Add CROSS_THREAD_OFF_R0_RELOCATION runtime symbol in ZeBinary builder if it is used in the program

CROSS_THREAD_OFF_R0_RELOCATION symbol was added in ZeBinaryBuilder only if EnableGlobalStateBuffer registry flag was set.
This relocation in the kernel is set when vISA_emitCrossThreadOffR0Reloc option is set. VC backend adds this option without checking for EnableGlobalStateBuffer registry setting for subtargets that have thread payload in memory.

This change makez ZeBinary builder add the runtime symbol if the relocation is detected in the program.

Author: PawelJurek

IGC/AdaptorOCL/OCL/sp/zebin_builder.cpp

@@ -160,7 +160,7 @@ void ZEBinaryBuilder::addProgramScopeInfo(const IGC::SOpenCLProgramInfo& program
 {
     addGlobalConstants(programInfo);
     addGlobals(programInfo);
-    addRuntimeSymbols();
+    addRuntimeSymbols(programInfo);
     addProgramSymbols(programInfo);
     addProgramRelocations(programInfo);
     addGlobalHostAccessInfo(programInfo);
@@ -375,10 +375,10 @@ void ZEBinaryBuilder::addSymbol(const vISA::ZESymEntry& sym, uint8_t binding,
         getSymbolElfType(sym), targetSect);
 }
 
-void ZEBinaryBuilder::addRuntimeSymbols()
+void ZEBinaryBuilder::addRuntimeSymbols(const IGC::SOpenCLProgramInfo& annotations)
 {
-    if (IGC_IS_FLAG_ENABLED(EnableGlobalStateBuffer))
-        mBuilder.addSymbol("__INTEL_PATCH_CROSS_THREAD_OFFSET_OFF_R0", /*addr*/0, /*size*/0,
+    if (annotations.m_hasCrossThreadOffsetRelocations)
+        mBuilder.addSymbol(vISA::CROSS_THREAD_OFF_R0_RELOCATION_NAME, /*addr*/0, /*size*/0,
             llvm::ELF::STB_GLOBAL, llvm::ELF::STT_NOTYPE, /*sectionId*/-1);
 }
 

IGC/AdaptorOCL/OCL/sp/zebin_builder.hpp

@@ -106,7 +106,7 @@ class ZEBinaryBuilder : DisallowCopy
     void addMiscInfoSection(std::string sectName, const uint8_t* data, uint32_t size);
 
     /// add runtime symbols
-    void addRuntimeSymbols();
+    void addRuntimeSymbols(const IGC::SOpenCLProgramInfo& annotations);
 
     /// add note section for IGC metrics
     void addMetrics(const uint8_t* data, uint32_t size);

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -6424,6 +6424,16 @@ namespace IGC
         if (ZEBinEnabled)
         {
             CreateRelocationTable(pMainKernel, pOutput->m_relocs);
+            for (const auto& reloc : pOutput->m_relocs)
+            {
+                if (reloc.r_symbol == vISA::CROSS_THREAD_OFF_R0_RELOCATION_NAME)
+                {
+                    IGC_ASSERT(context->type == ShaderType::OPENCL_SHADER);
+                    auto cl_context = static_cast<OpenCLProgramContext*>(context);
+                    cl_context->m_programInfo.m_hasCrossThreadOffsetRelocations = true;
+                    break;
+                }
+            }
         }
         else
         {

IGC/Compiler/CodeGenPublic.h

@@ -894,6 +894,7 @@ namespace IGC
         ZEBinProgramSymbolTable m_zebinSymbolTable;
         LegacySymbolTable m_legacySymbolTable;
         ZEBinGlobalHostAccessTable m_zebinGlobalHostAccessTable;
+        bool m_hasCrossThreadOffsetRelocations = false;
     };
 
     class CBTILayout

IGC/VectorCompiler/igcdeps/include/vc/igcdeps/cmc.h

@@ -129,6 +129,7 @@ class CGen8CMProgram : public iOpenCL::CGen8OpenCLProgramBase {
   void GetZEBinary(llvm::raw_pwrite_stream &programBinary,
                    unsigned pointerSizeInBytes) override;
   bool HasErrors() const { return !m_ErrorLog.empty(); };
+  bool HasCrossThreadOffsetRelocations();
   llvm::Error GetError() const;
 
   // CM kernel list.

IGC/VectorCompiler/igcdeps/src/PatchTokens.cpp

@@ -219,3 +219,14 @@ void CGen8CMProgram::GetZEBinary(llvm::raw_pwrite_stream &programBinary,
   }
   zebuilder.getBinaryObject(programBinary);
 }
+
+bool CGen8CMProgram::HasCrossThreadOffsetRelocations() {
+  for (const auto &kernel : m_kernels) {
+    for (const auto &reloc : kernel->getProgramOutput().m_relocs) {
+      if (reloc.r_symbol == vISA::CROSS_THREAD_OFF_R0_RELOCATION_NAME) {
+        return true;
+      }
+    }
+  }
+  return false;
+}

IGC/VectorCompiler/igcdeps/src/cmc.cpp

@@ -831,7 +831,8 @@ getDataAnnotation(const GenXOCLRuntimeInfo::DataInfo &Data) {
 
 static void
 fillOCLProgramInfo(IGC::SOpenCLProgramInfo &ProgramInfo,
-                   const GenXOCLRuntimeInfo::ModuleInfoT &ModuleInfo) {
+                   const GenXOCLRuntimeInfo::ModuleInfoT &ModuleInfo,
+                   bool HasCrossThreadOffsetRelocations) {
   auto ConstantAnnotation = getDataAnnotation<iOpenCL::InitConstantAnnotation>(
       ModuleInfo.Constant.Data);
   if (ConstantAnnotation)
@@ -866,16 +867,18 @@ fillOCLProgramInfo(IGC::SOpenCLProgramInfo &ProgramInfo,
   IGC_ASSERT_MESSAGE(
       ModuleInfo.ConstString.Relocations.empty(),
       "relocations inside constant string section are not supported");
+  ProgramInfo.m_hasCrossThreadOffsetRelocations = HasCrossThreadOffsetRelocations;
 };
 
 void vc::createBinary(
     vc::CGen8CMProgram &CMProgram,
     const GenXOCLRuntimeInfo::CompiledModuleT &CompiledModule) {
-  fillOCLProgramInfo(*CMProgram.m_programInfo, CompiledModule.ModuleInfo);
   for (const GenXOCLRuntimeInfo::CompiledKernel &CompKernel :
        CompiledModule.Kernels) {
     auto K = std::make_unique<CMKernel>(CMProgram.getPlatform());
     fillKernelInfo(CompKernel, *K);
     CMProgram.m_kernels.push_back(std::move(K));
   }
+  fillOCLProgramInfo(*CMProgram.m_programInfo, CompiledModule.ModuleInfo,
+      CMProgram.HasCrossThreadOffsetRelocations());
 }

visa/Optimizer.cpp

@@ -9027,7 +9027,7 @@ bool Optimizer::foldPseudoAndOr(G4_BB* bb, INST_LIST_ITER& ii)
             auto addInst = builder.createBinOp(G4_add, g4::SIMD1, dst, src0, src1,
                 InstOpt_WriteEnable | InstOpt_NoCompact, false);
             RelocationEntry::createRelocation(builder.kernel, *addInst,
-                1, "__INTEL_PATCH_CROSS_THREAD_OFFSET_OFF_R0", GenRelocType::R_SYM_ADDR_32);
+                1, CROSS_THREAD_OFF_R0_RELOCATION_NAME, GenRelocType::R_SYM_ADDR_32);
             instBuffer.push_back(addInst);
         };
 

visa/include/RelocationInfo.h

@@ -21,6 +21,7 @@ SPDX-License-Identifier: MIT
 namespace vISA {
 
 static const uint32_t MAX_SYMBOL_NAME_LENGTH = 1024;
+constexpr const char* CROSS_THREAD_OFF_R0_RELOCATION_NAME = "__INTEL_PATCH_CROSS_THREAD_OFFSET_OFF_R0";
 
 /// GenSymType - Specify the symbol's type
 enum GenSymType {