[ONNX] improve regex matching in onnx-importer name sanitization (llvm#3955)

Instead of adding unsupported characters on a case-by-case basis, we
should replace anything that isn't alphanumeric, `_`, or `.`.

Author: zjgarvey

python/torch_mlir/extras/onnx_importer.py

@@ -742,7 +742,7 @@ def _sanitize_name(self, name):
 
         # Remove characters that are invalid in MLIR identifier names.
         # https://mlir.llvm.org/docs/LangRef/#identifiers-and-keywords
-        return re.sub("[:/-]", "_", name)
+        return re.sub("[^\w\.]", "_", name)
 
     def tensor_proto_to_attr(self, tp: onnx.TensorProto) -> Attribute:
         tensor_type = self.tensor_proto_to_builtin_type(tp)

test/python/onnx_importer/BadName.onnx

@@ -0,0 +1,5 @@
+# The original constant name : "abz_.(1, 2)[$something, %anotherthing]"
+
+# RUN: %PYTHON -m torch_mlir.tools.import_onnx %S/BadName.onnx | FileCheck %s
+
+# CHECK: torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<_abz_._1__2___something___anotherthing_>