[Newcompiler] Optimizing compiled methods for performance [was: A day optimizing]

[bryce at kampjes.demon.co.uk]

Mathieu Suen writes:
 > What is tag checking?
 > 	Mth

Tag checking means checking that the argument is a SmallInteger.
SmallIntegers have 1 as their lowest bit while pointers to objects
have 0 as their lowest bit so to check an argument is an integer we
check to see if it's lowest bit is 1. To detag means to remove the 1
tag bit.

The fast case for a general SmallInteger addition is:

  bitTest arg1 0
  jumpIfUnset failureBlock
  sub 1 arg1
  bitTest arg2 0
  jumpIfUnset failureBlock
  sub 1 arg2
  add arg1 arg2
  jumpOverflow failureBlock
  add 1 arg2

Both arg1 and arg2 would need to be in registers that are not used
after the bytecode otherwise move instructions are nessisary.

It's possible to remove one of the subs used to detag and the
add at the end which retags. This is because the tag will transfer
through the addition.

So to add a 1 (which tagged is 3) constant just requires:

  bitTest arg1 0
  jumpIfUnset failureBlock
  add 3 arg1
  jumpOverflow failureBlock

If either argument is not a SmallInteger or the addition overflows
a full message send is performed which will deal with those cases.

When interpreting the branch mispredict overhead from decoding the
bytecodes is likely to dominate so the efficiency of the addition
sequence doesn't matter so long as it's tolerable. The VM uses a less
efficient method of checking for oveflow but it probably doesn't
matter on desktop out of order CPUs.

Bryce