[flang][openmp] target construct segfaults with amdgpu

[VeeEM]

Programs with a target construct without any additional directives segfault when compiled with flang for amdgpu. clang works fine, only flang is affected.

A program with just a target region is enough to reproduce:

program crash
!$omp target
!$omp end target
end program

$ flang -fopenmp --offload-arch=gfx1100 crash.f90 
flang-21: warning: OpenMP support in flang is still experimental [-Wexperimental-option]
flang-21: warning: OpenMP support in flang is still experimental [-Wexperimental-option]
flang-21: warning: OpenMP support in flang is still experimental [-Wexperimental-option]
$ ./a.out 
Segmentation fault

Stacktrace from gdb shows:

#0  0x00007ffff3d4fef5 in llvm::omp::target::plugin::AMDGPUKernelTy::launchImpl(llvm::omp::target::plugin::GenericDeviceTy&, unsigned int*, unsigned int*, KernelArgsTy&, KernelLaunchParamsTy, llvm::omp::target::plugin::AsyncInfoWrapperTy&) const
    () from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#1  0x00007ffff3d143a5 in llvm::omp::target::plugin::GenericKernelTy::launch(llvm::omp::target::plugin::GenericDeviceTy&, void**, long*, KernelArgsTy&, llvm::omp::target::plugin::AsyncInfoWrapperTy&) const ()
   from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#2  0x00007ffff3d1ae04 in llvm::omp::target::plugin::GenericDeviceTy::launchKernel(void*, void**, long*, KernelArgsTy&, __tgt_async_info*) () from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#3  0x00007ffff3d1e092 in llvm::omp::target::plugin::GenericPluginTy::launch_kernel(int, void*, void**, long*, KernelArgsTy*, __tgt_async_info*) () from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#4  0x00007ffff3becd17 in DeviceTy::launchKernel(void*, void**, long*, KernelArgsTy&, AsyncInfoTy&) () from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#5  0x00007ffff3bf847c in target(ident_t*, DeviceTy&, void*, KernelArgsTy&, AsyncInfoTy&) () from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#6  0x00007ffff3beff59 in __tgt_target_kernel () from /home/acc/install_llvm2/lib/x86_64-unknown-linux-gnu/libomptarget.so.21.0git
#7  0x000055555555675c in _QQmain ()
#8  0x0000555555556782 in main ()

Tested at f7f5aa2.

Machine specs: Ubuntu 24.04.1, rocm 6.3.0, Ryzen 9 9900X, RX 7900 GRE.

[VeeEM]

Looks like it's just plain target that crashes. target teams and distributed loops still work.

edit: Or actually, target teams can also crash, but at least every program does not crash always.

[VeeEM]

This pr might be the source: #131033

It's crashing here with ImplArgs = 0

llvm-project/offload/plugins-nextgen/amdgpu/src/rtl.cpp

Line 3389 in 96925fa

ImplArgs->BlockCountX = NumBlocks[0];

[jhuber6]

This is because apparently it's failing the following check. I suppose we should guard access to the implicit arguments, since this could show up if someone is doing something truly custom. However, for the flang case, it seems suspicious to me that these don't match up. The implicit arguments are basically mandatory now ever since COV5 moved the grid size lookup into there.

if (ArgsSize == LaunchParams.Size + getImplicitArgsSize())

I could make a simple patch to guard this for now, but I would recommend the flang team looks into why this doesn't match the clang behavior.