Avoid FP64 emulation related code if kernel doesn't use FP64 at all.

Change-Id: Ia68f9ff0d027d4f5b70099f3a41efea92a018f2a

Author: weiyu-chen

IGC/Compiler/CISACodeGen/CheckInstrTypes.cpp

@@ -97,6 +97,7 @@ CheckInstrTypes::CheckInstrTypes(IGC::SInstrTypes* instrList) : FunctionPass(ID)
     instrList->numInsts = 0;
     instrList->sampleCmpToDiscardOptimizationPossible = false;
     instrList->sampleCmpToDiscardOptimizationSlot = 0;
+    instrList->hasFP64Inst = false;
 }
 
 void CheckInstrTypes::SetLoopFlags(Function& F)
@@ -126,6 +127,22 @@ bool CheckInstrTypes::runOnFunction(Function& F)
     // check if module has debug info
     g_InstrTypes->hasDebugInfo = F.getParent()->getNamedMetadata("llvm.dbg.cu") != nullptr;
 
+    if (F.getType()->isDoubleTy())
+    {
+        g_InstrTypes->hasFP64Inst = true;
+    }
+    else
+    {
+        for (auto&& AI = F.arg_begin(), AE = F.arg_end(); AI != AE; ++AI)
+        {
+            if ((*AI).getType()->isDoubleTy())
+            {
+                g_InstrTypes->hasFP64Inst = true;
+                break;
+            }
+        }
+    }
+
     visit(F);
     SetLoopFlags(F);
     return false;
@@ -148,6 +165,11 @@ void CheckInstrTypes::visitInstruction(llvm::Instruction& I)
     {
         g_InstrTypes->hasGenericAddressSpacePointers = true;
     }
+
+    if (I.getType()->isDoubleTy())
+    {
+        g_InstrTypes->hasFP64Inst = true;
+    }
 }
 
 void CheckInstrTypes::visitCallInst(CallInst& C)

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -402,6 +402,7 @@ static void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSi
 
     bool needDPEmu = (IGC_IS_FLAG_ENABLED(ForceDPEmulation) ||
         (ctx.m_DriverInfo.NeedFP64(ctx.platform.getPlatformInfo().eProductFamily) && ctx.platform.hasNoFP64Inst()));
+    needDPEmu &= ctx.m_instrTypes.hasFP64Inst;
     uint32_t theEmuKind = (needDPEmu ? EmuKind::EMU_DP : 0);
     theEmuKind |= (ctx.m_DriverInfo.NeedI64BitDivRem() ? EmuKind::EMU_I64DIVREM : 0);
     theEmuKind |=

IGC/Compiler/CodeGenPublic.h

@@ -268,6 +268,7 @@ namespace IGC
         unsigned int numPsInputs;
         bool sampleCmpToDiscardOptimizationPossible;
         unsigned int sampleCmpToDiscardOptimizationSlot;
+        bool hasFP64Inst;
     };
 
     struct SSimplePushInfo

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

@@ -140,7 +140,7 @@ 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) ||
+        if (pCtx->m_instrTypes.hasFP64Inst && (IGC_IS_FLAG_ENABLED(ForceDPEmulation) ||
             (pCtx->m_DriverInfo.NeedFP64(pCtx->platform.getPlatformInfo().eProductFamily) && pCtx->platform.hasNoFP64Inst())))
         {
             m_hasPrivateMem = true;