reduce spatial compile-times in convolutions using a scf.for instead of materializing a huge number of instructions

src/PIM/Conversion/ONNXToSpatial/ONNXToSpatialPass.cpp

@@ -1,4 +1,5 @@
 #include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Pass/Pass.h"
 #include "mlir/Pass/PassManager.h"
@@ -75,7 +76,11 @@ void ONNXToSpatialPass::runOnOperation() {
   }
 
   ConversionTarget target(*ctx);
-  target.addLegalDialect<spatial::SpatialDialect, ONNXDialect, tensor::TensorDialect, arith::ArithDialect>();
+  target.addLegalDialect<spatial::SpatialDialect,
+                         ONNXDialect,
+                         tensor::TensorDialect,
+                         arith::ArithDialect,
+                         scf::SCFDialect>();
   target.addDynamicallyLegalOp<ONNXMatMulOp>(
     [](ONNXMatMulOp op) { return cast<ShapedType>(op.getY().getType()).getRank() != 2; });
   target.addIllegalOp<ONNXAddOp>();