Kernighan shared his thoughts on what he thinks of the world today — with its push away from C to more memory-safe programming languages, its hundreds of distributions of Linux — and with descendants of Unix powering nearly every cellphone.
And the compiler was slow, the code that came out was slow…
The compiler is slower because it has more to check for, but “the code that came out was slow” seems like nonsense, exaggeration, or PEBCAK. Rust code is highly performant and very close to C code.
The support mechanism that went with it — this notion of crates and barrels and things like that — was just incomprehensibly big and slow.
Dude what? C’s build systems like cmake are notoriously unfriendly to users. Crates make building trivial compared to the ridiculous hoops needed for C.
I have written only one Rust program, so you should take all of this with a giant grain of salt,” he said. “And I found it a — pain… I just couldn’t grok the mechanisms that were required to do memory safety, in a program where memory wasn’t even an issue!
He doesn’t say what the program was, and the borrow checker operates by a set of just a few extremely simple rules. There’s no idea of what he was trying to accomplish or how the borrow checker impeded that.
So my reaction as someone who cares deeply about how disastrously unsafe C is and the tangible havoc it creates in modern society:
I agree the compiler is slower. Honestly boo hoo. It’s slower for two very good reasons (better static analysis and better feedback).
The code being slower is such a minor issue as to effectively not be true. Benchmarks prove this.
I’m not going to take “big and slow” as a serious critique of Cargo from someone who idealizes C’s ridiculous, tedious, convoluted build system.
The borrow checker is trivial, and unlike C, the compiler actually gives you easy, intuitive feedback for why your code doesn’t build.
In my limited experience the speed a rust complied executable runs is highly dependent on compiler options. By default (from what I remember), rust includes a ton of debug info in the resulting program. With the correct compiler flags you can strip all that out and programs run very close to c speeds.
The default for cargo is debug builds why that would surprise anyone as being slower is beyond me, —release isn’t that much extra to type or alias. Do people not learn how their tools work any longer? This isn’t that far off from c/c++ where you set cflags etc to fit the final binaries purpose.
Gcc, clang, msvc, and all the other compilers also don’t optimize by default. It’s very normal and very expected for the default build to not include optimizations
Sure but you don’t normally run GCC or Clang directly; you make, and that normally does optimise. I think a closer example is CMake which doesn’t enable release mode by default.
MSVC is usually run from Visual Studio which makes it obvious which mode is being used so the default doesn’t matter so much.
As for “all the other compilers”, Go optimises by default. It does seem to be the exception though…
Yeah, cargo build produces a debug build and cargo build --release is for actually distributing to users. (It doesn’t add the debug symbols, but also spends more time optimizing.)
The support mechanism that went with it — this notion of crates and barrels and things like that — was just incomprehensibly big and slow.
Dude what? C’s build systems like cmake are notoriously unfriendly to users. Crates make building trivial compared to the ridiculous hoops needed for C.
I wouldn’t be surprised, if the guy does not normally use a build system to begin with. Professors don’t tend to have the time to write software that would require a build system (both in terms of complexity and being used by end users).
So, I’m guessing, all he wanted was rustc, but most Rust tutorials don’t bother explaining it, because cargo isn’t much harder to use.
Off the top of my head the compiler is slow because:
With C you can compile every file in parallel, in Rust compilation of a single crate is serial (hence splitting up large projects into many crates is important, but that makes development somewhat more difficult)
LLVM itself is pretty slow
Generic functions are monomorphized (there’s a unique machine code version of it for every concrete type combination they are called with) to improve runtime performance, but that gives LLVM a lot more work to do - see point 2
It’s slower for two very good reasons (better static analysis and better feedback).
Apparently that’s not really the reason. cargo check is usually quite fast.
I also wouldn’t say Rust code is slower than C. It wins in some places (e.g. strict aliasing) and loses in others (e.g. bounds checks) but in practice it’s usually much faster because it’s so much easier to use fast containers (not just linked lists everywhere), fast libraries, and multithreading.
I don’t know how else they could react:
The compiler is slower because it has more to check for, but “the code that came out was slow” seems like nonsense, exaggeration, or PEBCAK. Rust code is highly performant and very close to C code.
Dude what? C’s build systems like cmake are notoriously unfriendly to users. Crates make building trivial compared to the ridiculous hoops needed for C.
He doesn’t say what the program was, and the borrow checker operates by a set of just a few extremely simple rules. There’s no idea of what he was trying to accomplish or how the borrow checker impeded that.
So my reaction as someone who cares deeply about how disastrously unsafe C is and the tangible havoc it creates in modern society:
In my limited experience the speed a rust complied executable runs is highly dependent on compiler options. By default (from what I remember), rust includes a ton of debug info in the resulting program. With the correct compiler flags you can strip all that out and programs run very close to c speeds.
The default for cargo is debug builds why that would surprise anyone as being slower is beyond me, —release isn’t that much extra to type or alias. Do people not learn how their tools work any longer? This isn’t that far off from c/c++ where you set cflags etc to fit the final binaries purpose.
Tbf this mistake comes up so often I do wonder if cargo should have defaulted to release builds. It seems to be what beginners expect.
Gcc, clang, msvc, and all the other compilers also don’t optimize by default. It’s very normal and very expected for the default build to not include optimizations
Sure but you don’t normally run GCC or Clang directly; you
make, and that normally does optimise. I think a closer example is CMake which doesn’t enable release mode by default.MSVC is usually run from Visual Studio which makes it obvious which mode is being used so the default doesn’t matter so much.
As for “all the other compilers”, Go optimises by default. It does seem to be the exception though…
Yeah honestly this does smack of PEBKAC/RTFM
Yeah,
cargo buildproduces a debug build andcargo build --releaseis for actually distributing to users. (It doesn’t add the debug symbols, but also spends more time optimizing.)I wouldn’t be surprised, if the guy does not normally use a build system to begin with. Professors don’t tend to have the time to write software that would require a build system (both in terms of complexity and being used by end users).
So, I’m guessing, all he wanted was
rustc, but most Rust tutorials don’t bother explaining it, becausecargoisn’t much harder to use.Off the top of my head the compiler is slow because:
Apparently that’s not really the reason.
cargo checkis usually quite fast.I also wouldn’t say Rust code is slower than C. It wins in some places (e.g. strict aliasing) and loses in others (e.g. bounds checks) but in practice it’s usually much faster because it’s so much easier to use fast containers (not just linked lists everywhere), fast libraries, and multithreading.
No language guarantees high-speed code. Rust, like C and C++, is also perfectly suited for writing slow code