Promotion of bfloat is broken for strictfp functions

[arsenm]

b473864

This disabled a test in llvm/test/CodeGen/AMDGPU/llvm.is.fpclass.bf16 because promotion of bf16 fp_round is broken in strictfp functions. We are currently missing strict variants of BF16_TO_FP and FP_TO_BF16

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

https://github.com//pull/74405/commits/b473864900c68c84e37d68f0ea771ef42f13138a

This disabled a test in llvm/test/CodeGen/AMDGPU/llvm.is.fpclass.bf16 because promotion of bf16 fp_round is broken in strictfp functions. We are currently missing strict variants of BF16_TO_FP and FP_TO_BF16

[shiltian]

I'm working on a patch to add the two variants STRICT_FP_TO_BF16 and STRICT_BF16_TO_FP. For AMDGPU backend, how are we gonna handle the instruction selection? We don't have bfloat related conversion instructions, do we?

[arsenm]

I'm working on a patch to add the two variants STRICT_FP_TO_BF16 and STRICT_BF16_TO_FP. For AMDGPU backend, how are we gonna handle the instruction selection? We don't have bfloat related conversion instructions, do we?

They should get the standard expansions to convert to float

[shiltian]

I'm not sure how to move forward with #80056 because I can't design a valid test case.

[arsenm]

I'm not sure how to move forward with #80056 because I can't design a valid test case.

We can continue to fail to legalize if we know exceptions are enabled

[shiltian]

This issue has been partially solved by b0c158b. It doesn't hit the assertion now, though we still can't enable those tests because we don't have a valid lowering for AMDGPU.