Bug 5777 – Move semantics require full spec NRVO

Status
RESOLVED
Resolution
FIXED
Severity
major
Priority
P2
Component
dmd
Product
D
Version
D2
Platform
All
OS
All
Creation time
2011-03-23T18:40:00Z
Last change time
2011-07-16T12:58:17Z
Assigned to
nobody
Creator
k.hara.pg

Attachments

IDFilenameSummaryContent-TypeSize
935patches.zip3 Patchestext/plain3832
963patches.zip2 patchesapplication/x-zip-compressed2037

Comments

Comment #0 by k.hara.pg — 2011-03-23T18:40:27Z
DMD supports Named Return Value Optimization(NRVO), but now this doesn't work for struct that has @disabled postblit and user-defined destructor (e.g. std.typecons.scoped). This looks trivial, but serious problem. 1. Need of incorporating NRVO into the language specifications D does not contain rvalue seantinc in its type system. ('ref' is storage class, not part of type. This means casting is useless for this purpose, unlike C++.) Except for only one, This means D has no way that convert from lvalue to rvalue in current language spec. The one way is NRVO. This is not only trivial optimization, but also core of move semantics in D. ---- T f(){ T ret; // &ret == hidden pointer, and *hidden is 'move target'. // In this example, share memory of t. ret = ... return ret; // NRVO. } T t = f(); // t's memory address is taken hidden pointer ---- Moving object caller to callee is now D supports, but reversed flow is limitation. To full support moving callee to caller through return value, NRVO is required. 2. std.algorithm.move problem If Issue 4499 will be fixed, unary std.algorithm.move() will not work on noncopyable/has expplicit destructor object. To give correct behavior, NRVO spec update is need. 3. ABI problem In http://digitalmars.com/d/2.0/abi.html 'Return Value' term, struct its size is less than 8 byte is returned using registers. Issue 4500 is scoped!T and self-referenced class problem, but the way resolving it will need true NRVO. That has following spec: Any size object returned through hidden pointer, if the object has elaborate destructor at least. See http://d.puremagic.com/issues/show_bug.cgi?id=4500#c5 . Attachment code includes workaround for this issue, But it's too ugly. Do you think?
Comment #1 by k.hara.pg — 2011-03-25T19:06:29Z
Related issues: On return statement, copy construction should work correctly. - Issue 4437 - copy construction bug with "return this;" - Issue 4499 - calls to @disabled postblit are emitted Struct literal/constructor call should be rvalue. - Issue 5178 - StructLiteral should not be lvalue - Issue 5769 - struct elaborate constructor should make rvalue
Comment #2 by k.hara.pg — 2011-03-25T19:27:32Z
Created attachment 935 3 Patches These patches has 3 points. 1. Allow NRVO for user-defined destructor object. 2. Eliminate destructor call of return value that non-copyable object in NRVO function. 3. On assignment (e1 = e2) if e2 is non-copyable and prvalue, this expression generates following code: e1.__dtor(); e1 = e2; // bit copy This is much experimental, but I hope that it helps the language evolution.
Comment #3 by k.hara.pg — 2011-04-13T10:38:46Z
On trunk dmd, this code cannot compile because of fixing bug4499. ---- import std.algorithm; struct S { @disable this(this){} ~this(){} } void main() { S s1; S s2 = move(s1); } ----
Comment #4 by k.hara.pg — 2011-05-05T06:15:51Z
Created attachment 963 2 patches Update patches. Changes: Behavior changes on assignment (Comment#3 3.) makes bug, so I remove it.
Comment #5 by bugzilla — 2011-07-16T12:58:17Z