~icefox/garnet

Experiment with a lightweight systems programming language

#8 Thoughts on modules

~icefox commented on garnet todo

5 days ago

#30 Assertion fail in typeck

~icefox commented on garnet todo

a month ago
View project feed 

#Garnet

builds.sr.ht status

An experiment in a systems programming language along the lines of Rust but more minimal. Where Rust is a C++ that doesn't suck, it'd be nice for this to be a C that doesn't suck. Currently though, it's mostly just ideas, though a functioning compiler is slowly gaining more features.

Loosely based on https://wiki.alopex.li/BetterThanC, far more loosely based on https://wiki.alopex.li/GarnetLanguage which is an older set of concepts and has a different compiler entirely.

#Assumptions

Things where you go "it's a modern language, of COURSE it has this". If it doesn't have something like this, it's a hard error.

  • Unambiguous, context-free syntax
  • Good error messages
  • Cross-compile everywhere
  • Type inference
  • Sum types, no null, all that good jazz

#Runtime/language model goals

Things where you might need to need to make explicit design tradeoffs. It concerns the overlap of design and implementation. These are essentially directions explore rather than hard-and-fast rules, and may change with time.

  • Simplicity over runtime performance -- Rust and Go are very different places on this spectrum, but I think OCaml demonstrates you should be able to have a bunch of both. There needs to be more points on this spectrum. Investigate more.
  • Fast compiler -- This is a pain point for Rust for various reasons, and one of those things where having it work well is real nice.
  • Simplicity of compiler -- I'd rather have a GOOD compiler in 50k lines than a FANTASTIC compiler in 500k lines; investigate qbe for example.
  • I feel like these two things together should combine to (eventually) make compiler-as-library more of a thing, which seems like an overlooked field of study. It can be useful to aid JIT, metaprogramming, powerful dynamic linking, etc. It seems very silly that this remains Dark Magic outside of anything that isn't Lisp or Erlang. (That said, when you don't want this, you REALLY don't want it.)
  • As little undefined behavior as possible -- If the compiler is allowed to assume something can't happen, then the language should prevent it from happening if at all feasible. Let's stop calling it "undefined behavior" and call it an out of context problem, since it's often not undefineable, but rather it's something that the compiler doesn't have the information to reason about.
  • I am not CONVINCED that a linker is the best way to handle things. This has implications on things like distributing libraries, defining ABI's, using DLL's, and parallelizing the compiler itself. No solid thoughts here yet, but it is an area worth thinking about. Rust, C, Go and Swift present different points in this area to look at.

#C's advantages that I want to have

  • Easy to port to new systems
  • Easy to use on embedded systems
  • Easy to control code size bloat
  • Easy to get a (partial) mental model, which is low-level enough to teach you a lot
  • Simple and universal ABI for every platform -- easy for higher level stuff to call it, easy for it to call arbitrary stuff.
  • Compiles fast

#Pain points in Rust to think about

  • You can't be generic over mutability and ownership, so for example you end up with iter(), into_iter(), and iter_mut().
  • Related, the pile of AsRef, Deref, Borrow, ToOwned etc. traits.
  • Related, the pile of various things that look kinda like references/pointers but aren't, and all the hacks that go into making them work. Example: Box. Seems fine, right? Can't pattern match on it. See the box_pattern RFC.
  • Rust's hacky generic-ness over length of sequences/tuples is pretty lame
  • The slightly-magical relationship between String and &str, and &[] and [] and [T;N], is a little distressing
  • Magical AsRef and Deref behavior is a little distressing
  • std vs core vs alloc -- it'd be better if std didn't actually re-export core, because then more programs could be no_std implicitly. alloc is kinda a red-headed stepchild in this hierarchy; Zig's approach of explicit allocator objects everywhere may or may not be superior. Talk to some of the stdlib or embedded people about how they'd want to arrange it if they could; papering over weird platforms like wasm is a known annoyance. Maybe something like core for pure computational things, sys for platform-specific low-level stuff like threading and timekeeping primitives that may appear in a microcontroller or low-level VM without a full OS, then os or something for stuff like filesystems, processes, etc. Need better names though. I do like the idea of splitting out specific capabilities into specific parts that may or may not be present on all platforms though, instead of having a strictly additive model.
  • Syntax inconsistencies/nuisances: Fiddly match blocks, <>'s for generics (though the turbofish is wonderful), i32 is both a type and a module, -> and => being different is a PITA, you declare values with = in let statements but : in struct constructors,
  • Tail call optimization is not guarenteed -- Drop impl's get in the way, but it should be possible to avoid that, or at least make it so the compiler gives a warning if it can't guarentee that
  • Lack of construct-on-heap is occasionally kinda awful, though far more often totally unnoticable.
  • Rather mediocre support for data type reflection at either compile or run time, such as RTTI in general. Also bites us in trying to make C-like enums, separate enum discriminants from enums or vice versa (which makes them awkward to compose),
  • Rust's closures are awful.
  • On the note of boilerplate-y stuff, see https://github.com/rustwasm/walrus/blob/121340d3113e0102707b2b07cab3e764cea1ed6b/crates/macro/src/lib.rs for an example of a giant, complex, heavy proc macro that is used exactly once to generate a huge amount of entirely uninteresting --but nonetheless necessary-- code. It's good that you can use a macro for it, but it's kinda less good that you need to.
  • No function currying is rather a pain sometimes, especially when it's really just syntactic sugar for a trivial closure.
  • Rust's trait orphan rules are annoying, but may be too hard to be worth trying to solve.
  • Heckin gorram -> vs => still bleedin' trips me up after five years

#Glory points in Rust to exploit or even enhance

  • Move semantics everywhere
  • Derive traits
  • methods <-> functions
  • True, if conservative, constexpr's
  • Iterators just return Option
  • Math is checked by default
  • Stack unwinding without recovery -- very nice compromise of complexity
  • UTF-8 everywhere
  • Lack of magical constructors

#Functionality we sacrificed for simplicity

  • match blocks on function params, like Erlang -- just syntactic sugar
  • Monomorphized generics -- for now?
  • Cool arbitrary/rational number types -- can be a lib.
  • Though it is tempting, we will NOT do arbitrary-precision integer types such as being able to define an integer via an arbitrary range such as [-1, 572) or arbitrary size such as i23. Maybe later.
  • Like Rust, we don't need to target architectures smaller than 32 bits

#Wishlist items

#Goals

  • Being effectively finished someday.
  • A compilation model that doesn't necessitate a slow compiler
  • Being able to reason about what kind of code the compiler will actually output

#Antigoals

  • Async, promises, other fanciness for nonblocking I/O
  • Ultimate max performance in all circumstances
  • Anything requiring a proof solver as part of the type system

#Toolchain

  • rustc
  • logos lexer
  • custom parser (recursive descent + Pratt)
  • output Rust, just to make things work.
  • argh for command line opts
  • codespan for error reporting

Things to consider:

  • rustyline (for repl)
  • lasso or string-interner (for string interning)
  • ryu for parsing floats

Programs-as-separate-files tests:

#Backend thoughts

Something I need to consider a little is what I want in terms of a compiler backend, since emitting x86_64 opcodes myself basically sounds like the least fun thing ever.

Goals:

  • Not huge
  • Operates pretty fast
  • Outputs pretty good/fast/small code
  • Doesn't require binding to C/C++ code
  • Produces x86_64, ideally also Aarch64 and WASM, SPIR-V would be a nice bonus

Non-goals:

  • Makes best code evar
  • Super cool innovative research project
  • Supports every platform evar, or anything less than 32-bits (it'd be cool, but it's not a goal)

Options:

  • Write our own -- ideal choice in the long run, worst choice in the short run
  • LLVM -- Fails at "not huge", "operates fast" and "doesn't require C++ bindings"
  • Cranelift -- Might actually be a good choice, but word on the street (as of early 2020) is it's poorly documented and unstable. Investigate more.
  • QBE -- Fails at "doesn't require C bindings", but initially looks good for everything else. Its Aarch64 unit tests have some failures though, and it doesn't output wasm. Probably my top pick currently.
  • WASM -- Just output straight-up WASM and use wasmtime to run it. Cool idea in the short term, WASM is easy to output and doesn't need us to optimize it much in theory, and would work well enough to let us bootstrap the compiler if we want to. Much easier to output than raw asm, there's good libraries to output it, and I know how to do it.
  • C -- Just output C Code. The traditional solution, complicates build process, but will work.
  • Rust -- Rust compiles slow but that's the only downside, complicates build process, but will work. Might be useful if we can proof whatever borrow checking type stuff we implement against Rust's

Output Rust for right now, bootstrap the compiler, then think about it.

Trying out QBE and Cranelift both seem reasonable choices, and writing a not-super-sophisticated backend that outputs many targets seems semi-reasonable. Outputting WASM is probably the simplest low-level thing to get started with, but is a little weird since it is kinda an IR itself, so to turn an SSA IR into wasm you need a step such as LLVM's "relooper". So one might end up with non-optimized WASM that leans on the implementation's optimizer.

#License

MIT

#References

References on IR stuff:

References on backend stuff:

Technically-irrelelvant but cool papers:

#Random notes

#CI

  • Actual build takes ~1-2 minutes
  • Adding end-to-end unit tests it takes 5 minutes
  • Adding code coverage it takes ~15 minutes -- cargo-tarpaulin ain't instant but most of it is still spent in building it rather than running it.
  • Making a .deb package for cargo-tarpaulin would help a lot then. Talked to the Debian Rust packaging team and they're in favor, very helpful folks, but of course understaffed.

#Out Of Context Problems

nee "Undefined Behavior"

Some more thoughts on the lack of undefined behavior is... you COULD define "read an invalid pointer" to be "return unknown value, or crash the program". But only if you knew that pointer could never aim at memory-mapped I/O. Writing to an invalid pointer could literally do anything in terms of corrupting program state. Some slightly-heated discussion with devsnek breaks the problem down into two parts: For example, WASM does not have undefined behavior. If you look at a computer from the point of view of assembly language + OS, it MOSTLY lacks undefined behavior, though some things like data races still can result in it. If you smash a stack in assembly you can look at it and define what is going to happen. But from the point of view of the assumptions made by a higher-level language, especially one free to tinker with the ABI a little, there's no way you can define what will happen when a stack gets smashed. And even if you're writing in assembly on a microcontroller then you might still be able to do things that put the hardware in an inconsistent state by poking memory-mapped I/O. So, Undefined Behavior usually isn't really Undefined, rather it's defined by a system out of the scope of the language definition. So let's just stop calling it Undefined Behavior and call it an out of context problem.

Things that I think we CAN define context for:

  • Integer overflow either overflows or panics.
  • Constructing an undefined pointer just is a number in a register.
  • Reading an undefined pointer, in the absence of memmapped I/O, either gives a random result, panics, or causes the host system to produce an error (ie by segfaulting your program)
  • We may have well-defined pointers that are never valid, ie, null pointer. Reading and writing to these can do whatever we feel like. We should probably make them either panic or cause the host system to produce an error.
  • Reading uninitialized data should be a compile-time error. Manually eliding initialization for performance reasons just means your compiler isn't good enough at avoiding it itself. A good middle-ground might be some setup where in debug mode you can have runtime checks for reading uninitialized data before it's been written.
  • Order of evaluation of function arguments.
  • Constructing an impossible value, such as having an enum representing integer values 0-5 and stuffing the integer 6 into its slot. Not sure how to deal with this, but it's a case that needs handling.

Here's a list of things that I don't see a way of defining in any reasonable/performant way:

  • Writing an undefined pointer may do anything, ie by smashing the stack. If correctly executing a program requires assuming an un-smashed stack, well, that's tricky.
  • Reading a pointer pointing to mmapped I/O may similarly do anything. A notable recent experience was when reading a value from memory was required to acknowledge an interrupt.

Todo list of other common sources of UB in C, from https://stackoverflow.com/questions/367633/what-are-all-the-common-undefined-behaviours-that-a-c-programmer-should-know-a:

  • Converting pointers to objects of incompatible types
  • Left-shifting values by a negative amount (right shifts by negative amounts are implementation defined)
  • Evaluating an expression that is not mathematically defined (ie, div by 0)
  • Casting a numeric value into a value that can't be represented by the target type (either directly or via static_cast)
  • Attempting to modify a string literal or any other const object during its lifetime
  • A pile of other things that are mostly C++'s fault

Reading material: https://blog.regehr.org/archives/213 and the follow-on articles. Also, apparently Zig has opt-in undefined behavior for things like integer overflow, shift checking, etc which sounds pretty hot. Another option is to provide compiler-intrinsic escape hatches, similar to Rust's various unchecked_foo() functions.

Some of the optimizations that pointer UB enables are talked about here: https://plv.mpi-sws.org/rustbelt/stacked-borrows/paper.pdf Would be very interesting to have some emperical numbers about how much UB helps optimizations.

Again, note that we are NOT trying to make incorrect programs do something that could be considered correct, and we are NOT trying to define things that are inherently undefinable (such as writing to a truly random/unknown pointer). Sooner or later, defining what is defined is up to the programmer, and we are trying to make it so that there are as few rules and hidden gotchas as possible for the programmer to handle when dealing with these things.

Other reading material from people doing interesting unsafe Rust things: https://github.com/TimelyDataflow/abomonation/issues/32

Mandating a different pointer type for referring to mmapped I/O is probably not unreasonable, tbqh, and removes a source of semantic weirdness that compilers have problems dealing with. Basically have a MMIOPtr[T] type that is kinda similar to Rust's UnsafeCell<T>, but doesn't allow reads/writes to be reordered or optimized away, and which has different constraints on what values are allowed for it than a normal pointer does.

#On pointers/references

Evy
A downside of your #[offset] design is that you can create a lot of nonsense and (partial) overlaps all of which you have to check for.  May i suggest a different approach instead? Where you have nominally ordered struct layouts, have a field type padding. Padding is a parameterised type that can take either an alignment/offset, or a size. padding(size: 4) inserts 4 bytes of padding, padding(align: 16) rounds the offset of the next field up to the next multiple of 16 bytes, padding(offset: 0x100) rounds up to offset 256, may be zero-sized, and will error out of the fields before the padding exceed that offset. And padding cannot be used in types which's fields aren't nominally ordered.

icefox — Today at 12:03 PM
Oooo

Evy
So e.g.:
type PlicLayout = #[ordered] ${
  priority:             [U32; 1024]               padding(align:     0x1000),
  pending:              [U32;   32]               padding(align:     0x1000),
  enable_for_context:  [[U32;   32]      ; 15872] padding(offset: 0x20_0000),
  threshold_and_claim: [ThresholdAndClaim; 15872] padding(align:     0x1000),
}

You can even make align and size paddings available for arbitrary fields or variables. Like:
let mut x: Atomic(U32) align(cache) padding(align: cache) = 0;

let mut raw_u32_le: [u8; 4] align(4) = …;

icefox — Today at 12:06 PM
ooooOOOOOooooOOOOOooo
Downside is I'd have to figure out how all those options interact

Evy
In this example, the type of priority is actually ([U32; 1024], [Garbage; PaddingSize]), except that you can only safely touch the .0 payload. In this case, PaddingSize = 0. For enable_for_context it's however much padding is needed to round up to offset 0x20_0000, etc.  Now, two funny things happen now:

1. You could simply demand that types with padded fields will automagically be nominally ordered. Then you can throw the redundant #[ordered] away. If you didn't, well… let's just say that ordering padded fields optimally is not trivial, significantly more complex than a vanilla order by align, size desc.

2. For things like fast zero-copy serialisation, you can demand padding to have specific values instead of »undefined/uninit«. For example by annotating: x: T padding(align: N, fill: 0x00)
3. To create a paddingless C-like struct, all that's needed is to append padding(offset: 0) to the first field.

icefox — Today at 12:16 PM
1) If you need to have things ordered properly in memory then its up to you to define the ordering. That's fine.

Evy
Btw.: [Atomic(U32) align(cache); N] align(page)
Nicely flexible.

Evy
This raises the question of how to write something like Rust's
Box<[i32]>.
This one is tricky. Technically there are two competing pointer types: Single element pointers and pointers to arrays. The latter ones are fat pointers with a length attached. And actually you get even more pointer types. Like pointers to &dyn T things, where it's data-pointer and vtable-pointer packed together. Or function pointers, which on some platforms are magical.

If you have a more powerful type system than Rust 1.0 had, you wouldn't store a contents: *T, but instead a contents: impl Pointer, no second field needed.
Nice, when did Rust finally get these traits?
I feel like reading TWiR is useless, given how often awesome stuff went below my radar.

repnop, Resident RISC-V Shill — Today at 12:35 PM
Yeah it's definitely a toss up

Evy
#![feature(ptr_metadata)]

assert_eq!(std::ptr::metadata("foo"), 3_usize);

I could've used that a century ago! D:
[12:39 PM]
# Characters and strings

Do what Rust does.  No need to innovate here.
@icefox Do what Swift does. :<

icefox — Today at 12:42 PM
What does Swift do?

@Evy
I feel like reading TWiR is useless, given how often awesome stuff went below my radar.

icefox — Today at 12:43 PM
The problem is what is useful varies by person, and often starts small and slow
But yes

Evy
@icefox
# References and pointers
Having looked plenty at proglangs like Herb Sutter's simplified C++ or Ada or in parts C#, i came to have the opinion that this is the wrong approach when designing proglangs. What seems to be a better design is having in, out, in out, move, forward function args. If you think you explicitly need pointers or references, use generic library types like Ptr(T) or Ref(T). Even mutability wrappers like Ptr(Box(T)) (Lisp naming style.) or Ref(Mut(T)) (Rust naming style.). In all other scenarios, it's better to let the language and calling conventions decide whether to use copies or references or what. For example, for tiny args the proglang may decide to turn an in out T into pass-by-value-and-return-by-value behind the scenes. Plus, these annotations allow for way better static checking whether you actually used your function args accordingly to what you specified.

Evy
> @icefox
> What does Swift do?  
It doesn't just have UTF-8 slice plus Unicode Scalar Values, but also has built-in Grapheme Cluster support.(edited)

icefox — Today at 12:48 PM
You still need raw pointers sometimes, but that's an interesting idea. Do you have any references for the "simplified C++"?

Evy
Swift is the proglang with the currently best Unicode handling out there. »Best« being defined by how much spec is covered by the letter.

icefox — Today at 12:48 PM
The flip side of that is, how do you specify those things in struct fields?

@icefox
You still need raw pointers sometimes, but that's an interesting idea. Do you have any references for the "simplified C++"?

Evy
Lemme try find the paper.
https://github.com/hsutter/708


@icefox
The flip side of that is, how do you specify those things in struct fields?

Evy
What things?
Pointers?
Same thing, wrapper types.

icefox — Today at 12:52 PM
Like, how do you say &mut Foo or &Foo in that in/out/etc paradigm
Especially mutability

Evy
The sexy thing about in, out,… is that it gives you a lot of »dark magic« and other patterns for free and safely. in T automagically picks whether to use memcpy(T) or &T or &mut T behind the scenes. You you never have to overload functions or traits based on referenceness. And out T gives you safe access to uninitialised buffers, because the compiler can guarantee you never read before you write. forward is »in, until the last use, which is move«, etc.

@icefox
Like, how do you say &mut Foo or &Foo in that in/out/etc paradigm

Evy
&mut T etc. are impl-details.

If you take an argument in and don't modify it, use in T. If you take it in and observably modify it, you take in out T. If you only return something but never read what you got in (e.g. uninitialised buffers) you use out T. If you take ownership, you use move T. If you are allowed to inspect the data without modifying and then pass along however the caller intended (whether to move or borrow or whatever), you use forward T.

The compiler is free to decide whether or not to use pointers or pass by register or what not. You are 100% oblivious to whether pointers or copies are given.

icefox — Today at 12:57 PM
No no

Evy
So if you e.g. must have a pointer, e.g. to map a physical address to a virtual one, you take e.g. virt: in Pointer(T).

icefox — Today at 12:57 PM
Not for functions
For structs
Oh, the wrapper type

@icefox
For structs

Evy
Vec = struct(T: type) {
  ptr: Ptr(Mut(T)),
  len: USize,
  cap: USize,
}

Evy
Btw., a mutability wrapper would allow for some crazy code patterns. For example, you only need to impl iterators once, and they'll handily allow you to use the same generic iterator type to iterate over T, Ref(Mut(T)),  or just Ref(T). And it gives you dark magic interior mutability patterns like having a field Mut(Bool) for storing a dirty bit. And it does what Rust wants, getting rid of static mut by instead having a static wrapper with interior mutability.
I/O pointers
Ptr(T, volatile: true) or something-something. Maybe even Ptr(Volatile(T)). Where  stand-aloneVolatile(T) does its thing depending on in out etc. (i.e. standalone is useless)(edited)
fn, Fn, FnMut, FnOnce, sigh... can we do anything about that?
The problem is that upvalues, aka. closure captures, are magic hidden payloads, directly contradicting the explicit self design. If you want to get rid of the army of function traits, you have to make captures explicit. Then your »onceness« or »mutness« etc. emerge from whether your captures are in, in out, etc.
If any of your captures is out or in out, you got a FnMut. Got a move, it may be a FnOnce. Etc.
Ada goes a step further by distinguishing between procedures and functions, the latter being allowed to have magic, the former not. The former can even only return stuff via out parameters.

@icefox
You still need raw pointers sometimes, but that's an interesting idea. Do you have any references for the "simplified C++"?

Phlopsi — Today at 2:36 PM
only because we use programming languages, that are too close to the hardware.

icefox — Today at 3:17 PM
Sometimes you need to be close to the hardware.
I wait with bated breath for someone to write a UART driver in Haskell.

@Phlopsi
only because we use programming languages, that are too close to the hardware.  @Evy I agree with you on all the in out stuff. I still don't get the practical usefulness about forward, but I'm tired, so that's fine. 

Evy
Imagine things like debug-logging your stuff you pass right through. Your debug logger just says: fn debug(T: type)(x: forward T) -> forward T. If it came in as move, it comes out as move. If it came in as in out, it comes out as in out, etc. But within debug it is treated as an in.

Evy
@icefox
Sometimes you need to be close to the hardware.  
Ideally only when you're smarter than the compiler/language and making the compiler/language smarter than you is too hard of a problem to solve.

C++ is the perfect example for why e.g. in out is better in the 99% case than spelling out refs vs. moves vs. mutable refs. For an utterly extreme example, look at the amount of junk you need to program to implement a correct forwarding function using current C++. Further more, look at the amount of overload code bloat needed to fine-tune your code for all possible kinds of function args and returns. And that complexity grows exponentially with every added argument or return value. Having in out reduces hundreds of lines of »perfect« C++ code to what looks like a pretty naïve single function.

And i'm frequently wondering whether there are also better ways to abstract over data sharing and memory-mapped structures. I have not yet found the answer.
'cause like in function args, pointers only give you a »how«, not a »what«. And we apply patches and bandages to things like »volatile« and magical linker scripts etc. to add a »what« to the implemented »how«.

icefox — Today at 3:39 PM
Sure, and Rust performs the exact same kind of abstraction of arg representation you're talking about

Evy
Rust is like C++ just a »how«.
The only thing Rust has over C++ is that you can state what ref input a ref output belongs to.
Which is a biggie, no downplaying that.

icefox — Today at 3:49 PM
Not really, functions have no defined ABI
The compiler can and does turn references into copies and vice versa
And, there's always going to be times when I'm smarter than the language. The lang is an amplifier for my own reasoning capabilities, and a special purpose amplifier is going to give me a lot more leverage (for lower cost in its problem domain) than one that tries to be equally good at everything.

Evy
Very limitedly. Copies to refs is demanded by ABI if your struct size exceeds two registers, or if you have too many args. The reverse is a magical recent optimisation of LLVM. But it gives you none of the semantics and semantic guarantees. You still have to overload traits and shit for ref vs. move, you still cannot distinguish between out and in out with a &mut T (it's always treated as in out), etc.

icefox — Today at 3:55 PM
Yeah that part I'm definitely interested in

#On Motivation

A quote from Graydon, original creator of Rust, from https://github.com/graydon/rust-prehistory:

While reading this -- if you're foolish enough to try -- keep in mind that I was balanced between near-total disbelief that it would ever come to anything and miniscule hope that if I kept at experiments and fiddling long enough, maybe I could do a thing.

I had been criticizing, picking apart, ranting about other languages for years, and making doodles and marginalia notes about how to do one "right" or "differently" to myself for almost as long. This lineage representes the very gradual coaxing-into-belief that I could actually make something that runs

As such, there are long periods of nothing, lots of revisions of position, long periods of just making notes, arguing with myself, several false starts, digressions into minutiae that seem completely absurd from today's vantage point (don't get me started on how long I spent learning x86 mod r/m bytes and PE import table structures, why?) and self-important frippery.

The significant thing here is that I had to get to the point of convincing myself that there was something there before bothering to show anyone; the uptick in work in mid-to-late 2009 is when Mozilla started funding me on the clock to work on it, but it's significant that there were years and years of just puttering around in circles, the kind of snowball-rolling that's necessary to go from nothing to "well... maybe..."

I'd encourage reading it in this light: Delusional dreams very gradually coming into focus, not any sort of grand plan being executed.