Slashdot Mirror
Optimizing Stack Based Architectures?
An anonymous reader queries: "I'm currently writing a stack oriented interpreter (*ahem* Managed Code Environment) with complimentary compiler (that will be under MIT license), and was wondering if there have been any advances in stack architecture optimization? Some intense Googling turned up this paper, but it seems a bit dated, and focuses mainly on managing local variables, which is inapplicable to me because my interpreter directly supports local vars. Any thoughts or useful links on the topic would be appreciated."
Tail-call optimization turns a call followed by an exit into a jump.
If you haven't looked at it then I would recommend GNU Vmgen, a virtual machine interpreter generator. http://savannah.gnu.org/projects/vmgen/
- docs/
It comes with a nice manual that is an interesting read even if you are writing your interpreter by hand: http://www.complang.tuwien.ac.at/anton/vmgen/html
A german company even developed a microprocessor specifically designed for running forth.
Unfortunately there was not much internet in those days and no HTML, but I think that is a good timespan to look at in your research.
Look up Forth in MIT's library archives, it is more likely to throw up stuff than Google.
And if you thought that was boring you obviously havn't read my Journal ;-)
I have no first hand knowledge in the subject, but perhaps this link to citeseer will help.
-blue
four nine eighteen twenty-7 thirty-nine forty-7 fiftyeight sixty-nine seventy-9 eighty-8 one-hundred-and-nine one-twenty
I looked into stack machine optimization some time ago, and all the material seemed to fall into the category called "stack folding" which basicly amounts to combining loads with ALU ops rather than pushing onto the stack and then performing the op, just do the op as the data won't be used again on the stack. VM optimization as opposed to HW stack machine optimization is a slightly diferent ballgame as you don't have the direct bottleneck issues of the hardware. The benifits of stack machines are that they have one bottleneck and that can be optimized, but this also means that that bottleneck is always there you are really limited in pipelining as to keep the stack valid you need to wait for each operation to complete in sequence. The reason to VM a stack arch is the simplicity, but the stack becomes the bottleneck so it isn't the best solution for speed. I think that Stacks make a lot of sense for HW but I'm unsure if there is a benifit in VM SW, seems to me that a VM RISC arch would remove all the stack maintainence issues and all else being equal being that you are running your VM on a machine that probably has multiple registers that the code using the underlying arch would be faster. The VM wants to be the greatest common denominator of all target arch's which isn't going to be simple but without direct stack support on the uP you will have a fair amount of overhead.
You are asking the wrong question.
Don't get me wrong, stack machines are certainly the simplest and arguably the most elegant & compact format, and certainly nice to use if you final target is mostly JIT/native code.
If you are interpreting, your primary concern is _speed_. Speed in an interpreter on modern architectures (especially the P4!) is determined by almost constant branch mispredictions for every interpreted instruction (assuming a for(;;) switch(..) {..}; central interpreting loop, which sadly is still the fastest portable way of doing things). So your goal is to have the minimum amount of instructions generated for any particular language construct.
Generating code for a register machine can be just as easy as for a stack machine if you have the right infrastructure (assuming you don't work with a bounded set of registers, which you have no need to in an interpreter).
Register machines do much better here than stack machines, I would estimate about 1.5 times less instructions overal. Don't worry so much about the amount of work inside an instruction, aim to do as much as possible at once without branching as your VM design.
It may be even be worth it to integrate struct field and array indirection as part of your opcode, as once you switch to a register machine, "getfield" type instructions will become your most frequent ones. So having say an "add" instruction that can directly address struct fields for its 3 arguments is going to be a big win (i.e. compile a.b = c.d + e.f into a single instruction). By having a pointer in your stack frame that points to itself, you can even do this for single variables in the same instruction.
You need to read through comp.compilers.
In the paper referenced, the claim is that 1 in 4 instructions in a Forth VM is a DUP. That instantly makes me think that either the instruction set is horribly designed, or the compiler was generating massively inefficient code. Possibly both. The possible lesson to be learned from this is that, regardless of whether you have a stack or register based VM, other design decisions are likely to have at least as great an impact on performance.
There's also the possibility to use a hybrid machine that's mostly stack based, but if - for some reason - you find yourself DUP'ing or SWAP'ing a lot, having a few registers knocking around may solve that problem very nicely, without turning the whole thing into a register based machine. i.e. you can bolt it on in a mostly backwards compatible, low effort, fashion if you find you have the need.
Try giving your query at http://portal.acm.org/, they return quite a bunch of articles, dunno how many of them are relevant. Download of article text may cost, though...
- Hubert
Hahaha. Guess what? Yet another AC is writing a bytecode interpreter/compiler. Wheee.... I feel so unique!
:)
Something tells me I might know you, and I think you're just trying to get my secrets. I'm not going to tell you that I used a ton of inlining, tail call optimization and an opcode that combines load/operator/store to improve the speed of my interpreter by about 10x. Oh wait -- you tricked me! Drat! Well you're not going to get me to tell about how I'm getting ready to do JIT compilation to improve that by another 10x. Aaaaaaargh. Not again. *sigh* Outwitted by an anonymous coward. I feel so ashamed. *sob*
p.s. If I submitted my interpreter to the language shootout page today, it would debut at #10 -- without a JIT compiler (see list below). My goal is to debut at #5-6 with a JIT compiler. Wanna race?
1. C (gcc) [compiled]
2. Ocaml (ocaml) [compiled]
3. SML (mlton) [compiled]
4. C++ (g++) [compiled]
5. SML (smlnj) [compiled]
6. Common Lisp (cmucl) [compiled]
7. Scheme (bigloo) [compiled]
8. Ocaml (ocamlb) [interpreted]
9. Java (java) [compiled]
10. Pike (pike) [interpreted]
11. Forth (gforth) [interpreted]
12. Lua (lua) [interpreted]
13. Python (python) [interpreted]
14. Perl (perl) [interpreted]
15. Ruby (ruby) [interpreted]
Once you start interpreting an OO-like or procedural language on this register-based machine you'll find that you have to create a stack anyway in order to save all the registers between function calls. Stack based machine or register based machine - in the end it's a complete wash.
The MIT license sucks. Hard.
Such a license came from a time when software patents were just a glint in a lawyer's eye. You're writing an optimizing stack interpreter. This is something which could easily have patentable features in it. Even if you're not planning to patent anything, your licensees need to be protected from your (theoretical) future tyranny.
Thus you need to include a patent release in your license. Something along the lines of "the author of the software grants you a non-exclusive, non-revokable worldwide license to use any patents fully or partly owned or controlled by the author, but only to the extent that such use is necessary to operate the Software".
For what its worth I've been writing an interpreter for mathematical equations in Java. The principal problem here is handeling different types of vectors and matricies (i.e. 2,3 & 4 dimensional vectors, 2 by 2 matricies etc). For each datatype I've used a different stack, so three different stacks for the three different vector types, and 9 stacks for matricies. Theres also corresponding heeps for each data type which are allocated during the compilation stage, eliminating the need for any object creation during evaluation.
There are four sorts of people in the world: fools, lunatics, idiots and morons. - Umberto Eco, Foucaut's pendulum.
Use virtual register file, like Tao/Amiga VM. Much nicer, but probably patented in the US of Idiocy.
You should try citeseer or one of its mirror too.
d +virtual+machine
Most paper are free for download.
IEEE is another website or ACM queue.
http://citeseer.ist.psu.edu/cis?cs=1&q=stack+base
Lua switched from a stack based VM to a (semi-unbounded) register based VM for Lua 5. The speed improvement is quite noticeable (>30% if you do not account for time spent in C libraries). And there is still room for improvement if you are willing to drop portability and/or debugging/tracing.
The source code is very concise and reads nicely. The VM interpreter core loop is < 400 lines in src/lvm.c. Read this along with src/lopcodes.h.
Please read and understand both Lua 4 and Lua 5 VM core BEFORE designing your own VM. Go and download the code at www.lua.org/ftp/, it's less than 200K each.
BTW: It has tail calls, too.
From what I know today, the A Series emulator runs on big Intel clusters with very good performance.
You might want to extend your Google search to include the Burroughs/Unisys stack machines. A trip over to comp.sys.unisys with your question may get a response from people inside that have done stack based emulators for years and years.
If you want to use an interpreted stack-based architecture you're already 1.5 orders of magnitude slower than machine code. What's the point of spending a lot of time making that 1.4 orders of magnitude?