Disable TrivialLocalMemoryOpsElimination pass from the pipeline

The pass runs into multiple edge-case issues with aliased (and even direct) SLM accesses. Originally implemented within commit ac8f9d5, no clear performance benefit could be identified as of today.
intel · Dec 18, 2024 · 3ae1be7 · 3ae1be7
1 parent 0ea9c16
commit 3ae1be7
Show file tree

Hide file tree

Showing 2 changed files with 1 addition and 1 deletion.
diff --git a/IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp b/IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp
@@ -1852,7 +1852,6 @@ void OptimizeIR(CodeGenContext* const pContext)
             mpm.add(createSROAPass());
         }
 
-        mpm.add(new TrivialLocalMemoryOpsElimination());
         if (pContext->type == ShaderType::COMPUTE_SHADER &&
             (IGC_IS_FLAG_ENABLED(RemoveUnusedTGMFence) ||
                 pContext->getModuleMetaData()->enableRemoveUnusedTGMFence))

diff --git a/IGC/Compiler/CustomSafeOptPass.hpp b/IGC/Compiler/CustomSafeOptPass.hpp
@@ -108,6 +108,7 @@ namespace IGC
         llvm::Value* analyzeTreeForTrunc64bto32b(const llvm::Use& OperandUse, llvm::SmallVector<llvm::BinaryOperator*, 8>& OpsToDelete);
     };
 
+    // TODO: Remove this pass as unused
     class TrivialLocalMemoryOpsElimination : public llvm::FunctionPass, public llvm::InstVisitor<TrivialLocalMemoryOpsElimination>
     {
     public:
-Original file line number
+Diff line change
@@ Expand Up / @@ -108,6 +108,7 @@ namespace IGC @@
             llvm::Value* analyzeTreeForTrunc64bto32b(const llvm::Use& OperandUse, llvm::SmallVector<llvm::BinaryOperator*, 8>& OpsToDelete);
         };
+        // TODO: Remove this pass as unused
         class TrivialLocalMemoryOpsElimination : public llvm::FunctionPass, public llvm::InstVisitor<TrivialLocalMemoryOpsElimination>
         {
         public:
@@ Expand Down @@