[LLVMdev] Explicitly Managed Stack Frames
sabre at nondot.org
Wed Jan 11 22:42:02 PST 2006
On Wed, 11 Jan 2006, Ben Chambers wrote:
> I was wondering what the current state of this (explicitly managed stack frames)
> is. Is it being worked on? If not, how hard do you think it would be for me to
> add it?
I'm not sure what you mean. It depends on correct tail calls, but no
other LLVM-level support.
> I am more than willing to work on it, but don't have any experience
> with LLVM, so it might take a while. The reason I ask is because I am starting
> work on a project to make a language similar to ML with which to experiment, and
> hope to build it upon LLVM. Unfortunately, in order to implement things like
> lexical scoping and closures, I need to be able to not only access variables on
> the frame but also pass in a static link.
Ok. Are you planning on garbage collecting your stack frames, or do you
just need a static link? If you just need a static link (because of
nested functions like pascal) but don't need GC'd stack frames, there is
no need even for tail call support and CPS and other craziness, just pass
your static link or display as an argument: lower to LLVM just like you
would lower to machine code.
> As far as I can tell from reading the notes on the topic, the static link could
> be resolved by simply passing a pointer to the appropriate stack frame, so this
> isn't a serious issue.
See above: if all you need is a static link, you don't need anything fancy
> Also, if the continuation passing scheme were to be employed, how many new
> optimizations would have to be written to make it actually efficient? Is the
> only necessary change transforming the code to use CPS, or would other passes
> have to be written to make it efficient? How badly would CPS break the other
> optimizations that LLVM performs?
It is unclear. Noone has ported a performant language to use LLVM in this
way. My guess (from similar other experiences) is that it will be
relatively easy to get the first 80% of performance and relatively hard to
get the last 20%. It seems that's how things always go :)
> Are there ane resources on CPS that provide more insight? Most of what I could
> find online only talked about it from a theoretical perspective and did a lot of
> hand waving on the actual specifics.
I don't know of anything off-hand, but there are books that talk about it
and google can probably dig some stuff up.
More information about the llvm-dev