AMDGPU/TeraScale: Mesa's OpenCL implementation Clover crashes on `llvm::TargetInstrInfo::ReplaceTailWithBranchTo()`

[Mek101]

Originally posted here: https://github.com/void-linux/void-packages/issues/43921

Running some models causes Mesa's OpenCL implementation for TeraScale (Clover) to call llvm::TargetInstrInfo::ReplaceTailWithBranchTo() from /usr/lib/libLLVM-15.so, which itself is provided by libllvm version 15. This call causes a segmentation fault (verified via gdb).

To reproduce the bug:

• Use a TeraScale 3 GPU.
• Ensure Mesa uses Clover as an OpenCL implementation with llvm 15.
• Download an affected TensorFlow Lite model (squeezenet for example).
• Compile TensorFlow Lite's benchmark with GPU support.
• Run it with the --use_gpu=true --gpu_backend="cl" options.
• Observe the segmentation fault

In my particular case I'm using an old Radeon 8670D, a Terascale 3 iGPU with the radeon kernel driver. Since segmentation faults are typically caused by null-pointer dereferencing in CPU code I tried to solve the problem in the linked issue by patching llvm with a backport of https://github.com/llvm/llvm-project/commit/17069608940d22cd6266afb948443b11793f0a57, but the segmentation fault still occurred.

[Mek101]

For some reason GDB doesn't suspend execution even if I put breakpoints, however the segfault, according to gdb, seems to origin from this line.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

[llvmbot]

@llvm/issue-subscribers-opencl

[Mek101]

I tried building llvm in debug mode (-Og gcc's option), but I don't have enough ram/swap and my system locks up before finishing