Issue in macOS with the OpenCL atom_add function when array is of long type

[stratika]

Describe the bug

In macOS, I noticed that the TestAtomics unit-tests have one test failing. The error is related to the cl2Metal conversion which implies that the Metal clang implementation probably does not support the long type for atom_add.

This issue seems to be OS-specific. Hence, I am opening it in order to document it and provide a patch as workaround.

How To Reproduce

Build TornadoVM with OpenCL backend on macOS:

./bin/tornadovm-installer --jdk jdk21 --backend opencl

Then run:

tornado-test -V uk.ac.manchester.tornado.unittests.atomics.TestAtomics#testAtomic18_parallel_api_LongArray

WARNING: Using incubator modules: jdk.incubator.vector
UNSUPPORTED (log once): buildComputeProgram: cl2Metal failed

[ERROR] TornadoVM JIT Compiler - OpenCL Build Error Log:

program_source:21:13: error: call to 'atom_add' is ambiguous
    l_8  =  atom_add((volatile __global long *) ul_7, 1L);
            ^~~~~~~~
/System/Library/PrivateFrameworks/GPUCompiler.framework/Versions/32023/Libraries/lib/clang/32023.406/include/opencl-c.h:13111:12: note: candidate function
int __ovld atom_add(volatile __global int *p, int val);
           ^
/System/Library/PrivateFrameworks/GPUCompiler.framework/Versions/32023/Libraries/lib/clang/32023.406/include/opencl-c.h:13112:13: note: candidate function
uint __ovld atom_add(volatile __global uint *p, uint val);

Expected behavior

It seems that the Metal-based OpenCL implementation of the atom_add may not support long types.
If I change at the code gen the casted type to be int instead of long, it works:

l_8  =  atom_add((volatile __global int *) ul_7, 1L);

See the patch: ISSUE_637.patch

Computing system setup (please complete the following information):

• OS: macOS 15.3.2
• OpenCL and Driver versions: n/a
• TornadoVM commit id: 4d8be35
• TornadoVM version: 1.0.11-dev
• branch: develop

---