[Issue 8991] New: adding a __ctfe branch with return to a function breaks NRVO

[d-bugmail at puremagic.com]

http://d.puremagic.com/issues/show_bug.cgi?id=8991

           Summary: adding a __ctfe branch with return to a function
                    breaks NRVO
           Product: D
           Version: D2
          Platform: All
        OS/Version: All
            Status: NEW
          Severity: normal
          Priority: P2
         Component: DMD
        AssignedTo: nobody at puremagic.com
        ReportedBy: dmitry.olsh at gmail.com


--- Comment #0 from Dmitry Olshansky <dmitry.olsh at gmail.com> 2012-11-10 03:14:36 PST ---
Sample obtained while trying to make move work (at least making a copy) during
CTFE.
In the following snippet if __ctfe section is commented out, then return value
doesn't get copied. If it's present however there is a postblit call.

The expected result is that __ctfe should never affect code generation of
run-time code.

Tested on DMD 2.061 from git master.

import core.stdc.string;

T move(T)(ref T source)
{    
    if (__ctfe)
    {
    *cast(int*)0 = 0; //to demonstrate that no CTFE is attempted
        T result = source;    
        return result;   //must have broke NRVO
    }   

    T result = void;


    static if (is(T == struct))
    {
    static T empty;
        memcpy(&result, &source, T.sizeof);
    memcpy(&source, &empty, T.sizeof);
    }
    else
    {
        result = source;
    }
    return result;
}

unittest
{
    // Issue 5661 test(2)
    static struct S4
    {
        static struct X
        { 
            int n = 0; 
            this(this){n = 0;} 
        }
        X x;
    }
    S4 s41;
    s41.x.n = 1;
    S4 s42 = move(s41);
    assert(s41.x.n == 0); //ok, memcpy-ed T.init over source
    assert(s42.x.n == 1); //fails, postblit got called somewhere 
}

-- 
Configure issuemail: http://d.puremagic.com/issues/userprefs.cgi?tab=email
------- You are receiving this mail because: -------