A patch was previously made to the LoopUnroll pass to allow loop

unrolling to occur even if it contained blocks with 'noduplicate'
instructions.  Given that the newer 'convergent' attribute models
the barrier restriction better, we change that here.

Change-Id: I0f075cebfa74130f50e232a1bdc799ff16cf9470

Author: scottp101

IGC/BiFModule/CMakeLists.txt

@@ -852,7 +852,6 @@ endfunction()
 ### -include-pch OCL_64 -cl-enable-half -D__EXECUTION_MODEL_DEBUG=1 -D__OPENCL_C_VERSION__=120 -Dcl_khr_fp16 -I                     .\IGCBiFModule\PointerSize\IBiF_size_t.cl            -o .\IGCsize_t_64.bc
 
 # Other:
-#     %LLVM_AS_EXE% IGCBiFModule\shared\IGCBiF_Impl_ll.ll -o .\IBiF_Impl_ll.bc
 #     %LLVM_LINK_EXE% .\IBiF_Impl_int.bc .\IGCBiF_Impl_ll.bc -o .\OCLBiFImpl.bc
 
 
@@ -916,7 +915,6 @@ igc_bif_build_bc(
     OUTPUT               "${IGC_BUILD__BIF_DIR}/IBiF_Impl_int_spirv.bc"
     TRIPLE               spir64
     SOURCES              "${IGC_OPTION__BIF_SRC_OCL_DIR}/Implementation/IBiF_Impl.cl"
-                         "${IGC_OPTION__BIF_SRC_OCL_DIR}/Implementation/IGCBiF_Impl_ll.ll"
     FORCE_INCLUDE        ${IGC_BUILD__BIF_OCL_PCH_SRC}
     INCLUDE_DIRECTORIES  ${IGC_BUILD__BIF_SPIRV_COMMON_INC_DIRS}
     DEFINES              "__EXECUTION_MODEL_DEBUG=1" "__OPENCL_C_VERSION__=200" "__IGC_BUILD__" ${KHR_DEFINES} ${FLAG}

IGC/BiFModule/Implementation/IGCBiF_Impl_ll.ll

@@ -115,6 +115,9 @@ local uchar* __builtin_IB_AllocLocalMemPool(uint, uint);
 void     __builtin_IB_memfence(bool commitEnable, bool flushRW, bool flushConstant, bool flushTexture, bool flushIcache, bool isGlobal, bool invalidateL1);
 void     __builtin_IB_flush_sampler_cache(void);
 
+// Barrier
+void     __builtin_IB_thread_group_barrier(void) __attribute__((convergent));
+
 // Workitem functions
 uint     __builtin_IB_get_work_dim(void) __attribute__((const));
 uint     __builtin_IB_get_group_id(uint) __attribute__((const));

IGC/BiFModule/Implementation/IGCBiF_Intrinsics.cl

@@ -26,10 +26,6 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 // Barrier Instructions
 
-// Just empty prototype - implementation handwritten in .ll
-
-INLINE void __intel_Barrier( void );
-
 static void __intel_atomic_work_item_fence( Scope_t Memory, uint Semantics )
 {
     bool fence = Semantics & ( Acquire | Release | AcquireRelease | SequentiallyConsistent );
@@ -66,7 +62,7 @@ void __builtin_spirv_OpControlBarrier_i32_i32_i32(Scope_t Execution, Scope_t Mem
 
     if( Execution <= Workgroup )
     {
-        __intel_Barrier();
+        __builtin_IB_thread_group_barrier();
     }
 }
 

IGC/BiFModule/Implementation/barrier.cl

@@ -211,8 +211,6 @@ static uint get_spirv_mem_order( memory_order order )
 //*****************************************************************************/
 // Synchronization functions
 //*****************************************************************************/
-// Just empty prototype - implementation handwritten in .ll
-INLINE void __intel_Barrier( void );
 
 INLINE void OVERLOADABLE work_group_barrier(cl_mem_fence_flags flags)
 {

IGC/BiFModule/Languages/OpenCL/IBiF_Impl.cl

@@ -102,10 +102,10 @@
     "GenISA_icmpxchgatomictyped": ["anyint",["anyptr","int","int","int",0,0],"ReadWriteArgMem"],
     "GenISA_atomiccounterinc": ["int",["anyptr"],"ReadWriteArgMem"],
     "GenISA_atomiccounterpredec": ["int",["anyptr"],"ReadWriteArgMem"],
-    "GenISA_threadgroupbarrier": ["void",[],"NoDuplicate"],
-    "GenISA_memoryfence": ["void",["bool","bool","bool","bool","bool","bool","bool"],"NoDuplicate"],
+    "GenISA_threadgroupbarrier": ["void",[],"Convergent"],
+    "GenISA_memoryfence": ["void",["bool","bool","bool","bool","bool","bool","bool"],"Convergent"],
     "GenISA_flushsampler": ["void",[],"None"],
-    "GenISA_globalSync": ["void",[],"None"],
+    "GenISA_globalSync": ["void",[],"Convergent"],
     "GenISA_uavSerializeOnResID": ["void",["int"],"None"],
     "GenISA_uavSerializeAll": ["void",[],"None"],
     "GenISA_WorkGroupAny": ["int",["int"],"None"],