3 ms·

Can these tools reason about events, async, promises, threads, queues, all these kinds of things? I think the plain graph representation will tell just a minusc

• marktangotango 7y ago
  My understanding it just an AST over which one could implement the types of analysis you describe. I would like to have more detail as well.
• beliu 7y ago
  The internal representation depends on the tool and application. Here are a few examples: * For deeper analysis, we often use the same internal data structures that the compiler itself uses. This includes data structures that capture AST and type information. In the case of some languages, the compiler doesn't capture enough information (e.g., dynamically typed languages with no static typing). In those cases, the data structures are often enhanced/similar-but-not-quite-the-same as what the interpreter/compiler uses. * In "shallower" analysis, you can use a language-agnostic parser to generate a generic AST. There's various techniques for doing so in a language-independent fashion, depending on what tradeoffs you want to make with speed, accuracy, and generality. The tree-sitter library is one such example. You can also use tools like CTAGS + textual heuristics in the extremely shallow case. * Then, there's the Kythe approach, which constructs a taxonomy of source code graph nodes and edges that aims to capture every possible relationship in every language. Each language requires an integration to convert from the source code in that language to the Kythe schema; that integration can use either the "shallow" or "deep" approaches described above, and the right answer depends again on your tradeoff of speed vs. accuracy. As for combining dynamic information like profiling and runtime info, this is something we've been thinking about for awhile. We think there's plenty of low-hanging fruit that combines static information (types, AST, source code) with runtime info (stack traces, profiling info). We've a couple of efforts in motion toward this end. One is the Sourcegraph extension API (https://sourcegraph.com/extensions https://sourcegraph.com/extensions), which allows injection of third-party info, often from runtime tools, into wherever you view code, whether it's in your editor, on GitHub, or on Sourcegraph. Another is allowing a user to plugin stacktrace data to heuristically select which jump-to-def/find-references options are most likely to be useful given what your currently debugging. I'd love to hear what you're thinking of—care to share your ideas in this space?
  • Eli_P 7y ago
    I thought of making a tool for language-agnostic verification and/or refactoring advisory. That would be a modular system with a design looking something like: AST -[basis]-> Normalized AST -[to sequence]-> Common Representation [graph; embeddings] -> [Code Structure DB; Temporal DB(reflects what's a code actually doing, runtime info involved)] <-[query for anomalies] The idea is simple, we teach some database, feeding with lots of source codes. We may want also to define some rules, like, what's a deadlock, what's an event, etc. I've no clear idea how to implement the rules, but it should be possible to make an embeddings for the temporal data and map it onto a space where checking for correctness can be done in linear time, fingers crossed. But here comes the question, what's a temporal db should be? Should it be a separate data definition language, or code is the data? As for basis, we're supposed either to use some generic calculus and its subset for the particular programming language, or somehow make it possible to aggregate into the same database with different bases. Obviously this approaches ask for machine learning instead of traversing the graphs. I think there's good reason for that. For example, if I've written many chunks of code which actually do the same thing, it'd have taken enormous time to prove that using combinatoric approach (graph self-similarities, NP).