[llvm] [OpenMP][Offload] Fix flang offload test (PR #208829)

[Matt Arsenault via llvm-commits]

arsenm wrote:


> seems like the driver/toolchain didn't infer the target correctly.
> $BUILD/bin/flang -c -fopenmp --offload-arch=gfx90a \
>   $SRC/Inputs/declare-target-common-block-sub.f90 \
>   -o sub.bad.o

At this stage the driver should still be using the old amdgcn-amd-amdhsa

> 
> $BUILD/bin/flang -fopenmp \
>   -I $SRC \
>   -I $BUILD/runtimes/runtimes-bins/openmp/runtime/src \
>   -L $BUILD/lib/x86_64-unknown-linux-gnu \
>   -L $BUILD/lib \
>   -L $BUILD/runtimes/runtimes-bins/openmp/runtime/src \
>   -Wl,-rpath,$BUILD/lib/x86_64-unknown-linux-gnu \
>   -Wl,-rpath,$BUILD/runtimes/runtimes-bins/openmp/runtime/src \
>   -Wl,-rpath,$BUILD/lib/x86_64-unknown-linux-gnu \
>   -Wl,-rpath,$BUILD/lib \
>   -Xoffload-linker -L$BUILD/lib/amdgpu-amd-amdhsa \
>   -fopenmp-targets=amdgpu-amd-amdhsa \
>   $SRC/offloading/fortran/declare-target-common-block-main.f90 \
>   -o fail.out \
>   -Xoffload-linker -lompdevice \
>   sub.bad.o

The problem here is mixing and matching the old and the new name. This should work in principle, but the current compatibility expectation is a migration will be complete, so this is a test artifact. I have a patch about ready to make this work. 



> 
> $BUILD/bin/llvm-objcopy --dump-section .llvm.offloading=fail.offload fail.out
> $BUILD/bin/llvm-offload-binary fail.offload \
>   --image=file=fail.device.o,arch=gfx90a,triple=amdgpu-amd-amdhsa
> 
> 
> $BUILD/bin/llvm-readelf -s fail.device.o | grep foo || true
> 
> => no symbols about foo.

Missing symbol is a weird failure mode 

https://github.com/llvm/llvm-project/pull/208829