minor refactor

Change-Id: I0e15cae06331280e99622d95d10210e1e21c0b88

Author: Jay-Jiewu-Lu

IGC/Compiler/CISACodeGen/PatternMatchPass.cpp

@@ -3792,7 +3792,14 @@ namespace IGC
 
     bool CodeGenPatternMatch::MatchWaveShuffleIndex(llvm::GenIntrinsicInst& I)
     {
-        HandleSubspanUse(I.getArgOperand(0));
+        llvm::Value* helperLaneMode = I.getOperand(2);
+        assert(helperLaneMode);
+        if (int_cast<int>(cast<ConstantInt>(helperLaneMode)->getSExtValue()) == 1)
+        {
+            //only if helperLaneMode==1, we enable helper lane under some shuffleindex cases (not for all cases).
+            HandleSubspanUse(I.getArgOperand(0));
+            HandleSubspanUse(&I);
+        }
         return MatchSingleInstruction(I);
     }
 

IGC/Compiler/Optimizer/CodeAssumption.cpp

@@ -116,11 +116,12 @@ void CodeAssumption::uniformHelper(Module* M)
                     // The value must be uniform. Using shuffle with index=0 to
                     // enforce it. (This is entry BB, thus lane 0 must be active.)
                     Type* int32Ty = Type::getInt32Ty(M->getContext());
-                    Value* args[2];
+                    Value* args[3];
                     args[0] = CallI;
                     args[1] = ConstantInt::getNullValue(int32Ty);
+                    args[2] = ConstantInt::get(int32Ty, 0);
 
-                    Type* ITys[2] = { args[0]->getType(), int32Ty };
+                    Type* ITys[3] = { args[0]->getType(), int32Ty, int32Ty};
                     Function* shuffleIntrin = GenISAIntrinsic::getDeclaration(
                         M,
                         GenISAIntrinsic::GenISA_WaveShuffleIndex,

IGC/Compiler/Optimizer/OpenCLPasses/SubGroupFuncs/SubGroupFuncsResolution.cpp

@@ -562,9 +562,11 @@ void SubGroupFuncsResolution::visitCallInst(CallInst& CI)
         CheckSIMDSize(CI, "Shuffle not supported in SIMD32");
 
         // Creates intrinsics that will be lowered in the CodeGen and will handle the sub_group_shuffle function
-        Value* args[2];
+        IRBuilder<> IRB(&CI);
+        Value* args[3];
         args[0] = CI.getArgOperand(0);
         args[1] = CI.getArgOperand(1);
+        args[2] = IRB.getInt32(0);
 
         Function* simdShuffleFunc = GenISAIntrinsic::getDeclaration(CI.getCalledFunction()->getParent(),
             GenISAIntrinsic::GenISA_WaveShuffleIndex, args[0]->getType());

IGC/GenISAIntrinsics/Intrinsic_definitions.py

@@ -283,8 +283,9 @@
     "GenISA_WaveBallot": ["int",["bool"],"Convergent,InaccessibleMemOnly"],
     # For each active lane n, return value of n-th bit from the input bitfield. Signature: (bitfield)->bool
     "GenISA_WaveInverseBallot": ["bool",["int"],"Convergent,InaccessibleMemOnly"],
-    # Read from a specific lane. Signature: (value, lane)->value
-    "GenISA_WaveShuffleIndex": ["anyint",[0,"int"],"Convergent,NoMem"],
+    # Read from a specific lane. Signature: (value, lane, helperLaneMode)->value
+    # helperLaneMode is to tell if helper lane could be used under some situations. 0: not used; 1: used.
+    "GenISA_WaveShuffleIndex": ["anyint",[0,"int", "int"],"Convergent,NoMem"],
     # Accumulate all the active lanes. Signature: (value, op)->result; where op is one of IGC::WaveOps
     "GenISA_WaveAll": ["anyint",[0,"char"],"Convergent,InaccessibleMemOnly"],
     # Accumulate all active lanes within consecutive input clusters and broadcast the result to associated output clusters.

IGC/LLVM3DBuilder/BuiltinsFrontend.hpp

@@ -762,7 +762,7 @@ class LLVM3DBuilder : public llvm::IGCIRBuilder<T, Inserter>
     llvm::CallInst* create_countbits(llvm::Value* src);
     llvm::Value* create_waveBallot(llvm::Value* src);
     llvm::Value* create_waveInverseBallot(llvm::Value* src);
-    llvm::Value* create_waveshuffleIndex(llvm::Value* src, llvm::Value* index);
+    llvm::Value* create_waveshuffleIndex(llvm::Value* src, llvm::Value* index, llvm::Value* helperLaneMode = nullptr);
     llvm::Value* create_waveAll(llvm::Value* src, llvm::Value* type);
     llvm::Value* create_wavePrefix(
         llvm::Value* src, llvm::Value* type, bool inclusive,

IGC/LLVM3DBuilder/BuiltinsFrontendDefinitions.hpp

@@ -4373,14 +4373,14 @@ inline llvm::Value* LLVM3DBuilder<preserveNames, T, Inserter>::create_waveBallot
 }
 
 template<bool preserveNames, typename T, typename Inserter>
-inline llvm::Value* LLVM3DBuilder<preserveNames, T, Inserter>::create_waveshuffleIndex(llvm::Value* src, llvm::Value* index)
+inline llvm::Value* LLVM3DBuilder<preserveNames, T, Inserter>::create_waveshuffleIndex(llvm::Value* src, llvm::Value* index, llvm::Value* helperLaneMode)
 {
     llvm::Module* module = this->GetInsertBlock()->getParent()->getParent();
     llvm::Function* pFunc = llvm::GenISAIntrinsic::getDeclaration(
         module,
         llvm::GenISAIntrinsic::GenISA_WaveShuffleIndex,
         src->getType());
-    return this->CreateCall2(pFunc, src, index);
+    return this->CreateCall3(pFunc, src, index, (helperLaneMode ? helperLaneMode : this->getInt32(0)));
 }
 
 template<bool preserveNames, typename T, typename Inserter>