Pet peeve about compiler education that is not repeated by this article (yay, well done!): the way you have to wade through months of parsing bullshit to get to the really interesting stuff (optimization and static analysis).
If you look back in the late 70's, you'll discover something interesting: Almost everything was parsing of one form or another.
Really. Parsing of flowgraphs, etc. Dataflow was done not just by bitvector methods (which were fairly new), but by parsing of graphs using graph grammars.
Most optimization analysis are very closely related to parsing (and can often be solved by graph reachability). The complexity bounds of most of them are also driven by parsing complexity results.
For example, andersen-style points-to analysis is solvable as a context-free-language problem. The time bounds of it are driven by the time bounds on context-free-language parsing. In particular, all-pairs reachability on Context Free Languages (N^3).
etc Parsing is the basis of compilers for a very good reason. The fact that we have advanced other types of methods on top of analysis for optimization and analysis does not change the fact that, at the core, a lot of them are based on equivalence to parsing of graphs and graph reachability.
Another way to look at it: If you ever want to be able to be really good at compiler optimization, don't skip parsing. Otherwise you will miss the connections that are really important to being able to design efficient new algorithms, instead of just learning old ones.
Then you are doing it wrong :)
"(and I'm talking about the act of parsing, not the algorithms as applied elsewhere)"
I'm not sure what this is supposed to mean. It sounds like you are trying to separate parsing text into ASTs and parsing flowgraphs, so you can claim one is boring, and the other is "algorithms as applied elsewhere". There is no such separation. We just happen to name variables a little differently so it doesn't quite look like "figure out if this one parentheses matches this other one". But it is the same algorithm. If you are trying to separate parsing as the thing that happens that makes an AST, that is not the only thing that is parsing. Yes, some books colloquially separate phases like that, but it's silly.
"Think about this: you start learning compilers, you spend months on parsing, and after that what do you have? Nothing - you "compiled" stuff into an AST. "
You are definitely doing it wrong. The same parsing of a grammar you taught someone to make an AST, you can now say "we can apply a slightly different language grammar to the AST and generate machine code", or "apply a slightly different language grammar to the AST and parse out flowgraphs and statements"
etc parsing isn't just take a bunch of text, turn it into an AST. The fact that a ton of people think this is one reason there are so many crappy little compilers that never go anywhere :)
To make an analogy here (from Lockhart's lament: https://www.maa.org/external_archive/devlin/LockhartsLament....), you are saying that color theory is massively important throughout one's career as a artissti. I am saying that I just like painting things.
Sure, there's lots of cool shit in parsing theory no doubt, with tons of broad application. But the way that we are taught it ("here's how we get an AST from written text") turns away people who are into program analysis. As such, we shouldn't consider parsing a pre-requisite for program analysis, regardless of how interesting or applicable the theory of parsing is.
With the broader definition, a large number of what you consider the interesting parts of program analysis are actually parsing problems. Instruction selection, for example, involves tiling the AST (or more typically, the IR) with a minimal set of machine instructions. In many compilers, this is implemented with a machine-definition file that describes a grammar for transforming the IR into machine-language instructions. Peephole optimization involves recognizing certain patterns (eg. division by a constant power of 2) and then transforming them into more efficient constructs (eg. shift-right). This can also be described by a grammar on the IR.
I'm not sure if this is explicitly taught in compiler courses - it wasn't really in mine, but it was alluded to - but it's a really powerful way of thinking of the passes in a compiler. It lets you unify concepts, so that instead of memorizing all the specific algorithms used, you can think of a pass as "this is a transformation from this specific input language to this specific output language".
Compiler development = Finite State Automaton + Push Down Automaton + Context Free Grammar + Intermediate representation + Static Analysis/Optimiation + Stack Machine + ...
Hardware Architecture = Stack Machine/Turing Machine + electricity + creativity
Stack Machine == Turing Machine
Compiler development starts and ends with Turing Machines.
Hardware architecture ends with Turing Machines operating at the speed of electricity through metal.
As it turns out (and I've tried many times myself too), designing a parser generator that will retain a compact representation without losing good error reporting is very hard.
E.g. consider a BNF type grammar. Almost every symbol represents a potential error. But often deriving that error message automatically in a way that is helpful to a human is incredibly hard because the grammar lacks information about which mistakes are more likely for humans to make.
So we end up with bad error reporting, or you end up having to write code to analyse errors anyway, and then a lot of the benefits disappear - the actual parsing is usually easy to do compactly with a recursive descent parser and a tiny library of helpers anyway.
But speed ime also is a major issue. GCC sped up a good bit when switching, and we are seeing speed issues with bison generated parsers.
Ironically one of the benefits of bison is that you get feedback about whether your grammar is conflict free. Can be a significant advantage over a hand built recdec parser for a complex language like SQL.
The feedback certainly is useful, and I'm all for people using such tools to validate grammars for testing. Though people manage to abuse that by writing extra code too (Ruby being a prime example of using a parser generator and then abusing the hell out of it in ways that complicates the grammar).
I really wish more work would go into research on automating error reporting, though. As much as I handwrite most of my parsers now, I wish I didn't have a reason to.
Indeed.
In college, it took me a bit of time to understand what a formal language really was and what the difference between a parser and a simple string processor was and when I arrived in industry, I notice this were something a certain proportion of otherwise experienced programmers simply did not understand this.