[libclc] libclc: Use separate acquire and release fences in work_group_barrier (PR #185190)

[Wenju He via cfe-commits]

================
@@ -10,8 +10,22 @@
 #include <clc/synchronization/clc_work_group_barrier.h>
 
 _CLC_OVERLOAD _CLC_DEF void
-__clc_work_group_barrier(int memory_scope, int memory_order,
+__clc_work_group_barrier(int memory_scope,
----------------
wenju-he wrote:

shall we keep the memory_order parameter? This may not be needed for OpenCL. But it is probably allowed to enforce a specific memory order in SPIR-V ControlBarrier when Execution is Workgroup: https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpControlBarrier

https://github.com/llvm/llvm-project/pull/185190