Optimization barrier not working on Clang-CL
Summary
p_sin_cos function when using AVX returns bad/random results.
Only the AVX optimized part returns bad results (if the input is 18 float in size, the first 16 outputs are bad, but the last 2, the non-optim ones, are good).
The probable cause is EIGEN_OPTIMIZATION_BARRIER.
Environment
- Operating System : Windows
- Architecture : x64
- Eigen Version : main
- Compiler Version : Clang-CL
- Compile Flags :
- Vector Extension : AVX
Minimal Example
I have not been able to reproduce it with only the Eigen library, and it is only reproducible a custom verion of clan-cl/Llvm.
The example I have is using a Onnx model with Onnx-Runtime that has a Cosine operator. The model return random output dur to the Cosine operator calling cos function from Eigen.
Steps to reproduce
- first step
- second step
- ...
What is the current bug behavior?
What is the expected correct behavior?
Relevant logs
Anything else that might help
The macro EIGEN_OPTIMIZATION_BARRIER is not defined when using Clang-cl compiler. Even if this compiler's backend is similar to GNUC, it does not define the preprocessor macro __GNUC__, unlike other clang compiler. But it does define the preproc macro __clang__.
Before this commit (82d61af3), it used to work under clan-cl.
I suggest replacing the line here (https://gitlab.com/libeigen/eigen/-/blob/master/Eigen/src/Core/util/Macros.h?ref_type=heads#L996), From
#if EIGEN_COMP_GNUCTo
#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG-
Have a plan to fix this issue.