Still works on 2.064. This bug in particular makes it impossible to sort an array of SysTimes or an array of structs containing SysTime. Not sure if the root cause lies in a Phobos change or in a DMD change.
---
import std.algorithm;
void main() {
SysTime a, b;
swap(a, b);
}
---
bug_sort.d(4): Error: template std.algorithm.swap cannot deduce function from argument types !()(SysTime, SysTime), candidates are:
C:\D\dmd2\windows\bin\..\..\src\phobos\std\algorithm.d(1997): std.algorithm.swap(T)(ref T lhs, ref T rhs) if (allMutableFields!T && !is(typeof(T.init.proxySwap(T.init))))
C:\D\dmd2\windows\bin\..\..\src\phobos\std\algorithm.d(2042): std.algorithm.swap(T)(T lhs, T rhs) if (is(typeof(T.init.proxySwap(T.init))))
Comment #1 by sludwig — 2014-01-29T03:02:55Z
"import std.algorithm;" should read "import std.algorithm, std.datetime;"
Comment #2 by dlang-bugzilla — 2014-01-29T03:21:30Z
> Caused by https://github.com/D-Programming-Language/phobos/pull/1603
Indeed. However, as I explained there, I think it is legit behavior:
SysTime contains a rebindable, and Rebindable contains an immutable, and swap can't make the choice to mutate immutable.
That said, the issue could be in Rebindable to being with: It stores an immutable member, but obviously mutates it all the time. In that case, why bother storing an immutable at all? I think the fix is there.
Comment #4 by dlang-bugzilla — 2014-01-29T04:14:37Z
> SysTime contains a rebindable, and Rebindable contains an immutable, and swap
can't make the choice to mutate immutable.
Actually, it contains a union between mutable and immutable. And unions imply that all guarantees are off.
So, maybe the solution would be to make swap allow an immutable field if it's in an union with a mutable field?
Comment #5 by monarchdodra — 2014-01-29T04:41:14Z
(In reply to comment #4)
> Actually, it contains a union between mutable and immutable. And unions imply
> that all guarantees are off.
>
> So, maybe the solution would be to make swap allow an immutable field if it's
> in an union with a mutable field?
There is currently (AFAIK) no way to introspect that a member is part of an anonymous union. It *could* make for a partial solution though, yes.
Also, aren't there cases where an union contains *all* const members?
In any case, yes, there is room for improvement.
Comment #6 by dlang-bugzilla — 2014-01-29T04:59:47Z
Well, the regression needs to be solved one way or another. Not being able to sort an array of timestamps is pretty bad.
Comment #7 by monarchdodra — 2014-01-29T05:18:11Z
(In reply to comment #6)
> Well, the regression needs to be solved one way or another.
Absolutly.
> Not being able to sort an array of timestamps is pretty bad.
If your original issue is that of sorting timestamps, then you may also be interested in:
https://d.puremagic.com/issues/show_bug.cgi?id=11129
Because timestamps *are* assignable, so the fact that the elements aren't swappable should not be preventing you from sorting them.
I think it was when I was originally looking into that issue that I introduced that check. Small incremental improvements I guess.
Comment #8 by yebblies — 2014-01-29T05:20:31Z
(In reply to comment #5)
>
> There is currently (AFAIK) no way to introspect that a member is part of an
> anonymous union. It *could* make for a partial solution though, yes.
>
It should be possibly by comparing .offsetof with other members.
Comment #9 by k.hara.pg — 2014-01-29T05:56:38Z
(In reply to comment #3)
> > Caused by https://github.com/D-Programming-Language/phobos/pull/1603
>
> Indeed. However, as I explained there, I think it is legit behavior:
>
> SysTime contains a rebindable, and Rebindable contains an immutable, and swap
> can't make the choice to mutate immutable.
>
> That said, the issue could be in Rebindable to being with: It stores an
> immutable member, but obviously mutates it all the time. In that case, why
> bother storing an immutable at all? I think the fix is there.
If a non-mutable field has one or more overlapped union _mutable_ fields, the whole struct is treated as modifiable.
import std.traits : Unqual;
struct Rebindable(T)
{
union
{
T origin;
Unqual!T stripped; // overlapped union mutable field of 'origin'
}
}
void main()
{
Rebindable!int r1 = {origin:10};
Rebindable!int r2 = {origin:20};
r1 = r2; // Rebindable!int is modifiable (assignable)
// even if non-mutable field `T origin;` exists.
assert(r1.origin == 20);
}
If all of overlapped union fields are non-mutable, the whole struct is not also modifiable.
struct S
{
union
{
immutable int x;
immutable int y;
}
}
void main()
{
S s1 = {x:10};
S s2 = {x:20};
s1 = s2; // cannot modify struct s1 S with immutable members
}
Comment #10 by dlang-bugzilla — 2014-01-29T06:05:44Z
Interesting. So this should "just work":
diff --git a/std/algorithm.d b/std/algorithm.d
index 036b918..0ed5606 100644
--- a/std/algorithm.d
+++ b/std/algorithm.d
@@ -2054,6 +2054,9 @@ void swap(T)(ref T lhs, ref T rhs) if (is(typeof(lhs.proxySwap(rhs))))
private template allMutableFields(T)
{
alias OT = OriginalType!T;
+ static if (is(typeof({ T t = void; t = t; })))
+ enum allMutableFields = true;
+ else
static if (is(OT == struct) || is(OT == union))
enum allMutableFields = isMutable!OT && allSatisfy!(.allMutableFields, FieldTypeTuple!OT);
else
@@ -2072,6 +2075,9 @@ unittest
struct S2{const int i;}
static assert( allMutableFields!S1);
static assert(!allMutableFields!S2);
+
+ struct S3{union X{int m;const int c;}X x;}
+ static assert( allMutableFields!S3);
}
unittest
Comment #11 by dlang-bugzilla — 2014-01-29T06:07:02Z
Derp, unit test line should be:
struct S3{union{int m;const int c;}}
Comment #12 by monarchdodra — 2014-01-29T06:26:10Z
(In reply to comment #8)
> It should be possibly by comparing .offsetof with other members.
Smart. Unfortunatly, for a "recursive" implementation, it is a bit difficult to exploit: More often than not, you want to know before hand that you are about to process an aggregate in an union.
(In reply to comment #9)
> If a non-mutable field has one or more overlapped union _mutable_ fields, the
> whole struct is treated as modifiable.
>
> import std.traits : Unqual;
> struct Rebindable(T)
> {
> union
> {
> T origin;
> Unqual!T stripped; // overlapped union mutable field of 'origin'
> }
> }
Yes, absolutely. That's what I had in mind. But as I said, even detecting that you are in an (anonymous) union is a bit difficult. Do-able with your suggestion, just... difficult.
(In reply to comment #10)
> Interesting. So this should "just work":
>
> diff --git a/std/algorithm.d b/std/algorithm.d
> index 036b918..0ed5606 100644
> --- a/std/algorithm.d
> +++ b/std/algorithm.d
> @@ -2054,6 +2054,9 @@ void swap(T)(ref T lhs, ref T rhs) if
> (is(typeof(lhs.proxySwap(rhs))))
> private template allMutableFields(T)
> {
> alias OT = OriginalType!T;
> + static if (is(typeof({ T t = void; t = t; })))
> + enum allMutableFields = true;
> + else
> static if (is(OT == struct) || is(OT == union))
> enum allMutableFields = isMutable!OT && allSatisfy!(.allMutableFields,
> FieldTypeTuple!OT);
> else
> @@ -2072,6 +2075,9 @@ unittest
> struct S2{const int i;}
> static assert( allMutableFields!S1);
> static assert(!allMutableFields!S2);
> +
> + struct S3{union X{int m;const int c;}X x;}
> + static assert( allMutableFields!S3);
> }
I think that's wrong, because "static if (is(typeof({ T t = void; t = t; })))" can work in the pressence of an opAssign: Complex structs with immutable members but with a valid opAssign can't be swapped. The only case where it would work is if T didn't have an opAssign to begin with. However, that test would be mostly useless, since we'd have to deploy code for structs that do have opAssign regardless.
Comment #13 by k.hara.pg — 2014-01-29T06:36:51Z
(In reply to comment #12)
> (In reply to comment #8)
> > It should be possibly by comparing .offsetof with other members.
>
> Smart. Unfortunatly, for a "recursive" implementation, it is a bit difficult to
> exploit: More often than not, you want to know before hand that you are about
> to process an aggregate in an union.
>
> (In reply to comment #9)
> > If a non-mutable field has one or more overlapped union _mutable_ fields, the
> > whole struct is treated as modifiable.
> >
> > import std.traits : Unqual;
> > struct Rebindable(T)
> > {
> > union
> > {
> > T origin;
> > Unqual!T stripped; // overlapped union mutable field of 'origin'
> > }
> > }
>
> Yes, absolutely. That's what I had in mind. But as I said, even detecting that
> you are in an (anonymous) union is a bit difficult. Do-able with your
> suggestion, just... difficult.
I think adding std.traits.isBlitAssignable(T) for the purpose would be good.
And you can refer the code in the compiler.
https://github.com/D-Programming-Language/dmd/blob/master/src/mtype.c#L8588
int TypeStruct::isAssignable()