[HIP] -fmath-errno causes undefined symbol expf

[yxsamliu]

https://godbolt.org/z/nTfccd1Tr

after 5f1d383

currently HIP expf function is implemented as __builtin_expf. When -fmath-errno is specified, __builtin_expf is emitted as a call of external lib function expf by clang codegen. Since ROCm device libs does not implement expf. This results in undefined symbol.

Before 5f1d383, HIP expf function is implemented by __ocml_exp_f32, therefore did not have such issue.

I kind of think it is not good to go back to __ocml_exp_f32 since it does not handle errno any way. I think probably we should just do not pass -fmath-errno to clang -cc1 for amdgpu since we do not support it.

@arsenm @b-sumner

[b-sumner]

Personally I don't agree with calling __builtin_expf here. But I do agree that passing -fmath-errno for device compilation should be avoided.

[yxsamliu]

one concern is that FE defines macro __NO_MATH_ERRNO__ when math-errno is disabled. If users' code has struct layout depending on that macro, ODR may be violated if host and device have different math-errno settings. To avoid that, we could still let -fmath-errno being passed to clang -cc1 but do not emit lib function call for clang builtins.

[arsenm]

First we should actually define the platform and library functions, which includes providing math functions with the standard names.

But failing that, we should simply not support -fmath-errno. It is not required in C99+ if you claim to support FP exceptions

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Yaxun (Sam) Liu (yxsamliu)

https://godbolt.org/z/nTfccd1Tr

after 5f1d383

currently HIP expf function is implemented as __builtin_expf. When -fmath-errno is specified, __builtin_expf is emitted as a call of external lib function expf by clang codegen. Since ROCm device libs does not implement expf. This results in undefined symbol.

Before 5f1d383, HIP expf function is implemented by __ocml_exp_f32, therefore did not have such issue.

I kind of think it is not good to go back to __ocml_exp_f32 since it does not handle errno any way. I think probably we should just do not pass -fmath-errno to clang -cc1 for amdgpu since we do not support it.

@arsenm @b-sumner