Having done a fair degree of programming in Wirthwhile languages, I think the only main design decision that I think was a mistake was the variables at the top.
I'm not sure of the value of seeing all of the variables used listed in one place, it has certainly led to me encountering a identifier scrolling up to determine the type then scrolling back down. When the variable is only used on a few consecutive lines it's just adding work to read and adding work to create. I daresay I have written harder to read code than I intended because I didn't want to go up and declare another variable for short term use. The temptation to inline the expression is always there, because you know what all the parts mean when you write it. It's only when you come back later that you get the regret.
It's possible it could be mitigated by defining something like (not sure if this is a valid grammar)
and bring in scoping on statement sequences. maybe call it stmt_block so that stmt_sequence can be a component that really is just a sequence of statements.
> I'm not sure of the value of seeing all of the variables used listed in one place
It means the compiler knows how much memory the function's activation frame will take and the offset into that for every variable before it encounters any code in the function.
Basically, it makes it easier to write a single-pass compiler. That was important in the 70s but is less important these days.
A lot of people in this subthread are echoing this, but it's at most maybe very slightly easier. For example, TinyCC is a one-pass C compiler and yet C has sub-scopes int foo() { /.../ { int x; /.../ } }. The same could be said of Ken Thompson's C compiler which I believe was also one-pass.
Does Titania not have/intend to have lexically local subscopes? That would seem very un-Wirthian to me.
This design decision makes compiler implementation easier and especially enables single-pass compilation. Later Oberon versions at least supported more than one declaration section in arbitrary order, but still no in-place declarations.
"Functions should always be short" is also one of those guidelines that people treat like a hard rule. There are occasions when a 100 line function is easier to read than 5 20 line functions, or god forbid 20 5 line functions.
Stop being overly dogmatic, it ALSO leads to worse code.
One would assume that, but in practice, the predominant style is not one of many short procedures. Instead it feels that there's a preference to just inline the code unless the resulting procedure will have more than one caller.
Control structures are deeply nested and this goes on for 64 (very dense) lines. The low line count but is an artifact of how Oberon is conventionally formatted. When reformatted to mimic the conventions of languages like C, Java or Python it works out to more than 120 lines.
When I program in Oberon (recreationally) I tend to follow this style even though I would extract the same code into a separate method were I writing in Java.
When I write the backend (this repo isn't even 24 hours old yet), you'll find out why variable declarations are at the top of a procedure. (Hint: it has something to do with the stack).
I wondered why my university made us use C90 for Systems Programming class (circa 2010) until I took Compilers. This quirk specifically stood out to me when considering code generation from an AST - it's a lot easier to simply allocate all required memory at the top of a stack frame when you have the variable declarations at the top of the function.
I'm aware that it lets you do things in a single pass manner, but this is the instance where I think the cost for allowing that is too great.
I always thought there must be a better solution, like emitting the compiled function body first which just increments the offset whenever a space for a variable is required and emit the function entry after the function exit last, so you can set up the stack frame with full knowledge of it's use. Then the entry can jump to the body. Scoping to blocks would let you decrement the offset upon exiting the block which would result in less stack use which would almost always be more beneficial than the cost of the additional jump.
Remarkable that Bill is interested in a version of Oberon-07. It's even more minimalistic than the previous Oberon versions. I spent a lot of time with the original Oberon language versions and experimented with extensions to make the language more useful for system programming (e.g. https://oberon-lang.github.io/ and https://github.com/rochus-keller/oberon/). Eventually I had to give up backward compatibility to get to a language which I really consider suitable for system programming (and still minimal in the spirit of Oberon, see https://github.com/micron-language/specification/ and https://github.com/rochus-keller/micron/); it's still evolving though.
If I get it right, Bill's language is considered for teaching purpose, which is also a goal of Wirth's languages, and for which these languages are well suited (especially for compiler courses). Also note that the name "Oberon" was not inspired by Shakespeare, but by the Voyager space probe's flyby and photography of Uranus's moons during the mid-1980s when the language was being developed (see https://people.inf.ethz.ch/wirth/ProjectOberon/PO.System.pdf page 12).
Please note the project isn't even 24 hours old yet.
But I am using Oberon-07 as base, and I might deviate from it quite soon too. But I won't be going in the direction of things like Oberon+ (which adds generic and OOP programming) or Micron which adds the entire type system necessary to interact with foreign code. I just wanted something to explain to people how to do tokenizing, parsing, semantic checking (not just basic types), and machine code generation, and this seemed like the best language to choose.
n.b. I know the name does comes from from the Voyager space probe, but I wanted to keep it directly related somehow, and Titania was the best fit. It's also a moon of Uranus, and there is a story relation to Oberon (Fairy King).
I would say it's probably not necessary to explain what the connection between Titania and Oberon is in the README. It's probably evident to most people?
Of course! It wasn't really a criticism, just a cheeky observation that the documentation for every Wirth-style language I've ever seen begins with the EBNF grammar. Though it's rare for a new language to do that today, I appreciate you continuing the tradition.
I have teached Pascal 25 years ago. The idea was to teach the basic principles of programming (loops, variables, arrays, linked lists, sorting, etc.) without worrying about the technical details (C was too tricky, python was not there). Plus Pascal is quite simple and has very few pitfalls.
Once students where proficient in Pascal, we could introduce compiler classes and, when sufficiently advanced, show what the Pascal BNF grammar looked like. So students had a complete picture of a language. Pascal's BNF grammar is very simple.
Also, Pascal enforces strong program structures (BEGIN, END, PROCEDURE, FUNCTION, etc). which helps to frame practical work.
Oberon is a general-purpose programming language, not a DSL. Even though it is very minimal, you can still do quite a bit in it.
But the point of teaching compiler development is to teach people how to do the basic things from tokenizing, parsing, semantic checking, and code generation (directly to machine code).
I have found this is actually a skill most programmers don't even know how to do, especially just tokenizing and parsing, so I thought I'd use Oberon-07 as a base/inspiration for it.
n.b. at the time of this comment, the repo/project is not even 24 hours old yet.
He's the author of Odin, so he has experience writing compilers, so he also wrote a toy compiler in his language as a fun weekend project I guess. Of course it's only a good learning resource for people familiar with Odin. I don't know much about Odin, but from glancing at the code it looks like there are some memory management related features that he's using, which would look uglier in other languages.
Readit News
I'm not sure of the value of seeing all of the variables used listed in one place, it has certainly led to me encountering a identifier scrolling up to determine the type then scrolling back down. When the variable is only used on a few consecutive lines it's just adding work to read and adding work to create. I daresay I have written harder to read code than I intended because I didn't want to go up and declare another variable for short term use. The temptation to inline the expression is always there, because you know what all the parts mean when you write it. It's only when you come back later that you get the regret.
It's possible it could be mitigated by defining something like (not sure if this is a valid grammar)
and bring in scoping on statement sequences. maybe call it stmt_block so that stmt_sequence can be a component that really is just a sequence of statements.It means the compiler knows how much memory the function's activation frame will take and the offset into that for every variable before it encounters any code in the function.
Basically, it makes it easier to write a single-pass compiler. That was important in the 70s but is less important these days.
Does Titania not have/intend to have lexically local subscopes? That would seem very un-Wirthian to me.
This made me smile; going to use it in the future.
Variables are at the top because:
- you immediately see them (so, perhaps, easier to reason about a function? I dunno)
- the compiler is significantly simplified (all of Wirths' languages compile superfast and, if I'm not mistaken, all are single-pass compilers)
However, I feel that Wirth was overly dogmatic on his approaches. And "variables must always be at the top" is one of those.
Hint: This about this from a single pass compiler basis and how much memory needs to be reserved from the procedure's stack frame.
Stop being overly dogmatic, it ALSO leads to worse code.
For example, search for "PROCEDURE Scan" here: https://people.inf.ethz.ch/wirth/ProjectOberon/Sources/Texts...
Control structures are deeply nested and this goes on for 64 (very dense) lines. The low line count but is an artifact of how Oberon is conventionally formatted. When reformatted to mimic the conventions of languages like C, Java or Python it works out to more than 120 lines.
When I program in Oberon (recreationally) I tend to follow this style even though I would extract the same code into a separate method were I writing in Java.
I always thought there must be a better solution, like emitting the compiled function body first which just increments the offset whenever a space for a variable is required and emit the function entry after the function exit last, so you can set up the stack frame with full knowledge of it's use. Then the entry can jump to the body. Scoping to blocks would let you decrement the offset upon exiting the block which would result in less stack use which would almost always be more beneficial than the cost of the additional jump.
If I get it right, Bill's language is considered for teaching purpose, which is also a goal of Wirth's languages, and for which these languages are well suited (especially for compiler courses). Also note that the name "Oberon" was not inspired by Shakespeare, but by the Voyager space probe's flyby and photography of Uranus's moons during the mid-1980s when the language was being developed (see https://people.inf.ethz.ch/wirth/ProjectOberon/PO.System.pdf page 12).
But I am using Oberon-07 as base, and I might deviate from it quite soon too. But I won't be going in the direction of things like Oberon+ (which adds generic and OOP programming) or Micron which adds the entire type system necessary to interact with foreign code. I just wanted something to explain to people how to do tokenizing, parsing, semantic checking (not just basic types), and machine code generation, and this seemed like the best language to choose.
n.b. I know the name does comes from from the Voyager space probe, but I wanted to keep it directly related somehow, and Titania was the best fit. It's also a moon of Uranus, and there is a story relation to Oberon (Fairy King).
I am curious about your thoughts on var parameters (i.e. mutable references), as in:
I would say it's probably not necessary to explain what the connection between Titania and Oberon is in the README. It's probably evident to most people?
Also, in true Wirth style, the documentation mainly consists of the language grammar :)
And I might plan on making this a recorded series of explaining how to make compilers from scratch with this language as a reference.
It's a bit more than just the grammar, but I agree it's generally underspecified.
It'd be nice to see some discussion of the motivation for its departures from Oberon07.
Just wow....
If the latter, I wonder how you can manage another programming language alongside Odin — anyway, thank you and great respect for both!
Not trying to be confrontational, genuinely curious.. but why is this an area where you'd want a DSL?
My initial reaction is : When I'm learning a topic, the last thing I want to be worrying about is learning the ergonomics of a new language
I'm guessing there's a good rational I'm missing
it'd be nice to see some piece of compiler related code in this language that'd be ugly in a general purpose language
Once students where proficient in Pascal, we could introduce compiler classes and, when sufficiently advanced, show what the Pascal BNF grammar looked like. So students had a complete picture of a language. Pascal's BNF grammar is very simple.
Also, Pascal enforces strong program structures (BEGIN, END, PROCEDURE, FUNCTION, etc). which helps to frame practical work.
But the point of teaching compiler development is to teach people how to do the basic things from tokenizing, parsing, semantic checking, and code generation (directly to machine code).
I have found this is actually a skill most programmers don't even know how to do, especially just tokenizing and parsing, so I thought I'd use Oberon-07 as a base/inspiration for it.
n.b. at the time of this comment, the repo/project is not even 24 hours old yet.