sign-extension (JIT must sign-extend, not zero-extend).
+ */
+ volatile long val;
+
+ /* val will be positive, if JIT does zero-extension instead of
sign-extension */
+ val = a;
+ if (val >= 0)
+ return 1;
+
+ val = b;
+ if (val >= 0)
+ return 2;
+
+ val = c;
+ if (val >= 0)
+ return 3;
+
/* Provoke the compiler to assume that the caller has
sign-extended a,
* b and c on platforms where this is required (e.g. s390x).
+ *
+ * Original behavior: return sum for backward compatibility
*/
return (long)a + (long)b + (long)c + d;
}
- Hari
