Fix manageablebarrier hang

Move the markID after the barrier and fence on workgroup scope in getFreeID() to ensure that all of the threads will read correct value from IDPool,
before the threadID:0 will update it in function emitInit.

Author: lwesiers

IGC/Compiler/Optimizer/OpenCLPasses/ManageableBarriers/ManageableBarriersResolution.cpp

@@ -201,13 +201,12 @@ Value *ManageableBarriersResolution::prepareBarrierIDPoolPtr(Instruction *pInser
 // Binary form of IDPool 00000000 00000000 00000000 00010000
 // first free ID is 5
 
-void ManageableBarriersResolution::markID(Value *IDPool, Value *IDBarrier, Instruction *pInsertBefore) {
+void ManageableBarriersResolution::markID(Value *IDPool, Value *currentIDPoolState, Value *IDBarrier, Instruction *pInsertBefore) {
   IGCIRBuilder<> builder(pInsertBefore);
 
   Value *maskID = builder.CreateShl(builder.getInt32(1), IDBarrier);
   Value *notBit_maskID = builder.CreateXor(maskID, builder.getInt32(-1));
 
-  Value *currentIDPoolState = builder.CreateLoad(builder.getInt32Ty(), IDPool);
   Value *currentIDPoolStateUpdated = builder.CreateAnd(notBit_maskID, currentIDPoolState);
 
   builder.CreateStore(currentIDPoolStateUpdated, IDPool);
@@ -220,9 +219,8 @@ void ManageableBarriersResolution::releaseID(Value *IDPool, Value *IDBarrier, In
   ResolveOCLAtomics::CallAtomicSingleLane(AtomicOp::EATOMIC_OR, IDPool, maskID, pInsertBefore);
 }
 
-Value *ManageableBarriersResolution::getFreeID(Value *IDPool, Instruction *pInsertBefore) {
+Value *ManageableBarriersResolution::getFreeID(Value *currentIDPoolState, Instruction *pInsertBefore) {
   IGCIRBuilder<> builder(pInsertBefore);
-  Value *currentIDPoolState = builder.CreateLoad(builder.getInt32Ty(), IDPool);
 
   Function *func_llvm_GenISA_firstbitLo = GenISAIntrinsic::getDeclaration(mModule, GenISAIntrinsic::GenISA_firstbitLo);
   Value *freeIDNumber = builder.CreateCall(func_llvm_GenISA_firstbitLo, {currentIDPoolState});
@@ -270,14 +268,13 @@ void ManageableBarriersResolution::emitInit(CallInst *pInsertPoint) {
     // Get current free ID for named barrier
     Value *barrierIDPoolPtr = getBarrierIDPoolPtr(pInsertPoint);
 
-    Value *getFirstFreeID = getFreeID(barrierIDPoolPtr, chekForSingleLane);
+    LoadInst *currentIDPoolState = new LoadInst(builder.getInt32Ty(), barrierIDPoolPtr, "", chekForSingleLane);
+    Value *getFirstFreeID = getFreeID(currentIDPoolState, chekForSingleLane);
     Value *ptrToBarrierSlot = getManageableBarrierstructDataPtr(pInsertPoint, getFirstFreeID, chekForSingleLane);
 
     // Fill the basic block section for the Init function of barriers
     Instruction *instrJump = BranchInst::Create(bbAfter, bbInitSection);
 
-    markID(barrierIDPoolPtr, getFirstFreeID, instrJump);
-
     storeManageableBarrierstructData(ptrToBarrierSlot, MBDynamicStructFields::BarrierID, getFirstFreeID, instrJump);
 
     storeManageableBarrierstructData(ptrToBarrierSlot, MBDynamicStructFields::ProducerCount,
@@ -297,6 +294,10 @@ void ManageableBarriersResolution::emitInit(CallInst *pInsertPoint) {
     GenIntrinsicInst::Create(GenISAIntrinsic::getDeclaration(mModule, GenISAIntrinsic::GenISA_threadgroupbarrier), {},
                              "", pInsertPoint);
 
+    // Move the markID after the barrier and fence on workgroup scope to ensure that all of the threads will
+    // read correct value from IDPool, before the threadID:0 will update it.
+    markID(barrierIDPoolPtr, currentIDPoolState, getFirstFreeID, pInsertPoint);
+
     pInsertPoint->replaceAllUsesWith(ptrToBarrierSlot);
   }
 }

IGC/Compiler/Optimizer/OpenCLPasses/ManageableBarriers/ManageableBarriersResolution.hpp

@@ -99,9 +99,9 @@ class ManageableBarriersResolution : public llvm::ModulePass,
   llvm::Value *prepareBarrierIDPoolPtr(llvm::Instruction *pInsertBefore);
   llvm::Value *getBarriersDataPoolPtr(llvm::Instruction *pCallInst);
   llvm::Value *getBarrierIDPoolPtr(llvm::Instruction *pCallInst);
-  void markID(llvm::Value *IDPool, llvm::Value *IDBarrier, llvm::Instruction *pInsertBefore);
+  void markID(llvm::Value *IDPool, llvm::Value *currentIDPoolState, llvm::Value *IDBarrier, llvm::Instruction *pInsertBefore);
   void releaseID(llvm::Value *IDPool, llvm::Value *IDBarrier, llvm::Instruction *pInsertBefore);
-  llvm::Value *getFreeID(llvm::Value *IDPool, llvm::Instruction *pInsertBefore);
+  llvm::Value *getFreeID(llvm::Value *currentIDPoolState, llvm::Instruction *pInsertBefore);
 
   void clearData();
 

IGC/ocloc_tests/manageablebarriers/manageablebarriers_simple_case.cl

@@ -29,13 +29,6 @@
 // CHECK: [[MB_DATA_OFFSET_IntPTR:%[0-9]+]] = add i32 [[BarrierIDPool_FirstFreeID_Offset]], [[MB_DATA_GetIntPTR]]
 // CHECK: [[MB_DATA_OFFSET_PTR:%[0-9]+]] = inttoptr i32 [[MB_DATA_OFFSET_IntPTR]] to ptr addrspace(3)
 
-//// Mark the FreeID in the ManageableBarriers ID Pool (that this ID is busy)
-// CHECK: [[FreeID_InBit:%[0-9]+]] = shl i32 1, [[BarrierIDPool_FirstFreeID]]
-// CHECK: [[FreeID_InBitNeg:%[0-9]+]] = xor i32 [[FreeID_InBit]], -1
-// CHECK: [[BarrierIDPool_Load2:%[0-9]+]] = load i32, ptr addrspace(3) [[BarrierIDPool_Ptr]], align 4
-// CHECK: [[BarrierIDPool_UpdateBits:%[0-9]+]] = and i32 [[FreeID_InBitNeg]], [[BarrierIDPool_Load2]]
-// CHECK: store i32 [[BarrierIDPool_UpdateBits]], ptr addrspace(3) [[BarrierIDPool_Ptr]], align 4
-
 //// Fill with data for the ManageBarrier struct in SLM (for the particular single ManageableBarrier)
 //// Fill the FreeID in the ManageableBarriers Data ID offset
 // CHECK: [[MB_DATA_OFFSET_ID_GetIntPTR:%[0-9]+]] = ptrtoint ptr addrspace(3) [[MB_DATA_OFFSET_PTR]] to i32
@@ -65,6 +58,12 @@
 // CHECK: call void @llvm.genx.GenISA.memoryfence(i1 true, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false, i32 0)
 // CHECK: call void @llvm.genx.GenISA.threadgroupbarrier()
 
+//// Mark the FreeID in the ManageableBarriers ID Pool (that this ID is busy)
+// CHECK: [[FreeID_InBit:%[0-9]+]] = shl i32 1, [[BarrierIDPool_FirstFreeID]]
+// CHECK: [[FreeID_InBitNeg:%[0-9]+]] = xor i32 [[FreeID_InBit]], -1
+// CHECK: [[BarrierIDPool_UpdateBits:%[0-9]+]] = and i32 [[FreeID_InBitNeg]], [[BarrierIDPool_Load]]
+// CHECK: store i32 [[BarrierIDPool_UpdateBits]], ptr addrspace(3) [[BarrierIDPool_Ptr]], align 4
+
 /////////////////////////////////////////////
 //// ManageableBarriersArriveINTEL resolution
 /// Get the data for this barrier