Specify when use private memory for fp64

This change determines when the privMem should be used
for cases with fp64.

Author: matborzyszkowski

IGC/Compiler/Optimizer/OpenCLPasses/PrivateMemory/PrivateMemoryUsageAnalysis.cpp

@@ -42,6 +42,8 @@ bool PrivateMemoryUsageAnalysis::runOnModule(Module& M)
 
     bool hasStackCall = false;
 
+    m_hasDPDivSqrtEmu = !pCtx->platform.hasNoFP64Inst() && !pCtx->platform.hasCorrectlyRoundedMacros() && pCtx->m_DriverInfo.NeedFP64DivSqrt();
+
     // Run on all functions defined in this module
     for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
     {
@@ -125,7 +127,6 @@ bool PrivateMemoryUsageAnalysis::runOnFunction(Function& F)
         CodeGenContext* pCtx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
         // This is the condition that double emulation is used.
         if ((IGC_IS_FLAG_ENABLED(ForceDPEmulation) ||
-            (!pCtx->platform.hasNoFP64Inst() && !pCtx->platform.hasCorrectlyRoundedMacros() && pCtx->m_DriverInfo.NeedFP64DivSqrt()) ||
             (pCtx->m_DriverInfo.NeedFP64(pCtx->platform.getPlatformInfo().eProductFamily) && pCtx->platform.hasNoFP64Inst())))
         {
             m_hasPrivateMem = true;
@@ -171,4 +172,26 @@ void PrivateMemoryUsageAnalysis::visitBinaryOperator(llvm::BinaryOperator& I)
             break;
         }
     }
+
+    // Check if an instruction is fp64 div to enable privMem
+    if (m_hasDPDivSqrtEmu)
+    {
+        if (I.getOpcode() == Instruction::FDiv)
+        {
+            m_hasPrivateMem = true;
+        }
+    }
+}
+
+void PrivateMemoryUsageAnalysis::visitCallInst(llvm::CallInst& CI)
+{
+    // Check if a sqrtd builtin is called to enable privMem
+    if (m_hasDPDivSqrtEmu)
+    {
+        Function* calledFunc = CI.getCalledFunction();
+        if (calledFunc->getName().startswith("__builtin_IB_native_sqrtd"))
+        {
+            m_hasPrivateMem = true;
+        }
+    }
 }

IGC/Compiler/Optimizer/OpenCLPasses/PrivateMemory/PrivateMemoryUsageAnalysis.hpp

@@ -61,6 +61,11 @@ namespace IGC
         /// @param  I The binary op
         void visitBinaryOperator(llvm::BinaryOperator& I);
 
+        /// @brief  CallInst instructions visitor.
+        ///         Analyzes if there are private memory allocation.
+        /// @param  CI The binary op
+        void visitCallInst(llvm::CallInst& CI);
+
     private:
         /// @brief  Function entry point.
         ///         Finds all alloca instructions in this function, analyzes them and adds
@@ -72,6 +77,9 @@ namespace IGC
         /// @brief  A flag signaling if the current function uses private memory
         bool m_hasPrivateMem;
 
+        /// @brief A flag signaling if the platform has partial fp64 emulation
+        bool m_hasDPDivSqrtEmu = false;
+
         /// @brief  MetaData utils used to generate LLVM metadata
         IGCMD::MetaDataUtils* m_pMDUtils;
     };