http://d.puremagic.com/issues/show_bug.cgi?id=5546

           Summary: Assigning and initializing structs from functions make
                    more copies than necessary
           Product: D
           Version: D2
          Platform: x86_64
        OS/Version: Linux
            Status: NEW
          Severity: enhancement
          Priority: P2
         Component: DMD
        AssignedTo: nobody@puremagic.com
        ReportedBy: akb825@gmail.com


--- Comment #0 from akb825@gmail.com 2011-02-07 23:31:56 PST ---
When initializing a struct from a function that returns by value, more copies (calling the post blit and destructors) are called than are necessary. For example, see the attached source file.

The output for TestCopy.d is currently:
Creating temp
Copying temp
Deleting temp
Creating copy
Copying copy
Deleting copy
Forwarding copy
Copying copy
Deleting copy
Returning global
Copying global
Deleting global
Deleting copy
Deleting temp

Ideally, the output should look like this:
Creating temp
Creating copy
Forwarding copy
Returning global
Copying global
Deleting global
Deleting copy
Deleting temp

When a struct is being initialized by the return value of a function, apart from the memory being blitted over, no post blit or destructor should need to be called, since semantically it's equivalent to directly initializing the struct in the called function. This can be achieved by always returning a local object and not destructing the local object being returned. In the case of globalFunc(), which is returning a non-local object, a temporary would be made before returning from globalFunc().

When assigning the returning value of a function to a struct that's already initialized, additional optimizations can be made if no custom assignment operator exists.

For example:
Test testVal;
testVal = function();

This will create the copy for the return value of function(), post blit testVal, destroy the previous value of testVal, then destroy the return value of function(). If Test or any of its members have an overridden assignment operator, they must be called. However, in the case where there is no custom assignment operator, the post blit of testVal and destruction of the return value of function() can be omitted, since you are semantically moving the value from the return value to testVal.

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

http://d.puremagic.com/issues/show_bug.cgi?id=5546



--- Comment #1 from akb825@gmail.com 2011-02-07 23:32:33 PST ---
Created an attachment (id=901)
Source file that demonstrates the behavior.

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

http://d.puremagic.com/issues/show_bug.cgi?id=5546



--- Comment #2 from akb825@gmail.com 2011-02-08 00:42:20 PST ---
After thinking about the problem a bit more, I have a couple things to add.

First, I should mention my second point for assigning the return value of a function to an existing struct applies to all temporaries.

Test testVal;
testVal = Test("str");

can benefit from the same optimizations as

Test testVal;
testVal = function();

Second, it would be very useful to be able to mark the custom assignment operator to have the same behavior on the above two cases as if there is no overloaded assignment operator. (aka: skip the assignment and reduce it to a blit + post blit and destruction of the old value when assigning from a temporary) I would suggest something like putting "@ignoreIfTemp" before the opAssign definition, and would be ignored if members of the struct have custom assignment operators without that property. This would allow one of the most useful benefits of rvalue references from C++0x to be used without having to add a whole new type qualifier.

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