selftests/bpf: Keep the loop in bpf_testmod_loop_test

Some compilers (for example clang-15) optimize bpf_testmod_loop_test and
remove the loop:

gcc version
(gdb) disassemble bpf_testmod_loop_test
Dump of assembler code for function bpf_testmod_loop_test:
   0x0000000000000570 <+0>:     callq  0x575 <bpf_testmod_loop_test+5>
   0x0000000000000575 <+5>:     xor    %eax,%eax
   0x0000000000000577 <+7>:     test   %edi,%edi
   0x0000000000000579 <+9>:     jle    0x587 <bpf_testmod_loop_test+23>
   0x000000000000057b <+11>:    xor    %edx,%edx
   0x000000000000057d <+13>:    add    %edx,%eax
   0x000000000000057f <+15>:    add    $0x1,%edx
   0x0000000000000582 <+18>:    cmp    %edx,%edi
   0x0000000000000584 <+20>:    jne    0x57d <bpf_testmod_loop_test+13>
   0x0000000000000586 <+22>:    retq
   0x0000000000000587 <+23>:    retq

clang-15 version
(gdb) disassemble bpf_testmod_loop_test
Dump of assembler code for function bpf_testmod_loop_test:
   0x0000000000000450 <+0>:     nopl   0x0(%rax,%rax,1)
   0x0000000000000455 <+5>:     test   %edi,%edi
   0x0000000000000457 <+7>:     jle    0x46b <bpf_testmod_loop_test+27>
   0x0000000000000459 <+9>:     lea    -0x1(%rdi),%eax
   0x000000000000045c <+12>:    lea    -0x2(%rdi),%ecx
   0x000000000000045f <+15>:    imul   %rax,%rcx
   0x0000000000000463 <+19>:    shr    %rcx
   0x0000000000000466 <+22>:    lea    -0x1(%rdi,%rcx,1),%eax
   0x000000000000046a <+26>:    retq
   0x000000000000046b <+27>:    xor    %eax,%eax
   0x000000000000046d <+29>:    retq

Note: The jne instruction is removed in clang-15 version.

Force the compile to keep the loop by making sum volatile.
Signed-off-by: Song Liu <song@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20230412210423.900851-4-song@kernel.org

tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c

@@ -130,7 +130,11 @@ bpf_testmod_test_btf_type_tag_percpu_2(struct bpf_testmod_btf_type_tag_3 *arg) {
 
 noinline int bpf_testmod_loop_test(int n)
 {
-	int i, sum = 0;
+	/* Make sum volatile, so smart compilers, such as clang, will not
+	 * optimize the code by removing the loop.
+	 */
+	volatile int sum = 0;
+	int i;
 
 	/* the primary goal of this test is to test LBR. Create a lot of
 	 * branches in the function, so we can catch it easily.