[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


From: Stefan Israelsson Tampe
Subject: compile-rtl
Date: Sun, 14 Oct 2012 17:13:05 +0200


I'm right now trying to port compile-glil to compile-rtl and I would say that what's hindering me is
what design choices to take?

The main problem is that we do not have a stack pointer the same way as before. Of cause we still need
to store temporary values and we will simply have to use predefined local register indices, that's not hard to implement though.
The problem is that we will have memory not cleared that's not going to be GC:ed until the
frame is changed. This will cause some potential memory leaks. To let it be as it is designed right now, will mean that it is
very difficult looking at the scheme code to see what will be protected from gc or not. And I fear that we will need to handle
this in some smart way. The solutions are either, keep a stack pointer in a register and updated it for each push and pop
as before. Or do what we do after let forms, clear the variable slots  - expensive! What I think is
that we need to have the equivalent of a stack pointer in order to properly handle gc'ing. We could still have a register approach
and gain quite some speed bump (2x) from using direct addressing in stead of using the middle layer of the stack for everything.

My vote is introduce a register sp again!

Now when we are tail-calling in rtl, we just fill the argument slots with the new function arguments directly and then tail-call by filling in
number of arguments and function. This is very smart and just some simple features added would mean that a lot of translation
from one memory location to the other is skipped. I really like how the rtl code handle this but there is an expense. It complicates life a little
when we overwrite arguments that are used in the calculation of other arguments. I'm working on how to handle this but I just wanted to point
out how nice this design choice is.

call's, here currently we send a list of register positions to the call function and the call function itself will copy those to the argument fields. This is the opposite
of the tail-call, in a case like (f 1 3) you will create two temporary variables and copy them later on to the argument position. Here I would like to change the
semantic so that we fill in the arguments directly just like in the tail-call. Also, I would like to use the end of the stack region to compute the function frame. Well even if we use the end of the full frame we will save one move per argument which is nice.

Return values.
We talked about this before and there is some concerns how to make this fast. I would like to skip the complication by simply put the return values in the
argument positions to the function. Also the number of arguments is mediated to the reciever. I would like to skip the mv-return slot and add a rmove function

(call     pos proc)
(rmove pos (i ...))

And it is rmove's responsibility to move the arguments to it's return positions, also it's possible for functions calling functions
to just clear the function slot and keep the values for later use by simply increasing the stack to contain the return value(s).
this means that we can keep the copying to the minimal (which is one of the big costs). Also keeping it this way lead to quite some smooth
handling of code that evaluates a function on the return values, one just need to fill in the function and return ip and evaluate. cool.
The downside is for native code where we may transfer things back in computer registers. But the overhead of function call's are of such dignity that
the copying of one value is not of importance even for natively compiled functions, The upfront cost of handling functions is pretty high. Note that the improvement's that I suggest have two properties, 1. they are composable, 2. they scale well in the number of arguments and return values.
My feeling is that speed bumps due to function overhead can be adressed by
a) Inlining
b) A well designed interface of user defined instructions
c) Develop separate function call mechanisms that a compiler can deduce and use.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]