02-Jun-2013 22:54, Walter Bright пишет:
On 6/2/2013 10:58 AM, Dmitry Olshansky wrote:
Anyhow cramming the whole interpreter in one huge function too has
downsides
(and hopelessly inefficent register allocation is one).
This should not be the case with register allocation using live ranges.
The problem is that if every goto *get_me_new_opcode compile may have no
idea of where it will through it with computed labels. Hence some
conservativeness of what registers contain and reloads and in effect all
labels end up as isolated "functions". Plus a big stack frame is liability.
And then there is a fair share of obstacles in making the kind of
threaded interpreter code optimized properly, this post I find interesting:
http://article.gmane.org/gmane.comp.lang.lua.general/75426
You can also help with this by organizing declarations so the variables
have the shortest scope.
Then splitting opcodes as functions separates their timespan even more.
--
Dmitry Olshansky
