Open DataCorrupted opened 1 year ago
@llvm/issue-subscribers-backend-amdgpu
I'm surprised any cases select for this. f64 fpow is not implemented
I think the problem is... should it be implemented? If we should, I'll leave it here for future reference. If it shouldn't because the hardware specification says its not supported, shouldn't we at least lower it so the compiler doesn't crash?
I would like if we could directly handle all the math intrinsics. We currently rely on an IR library implementation here https://github.com/RadeonOpenCompute/ROCm-Device-Libs/blob/amd-stg-open/ocml/src/powD_base.h
To implement the intrinsic, we would need to write an expansion that produces equivalent code (which could be a lot), or somehow support libcall legalization
Ah, thanks for the clarification.
When compiling the following code with amdgcn (godbolt), the compiler fails. It seems as long as one of the argument is non constant, instruction selection will fail.
I am wondering if this is a feature in AMDGCN, or a missing pattern?
Here is the stack trace: