Re: [PATCH bpf-next v3 0/5] bpf: tools: support build selftests/bpf with clang

[Alexei Starovoitov <ast@fb.com>]

On 4/13/21 8:34 AM, Yonghong Song wrote:
> To build kernel with clang, people typically use
>    make -j60 LLVM=1 LLVM_IAS=1
> LLVM_IAS=1 is not required for non-LTO build but
> is required for LTO build. In my environment,
> I am always having LLVM_IAS=1 regardless of
> whether LTO is enabled or not.
> 
> After kernel is build with clang, the following command
> can be used to build selftests with clang:
>    make -j60 -C tools/testing/selftests/bpf LLVM=1 LLVM_IAS=1
> 
> I am using latest bpf-next kernel code base and
> latest clang built from source from
>    https://github.com/llvm/llvm-project.git
> Using earlier version of llvm may have compilation errors, see
>    tools/testing/selftests/bpf
> due to continuous development in llvm bpf features and selftests
> to use these features.
> 
> To run bpf selftest properly, you need have certain necessary
> kernel configs like at:
>    bpf-next:tools/testing/selftests/bpf/config
> (not that this is not a complete .config file and some other configs
>   might still be needed.)
> 
> Currently, using the above command, some compilations
> still use gcc and there are also compilation errors and warnings.
> This patch set intends to fix these issues.
> Patch #1 and #2 fixed the issue so clang/clang++ is
> used instead of gcc/g++. Patch #3 fixed a compilation
> failure. Patch #4 and #5 fixed various compiler warnings.
> 
> Changelog:
>    v2 -> v3:
>      . more test environment description in cover letter. (Sedat)
>      . use a different fix, but similar to other use in selftests/bpf
>        Makefile, to exclude header files from CXX compilation command
>        line. (Andrii)
>      . fix codes instead of adding -Wno-format-security. (Andrii)

I struggled to tweak my llvm setup, but at the end it compiled and
selftests/bpf/test_progs passed compiled by clang,
so I've applied to bpf-next.

The things I've seen:
1.
include <iostream> not found due to my setup quirks.
2.
diff selftests/bpf/tools/build/libbpf/libbpf_global_syms.tmp	
diff selftests/bpf/tools/build/libbpf/libbpf_versioned_syms.tmp	
  btf__set_pointer_size
  btf__str_by_offset
  btf__type_by_id
+LIBBPF_0.0.1
+LIBBPF_0.0.2
and this was happening with packaged llvm builds,
but my own llvm build was fine, so I didn't debug further.

3.
clang-12: error: unsupported option '-mrecord-mcount' for target 
'x86_64-unknown-linux-gnu'
due to kernel not built with clang.

I suspect followups will be needed to make it bulletproof.