[clang] [llvm] [BPF] Make -mcpu=v3 as the default (PR #107008)

[via llvm-commits]

yonghong-song wrote:

@yuvald-sweet-security I did one example with your above example,
```
sudo ~/veristat /tmp/test-v1.o --filter=trace_ret_vfs_writev_tail
sudo ~/veristat /tmp/test-v3.o --filter=trace_ret_vfs_writev_tail
```

I can reproduce the issue. The v3 failure reason is actually due to your have inline asm code which prevents compiler from doing a good job for v3. I did the following change to the source
```
diff --git a/pkg/ebpf/c/common/filesystem.h b/pkg/ebpf/c/common/filesystem.h
index f1f042e35..27dffb431 100644
--- a/pkg/ebpf/c/common/filesystem.h
+++ b/pkg/ebpf/c/common/filesystem.h
@@ -487,6 +487,8 @@ statfunc void fill_vfs_file_bin_args(u32 type,
 statfunc void fill_file_header(u8 header[FILE_MAGIC_HDR_SIZE], io_data_t io_data)
 {
     u32 len = (u32) io_data.len;
+    // The following is needed only for -mcpu=v1.
+    asm volatile("" : "+r"(len));
     if (io_data.is_buf) {
         // inline bounds check to force compiler to use the register of len
         asm volatile("if %[size] < %[max_size] goto +1;\n"
@@ -499,10 +501,15 @@ statfunc void fill_file_header(u8 header[FILE_MAGIC_HDR_SIZE], io_data_t io_data
         __builtin_memset(&io_vec, 0, sizeof(io_vec));
         bpf_probe_read(&io_vec, sizeof(struct iovec), io_data.ptr);
         // inline bounds check to force compiler to use the register of len
+#if 0
         asm volatile("if %[size] < %[max_size] goto +1;\n"
                      "%[size] = %[max_size];\n"
                      :
                      : [size] "r"(len), [max_size] "i"(FILE_MAGIC_HDR_SIZE));
+#else
+        if (len >= FILE_MAGIC_HDR_SIZE)
+            len = FILE_MAGIC_HDR_SIZE;
+#endif
         bpf_probe_read(header, len, io_vec.iov_base);
     }
 }
```
The barrier_var(len) (i.e., asm volatile("" : "+r"(len))) is only needed for cpu=v1 but also works for cpu=v3.

The original asm code prevents a good compiler transformation at cpu=v3.

I suggest you try to replace asm code with proper C code and if necessary barrier_var.
That should resolve your issue in most cases and it is also portable between different cpu versions.

https://github.com/llvm/llvm-project/pull/107008