Special thanks to Juan Benet for mentioning this idea at Devcon 1.
Recently I wrote a draft of EIP 105 which propose using a subset of webassembly as Ethereum’s VM. If you aren’t aware webassembly “is a new, portable, size- and load-time-efficient format suitable for compilation to the web.” One interesting note about Webassembly doesn’t compile to linear byte code. Instead it uses an Abstract Syntax Tree (AST). This might not surprise you if you have an experience with LLVM IR. But for me it was a new concept to have bytecode in this form.
What is an AST?
It is just a tree repesentation of some code. Each node of the AST represents an expression. Each function body consists of exactly one expression. Compared to the source code, an AST does not include certain elements, such as inessential punctuation and delimiters (braces, semicolons, parentheses, etc.).
To give you a better idea here is a textual representation of some webassembly code using s-expressions.
;; Recursive factorial (func $factorial (param $i i64) (result i64) (if_else (i64.eq (get_local $i) (i64.const 0)) (i64.const 1) (i64.mul (get_local $i) (call $factorial (i64.sub (get_local $i) (i64.const 1))))))
The AST of the above could be displayed like
figure 1 - an AST
This can then be serialized and sent across the wire.
Why use an AST?
The rationale that webassembly gives is
Trees allow a smaller binary encoding: JSZap.
Polyfill prototype shows simple and efficient translation to asm.js.
Some auxiliary reason might be:
- Effective to JIT
- Code Deduplication
The idea has a bit of an interesting history. It appears that Michael Franz first used the idea to compress java bytecode in a paper on Slim Binaries. The slim binaries were also implemented in the Oberon OS
In addition to the Slim Binaries paper here are some more papers if you are interested in the subject.
- Adaptive Compression of Syntax Trees andIterative Dynamic Code Optimization:Two Basic Technologies for Mobile-Object Systems
- A Tree-Based Alternative to Java Byte-Codes
So another fun thing to do with AST is to merklize them! To review; a merkle tree is just a tree which links its nodes together by using the cryptographic hashes of the nodes.
The result is one root hash that points to the root node. This allows for efficient and secure verification of the contents of large data structures. To turn a AST into Merkle AST all you have walk from the leaf nodes up to the root hashing each node along the path.
If you look at figure one you can count 19 nodes. This is a fairly small program and the number of nodes in a larger program can add up fast. The computation time needed to merklize larger AST might start to add up rather fast. One method to increase efficiency would be to store entire subroutines in a single node and only create merkle tree branches at the point where the AST has a branch condition. For example see figure 3.
figure 3 - nodes contain entire subroutines
In figure 3 the green blocks represent subroutines that are in a single node. Note how the
if else still forms a block by itself since it is a branch condition. This also provide an nice opportunity for parallelization. The interpreter or JIT can run the first branch while the next branch is still being fetched.
Why Merkle ASTs?
One reason is code security. It would make DLL hijack impossible. Of course how you would use DLL’s would be a bit different. Instead of open a dynamic link by a name you would reference the root hash of the routine you wanted to use. This way you have 100% confidence that you're getting the code that you want.
Have you ever had a software problem that you googled and found that 100s of forum posts with the same problem as you but none of their solutions worked for you? This is maybe in part caused by the fact you computer is in a different configuration or ‘state’ then thiers. If we had secure merklized code this would a lot less common since the state of any given piece of software could be immediately determined or set by the root hash.
Perhaps one the most convincing reason is bandwidth saving and massive code deduplication. How many nearly-the-same versions of libraries you have on your harddrive? Or how many time the same function and routine is duplicated through programs?
Let's say you have you already have the green nodes since they are common subroutines. You would only have to download the orange nodes. Where this bandwidth saving could be very important is things like Ethereum light clients but also for general computation. As Web Pages begin to more and more resemble apps the larger their code size becomes. In the age of ephemeral webapps code is repeatedly downloaded many times. How many times do you think you have downloaded jquery? Couple this with a peer-to-peer distribution method like IPFS and I think you would have a very efficient system.