[mlir][SparseTensor] Simplify pipeline (#152908)

This refactoring improves compilation time.

Author: matthias-springer

mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp

@@ -71,20 +71,13 @@ void mlir::sparse_tensor::buildSparsifier(OpPassManager &pm,
   pm.addPass(createLowerAffinePass());
   pm.addPass(
       createConvertVectorToLLVMPass(options.convertVectorToLLVMOptions()));
-  pm.addPass(createFinalizeMemRefToLLVMConversionPass());
   pm.addNestedPass<func::FuncOp>(createConvertComplexToStandardPass());
   pm.addNestedPass<func::FuncOp>(arith::createArithExpandOpsPass());
   pm.addNestedPass<func::FuncOp>(createConvertMathToLLVMPass());
   pm.addPass(createConvertMathToLibmPass());
   pm.addPass(createConvertComplexToLibm());
   pm.addPass(
       createConvertVectorToLLVMPass(options.convertVectorToLLVMOptions()));
-  pm.addPass(createConvertComplexToLLVMPass());
-  pm.addPass(
-      createConvertVectorToLLVMPass(options.convertVectorToLLVMOptions()));
-  pm.addPass(createConvertFuncToLLVMPass());
-  pm.addPass(createArithToLLVMConversionPass());
-  pm.addPass(createConvertControlFlowToLLVMPass());
 
   // Finalize GPU code generation.
   if (gpuCodegen) {
@@ -99,8 +92,8 @@ void mlir::sparse_tensor::buildSparsifier(OpPassManager &pm,
     pm.addPass(createGpuModuleToBinaryPass(gpuModuleToBinaryPassOptions));
   }
 
-  // Convert poison values.
-  pm.addPass(createUBToLLVMConversionPass());
+  // Convert to LLVM.
+  pm.addPass(createConvertToLLVMPass());
 
   // Ensure all casts are realized.
   pm.addPass(createReconcileUnrealizedCastsPass());