[OCL]Need some cleanup before retry.

CShaderProgram has llvm::Function ptr as its field. Right before retry,
the existing module is deleted, so all of its functions are deleted as well.
This is to make sure those function ptr is set to null to avoid dangling
function ptr.

Also, for kernels that need retry,  its CShaderPrograms need to be freed
to avoid memory leak.

Don't exxpect any functional change.

Author: jgu222

IGC/AdaptorOCL/OCL/sp/spp_g8.cpp

@@ -137,6 +137,13 @@ CGen8OpenCLProgram::~CGen8OpenCLProgram()
     m_ShaderProgramList.clear();
 }
 
+void CGen8OpenCLProgram::clearBeforeRetry()
+{
+    for (auto P : m_ShaderProgramList) {
+        P->clearBeforeRetry();
+    }
+}
+
 RETVAL CGen8OpenCLProgramBase::GetProgramDebugData(Util::BinaryStream& programDebugData)
 {
     // Used by VC only

IGC/AdaptorOCL/OCL/sp/spp_g8.h

@@ -98,6 +98,9 @@ class CGen8OpenCLProgram : public CGen8OpenCLProgramBase
     // Used to track the kernel info from CodeGen
     std::vector<IGC::CShaderProgram*> m_ShaderProgramList;
 
+    // invoked when the current module is deleted to avoid dangling Function ptr.
+    void clearBeforeRetry();
+
 private:
 
     class CLProgramCtxProvider : public CGen8OpenCLStateProcessor::IProgramContext {

IGC/AdaptorOCL/dllInterfaceCompute.cpp

@@ -1452,6 +1452,7 @@ bool TranslateBuildSPMD(
             {
                 splitter.retry();
                 kernelFunctions.clear();
+                oclContext.clearBeforeRetry();
                 oclContext.clear();
 
                 // Create a new LLVMContext

IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp

@@ -2787,6 +2787,8 @@ namespace IGC
         // Clear the retry set and collect kernels for retry in the loop below.
         ctx->m_retryManager.kernelSet.clear();
 
+        // If kernel needs retry, its shaderProgram should be deleted
+        SmallVector<CShaderProgram*, 8> toBeDeleted;
         // gather data to send back to the driver
         for (const auto& k : shaders)
         {
@@ -2817,6 +2819,20 @@ namespace IGC
                 else if (COpenCLKernel::IsValidShader(simd8Shader))
                     GatherDataForDriver(ctx, simd8Shader, pKernel, pFunc, pMdUtils, SIMDMode::SIMD8);
             }
+
+            if (ctx->m_programOutput.m_ShaderProgramList.empty() ||
+                (ctx->m_programOutput.m_ShaderProgramList.back() != pKernel))
+            {
+                // This CShaderProgram needs rebuilding in retry run.
+                toBeDeleted.push_back(pKernel);
+            }
+        }
+
+        for (int i=0, sz = (int)toBeDeleted.size(); i < sz; ++i)
+        {
+            CShaderProgram* SP = toBeDeleted[i];
+            shaders.erase(SP->getLLVMFunction());
+            delete SP;
         }
 
         // The skip set to avoid retry is not needed. Clear it and collect a new set

IGC/Compiler/CISACodeGen/ShaderCodeGen.hpp

@@ -732,8 +732,19 @@ class CShaderProgram
     void DeleteShader(SIMDMode simd, ShaderDispatchMode mode = ShaderDispatchMode::NOT_APPLICABLE);
     CodeGenContext* GetContext() { return m_context; }
 
+    llvm::Function* getLLVMFunction() const { return m_kernel; }
     ShaderStats* m_shaderStats;
 
+    // invoked to clear Func ptr when the current module is deleted (so is func within it).
+    void clearBeforeRetry() {
+        m_kernel = nullptr;
+        for (auto S : m_SIMDshaders) {
+            if (S != nullptr) {
+                S->entry = nullptr;
+            }
+        }
+    }
+
 protected:
     CShader*& GetShaderPtr(SIMDMode simd, ShaderDispatchMode mode);
     CShader* CreateNewShader(SIMDMode simd);

IGC/Compiler/CodeGenPublic.h

@@ -1936,6 +1936,10 @@ namespace IGC
         uint32_t getIntelScratchSpacePrivateMemoryMinimalSizePerThread() const override;
         void failOnSpills();
         bool needsDivergentBarrierHandling() const;
+
+        void clearBeforeRetry() {
+            m_programOutput.clearBeforeRetry();
+        }
     private:
         llvm::DenseMap<llvm::Function*, std::string> m_hashes_per_kernel;
     };