Support pointer argument of built-in functions

[mahmood82]

No description provided.

[bcardosolopes]

You also need to add a testcase

[bcardosolopes]

Still needs a testcase

[mahmood82]

Still needs a testcase

Hi, It's reproducible on my local project, I need to make a unit test for incubator. I'll provide it ASAP.

[mahmood82]

Unfortunately I can't add a proper test for this change. The behavior depends on Clang's handling of OpenCL built-ins.
Reproducing the scenario would require adding a private. The issue only manifests with my local specific setup, so a standard test can't capture it.

[bcardosolopes]

Because this doesn't exist in OG, let's just add a few asserts to clean this up:

auto ptrType = mlir::dyn_cast<cir::PointerType>(argType);
assert(ptrType && "expected pointer type");
assert(ptrType.getPointee() != expectedPtrType.getPointee() && "types should not match");

[bcardosolopes]

The behavior depends on Clang's handling of OpenCL built-ins

This is something that should be added at some point right? Can you perhaps just add a commented testcase to one of the testcases so that when you get to it we can enable it?

[mahmood82]

The behavior depends on Clang's handling of OpenCL built-ins

This is something that should be added at some point right? Can you perhaps just add a commented testcase to one of the testcases so that when you get to it we can enable it?

What I had in mind is something like:

__kernel void test(int size) {
  __my_private_check(size == 1024, "size check!");
}

For this to be testable upstream, __my_private_check (or an equivalent) would need to be a user-visible OpenCL builtin. Today it's a private, internal helper in Clang, and there is no language-level construct or intrinsic that can express this behavior. Without introducing such a builtin on the Clang side, there's unfortunately nothing meaningful I can encode as a test, even as a commented one.