Copyless call-by-value [RTC 32245]

Copyless call-by-value
In the current implementation, call-by-value allocates memory for function parameters and call-by-reference does not. Call-by-value must be used if the caller must be sure that the callee does not modify the argument. Moreover, call-by-value has nicer semantics; it is much easier to understand the workings of a program that uses by call-by-value rather than call-by-reference because of the no-change guarantee. However, call-by-value can negatively impact performance as was clearly observed with logging in the BCD application.
KBC therefore submits a request for enhancement to IBM to combine the nice semantics of call-by-value with the superior performance of call-by-reference. This would mean that the argument would not be copied and the caller would have the guarantuee that the object passed into the function is not modified. The immutability guarantuee should be the same as if a copy had been made: arbitrarily deeply nested components of variables of complex data types, e.g. fields of records, must remain unchanged when a call returns.
EGL assignment rules are not very strict and assignment often relies on implicit type conversions. When such implicit conversion is needed, it may be difficult to avoid a copy.
At present, KBC does not have firm views on the details of the syntax that would trigger this new parameter passing mechanism. Obviously, backward compatibility with the current language definition prevents the default parameter passing mechanism and the inOut parameter modifier to trigger the new behavior. An additional parameter modifier for copyless pass-by-value would be acceptable, as would altered behavior of the in parameter modifier.
Nonetheless, KBC would appreciate the opportunity to review IBM's detailed proposals.



This RFE has already been discussed with Matthew Heitz.