On Sun, Aug 25, 2024 at 01:23:51PM -0700, Yonghong Song wrote:
>
> On 8/24/24 4:21 PM, Tony Ambardar wrote:
[snip]
> >
> > Test '#525 verif_scale_pyperf600:FAIL' was caused by clang miscompilation
> > exposed by my use of clang-19 and clang-20. The test passes when built
> > with clang-17 (used by BPF CI) or clang-18 which I switched to use.
>
> x86 has the same issue where clang19 generated code will cause verification
> failure. Eduard is working on this.
>
> >
> > One symptom of the problem is easily seen by manually compiling:
> >
> > $ clang-18 -g -Wall -Werror -D__TARGET_ARCH_s390 -mbig-endian -Itools/testing/selftests/bpf/tools/include -Itools/testing/selftests/bpf -Itools/include/uapi -Itools/testing/selftests/usr/include -Wno-compare-distinct-pointer-types -idirafter /usr/lib/llvm-18/lib/clang/18/include -idirafter /usr/local/include -idirafter /usr/lib/gcc-cross/s390x-linux-gnu/11/../../../../s390x-linux-gnu/include -idirafter /usr/include/s390x-linux-gnu -idirafter /usr/include -DENABLE_ATOMICS_TESTS -O2 --target=bpfeb -c tools/testing/selftests/bpf/progs/pyperf600.c -mcpu=v3 -o pyperf600.clang18.bpf.o
> >
> > $ clang-19 -g -Wall -Werror -D__TARGET_ARCH_s390 -mbig-endian -Itools/testing/selftests/bpf/tools/include -Itools/testing/selftests/bpf -Itools/include/uapi -Itools/testing/selftests/usr/include -Wno-compare-distinct-pointer-types -idirafter /usr/lib/llvm-19/lib/clang/19/include -idirafter /usr/local/include -idirafter /usr/lib/gcc-cross/s390x-linux-gnu/11/../../../../s390x-linux-gnu/include -idirafter /usr/include/s390x-linux-gnu -idirafter /usr/include -DENABLE_ATOMICS_TESTS -O2 --target=bpfeb -c tools/testing/selftests/bpf/progs/pyperf600.c -mcpu=v3 -o pyperf600.clang19.bpf.o
> >
> > $ llvm-readelf-18 -S pyperf600.clang{18,19}.bpf.o |grep .symtab
> > [27] .symtab SYMTAB 0000000000000000 1739d0 01ad60 18 1 4572 8
> > [27] .symtab SYMTAB 0000000000000000 14f048 0001e0 18 1 12 8
> >
> > Notice that the .symtab has shrunk by ~200X for example going to clang-19!
> > (CCing llvm maintainers)
>
> This is a known issue. In llvm18, all labels (to identify basic blocks) are in symbol table.
> Those labels are removed from symbol table in llvm19.
Glad to hear this a known issue being looked at now. A quick search on my part found nothing, so sorry for the noise and thanks for clarifying.
>
> >
> >
> > Kind regards,
> > Tony
> >
