Check: use of 'short' in TextLayoutManager?

[Andreas Delmelle <an...@telenet.be>]

Hi all

Just checking if anyone would mind terribly if the AreaInfo inner  
class in TextLayoutManager were changed to use 'int' indices rather  
than 'short'.

AFAIK, this is the only reason we currently need to split up FOText  
instances larger than 32K characters. At the time I introduced this  
fix, I opted to split the FOText instances, since I was a bit  
concerned about a rise in memory consumption. Upon closer inspection,  
and some browsing around, this seems to have far less of an impact  
than I thought...

I did a quick test, allocating space for an array of 100.000  
instances of:
- a wrapper around three short values (values Short.MAX_VALUE)
- a wrapper around three int values (values Integer.MAX_VALUE)

The results:
- 3 shorts -> 100.000 instances = (2200K - 800K) = 1600K
- 3 ints -> 100.000 instances = (2980K - 800K) = 2180K

The 800K is subtracted to account for the instance size (8 bytes *  
100K).

So three ints do not take up 100% as much space (as I initially  
expected), but only roughly between 30% and 50%...


Opinions?

Cheers

Andreas

Re: Check: use of 'short' in TextLayoutManager?

[Andreas Delmelle <an...@telenet.be>]

On Jun 20, 2008, at 09:34, Max Berger wrote:


> The JVM converts all shorts to ints before doing any arithmetic, so
> shorts are actually a little bit slower than ints.
>
> As for the size: If you have primitive values, an int takes up 4 byte,
> whereas a short takes 2 bytes (+java overhead), meaning 100.000 * 3  
> int
> rather than shorts should take up 600kbyte of memory more - hardly
> noticeable compared to fops overall memory usage.

Thanks for the info.

FYI: Change has been committed with r670422

Since TextLayoutManager now no longer duplicates the char array, the  
increase due to switching to ints is even compensated somewhat...

Greetz

Andreas

Re: Check: use of 'short' in TextLayoutManager?

[Max Berger <ma...@berger.name>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Andreas,

The JVM converts all shorts to ints before doing any arithmetic, so
shorts are actually a little bit slower than ints.

As for the size: If you have primitive values, an int takes up 4 byte,
whereas a short takes 2 bytes (+java overhead), meaning 100.000 * 3 int
rather than shorts should take up 600kbyte of memory more - hardly
noticeable compared to fops overall memory usage.

When you use the non-primitive types, java >= 1.5 uses the flyweight
pattern, meaning there is one instance for every number used, you would
only store a pointer, which has the same size for Integer and Short.

So +1 for making this change.

Max

Andreas Delmelle schrieb:
> 
> Hi all
> 
> Just checking if anyone would mind terribly if the AreaInfo inner class
> in TextLayoutManager were changed to use 'int' indices rather than 'short'.
> 
> AFAIK, this is the only reason we currently need to split up FOText
> instances larger than 32K characters. At the time I introduced this fix,
> I opted to split the FOText instances, since I was a bit concerned about
> a rise in memory consumption. Upon closer inspection, and some browsing
> around, this seems to have far less of an impact than I thought...
> 
> I did a quick test, allocating space for an array of 100.000 instances of:
> - a wrapper around three short values (values Short.MAX_VALUE)
> - a wrapper around three int values (values Integer.MAX_VALUE)
> 
> The results:
> - 3 shorts -> 100.000 instances = (2200K - 800K) = 1600K
> - 3 ints -> 100.000 instances = (2980K - 800K) = 2180K
> 
> The 800K is subtracted to account for the instance size (8 bytes * 100K).
> 
> So three ints do not take up 100% as much space (as I initially
> expected), but only roughly between 30% and 50%...
> 
> 
> Opinions?
> 
> Cheers
> 
> Andreas

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFIW12e+Gr+4pk71JwRAgWoAJwOIDLdgcYEBZQ4Tx0gPZaJe4KAXQCfc+Cs
u2CYsAOxh6HnzLzu81iOH7c=
=WlBi
-----END PGP SIGNATURE-----