Revert cx_Oracle WITH_UNICODE change under > 5.0

Fixed performance regression caused by the fix for :ticket:`3937` where
cx_Oracle as of version 5.3 dropped the ``.UNICODE`` symbol from its
namespace,  which was interpreted as cx_Oracle's "WITH_UNICODE" mode being
turned on unconditionally, which invokes functions on the SQLAlchemy
side which convert all strings to unicode unconditionally and causing
a performance impact.  In fact, per cx_Oracle's author the
"WITH_UNICODE" mode has been removed entirely as of 5.1, so the expensive unicode
conversion functions are no longer necessary and are disabled if
cx_Oracle 5.1 or greater is detected under Python 2.  The warning against
"WITH_UNICODE" mode that was removed under :ticket:`3937` is also restored.

Change-Id: Iddd38d81a5adb27c953a5ee2eae5529a21da16e1
Fixes: #4035

2 files changed, 28 insertions, 4 deletions

diff --git a/doc/build/changelog/unreleased_10/4035.rst b/doc/build/changelog/unreleased_10/4035.rst
new file mode 100644
index 000000000..74acdfe71
--- /dev/null
+++ b/doc/build/changelog/unreleased_10/4035.rst
@@ -0,0 +1,15 @@
+.. change::
+	:tags: bug, oracle, performance, py2k
+	:tickets: 4035
+	:versions: 1.0.19, 1.1.13, 1.2.0b3
+
+	Fixed performance regression caused by the fix for :ticket:`3937` where
+	cx_Oracle as of version 5.3 dropped the ``.UNICODE`` symbol from its
+	namespace,  which was interpreted as cx_Oracle's "WITH_UNICODE" mode being
+	turned on unconditionally, which invokes functions on the SQLAlchemy
+	side which convert all strings to unicode unconditionally and causing
+	a performance impact.  In fact, per cx_Oracle's author the
+	"WITH_UNICODE" mode has been removed entirely as of 5.1, so the expensive unicode
+	conversion functions are no longer necessary and are disabled if
+	cx_Oracle 5.1 or greater is detected under Python 2.  The warning against
+	"WITH_UNICODE" mode that was removed under :ticket:`3937` is also restored.
diff --git a/lib/sqlalchemy/dialects/oracle/cx_oracle.py b/lib/sqlalchemy/dialects/oracle/cx_oracle.py
index f6c3c97b7..4e9f6314b 100644
--- a/lib/sqlalchemy/dialects/oracle/cx_oracle.py
+++ b/lib/sqlalchemy/dialects/oracle/cx_oracle.py
@@ -723,19 +723,28 @@ class OracleDialect_cx_oracle(OracleDialect):
             self._cx_oracle_string_types = set()
             self._cx_oracle_with_unicode = False
         elif util.py3k or (
-                self.cx_oracle_ver >= (5,) and not \
+                self.cx_oracle_ver >= (5,) and
+                self.cx_oracle_ver < (5, 1) and not
                 hasattr(self.dbapi, 'UNICODE')
         ):
             # cx_Oracle WITH_UNICODE mode.  *only* python
-            # unicode objects accepted for anything
+            # unicode objects accepted for anything.  This
+            # mode of operation is implicit for Python 3,
+            # however under Python 2 it existed as a never-used build-time
+            # option for cx_Oracle 5.0 only and was removed in 5.1.
             self.supports_unicode_statements = True
             self.supports_unicode_binds = True
             self._cx_oracle_with_unicode = True
 
             if util.py2k:
                 # There's really no reason to run with WITH_UNICODE under
-                # Python 2.x.  However as of cx_oracle 5.3 it seems to be
-                # set to ON for default builds
+                # Python 2.x.  Give the user a hint.
+                util.warn(
+                    "cx_Oracle is compiled under Python 2.xx using the "
+                    "WITH_UNICODE flag.  Consider recompiling cx_Oracle "
+                    "without this flag, which is in no way necessary for "
+                    "full support of Unicode and causes significant "
+                    "performance issues.")
                 self.execution_ctx_cls = \
                     OracleExecutionContext_cx_oracle_with_unicode
         else: