fix FuseResourceLoop 3rd try

cannot reuse source variable for inst inside a fused resource-loop

Author: Gang Y Chen

IGC/Compiler/CISACodeGen/CShader.cpp

@@ -42,6 +42,7 @@ CShader::CShader(Function* pFunc, CShaderProgram* pProgram)
     m_DL = nullptr;
     m_FGA = nullptr;
     m_VRA = nullptr;
+    m_RLA = nullptr;
     m_EmitPass = nullptr;
     m_HW_TID = nullptr;
 
@@ -3277,7 +3278,12 @@ CVariable* CShader::GetSymbol(llvm::Value* value, bool fromConstantPool)
         {
             if (auto Inst = dyn_cast<Instruction>(value))
             {
-                var = GetSymbolFromSource(Inst, preferredAlign);
+                // cannot reuse source variable
+                // when the instruction is inside a resource-loop
+                if (m_RLA->GetResourceLoopMarker(Inst) ==
+                    ResourceLoopAnalysis::MarkResourceLoopOutside) {
+                    var = GetSymbolFromSource(Inst, preferredAlign);
+                }
             }
         }
     }

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -687,6 +687,7 @@ bool EmitPass::runOnFunction(llvm::Function& F)
     m_currShader->SetFunctionGroupAnalysis(m_FGA);
     m_currShader->SetPushInfoHelper(&(m_moduleMD->pushInfo));
     m_currShader->SetVariableReuseAnalysis(m_VRA);
+    m_currShader->SetResourceLoopAnalysis(m_RLA);
     m_currShader->SetDeSSAHelper(m_deSSA);
     m_currShader->SetEmitPassHelper(this);
 

IGC/Compiler/CISACodeGen/ResourceLoopAnalysis.cpp

@@ -175,6 +175,16 @@ bool ResourceLoopAnalysis::runOnFunction(Function &F) {
         }
 
         if (!LoopEnd) {
+          if (prevMemIter != BB->end()) {
+            // mark instructions in between two mem-op as inside
+            auto III = prevMemIter;
+            ++III;
+            while (III != II) {
+              auto InBetween = &*III;
+              LoopMap[InBetween] = MarkResourceLoopInside;
+              ++III;
+            }
+          }
           LoopMap[I] = MarkResourceLoopInside;
           prevMemIter = II;
           DefSet.insert(I);

IGC/Compiler/CISACodeGen/ShaderCodeGen.hpp

@@ -48,6 +48,7 @@ class DeSSA;
 class CoalescingEngine;
 class GenXFunctionGroupAnalysis;
 class VariableReuseAnalysis;
+class ResourceLoopAnalysis;
 class EmitPass;
 
 struct PushInfo;
@@ -292,6 +293,7 @@ class CShader
     void        SetDominatorTreeHelper(llvm::DominatorTree* DT) { m_DT = DT; }
     void        SetDataLayout(const llvm::DataLayout* DL) { m_DL = DL; }
     void        SetVariableReuseAnalysis(VariableReuseAnalysis* VRA) { m_VRA = VRA; }
+    void        SetResourceLoopAnalysis(ResourceLoopAnalysis *RLA) { m_RLA = RLA; }
     void        SetMetaDataUtils(IGC::IGCMD::MetaDataUtils* pMdUtils) { m_pMdUtils = pMdUtils; }
     void        SetScratchSpaceSize(uint size) { m_ScratchSpaceSize = size; }
 
@@ -667,6 +669,7 @@ class CShader
     const llvm::DataLayout* m_DL;
     GenXFunctionGroupAnalysis* m_FGA;
     VariableReuseAnalysis* m_VRA;
+    ResourceLoopAnalysis* m_RLA;
 
     uint m_numBlocks;
     IGC::IGCMD::MetaDataUtils* m_pMdUtils;