SWDEV-102314 "[OCL-LC-ROCm] AMD lower kernel calls" #35
Conversation
dfukalov
changed the title
| F.empty() ? copyDeclaration(F) : CloneFunction(&F, ignored, false); | ||
| F.getParent()->getOrInsertFunction(NewF->getName(), NewF->getFunctionType(), | ||
| NewF->getAttributes()); | ||
| NewF->setCallingConv(CallingConv::C); |
There was a problem hiding this comment.
Why not just getOrInsertFunction with the new name from setNameForBody?
|
LGTM. |
|
Are you sure this is necessary? I thought the compiler already did this. |
|
There is code in AMDGPUAlwaysInlinePass, which is meant to do this. If that code isn't working, I would prefer to fix that rather than add a new pass. |
|
There is and ability in OpenCL to call kernel function from other kernel or function. The pass finds such calls, create clones for callees (with mangled name and removed amdgpu_kernel calling convention) and patches calls to such functions. |
|
Ok, I forgot we would need this for when function calls were supported. Did you test this with the OCL conformance tests? |
Ported lower kernels calls pre-link pass from hsail compiler