D issues are now tracked on GitHub. This Bugzilla instance remains as a read-only archive.
Issue 6194 - [GSoC] Destructor gets called on object before it is copied when calling writeln()
Summary: [GSoC] Destructor gets called on object before it is copied when calling writ...
Status: RESOLVED FIXED
Alias: None
Product: D
Classification: Unclassified
Component: dmd (show other issues)
Version: D2
Hardware: All All
: P2 major
Assignee: No Owner
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2011-06-22 08:46 UTC by Cristi Cobzarenco
Modified: 2011-07-31 13:44 UTC (History)
2 users (show)

See Also:


Attachments

Note You need to log in before you can comment on or make changes to this issue.
Description Cristi Cobzarenco 2011-06-22 08:46:01 UTC
import std.stdio;
import std.conv;

struct Test {
	int x;
	
	this( int m )     { x = m; }
	this( this )      { writeln( "Postblit: ", x ); }
	
	~this()           { x = 42; }
	
	string toString() { return to!string( x ); }
}

int main(string[] argv) {
	auto a = Test(3);
	writeln( a );
	return 0;
}

When the running the code above, 42 gets printed instead of 3. Looking at the "Postblit: " writes, one can see that the postblit ctor gets called 6 times and only the last copy sets x to 42. It seems that at the last copy the destructor gets called before the object gets blitted.

Using a debugger, I think I found the offending copy at line 1599 in file format.d. Strangely, as much as I tried to reproduce the same kind of copying, I could never get the bug to appear without using writeln.
Comment 1 kennytm 2011-06-22 09:51:21 UTC
Reduced:

---------------------------------------
struct Test {
    int x;
    ~this() { x = 42; }
}
void formattedWrite(A...)(A args) {
    Test* argsAddresses;
    foreach (a; args) {
        argsAddresses = &a;
        assert(argsAddresses.x != 42, "line 9: unexpected: x == 42");
    }
    assert(argsAddresses.x != 42, "line 11: unexpected: x == 42");
}
int main(string[] argv) {
    auto a = Test(3);
    formattedWrite(a);
    return 0;
}
---------------------------------------
core.exception.AssertError@y.d(11): line 11: unexpected: x == 42
----------------
5   y                                   0x000087e9 onAssertErrorMsg + 73
6   y                                   0x00011be6 _d_assert_msg + 26
7   y                                   0x00001ab6 void y.formattedWrite!(y.Test).formattedWrite(y.Test) + 138
8   y                                   0x00001944 _Dmain + 24
9   y                                   0x000122e3 extern (C) int rt.dmain2.main(int, char**).void runMain() + 23
10  y                                   0x00011e8d extern (C) int rt.dmain2.main(int, char**).void tryExec(scope void delegate()) + 29
11  y                                   0x0001232b extern (C) int rt.dmain2.main(int, char**).void runAll() + 59
12  y                                   0x00011e8d extern (C) int rt.dmain2.main(int, char**).void tryExec(scope void delegate()) + 29
13  y                                   0x00011e27 main + 179
14  y                                   0x00001921 start + 53
----------------
---------------------------------------
Comment 2 kennytm 2011-06-22 10:33:34 UTC
I'm not sure if it's a bug in DMD or Phobos. I'm inclined to it's Phobos bug. The loop variable 'a' in the reduced 'formattedWrite' is obviously escaping its scope, so calling the destructor is reasonable. However, this is essentially the implementation of std.format.formattedWrite:

1. the arguments are taken address 

    foreach (i, arg; args)
    {
        funs[i] = &formatGeneric!(Writer, typeof(arg), Char);
        // We can safely cast away shared because all data is either
        // immutable or completely owned by this function.
        argsAddresses[i] = cast(const(void*)) &arg;
    }

2. and then they are referred later for actual formatting.

        {
            funs[currentArg](w, argsAddresses[currentArg], spec);
            ++currentArg;
        }

https://github.com/D-Programming-Language/phobos/blob/master/std/format.d#L302
Comment 3 Cristi Cobzarenco 2011-06-24 04:23:28 UTC
(In reply to comment #2)
> I'm not sure if it's a bug in DMD or Phobos. I'm inclined to it's Phobos bug.
> The loop variable 'a' in the reduced 'formattedWrite' is obviously escaping its
> scope, so calling the destructor is reasonable. However, this is essentially
> the implementation of std.format.formattedWrite:
> 
> 1. the arguments are taken address 
> 
>     foreach (i, arg; args)
>     {
>         funs[i] = &formatGeneric!(Writer, typeof(arg), Char);
>         // We can safely cast away shared because all data is either
>         // immutable or completely owned by this function.
>         argsAddresses[i] = cast(const(void*)) &arg;
>     }
> 
> 2. and then they are referred later for actual formatting.
> 
>         {
>             funs[currentArg](w, argsAddresses[currentArg], spec);
>             ++currentArg;
>         }
> 
> https://github.com/D-Programming-Language/phobos/blob/master/std/format.d#L302


This works for me:
 foreach (i, arg; args)
    {
        funs[i] = &formatGeneric!(Writer, typeof(arg), Char);

        static if(hasAliasing!(typeof(arg)))
        {
-            argsAddresses[i] = &arg;
+            argsAddresses[i] = &args[i];

        }
        else
        {
            // We can safely cast away shared because all data is either
            // immutable or completely owned by this function.
-            argsAddresses[i] = cast(const(void*)) &arg;
+            argsAddresses[i] = cast(const(void*)) &args[i];
        }
    }

Unfortunately we can't replace the foreach with a for loop because we can't do typeof( args[ i ] ). Does any one have any idea how we could avoid copying the arguments needlessly?
Comment 4 Cristi Cobzarenco 2011-07-19 06:24:20 UTC
Here's a pull request with my fix.
https://github.com/D-Programming-Language/phobos/pull/151