- defer dynamic GAS resolution. - invoke static GAS resolution

in optimizeIR. - check if a generic pointer is present or not.

Change-Id: Ie5aed19e899e50421a55421ee2201843fea09cdb

Author: wpan

IGC/AdaptorOCL/UnifyIROCL.cpp

@@ -324,7 +324,7 @@ static void CommonOCLBasedPasses(
 
     mpm.add(new BreakConstantExpr());
 
-    if(IGC_IS_FLAG_ENABLED(EnableGASResolver))
+    if (IGC_IS_FLAG_ENABLED(EnableGASResolver))
     {
         // Add fix up of illegal `addrspacecast` in respect to OCL 2.0 spec.
         mpm.add(createFixAddrSpaceCastPass());
@@ -416,9 +416,6 @@ static void CommonOCLBasedPasses(
     // forget setting math flags. We expect the generic llvm optimizations after the unification
     // (optimizeIR()) will fully take advantage of the flags.
     mpm.add(new SetFastMathFlags());
-
-    mpm.add(createGenericAddressDynamicResolutionPass());
-
     mpm.add(new FixResourcePtr());
 
     bool isOptDisabled = pContext->getModuleMetaData()->compOpt.OptDisable;

IGC/Compiler/CISACodeGen/CheckInstrTypes.cpp

@@ -53,7 +53,7 @@ char CheckInstrTypes::ID = 0;
 CheckInstrTypes::CheckInstrTypes(IGC::SInstrTypes* instrList) : FunctionPass(ID), g_InstrTypes(instrList)
 {
     initializeLoopInfoWrapperPassPass(*PassRegistry::getPassRegistry());	
-	initializeCheckInstrTypesPass(*PassRegistry::getPassRegistry());
+    initializeCheckInstrTypesPass(*PassRegistry::getPassRegistry());
 
     instrList->CorrelatedValuePropagationEnable = false;
     instrList->hasLoop = false;
@@ -68,6 +68,7 @@ CheckInstrTypes::CheckInstrTypes(IGC::SInstrTypes* instrList) : FunctionPass(ID)
     instrList->hasFunctionAddressTaken = false;
     instrList->hasSel = false;
     instrList->hasPointer = false;
+    instrList->hasGenericAddressSpacePointers = false;
     instrList->hasLocalLoadStore = false;
     instrList->hasSubroutines = false;
     instrList->hasPrimitiveAlloca = false;
@@ -79,13 +80,13 @@ CheckInstrTypes::CheckInstrTypes(IGC::SInstrTypes* instrList) : FunctionPass(ID)
     instrList->hasBarrier = false;
     instrList->hasDiscard = false;
     instrList->mayHaveIndirectOperands = false;
-	instrList->numSample = 0;
-	instrList->numBB = 0;
-	instrList->numLoopInsts = 0;
+    instrList->numSample = 0;
+    instrList->numBB = 0;
+    instrList->numLoopInsts = 0;
     instrList->numOfLoop = 0;
     instrList->numInsts = 0;
-	instrList->sampleCmpToDiscardOptimizationPossible = false;
-	instrList->sampleCmpToDiscardOptimizationSlot = 0;
+    instrList->sampleCmpToDiscardOptimizationPossible = false;
+    instrList->sampleCmpToDiscardOptimizationSlot = 0;
 }
 
 void CheckInstrTypes::SetLoopFlags(Function &F)
@@ -116,15 +117,22 @@ bool CheckInstrTypes::runOnFunction(Function &F)
     g_InstrTypes->hasDebugInfo = F.getParent()->getNamedMetadata("llvm.dbg.cu") != nullptr;
 
     visit(F);
-	SetLoopFlags(F);
+    SetLoopFlags(F);
     return false;
 }
 
-void CheckInstrTypes::visitInstruction(llvm::Instruction &I) {
+void CheckInstrTypes::visitInstruction(llvm::Instruction &I)
+ {
     if (!llvm::isa<llvm::DbgInfoIntrinsic>(&I))
     {
         g_InstrTypes->numInsts++;
     }
+
+    auto PT = dyn_cast<PointerType>(I.getType());
+    if (PT && PT->getPointerAddressSpace() == ADDRESS_SPACE_GENERIC)
+    {
+        g_InstrTypes->hasGenericAddressSpacePointers = true;
+    }
 }
 
 void CheckInstrTypes::visitCallInst(CallInst &C)
@@ -245,6 +253,12 @@ void CheckInstrTypes::visitAllocaInst(AllocaInst &I)
     {
         g_InstrTypes->hasPrimitiveAlloca = true;
     }
+
+    auto PT = dyn_cast<PointerType>(I.getAllocatedType());
+    if (PT && PT->getPointerAddressSpace() == ADDRESS_SPACE_GENERIC)
+    {
+        g_InstrTypes->hasGenericAddressSpacePointers = true;
+    }
 }
 
 void CheckInstrTypes::visitLoadInst(LoadInst &I)
@@ -255,21 +269,29 @@ void CheckInstrTypes::visitLoadInst(LoadInst &I)
     {
         g_InstrTypes->hasLocalLoadStore = true;
     }
+    if (I.getPointerAddressSpace() == ADDRESS_SPACE_GENERIC)
+    {
+        g_InstrTypes->hasGenericAddressSpacePointers = true;
+    }
 }
 
 void CheckInstrTypes::visitStoreInst(StoreInst &I)
 {
     g_InstrTypes->numInsts++;
     g_InstrTypes->hasLoadStore = true;
     uint as = I.getPointerAddressSpace();
-    if(as != ADDRESS_SPACE_PRIVATE)
+    if (as != ADDRESS_SPACE_PRIVATE)
     {
         g_InstrTypes->psHasSideEffect = true;
     }
-    if (I.getPointerAddressSpace() == ADDRESS_SPACE_LOCAL)
+    if (as == ADDRESS_SPACE_LOCAL)
     {
         g_InstrTypes->hasLocalLoadStore = true;
     }
+    if (as == ADDRESS_SPACE_GENERIC)
+    {
+        g_InstrTypes->hasGenericAddressSpacePointers = true;
+    }
 }
 
 void CheckInstrTypes::visitPHINode(PHINode &PN)
@@ -283,3 +305,12 @@ void CheckInstrTypes::visitSelectInst(SelectInst &I)
     g_InstrTypes->numInsts++;
     g_InstrTypes->hasSel = true;
 }
+
+void CheckInstrTypes::visitGetElementPtrInst(llvm::GetElementPtrInst &I)
+{
+    g_InstrTypes->numInsts++;
+    if (I.getPointerAddressSpace() == ADDRESS_SPACE_GENERIC)
+    {
+        g_InstrTypes->hasGenericAddressSpacePointers = true;
+    }
+}

IGC/Compiler/CISACodeGen/CheckInstrTypes.hpp

@@ -72,6 +72,7 @@ namespace IGC
         void visitAllocaInst(llvm::AllocaInst &I);
         void visitLoadInst(llvm::LoadInst &I);
         void visitStoreInst(llvm::StoreInst &I);
+        void visitGetElementPtrInst(llvm::GetElementPtrInst &I);
         void visitPHINode(llvm::PHINode &PN);
         void visitSelectInst(llvm::SelectInst &I);
         void SetLoopFlags(llvm::Function &F);

IGC/Compiler/CISACodeGen/FixAddrSpaceCast.cpp

@@ -73,7 +73,7 @@ FunctionPass *IGC::createFixAddrSpaceCastPass() {
 char AddrSpaceCastFixing::ID = 0;
 
 #define PASS_FLAG     "igc-addrspacecast-fix"
-#define PASS_DESC     "Fix invalid addrspacecast-relevant patterns due to incompatible OCL FE, i.e. clang"
+#define PASS_DESC     "Fix invalid addrspacecast-relevant patterns"
 #define PASS_CFG_ONLY false
 #define PASS_ANALYSIS false
 namespace IGC {

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -71,6 +71,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #include "Compiler/CISACodeGen/POSH_RemoveNonPositionOutput.h"
 
 #include "Compiler/CISACodeGen/SLMConstProp.hpp"
+#include "Compiler/Optimizer/OpenCLPasses/GenericAddressResolution/GenericAddressDynamicResolution.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/PrivateMemory/PrivateMemoryUsageAnalysis.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/PrivateMemory/PrivateMemoryResolution.hpp"
 #include "Compiler/Optimizer/OpenCLPasses/ProgramScopeConstants/ProgramScopeConstantResolution.hpp"
@@ -435,6 +436,18 @@ inline void AddLegalizationPasses(CodeGenContext &ctx, const CShaderProgram::Ker
     // Promotes indirect resource access to direct
     mpm.add(new PromoteResourceToDirectAS());
 
+    if (ctx.m_instrTypes.hasGenericAddressSpacePointers)
+    {
+        if (!isOptDisabled &&
+            IGC_IS_FLAG_ENABLED(EnableGASResolver))
+        {
+            mpm.add(createFixAddrSpaceCastPass());
+            mpm.add(createResolveGASPass());
+            mpm.add(createSROAPass());
+        }
+        mpm.add(createGenericAddressDynamicResolutionPass());
+    }
+
     // Resolve the Private memory to register pass
     if (!isOptDisabled)
     {
@@ -457,8 +470,8 @@ inline void AddLegalizationPasses(CodeGenContext &ctx, const CShaderProgram::Ker
     if (ctx.m_enableSubroutine)
     {
         // Sort functions if subroutine is enabled.
-		mpm.add(llvm::createGlobalDCEPass());
-		mpm.add(new PurgeMetaDataUtils());
+        mpm.add(llvm::createGlobalDCEPass());
+        mpm.add(new PurgeMetaDataUtils());
         mpm.add(createGenXCodeGenModulePass());
     }
 
@@ -1160,7 +1173,15 @@ void OptimizeIR(CodeGenContext* pContext)
         mpm.add(llvm::createInstructionCombiningPass());
         mpm.add(llvm::createDeadCodeEliminationPass());       // this should be done both before/after constant propagation
 
-		if( pContext->m_instrTypes.hasMultipleBB )
+        if (pContext->m_instrTypes.hasGenericAddressSpacePointers &&
+            IGC_IS_FLAG_ENABLED(EnableGASResolver))
+        {
+            mpm.add(createFixAddrSpaceCastPass());
+            mpm.add(createResolveGASPass());
+            mpm.add(createSROAPass());
+        }
+
+        if (pContext->m_instrTypes.hasMultipleBB)
         {
             // disable loop unroll for excessive large shaders
             if( pContext->m_instrTypes.hasLoop )