A Week with Mozilla's Rust
relistan.com
relistan.com
let computed_key = match (key.len() > self.block_size, key.len() < self.block_size) {
(true, false) => self.zero_pad(self.hash(key).digest),
(false, true) => self.zero_pad(key),
(false, false) => key
}
Also, does the compiler complain about my example above being not exhaustive? I think in OCaml it would, which is sometimes nice.
Yes, although the error message isn't very good:
crypto.rs:55:23: 59:5 error: non-exhaustive patterns: true not covered
crypto.rs:55 let computed_key = match (key.len() > self.block_size, key.len() < self.block_size) {
crypto.rs:56 (true, false) => self.zero_pad(self.hash(key).digest),
crypto.rs:57 (false, true) => self.zero_pad(key),
crypto.rs:58 (false, false) => key
crypto.rs:59 };
error: aborting due to previous error
match key.len() {
k if k > self.block_size => ..
k if k < ... => ..
k => ...
match key.cmp(self.block_size) {
Equal => ..
Greater => ..
Less => ...Idris [3] is another language with dependent types (and is more Rust-like) designed for systems programming. I don't think Idris has type level presberger arithmetic though making some things a bit verbose.
[1] http://www.ats-lang.org/ [2] http://bluishcoder.co.nz/2010/09/01/dependent-types-in-ats.h... [3] http://www.idris-lang.org/
I am familiar with ATS and I don't like that I am only familiar with it from the Language Shootout contest website. I haven't heard of it used in production, no blogs, no news on HN.
It is a bad excuse and rather sad but my main charge against it is that it is not popular enough.
[1] http://bluishcoder.co.nz/tags/ats/ [2] http://www.reddit.com/r/ATS/ [3] http://mmpool.bitparking.com
Oh, but it does! There's the cond! macro:
let computed_key = cond!(
(key.len() > self.block_size) { self.zero_pad(self.hash(key).digest) }
(key.len() < self.block_size) { self.zero_pad(key) }
_ { key }
); match key.cmp(self.block_size) {
Equal => ..
Greater => ..
Less => ...
}
The second sentence:
> I want a language where I can be much more productive than in C: one in which I do not fear the correctness of my memory management, and don’t have to go through major gyrations to write concurrent code.
And we see no explicit demonstrations of the memory management and no demonstrations at all of the concurrency model. It feels like the author had a more fleshed out post in mind, but either decided it was running too long or got bored before reaching the conclusion. :\
Or it could be that this post is just the beginning of his journey of learning & blogging about Rust. He did post another one 4 days later, albeit not yet touching on the topics you mention:
http://relistan.com/a-pattern-for-wrapping-c-function-calls-...
I've been looking into Rust during the past few days and I must say I got stuck on the memory model - all the different pointers and their interaction with ownership and mutability, closures capturing their environment (variables outside the closure in the scope where the closure is defined), reference binding, and then things like std::Cell which on the one hand seems to be an alternative to std::Option but on the other hand it's using some unsafe operations in its implementation to change mutability or something... It's quite confusing and it stopped me from writing a simple example application I wanted to write because I wanted to understand what's going on a bit better first.
@ is going to sink into the library and will be deemphasized in the language to help with this. That leaves only ~ (pointers) and & (references).
> closures capturing their environment (variables outside the closure in the scope where the closure is defined)
Right, this is one of the difficult pieces. This is going away as well in favor of Java-like Runnables to make the variable capture easier to understand (with macro sugar to make them just as convenient as closures). The closures that remain will work just like the closures you know from other languages regarding variable capture, and the Rust-specific stuff regarding variable capture will be gone.
> std::Cell
Yeah, this needs to be better documented. Part of the problem is that we haven't totally nailed down what Cell will be used for yet.
1. avoid dynamic allocations (limit yourself to stack-only objects) and you'll get rid of those fears for memory management
2. limit yourself to functional programming and you'll get a pretty good concurrent-ready code
For this "I need a new language" stance you'll need some better reasons.
The situation with Go is completely different. Go is just... meh. Its concurrency model is worse than Erlang's, which predates it. Its type system is worse than Standard ML's, which predates it. Its runtime efficiency is noticeably worse than C++'s, especially so when using customized allocators and avoiding exceptions and RTTI. In short, Go is mediocre.
defer is quite a bit more complex than C++ stack unwinding, because it's dependent on implicit mutation of per-function state and can't really be implemented any other way in the general case. This shows up when you call defer repeatedly in a loop, for example.
Rust is a longer-term, more ambitious effort.
Modula-2, Ada, Turbo/Apple Pascal were there first, but sadly lost developer hearts to C.
let computed_key = if key.len() > self.block_size {
self.zero_pad(self.hash(key).digest)
} else if key.len() < self.block_size {
self.zero_pad(key)
} else {
key
}The idea is that while you may be able to traverse code-paths with side effects in some of these languages, they encourage treating a line of code (term used loosely, as an expression can span multiple lines) as a computable value. It may take a bit of getting used to at first, but from my own experience, it now feels weird and awkward when a language doesn't.
Reminds me of looking at a nasty language like MUMPS or something, where persistence is built into the core language; statement-oriented languages tend not to be composable.
However, I want to question the claim:
> Go is also not particularly friendly to interface to external libraries written in C
which doesn't seem to be explained anywhere in the article, and jars with my understanding.
Go does play very well with C; cgo[0] makes this pretty straightforward.
Maybe there are specific things that the author is trying to do that they found difficult in Cgo, but I would say that Go does play fairly nicely with C; that was a design goal.
Here's a simple example of cgo in action: http://golang.org/doc/articles/c_go_cgo.html. As you can see, it's relatively straightforward and easy to use.
So Go applications are something like "compiled, fast Ruby or Python." As far as I understand, Rust should theoretically be a way to write a piece of your application (even mainly C application) in something that gives you a different kind of expressibility. If there's need for that is another question, specifically, I'd like to see some examples that would really impress me. The examples the article author provided give me only the impression "OK it's weirder syntax but it's not so far from what would have to be written in C anyway, so where's the advantage?"
What I consider missing for "properly" interfacing with C is for Rust to be able to slurp C headers as they are, without the need for additional "translation" files that are presented in the article. At least the headers without the definitions of the functions. But the declarations of the structures and functions and even basic preprocessing were really, really convenient.
Memory safety.
That is correct, but the GP is correct as well. Using C code in Go is easy(just use cgo etc.), vice versa not so much(no such thing as goc etc.).
36 // The above description skipped over the possibility of the gcc-compiled
37 // function f calling back into Go. If that happens, we continue down
38 // the rabbit hole during the execution of f.
39 //
40 // To make it possible for gcc-compiled C code to call a Go function p.GoF,
41 // cgo writes a gcc-compiled function named GoF (not p.GoF, since gcc doesn't
42 // know about packages). The gcc-compiled C function f calls GoF.
43 //
44 // GoF calls crosscall2(_cgoexp_GoF, frame, framesize). Crosscall2
45 // (in cgo/gcc_$GOARCH.S, a gcc-compiled assembly file) is a two-argument
46 // adapter from the gcc function call ABI to the 6c function call ABI.
47 // It is called from gcc to call 6c functions. In this case it calls
48 // _cgoexp_GoF(frame, framesize), still running on m->g0's stack
49 // and outside the $GOMAXPROCS limit. Thus, this code cannot yet
50 // call arbitrary Go code directly and must be careful not to allocate
51 // memory or use up m->g0's stack.
52 //
53 // _cgoexp_GoF calls runtime.cgocallback(p.GoF, frame, framesize).
54 // (The reason for having _cgoexp_GoF instead of writing a crosscall3
55 // to make this call directly is that _cgoexp_GoF, because it is compiled
56 // with 6c instead of gcc, can refer to dotted names like
57 // runtime.cgocallback and p.GoF.)
58 //
59 // runtime.cgocallback (in asm_$GOARCH.s) switches from m->g0's
60 // stack to the original g (m->curg)'s stack, on which it calls
61 // runtime.cgocallbackg(p.GoF, frame, framesize).
62 // As part of the stack switch, runtime.cgocallback saves the current
63 // SP as m->g0->sched.sp, so that any use of m->g0's stack during the
64 // execution of the callback will be done below the existing stack frames.
65 // Before overwriting m->g0->sched.sp, it pushes the old value on the
66 // m->g0 stack, so that it can be restored later.
67 //
68 // runtime.cgocallbackg (below) is now running on a real goroutine
69 // stack (not an m->g0 stack). First it calls runtime.exitsyscall, which will
70 // block until the $GOMAXPROCS limit allows running this goroutine.
71 // Once exitsyscall has returned, it is safe to do things like call the memory
72 // allocator or invoke the Go callback function p.GoF. runtime.cgocallbackg
73 // first defers a function to unwind m->g0.sched.sp, so that if p.GoF
74 // panics, m->g0.sched.sp will be restored to its old value: the m->g0 stack
75 // and the m->curg stack will be unwound in lock step.
76 // Then it calls p.GoF. Finally it pops but does not execute the deferred
77 // function, calls runtime.entersyscall, and returns to runtime.cgocallback.
78 //
79 // After it regains control, runtime.cgocallback switches back to
80 // m->g0's stack (the pointer is still in m->g0.sched.sp), restores the old
81 // m->g0.sched.sp value from the stack, and returns to _cgoexp_GoF.
82 //
83 // _cgoexp_GoF immediately returns to crosscall2, which restores the
84 // callee-save registers for gcc and returns to GoF, which returns to f.Again, not an expert, I could be mistaken, but this seems to not be addressing the original issue raised (that C/Go interop is one-way only).
- Have some C code in your Go app that calls a Go function.
- Have some C code in your C app that calls a Go function.[0] http://stackoverflow.com/questions/6125683/call-go-functions...
https://github.com/mozilla/rust/blob/master/src/libextra/jso...
I know it's the least interesting thing, but for me the simplest rule to recognize Rust vs Go is that at the moment Go has the nice and minimalistic := and Rust the clumsy (for my aesthetics at least) "let."
Still, I understand that Rust attempts to solve more than Go in some aspects, and I really hope it can succeed. For that, yay Rust!
Can somebody explain me the reason for constructs like this:
self.error(~"trailing characters")
self.parse_ident("ull", Null),
It is not possible for us to have ":=" because we don't know whether to parse a pattern or an expression without a prefix token like "let". Rust's pattern language is far more expressive than Go's and one grammar will not cover both.
> Why should something that are facto string literals be owned and on heap? Isn't the need to allocate and box string literals on the heap something that makes the language unnecessarily slow?
Probably the `error` method explicitly asked for a heap-allocated string, perhaps because it wants to pass it to someone who will later free it. (Unlike in C, it would be a type error to attempt to free a constant string in read-only memory, which is what a plain string literal is in Rust.)
self.error(~"trailing characters")
There are also other possibilities -- by knowing that the heap allocated things have lower bits of addresses 0 you can mark the stuff using the lowest bits of pointers.
I know, I have maybe too low level approach. But why not, as soon as you want to be more convenient than C, you have to think low level too.
Still what I like is that at least at the moment this explicit declaration of boxing gives a nice feeling that nothing is done "behind the back" of the programmer, which is good -- the main problem of C++ is that you can't know if somebody hid something very nasty behind some plain innocent looking construct, or even what the compiler will implicitly do or won't.
EDIT-addition: Regarding "pattern" if the := would be a single token, not allowed in patterns, wouldn't it be OK then? Do you have such a valid sequence in the patterns as the combination of the single operators or whatever?
You can implement that yourself, by using an enum for example or by using a custom smart pointer. The moment you start adding more magic to "free" beyond "call free" you become less low level of a language.
The thing I miss most in C is the possibility for the some kind of"introspection" -- reaching out to the info that the compiler or linker has to know anyway. As far as I know D language is very good for such things.
It'd require some OS-specific magic. But you could probably do it.
enum StringRef {
Static(&'static str),
Owned(~str)
}
self.error(Static("trailing characters"))
// or
self.error(Owned(fmt!("%u trailing characters", count)))
The "error" function returns a struct containing the error string. I suspect the intention is to allow dynamically constructed error messages like "syntax error at line 4, char 3" (although it doesn't seem to be doing this anywhere). Since the string might be dynamically allocated, it has to be freed when the struct is freed, which obviously can't happen if the struct just has a pointer to the literal in static memory. (You could flag the string somehow as "should be freed/should not be freed", as you suggest in your other comment, but this has its own tradeoffs)
If you wanted an "error" function that would take string literals directly (and only string literals) you would do something like this: https://gist.github.com/bct/6300740
The "parse_ident" function doesn't return any portion of the input string, so it can just use a borrowed reference.
And how about some convenient "either or"?
The reason it's used here is so that the error messages can be constructed at runtime e.g. `fmt!("line %u: trailing characters", line_number)`, if it was using &'static str (i.e. a compile-time literal) then one could only use hard-coded error messages.
> What I consider missing for "properly" interfacing with
> C is for Rust to be able to slurp C headers as they are,
> without the need for additional "translation" files that
> are presented in the article.
https://github.com/crabtw/rust-bindgen/
C header files go in, the proper Rust definitions come out. There are long-term plans to adopt this as an official tool to be distributed with the compiler.
For other things, like some of the constants ncurses defines, it's much less clear what's going on (specifically, I had a hard time with any macro defined with the NCURSES_ACS macro, though that was pre-Go1 -- I believe cgo can now use these macros directly).
There's an additional impedance mismatch when the C headers use macros for conditional compilation (for, e.g., cipher suites that can be optionally compiled out), but it's not a worry if you either are conservative with the functionality you use, or control the packages installed on a system.
I haven't had a chance to play with Rust yet, but since it uses C linkage directly for it's FFI layer, I suspect these problems are less easily addressable.
(In fact, rustc used to build clang as well as LLVM, but this was disabled since it wasn't being used (yet).)
Rust can operate without garbage collection if you don't use shared pointers.
> "OK it's weirder syntax but it's not so far from what would have to be written in C anyway, so where's the advantage?"
Apart from memory safety, a lot of exciting, modern features: pattern matching, typeclasses, immutability per default, built-in concurrency and parallelism, a module system, a packaging system...
> What I consider missing for "properly" interfacing with C is for Rust to be able to slurp C headers as they are, without the need for additional "translation" files that are presented in the article.
I don't know any language with this kind of FFI...
Does Objective C count?
Note that this is not just a matter of Go's GC and runtime requirement, but simply the fact that they do not support cdecl / stdcall conventions.
There are plenty of garbage collected languages that can be linked into a C program (Lua, Python, even Haskell) - you just have to make extra calls to start-up the runtime.
"Why for instance is it not a good idea to write a library in Go? The reason for this is that Go for needs quite a heavy runtime that does garbage collection and provides a scheduler for it's coroutines."
[0] http://lucumr.pocoo.org/2013/8/18/beautiful-native-libraries...
Eventually once all of our LLVM patches are upstream and the versions including the patches make it into common repositories you may be able to use your system LLVM to compile the Rust compiler, at which point the compile times will drop dramatically.
We continue to work on compile speed of Rust code all the time.
Call me spoiled but I expect a one-click installer in this day and age. I still haven't tried to actually use Rust because of that and I am sure I am not the only one.
It will come in its due time.
But, windows is (unfortunately) a bit of a second class platform at the moment (although there have been some very big strides in the last week or so). So part of the reason for poor windows packaging is Rust isn't ready on windows yet (even more so that its normal pre-alpha-ness on Linux and Mac).
Every time I update my netbook with a new Rust release it takes a few hours, it almost feels like I am compiling KDE or something.
- it has to build LLVM: Rust currently has to use a custom version; the goal is for this to be unnecessary so that the system LLVM is ok.
- it has to build itself and the standard libraries 3 times: rustc is written in Rust; so it has to bootstrap.
- it eats a lot of memory, so bootstrapping on low memory systems is bad. Much of this is historical; the bootstrapped code means that large portions of the compiler are written when the language was less developed, and so is very non-idiomatic. It's progessively being modernised
The first and last can be fixed; the second can't, at least until there are binary releases.
Given my experience with compiler development, back when I was in the university, my expectations is that the second point would be pretty fast anyway, if the first and last issues are solved.
I have good experience with languages that use modules, some of them had bootstraped compilers like the Oberon family.
That is why I said I am hoping for these issues to be solved, as I think building LLVM is the main culprit.
struct Digest { digest: ~[u8] } // what's ~ here?
for self.digest.iter().advance |&byte| { acc = acc.append(fmt!("%02x", byte as uint)); } acc
What does the above mean? acc as a separate line and nothing else
I hate it when people reinvent the same construct that can be found in other languages but differently. I guess they want to do something different in their language, is it?
The `~[u8]` would be written `std::unique_ptr<u8>` in C++. They're used everywhere in Rust, and having to write out unique_ptr would get awfully tiresome.
The `acc` on its own line is a result of "the last statement gets returned" like many functional languages do. If you were writing C++, you'd write `return acc;` instead. This one is more of a convenience, and you could get away without it.
Wouldn't it be `std::unique_ptr<std::vector<u8>>`?
(1) Still being worked on.
There's also the advantage that there's no way to conflict with the * multiplication operator.
~ is an owned pointer (like unique_ptr in C++); & is a "reference" or borrowed pointer (allowing access to data without taking ownership); * is a raw/unsafe pointer, and is identical to C's pointers.
Also, the * pointer doesn't conflict with multiplication because they can never occur in an ambiguous context:
foo * bar = NULL; // C
let bar: *Foo = std::ptr::null(); // Rust
let bar = std::ptr::null(); // Rust, using type inference~[T] is just a vector.
fn hexdigest(&self) -> ~str {
let mut acc = ~"";
for &byte in self.digest.iter() {
acc.push_str(fmt!("%02x", byte as uint));
}
acc
}