Interface method calls are slow.

[Godmar Back]

If time is a concern, implementing the faster interface dispatch Per
suggested could be done in two steps. 

The first step would be to implement the necessary data structures
in the run-time system, but not inline invocations yet.

That's what we do in Kaffe.  For Bryce's benchmark,
I'm seeing on a different machine:

IBM:	interface:        932ms [93ns / call]
Kaffe:	interface:       2143ms [214ns / call]

So it's about 2.3 times slower compared to 3.9 times.
In other words, even w/o inlining, a constant-time lookup 
scheme helps.

	- Godmar

> 
> I wrote a small benchmark to time different types of method calls. The code is
> available from http://waitaki.otago.ac.nz/~bryce/gcj/meth-inv-bench.tar.gz
> 
> The IBM JDK 1.1.8 gets:
> 
> no call:          181ms [18ns / call]
> local:            287ms [28ns / call]
> private:          163ms [16ns / call]
> interface:       1506ms [150ns / call]
> i/f after cast:   324ms [32ns / call]
> abstract:         951ms [95ns / call]
> abstract w/cast:  316ms [31ns / call]
> native:          4645ms [464ns / call]
> 
> While gcj (compiled from source with -O3) gets:
> 
> no call:           97ms [9ns / call]
> local:            734ms [73ns / call]
> private:          606ms [60ns / call]
> interface:       5890ms [589ns / call]
> i/f after cast:   688ms [68ns / call]
> abstract:         686ms [68ns / call]
> abstract w/cast:  688ms [68ns / call]
> native:           683ms [68ns / call]
> 
> I'm not too concerned about the differences between gcj and the VM here -
> obviously IBM's JITC is doing a lot of clever optimization that doesn't
> neccessarily apply to real-world situations, and maybe cannot even be done by an
> ahead-of-time compiler.
> 
> BUT...
> 
> Why are interface calls so expensive compared to other calls in gcj? I remember
> that there was talk a while back about changing the way interface calls are
> compiled by gcj - were these change ever made? Slow interface calls will have a
> significant impact on the performance of AWT, where every call to a component's
> native peer is sent through the interfaces in java.awt.peer. Many other newwer
> Java APIs are fairly interface-dependent as well. I am using a current libgcj.
> 
> It also appears that the private method call is not being inlined. I thought
> that gcj was inlining private methods now?
> 
> regards
> 
>   [ bryce ]
> 
>