Re: [Squeak-VMdev] Versiojn 4 changes

[Bryce K. <br...@ka...>]

Ned Konz writes:
 > On a related note, does it seem wasteful to anyone but me that we do the 
 > following in primBytecodeAdd:

Yes, that sequence does seem wasteful.

Ideally I'd like to generate:

	and arg1, arg2, andResult
	and andResult, 2r11, dummy1
	branch-if-not-zero overflowBlock

	add addResult, arg1, arg2
	branch-on-overflow overflowBlock

This sequence is possible if we change from integers having tag 1 to
integers having tag 0 leaving the tag in the lower bits. The Self
community wrote some nice papers on tagging and high performance
simple garbage collection.

The above sequence only has the add on the critical path during the
common case where the arguments are SmallIntegers and the result is
also a SmallInteger. The other four instructions do not matter so long
as both branches are correctly predicted as not taken.

Looking up oops can be done by using the small offset provided in most
load or store instructions. This means that a non-zero oop tag has no
overhead.

But interpreted code probably spends most of it's time recovering from
branch mispredicts. Compiled code should be sufficiently faster that
the garbage collector will be the major bottle-neck. Realistically any
changes to tagging should be looking to provide a large improvement to
garbage collector performance.
	
It would probably be more worthwhile looking at at: and at:put:. Could
Tim's improvements to primitives be used to create simple at:
primitives that are quick to execute?

My limited benchmarking indicates that code using arrays is dominated
by at: and at:put: overhead. We're not yet at the point where our
inefficient SmallInteger format matters. Once we're at the point where
SmallInteger addition is worth improving our garbage collector will be
a serious bottleneck. Currently the garbage collector uses about 10%
of the time reduce the 90% by enough to make these few instructions
matter and garbage collection will be closer to 50%.

So, yes it does seem wasteful, but first let's improve things so that
the waste really matters. Even then other things (the garbage
collector) will be more important which require image
changes. Anything that simplifies and speeds up at: and at:put:
will matter much more. Tim, that primitive work?

Bryce