I'm pretty sure this used to work:

import std.algorithm;

int[] ra;
copy(ra[5 .. 10], ra[4 .. 9]);

and would correctly shift the range [5 .. 10] down one
(unlike ra[4 .. 9] = ra[5 .. 10], which is prohibited by spec)

In dmd 2.057 of course it doesn't work; is this a bug or should I be looking elsewhere for this functionality?

On Saturday, February 11, 2012 12:09:32 Ellery Newcomer wrote:
> I'm pretty sure this used to work:
> 
> import std.algorithm;
> 
> int[] ra;
> copy(ra[5 .. 10], ra[4 .. 9]);
> 
> and would correctly shift the range [5 .. 10] down one (unlike ra[4 .. 9] = ra[5 .. 10], which is prohibited by spec)
> 
> In dmd 2.057 of course it doesn't work; is this a bug or should I be looking elsewhere for this functionality?

copy uses array copying if it can, which is what I would expect it to do. In fact, a comment in copy's source states that doing so results in a 10x - 20x improvement in speed over the generic implementation. So, I would say that it's safe to say that you can't use copy to copy overlapping arrays. I'm not even sure it's a good idea to use overlapping ranges. I haven't dealt with output ranges enough to say without studying up on them again.

- Jonathan M Davis

On Sat, Feb 11, 2012 at 10:30:43AM -0800, Jonathan M Davis wrote:
> On Saturday, February 11, 2012 12:09:32 Ellery Newcomer wrote:
> > I'm pretty sure this used to work:
> > 
> > import std.algorithm;
> > 
> > int[] ra;
> > copy(ra[5 .. 10], ra[4 .. 9]);
> > 
> > and would correctly shift the range [5 .. 10] down one (unlike ra[4 .. 9] = ra[5 .. 10], which is prohibited by spec)
> > 
> > In dmd 2.057 of course it doesn't work; is this a bug or should I be looking elsewhere for this functionality?
> 
> copy uses array copying if it can, which is what I would expect it to do. In fact, a comment in copy's source states that doing so results in a 10x - 20x improvement in speed over the generic implementation. So, I would say that it's safe to say that you can't use copy to copy overlapping arrays. I'm not even sure it's a good idea to use overlapping ranges. I haven't dealt with output ranges enough to say without studying up on them again.
[...]

This brings up an interesting point: what does the GC do if you have an array that's continually appended to, and also shrunk from the front? That is:

	ubyte[] buf;
	while ( ... ) {
		consumeData(buf[0]);
		buf = buf[1 .. $];

		ubyte x = getMoreData();
		buf ~= x;
	}

Will such a program "leak memory" in the sense that the memory chunk allocated to buf will grow larger and larger, even though its initial segment is never accessed again? Or is the GC smart enough to only reallocate the needed size when the memory chunk is moved (at some point when it has no more space to append x)?

If the GC is smart enough, then this kind of construct could be used instead of copying overlapping ranges.


T

-- 
If lightning were to ever strike an orchestra, it'd always hit the conductor first.

On Saturday, February 11, 2012 10:51:36 H. S. Teoh wrote:
> This brings up an interesting point: what does the GC do if you have an array that's continually appended to, and also shrunk from the front? That is:
> 
> 	ubyte[] buf;
> 	while ( ... ) {
> 		consumeData(buf[0]);
> 		buf = buf[1 .. $];
> 
> 		ubyte x = getMoreData();
> 		buf ~= x;
> 	}
> 
> Will such a program "leak memory" in the sense that the memory chunk allocated to buf will grow larger and larger, even though its initial segment is never accessed again? Or is the GC smart enough to only reallocate the needed size when the memory chunk is moved (at some point when it has no more space to append x)?
> 
> If the GC is smart enough, then this kind of construct could be used instead of copying overlapping ranges.

The memory blocks don't grow. When an array no longer has enough room to increase its size, and you append to it, a new block is allocated, and the array is moved to there. Any slices which refer to the original block still refer to it. And when a garbage collection cycle is run and a memory block has no references to it, it's collected. The only "leaking" that occurs is that once no more slices refer to a particular portion of a memory block, that portion is no longer accessible until the block has been recycled.

If you haven't read this article yet, you really should:

http://www.dsource.org/projects/dcollections/wiki/ArrayArticle

- Jonathan M Davis

On Sat, Feb 11, 2012 at 11:02:43AM -0800, Jonathan M Davis wrote:
> On Saturday, February 11, 2012 10:51:36 H. S. Teoh wrote:
> > This brings up an interesting point: what does the GC do if you have an array that's continually appended to, and also shrunk from the front? That is:
> > 
> > 	ubyte[] buf;
> > 	while ( ... ) {
> > 		consumeData(buf[0]);
> > 		buf = buf[1 .. $];
> > 
> > 		ubyte x = getMoreData();
> > 		buf ~= x;
> > 	}
> > 
> > Will such a program "leak memory" in the sense that the memory chunk allocated to buf will grow larger and larger, even though its initial segment is never accessed again? Or is the GC smart enough to only reallocate the needed size when the memory chunk is moved (at some point when it has no more space to append x)?
> > 
> > If the GC is smart enough, then this kind of construct could be used instead of copying overlapping ranges.
> 
> The memory blocks don't grow. When an array no longer has enough room to increase its size, and you append to it, a new block is allocated, and the array is moved to there. Any slices which refer to the original block still refer to it. And when a garbage collection cycle is run and a memory block has no references to it, it's collected. The only "leaking" that occurs is that once no more slices refer to a particular portion of a memory block, that portion is no longer accessible until the block has been recycled.

I know that. What I was asking is how the GC handles memory block reallocations: does it copy the original block in its entirety, or just the part referenced by the slice being appended to.


> If you haven't read this article yet, you really should:
> 
> http://www.dsource.org/projects/dcollections/wiki/ArrayArticle
[...]

After reading this article, I think it's clear that only the portion of the block referenced by buf gets copied, correct? And it gets copied to the beginning of the new block, so there is no wasted space in the new block and the old block will simply be GC'd. Cool.


T

-- 
Error: Keyboard not attached. Press F1 to continue. -- Yoon Ha Lee, CONLANG

http://d.puremagic.com/issues/show_bug.cgi?id=7484

On 02/11/2012 12:09 PM, Ellery Newcomer wrote:
> I'm pretty sure this used to work:
>
> import std.algorithm;
>
> int[] ra;
> copy(ra[5 .. 10], ra[4 .. 9]);
>
> and would correctly shift the range [5 .. 10] down one
> (unlike ra[4 .. 9] = ra[5 .. 10], which is prohibited by spec)
>
> In dmd 2.057 of course it doesn't work; is this a bug or should I be
> looking elsewhere for this functionality?

On Sat, 11 Feb 2012 14:24:53 -0500, H. S. Teoh <hsteoh@quickfur.ath.cx> wrote:

> On Sat, Feb 11, 2012 at 11:02:43AM -0800, Jonathan M Davis wrote:
>> On Saturday, February 11, 2012 10:51:36 H. S. Teoh wrote:
>> > This brings up an interesting point: what does the GC do if you have  
>> an
>> > array that's continually appended to, and also shrunk from the front?
>> > That is:
>> >
>> > 	ubyte[] buf;
>> > 	while ( ... ) {
>> > 		consumeData(buf[0]);
>> > 		buf = buf[1 .. $];
>> >
>> > 		ubyte x = getMoreData();
>> > 		buf ~= x;
>> > 	}
>> >
>> > Will such a program "leak memory" in the sense that the memory chunk
>> > allocated to buf will grow larger and larger, even though its
>> > initial segment is never accessed again? Or is the GC smart enough
>> > to only reallocate the needed size when the memory chunk is moved
>> > (at some point when it has no more space to append x)?
>> >
>> > If the GC is smart enough, then this kind of construct could be used
>> > instead of copying overlapping ranges.
>>
>> The memory blocks don't grow. When an array no longer has enough room
>> to increase its size, and you append to it, a new block is allocated,
>> and the array is moved to there. Any slices which refer to the
>> original block still refer to it. And when a garbage collection cycle
>> is run and a memory block has no references to it, it's collected. The
>> only "leaking" that occurs is that once no more slices refer to a
>> particular portion of a memory block, that portion is no longer
>> accessible until the block has been recycled.
>
> I know that. What I was asking is how the GC handles memory block
> reallocations: does it copy the original block in its entirety, or just
> the part referenced by the slice being appended to.
>
>
>> If you haven't read this article yet, you really should:
>>
>> http://www.dsource.org/projects/dcollections/wiki/ArrayArticle
> [...]
>
> After reading this article, I think it's clear that only the portion of
> the block referenced by buf gets copied, correct? And it gets copied to
> the beginning of the new block, so there is no wasted space in the new
> block and the old block will simply be GC'd. Cool.

Yes, most likely.  However, there is one case where it doesn't -- extending pages.

If an array slice can be extended by consuming the next allocated page (this is only for page-sized blocks and above), then instead of copying the data, it will just mark the next page as part of the block, and continue on.  In this case, the block is not "copied" at all, so it cannot deallocate the no-longer referenced data.

This is probably something that can be detected.  For instance, if you are appending to a slice of length 5 and its at the end of a 10,000 element array, it makes little sense to not copy that data (allowing the original block to be freed by the GC).  But the situations are always different.

-Steve