[gclist] Precise GC's performance

[David Chase]

I think you misunderstood the purpose of the experiment. The aim 
was to measure the cost of the additional code required, in a C-generating 
"compiler", to *enable* precise collection. All else is being held 
equal -- the GC is not running, so there is no GC cost, and the GC 
is not running, so the VM will grow to roughly the same size in both 
experiments (bigger text in the precise code, but not by much).  
The inflated VM may artificially increase both running times (thus 
enlarging the base against which a relative slowdown is measured), 
but life is hard, and this is (as I understand it) a relatively informal 
experiment that could still yield interesting numbers. 

This will also measure the cost of missed optimizations, if the 
precise-pointer-tracking code has the side-effect of inhibiting register 
allocation and/or forcing the optimizer to make conservative assumptions 
about aliasing.

yours,

David