LLVM – How to Add Link-Time Optimization to LLVM-Based Compiler?

I would say it is a safe bet that, in the vast majority of the cases, it is your code, not the compiler, that is broken. And even in the extraordinary case when it is the compiler, you are probably using some obscure language feature in an unusual way, for which the specific compiler is not prepared; in other words, you could most likely change your code to be more idiomatic and avoid the weak spot of the compiler.

At any rate, if you can prove that you found a compiler bug (based on the language spec), report it the compiler developers, so that they may get it fixed some time.

I've been working on a lot of parsing recently, and some of the key features are:

• A programmatic API -- so it can be used from within a programming language, ideally by simply importing a library. It can have a GUI or BNF-like interface as well, but the programmatic one is the key, because you can re-use your tooling, IDE, static analysis, testing, language abstraction facilities, programmer familiarity, documentation generator, build process, etc. Plus, you can interactively play with tiny parsers, which dramatically reduces the learning curve. These reasons place it at the top of my "important features" list.

• Error reporting, as @guysherman mentioned. When an error is found, I want to know where the error was and what was going on when it happened. Unfortunately, I haven't been able to find good resources for explaining how to generate decent errors when backtracking comes in to play. (Although note @Sk-logic's comment below).

• Partial results. When parsing fails, I want to be able to see what was successfully parsed from the part of the input that was before the location of the error.

• Abstraction. You can never build in enough functions, and the user will always need more, so trying to figure out all the possible functions up-front is doomed to failure. Is this what you mean by templates?

• I agree with your #2 (look-ahead prediction). I think it helps to generate good error reports. Is it useful for anything else?

• Support for building a parse tree as parsing occurs, perhaps:

  • a concrete syntax tree, for which the structure of the tree corresponds directly to the grammar and includes layout information for error reporting of later phases. In this case, the client shouldn't have to do anything to get the right tree structure -- it should depend directly on the grammar.
  • An abstract syntax tree. In this case, the user is able to fiddle with any and all parse trees

• Some kind of logging. I'm not sure about this one; maybe to show a trace of the rules the parser has tried, or to keep track of junk tokens such as whitespace or comments, in case (for instance) you want to use the tokens to generate HTML documentation.

• Ability to deal with context sensitive languages. Not sure how important this one is, either -- in practice, it seems cleaner to parse a superset of a language with a context-free grammar, then to check context-sensitive constraints in additional later passes.

• Custom error messages, so that I can tune the error reports in specific situations and perhaps more quickly understand and fix problems.

On the other hand, I don't find error-correction particularly important -- although I'm not up to date on current progress. The problems I've noticed are that the potential corrections that the tools supply are: 1) too numerous, and 2) don't correspond to the actual mistakes made, and so aren't all that helpful. Hopefully this situation will improve (or perhaps has already done so).