llvm: is it possible to merge validation and compi

Question

Generally speaking, when writing a llvm frontend, one will take an AST and first check that its semantics is well-defined. After this, one will take the AST and perform the IR build phase. I was wondering, how realistic is to perform directly the IR build phase onto the AST, and if errors are found during the build process, revert any partial changes to the module object?

I assume something like this would be required:

• remove defined Types
• remove defined Globals
• anything else i'm missing?

Any ideas about this? what are the general guidelines of what needs to done for a clean revert of module changes after a failed build phase?

Now, this is thinking in terms of optimistically compiling, and failing gracefully it somethings goes wrong. It might very well be that this is completely impossible or discouraged under the current LLVM model. A clear and well-documented answer in this regard is also completely acceptable

Edit In the end, I just want a reasonable way to add functions incrementally but revert gracefully to previous state of module and/or LLVMContext if a function build fails. Whatever is the preferred approach for that will be entirely satisfactory.

thanks!

Answer 1

Many compilers (not necessarily LLVM-related) mix semantic analysis with code generation, so it can definitely be done. However, I'm puzzled by your reference to "revert any partial changes to the module object". When you start building an IR module and encounter a semantic error in the AST, what is your plan? Do you want to spit an incomplete module? Why? Thinking about the way any regular compiler works, if there are semantic errors in the code (i.e. reference to an undefined variable), no output is created. Would you like something different?