The other answers are great, but let me just add that C++ cannot be parsed with conventional LL/LALR/LR parsers, because the syntax is ambiguous and requires disambiguation via type checking (i.e., there may be multiple parse trees but at most one will type check).

There was some research on parsing C++ with GLR but I don't think it ever made it into production compilers.

Other, more sane languages with unambiguous grammars may still choose to hand-write their parsers for all the reasons mentioned in the sibling comments. However, I would note that, even when using a parsing library, almost every compiler in existence will use its own AST, and not reuse the parse tree generated by the parser library. That's something you would only ever do in a compiler class.

Also I wouldn't say that frontend/backend is an evolution of previous terminology, it's just that parsing is not considered an "interesting" problem by most of the community so the focus has moved elsewhere (from the AST design through optimization and code generation).

Table-driven parsers with custom per-statement tokenizers are still common in surviving Fortran compilers, with the exception of flang-new in LLVM. I used a custom parser combinator library there, inspired by a prototype in Haskell's Parsec, to implement a recursive descent algorithm with backtracking on failure. I'm still happy with the results, especially with the fact that it's all very strongly typed and coupled with the parse tree definition.

Not sure about GCC, but in general there has been a big move away from using parser generators like flex/bison/ANTLR/etc, and towards using handwritten recursive descent parsers. Clang (which is the C/C++ frontend for LLVM) does this, and so does rustc.

This was in the olden days when your language's type system would maybe look like C's if you were serious and be even less of a thing when you were not.

The hard part about compiling Rust is not really parsing, it's the type system including parts like borrow checking, generics, trait solving (which is turing-complete itself), name resolution, drop checking, and of course all of these features interact in fun and often surprising ways. Also macros. Also all the "magic" types in the StdLib that require special compiler support.

This is why e.g. `rustc` has several different intermediate representations. You no longer have "the" AST, you have token trees, HIR, THIR, and MIR, and then that's lowered to LLVM or Cranelift or libgccjit. Each stage has important parts of the type system happen.

Compiler theory a) doesn't seem to have much to do with production compilers b) is unnecessarily heavyweight and scary about everything.

In particular, it makes parsing everything look like a huge difficult problem. This is my main problem with the Dragon Book.

In practice everyone uses hacky informal recursive-descent parsers because they're the only way to get good error messages.

Not really. Here’s a comparison of different languages: https://notes.eatonphil.com/parser-generators-vs-handwritten...

Most roll their own for three reasons: performance, context, and error handling. Bison/Menhir et al. are easy to write a grammar and get started with, but in exchange you get less flexibility overall. It becomes difficult to handle context-sensitive parts, do error recovery, and give the user meaningful errors that describe exactly what’s wrong. Usually if there’s a small syntax error we want to try to tell the user how to fix it instead of just producing “Syntax error”, and that requires being able to fix the input and keep parsing.

Menhir has a new mode where the parser is driven externally; this allows your code to drive the entire thing, which requires a lot more machinery than fire-and-forget but also affords you more flexibility.

Mostly because that's the part that had the best developed theory so that's what tended to be taught.

The rest of the f*cking owl is the interesting part.

"Frontend" as used by mainstream compilers is slightly broader than just lexing/parsing.

In typical modern compilers "frontend" is basically everything involving analyzing the source language and producing a compiler-internal IR, so lexing, parsing, semantic analysis and type checking, etc. And "backend" means everything involving producing machine code from the IR, so optimization and instruction selection.

In the context of Rust, rustc is the frontend (and it is already a very big and complicated Rust program, much more complicated than just a Rust lexer/parser would be), and then LLVM (typically bundled with rustc though some distros package them separately) is the backend (and is another very big and complicated C++ program).