libera/#sicl - IRC Chatlog
Search
14:49:25
pjb
And if you have continuation you will want to copy the PC from the hardware stack to the activation records anyways, so…
16:16:44
pjb
beach: it's a problem only for the allocation: we need to have a part of the heap that can be allocated as a stack with minimal overhead, but that can later be garbage collected as usual (and with the frame possibly moved elsewhere if needed).
16:17:48
pjb
So instead of using a link instruction that reserves space on the processor SP, you'd do that on the heap frame stack pointer.
16:19:49
pjb
It could be less, even since on return you wouldn't need to unlink, you'd just let the garbage collector do its work.
18:31:13
pjb
beach: for short-lived processes, it's basically 0 (no GC needed). ;-) it's only a concern for long-lived processes, or processes using a lot of memory (ie. probably not short-lived)... Of course, this answer assume something like unix :-/
22:52:16
hayley
Cliff replied, saying "At this point, It All Depends. I tweaked the Java inlining heuristics in tandem with the register allocator, and a lot of dedicated profiling. In the end, a mix of callee and caller-save registers was best... BUT it was a lot more important that the allocator not spill excessively when over-inlining. i.e. the cost of the extra spills had to be close to the cost of the callee/caller saves you would have paid anyhow if you had