EndOfStream - cruft?

Well, I dug around a bit more. Seems Andreas was concerned about performance, although I did not see an issue. A week later it was rolled back.

Also in question was the correct superclass for EndOfStream. I just made the change in a 4.2 image and TestRunner is not happy with it being an Error. Make it an Exception and that problem goes away.

Cheers,
Bob

5/24/00     Re: RE: EndOfStream an Error???

On Wed, 24 May 2000 01:32:44 -0700 "Raab, Andreas" [hidden email] wrote:
>As it is right now, we're paying the
>price of an error recovery for something that happens quite often. There are
>many places where senders of #next rely returning nil quickly if at the end
>of the collection. Such as in #next:into:startingAt:, #nextChunk,
>#nextDelimited:, #peekFor: etc. With the changes we have a slowdown of
>approx. 500 for a simple test case like
>
>    | s |
>    s _ ReadStream on: Array new.
>    [1 to: 100000 do:[:i| s next]] timeToRun.
>
>While the slowdown is okay for an error recovery it is most definitely not
>okay for those users who do in fact rely on the opposite - which is a quick
>return of nil for a past end read operation. And as it stands ALL those
>users are affected by the modification.

Andreas,

Do you have an example that's a little more realistic? I was wondering what the effect would be on typical code, so I made a change to count the number of times that EndOfStream was actually signalled. Then I browsed around for a while and when I checked the number, it was still zero! So, I thought, filing in some code ought to do it. Nope - still zero. It wasn't until I evaluated "(ReadStream on: '1') nextChunk" that I actually got a hit. Hmm...

Cheers,
Bob
======================
5/24/00    Re: RE: EndOfStream an Error???

On Wed, 24 May 2000 12:32:32 -0700 "Raab, Andreas" [hidden email] wrote:
>What I do know, however, is that not too long ago we've integrated a change
>set (I think from Bert) which modified quite some places to use the
>#next/#isNil test instead of #atEnd. Which leads me to the assumption that
>there ought to be places in which a quick return might be really critical.

I think there is an important distinction here. If you have code like:

[s atEnd] whileFalse: [el _ s next. el whatEver]

then you will be sending #atEnd for each element in the stream and sending #next one less time. By changing that to

[(el _ s next) isNil] whileFalse: [el whatEver]

you are essentially substituting #isNil for #atEnd for every element. That can add up to substantial savings. The change to EndOfStream changes only the performance of the *last* #next while all the others (assuming there are elements in the stream) remain as fast as ever. There will clearly be some performance penalty, but the magnitude will depend on the number of elements processed before end of stream. If the typical case involves empty streams, then the cost will be quite high. If the typical stream processed as above has tens or hundreds (or more) of elements, then the cost of EndOfStream will be a much smaller fraction of the total processing time.

I think, as has already been suggested, it might be wise to back off this change until we resolve whether Error or Exception is a better place for it (I lean toward Exception, Notification or perhaps ErrorsForWhichTheDefaultActionIsUsuallyAcceptable rather than Error). Once that is decided, we can see if there are any noticeable performance hits, especially once EndOfStream is added to the other places where it can be detected (StandardFileStream>>next, e.g.).

Cheers,
Bob

==============================================

On 7/22/13 2:31 PM, Chris Muller wrote:

Interesting, I wonder what the consequences were..

On Mon, Jul 22, 2013 at 11:22 AM, Bob Arning [hidden email] wrote:

For the historically-obsessed,

Change Set:        069EOS-hh (really2192EOS-hh)
Date:            18 May 2000
Author:            Helge Horch
Requires:        Squeak 2.8a

Modifies ReadStream>>next to signal an Error (EndOfStream) on attempts to
read
at or beyond the end of the stream.  Thanks to Bob Arning (who is silently
credited
in the stamps <g>), existing applications relying on seeing nils should not
be
affected.

Change Set:        102ReadStreamUndo-sma (really 2272ReadStreamUndo-sma)
Date:            1 June 2000
Author:            Stefan Matthias Aust

Undo #2192. The consequences of raising an exception for EOF during
ReadStream>>next should be worked out in 2.9a but not for 2.8.


and, AFAIK, that was the end of that.

Cheers,
Bob

On 7/22/13 12:09 PM, Chris Muller wrote:

What if a system needs to allow nil objects in a Stream of objects and
have it be treated as just another object in the stream rather than
the EOS marker?

So, maybe ReadStream>>#next _should_ be signaling EndOfStream, so the
sender of #next can know the difference in that scenario.  Since the
defaultAction is ^nil, presumably we could simply change #next to do
that with no breakage...

On Mon, Jul 22, 2013 at 9:02 AM, Frank Shearar [hidden email]
wrote:

EndOfStream has no references in the base image. It claims to be used
by ReadStream >> #next, but that's not true.

Time to die/deprecate?

frank

Given that there might be a performance issue in some (rare) cases, but especially because of the potential for the wrong handler seeing the exception, another approach might be more acceptable. Some possibilities:

=======================
next

    <primitive: 65>
    position >= readLimit
        ifTrue: [^eos]    "a new instance variable the user can leave nil or set to preferred marker"
        ifFalse: [^collection at: (position := position + 1)]
=======================
nextIfNone: aBlock

    <primitive: 65>
    position >= readLimit
        ifTrue: [^aBlock value]
        ifFalse: [^collection at: (position := position + 1)]

=======================
nextOrEOS: endOfStreamMarker

    <primitive: 65>
    position >= readLimit
        ifTrue: [^endOfStreamMarker]
        ifFalse: [^collection at: (position := position + 1)]
=======================
nextOrSelf

    <primitive: 65>
    position >= readLimit
        ifTrue: [^self]    "if the stream is known not to contain itself, then self could be the marker"
        ifFalse: [^collection at: (position := position + 1)]
=======================

Cheers,
Bob

On 7/22/13 8:59 PM, Levente Uzonyi wrote:

On Mon, 22 Jul 2013, Frank Shearar wrote:

I'm half-tempted to say "in that case the user should use Xtreams".
Because the weird behaviours around #next and #atEnd were pretty much
the main drivers behind it.

Do you use EndOfStream like this (to allow nil objects in a Stream of objects)?

Your question is misleading. nil can be in any stream, but #next will also return nil, when the stream is #atEnd. If you know that your stream doesn't have nils in it, then using #next is safe (and fast), otherwise you should not rely on #next's return value to decide if it was a valid read or not. That's what #atEnd is for.

Signaling an Exception when a ReadStream is #atEnd should not really cause performance issues. It can only be a problem, when you are repeatedly reading from short streams. Xtreams also uses this technique.


Levente

frank

On 22 July 2013 17:09, Chris Muller [hidden email] wrote:

What if a system needs to allow nil objects in a Stream of objects and
have it be treated as just another object in the stream rather than
the EOS marker?

So, maybe ReadStream>>#next _should_ be signaling EndOfStream, so the
sender of #next can know the difference in that scenario.  Since the
defaultAction is ^nil, presumably we could simply change #next to do
that with no breakage...

On Mon, Jul 22, 2013 at 9:02 AM, Frank Shearar [hidden email] wrote:

EndOfStream has no references in the base image. It claims to be used
by ReadStream >> #next, but that's not true.

Time to die/deprecate?

frank