xesite/blog/v-vaporware-2019-06-23.markdown

title	date	series	tags
V is for Vaporware	2019-06-23	v	rant

V is a programming language that has been hyped a lot. As it's recently had its first alpha release, I figured it would be a good idea to step through it and see if it lives up to the promises that the author has been claiming for months.

The V website claims the following on the front page:

• The compiler compiles 1.2 million lines of code compiled per CPU core per second
• The resulting code is as fast as C
• Built-in serialization without runtime reflection
• Minimal amount of allocations
• Zero dependencies
• Requires only 0.4 MB of space to build
• Able to translate arbitrary C/C++ code to V and build it faster than C/C++
• Hot code reloading
• 2d/3d graphics support in the standard library
• Effortless cross-compilation
• A powerful built-in web framework
• The compiler generates direct machine code

As far as I can tell, all of the above features are either "work-in-progress" or completely absent from the source repository.

Speed

The author mentions that the compiler is fast, stating the following:

Fast compilation

V compiles ≈1.2 million lines of code per second per CPU core. (Intel i5-7500 @ 3.40GHz, SM0256L SSD, no optimization)

Such speed is achieved by direct machine code generation [wip] and a strong modularity.

V can also emit C, then the compilation speed drops to ≈100k lines/second/CPU.

Direct machine code generation is at a very early stage. Right now only x64/Mach-O is supported. This means that for now emitting C has to be used. By the end of this year x64 generation should be stable enough.

This has a few pretty fantastic claims. Let's see if they can be replicated. Creating a 1.2 million line of code file should be pretty easy:

-- lua
print "fn main() {"

for i = 0, 1200000, 1
do
  print "println('hello, world ')"
end

print "}"

Then let's run this script to generate the 1.2 million lines of code:

$ time lua5.3 ./gencode.lua > 1point2mil.v
        4.29 real         0.83 user         3.27 sys

And compile the resulting file:

$ time v 1point2mil.v
pass=2 fn=`main`
panic: 1point2mil.v:50003
more than 50 000 statements in function `main`
        2.43 real         2.13 user         0.15 sys

Oh boy. It's also worth noting that it was more than 2 seconds to only compile 50,000 lines of code on my Core m7 12" MacBook.

No Dependencies

V claims to have zero dependencies. Again quoting from the website:

400 KB compiler with zero [wip] dependencies

The entire language and its standard library are less than 400 KB. V is written in V, and you can build it in 0.4 seconds.

(By the end of this year this number will drop to ≈0.15 seconds.)

...

Right now the V compiler does have one dependency: a C compiler. But it's needed to bootstrap the language anyway, and if you are doing development, chances are you already have a C compiler installed.

It's a small dependency, and it's not going to be needed once x64 generation is mature enough.

AMD64 is not the only CPU architecture that exists, but okay I'll take that you are only targeting the most common one.

Digging through the readme, its graphics library and HTTP support require some dependencies:

In order to build Tetris and anything else using the graphics module, you will need to install glfw and freetype.

If you plan to use the http package, you also need to install libcurl.

glfw and libcurl dependencies will be removed soon.

Ubuntu:
sudo apt install glfw libglfw3-dev libfreetype6-dev libcurl3-dev

macOS:
brew install glfw freetype curl

I'm sorry, but this combined with the explicit dependency on a C compiler means that V has dependencies. Now, breaking the grammar down pretty literally it says the compiler has zero dependencies. Let's see what ldd says about the compiler when built on Linux:

$ ldd v
        linux-vdso.so.1 (0x00007ffc0f02e000)
        libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f356c6cc000)
        libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f356c2db000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f356cb25000)

So the compiler with "zero dependencies" is a dynamically linked binary with dependencies on libpthread and libc (the other two are glibc-specific).

Also of note, I had to modify the Makefile in order to get it to build on Linux without segfaulting every time it tried to compile code:

$ git diff
diff --git a/compiler/Makefile b/compiler/Makefile
index e29d30d..353824d 100644
--- a/compiler/Makefile
+++ b/compiler/Makefile
@@ -4,7 +4,7 @@ v: vc
        ./vc -o v .

 vc: v.c
-       cc -std=c11 -w -o vc v.c
+       clang -Dlinux -std=c11 -w -o vc v.c

 v.c:
        wget https://vlang.io/v.c

Otherwise it would segfault every time I tried to run it with:

$ ./v --help
fish: “./v --help” terminated by signal SIGSEGV (Address boundary error)

Before I added the -Dlinux flag, it also failed compile with the following error:

$ make
clang -std=c11 -w -o vc v.c
./vc -o v .
cc: error: unrecognized command line option ‘-mmacosx-version-min=10.7’
V panic: clang error
Makefile:4: recipe for target 'v' failed
make: *** [v] Error 1

Implying that the compiler was falsely detecting Linux as macOS.

Memory Safety

V claims to be memory-safe:

Memory management

There's no garbage collection or reference counting. V cleans up what it can during compilation.

So I made a simple "hello world" program:

fn main() {
  println('hello world!') // V only supports single quoted strings
}

and built it on my Linux box with valgrind installed. Surely a "hello world" program has no good reason to leak memory, right?

$ time v hello.v
0.02user 0.00system 0:00.32elapsed 9%CPU (0avgtext+0avgdata 6196maxresident)k
0inputs+104outputs (0major+1162minor)pagefaults 0swaps

$ valgrind ./hello
==5860== Memcheck, a memory error detector
==5860== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
==5860== Using Valgrind-3.13.0 and LibVEX; rerun with -h for copyright info
==5860== Command: ./hello
==5860==
hello, world
==5860==
==5860== HEAP SUMMARY:
==5860==     in use at exit: 1,000 bytes in 1 blocks
==5860==   total heap usage: 2 allocs, 1 frees, 2,024 bytes allocated
==5860==
==5860== LEAK SUMMARY:
==5860==    definitely lost: 0 bytes in 0 blocks
==5860==    indirectly lost: 0 bytes in 0 blocks
==5860==      possibly lost: 0 bytes in 0 blocks
==5860==    still reachable: 1,000 bytes in 1 blocks
==5860==         suppressed: 0 bytes in 0 blocks
==5860== Rerun with --leak-check=full to see details of leaked memory
==5860==
==5860== For counts of detected and suppressed errors, rerun with: -v
==5860== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)

Looking at the generated C code it's plainly obvious to see this memory leak. init_consts creates a 1000 byte allocation and never frees it. This is a memory leak that is unavoidable in any program compiled with V. This is potentially confusing for people who are trying to debug memory leaks in their V code. They will always be off by 1 allocation and 1000 bytes leaked without an easy way to tell why that is the case. The compiler itself also leaks memory:

$ valgrind v hello.v
==9096== Memcheck, a memory error detector
==9096== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
==9096== Using Valgrind-3.13.0 and LibVEX; rerun with -h for copyright info
==9096== Command: v hello.v
==9096==
==9096==
==9096== HEAP SUMMARY:
==9096==     in use at exit: 3,861,785 bytes in 24,843 blocks
==9096==   total heap usage: 25,588 allocs, 745 frees, 4,286,917 bytes allocated
==9096==
==9096== LEAK SUMMARY:
==9096==    definitely lost: 778,354 bytes in 18,773 blocks
==9096==    indirectly lost: 3,077,104 bytes in 6,020 blocks
==9096==      possibly lost: 0 bytes in 0 blocks
==9096==    still reachable: 6,327 bytes in 50 blocks
==9096==         suppressed: 0 bytes in 0 blocks
==9096== Rerun with --leak-check=full to see details of leaked memory
==9096==
==9096== For counts of detected and suppressed errors, rerun with: -v
==9096== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)

Space Required to Build

V also claims to only require 400-ish kilobytes of disk space to build itself. Let's test this claim with a minimal Dockerfile:

FROM xena/alpine

RUN apk --no-cache add build-base libexecinfo-dev clang git \
 && git clone https://github.com/vlang/v /root/code/v \
 && cd /root/code/v/compiler \
 && wget https://vlang.io/v.c \
 && clang -Dlinux -std=c11 -w -o vc v.c \
 && ./vc -o v . \
 && du -sh /root/code/v /root/.vlang0.0.12 \
 && apk del clang

Except it doesn't build on Alpine:

/usr/bin/ld: /tmp/v-c9fb07.o: in function `os__print_backtrace':
v.c:(.text+0x84d9): undefined reference to `backtrace'
/usr/bin/ld: v.c:(.text+0x8514): undefined reference to `backtrace_symbols_fd'
clang-8: error: linker command failed with exit code 1 (use -v to see invocation)

It looks like backtrace() is a glibc-specific addon. Let's link against libexecinfo to fix this:

 && clang -Dlinux -lexecinfo -std=c11 -w -o vc v.c \

Cloning into '/root/code/v'...
Connecting to vlang.io (3.91.188.13:443)
v.c                  100% |********************************|  310k  0:00:00 ETA
Segmentation fault (core dumped)

Annoying, but we can adjust to Ubuntu fairly easily:

FROM ubuntu:latest

RUN apt update \
 && apt -y install wget build-essential clang git \
 && git clone https://github.com/vlang/v /root/code/v \
 && cd /root/code/v/compiler \
 && wget https://vlang.io/v.c \
 && clang -Dlinux -std=c11 -w -o vc v.c \
 && ./vc -o v . \
 && du -sh /root/code/v /root/.vlang0.0.12 \
 && apt -y remove clang

As of the time of writing this article, the image ubuntu:latest has an uncompressed size of 64.2MB. If the V compiler only requires 400 KB to build like it claims, the resulting image size for this Dockerfile should be around 65 MB at worst, right? the resulting du command should show 400 KB in total, right?

3.4M    /root/code/v
304K    /root/.vlang0.0.12

3.7 MB. That means the 400 KB claim is either a lie or "work-in-progress". Coincidentally, the compiler uses about as much disk space as it leaks during the compilation of "Hello, world".

HTTP Module

V has a http module. It leaves a lot to be desired. My favorite part is the implementation of download_file on macOS:

fn download_file(url, out string) {
	// println('\nDOWNLOAD FILE $out url=$url')
	// -L follow redirects
	// println('curl -L -o "$out" "$url"')
	os.system2('curl -s -L -o "$out" "$url"')
	// res := os.system('curl -s -L -o "$out" "$url"')
	// println(res)
}

This has no error checking (the function os.system2 returns the exit code of curl) and it shells out to curl instead of using libcurl. Other parts of the http module use libcurl correctly (though the HTTP status code, headers and other important metadata are not returned). There is also no support for overriding the HTTP transport, setting a custom TLS configuration or many other basic features that libcurl provides for free.

I wasn't expecting it to have HTTP support out of the box, but even then I still feel disappointed.

Suggestions for Improvement

I would like to see V be a tool for productive development. I can't see it doing that in the near future though. I would like to suggest the following to the V developer in order for them to be able to improve in the future:

Firstly, do not make claims about disk space, speed or dependencies without explaining what you mean by that in detail.

Do not shell out to arbitrary commands in the standard library for any reason. If an attacker can somehow run code on a server with a V binary that uses the download_file function, they can replace curl with a malicious binary that is able to do anything the attacker wants. This feels like a huge vulnerability, especially given that the playground allows you to run this function.

AMD64 is not the only processor architecture that exists. It's nice that you're supporting it, but this means that any program compiled with V will be stuck on that architecture. This also means that V cannot currently be used for systems programming like building a system-level package manager.

Do not leak memory in "Hello world". You could solve the 1000 kilobyte leak by adding the following generated C code and calling it after the user-written main() function:

void destroy_consts() { free(g_str_buf); }

If you claim your compiler can support 1.2 million lines of code, do not make it have a limit of 50,000 statements in one function. Yes it is somewhat crazy to have 1.2 million statements in a single function, but as a compiler author it's generally not your position to make these kinds of judgments. If the user wants to have 1.2 million statements in a function, let them.

Do not give code examples for libraries that you have not released. This means don't show anything about the "built-in web framework" until you have code to back your claim. If there is no code to back it up, you have backed yourself into a corner where you are looking like you are lying. I would have loved to benchmark V's web framework against Nim's Jester and Go's net/http, but I can't.

Thanks for reading this far. I hope this feedback can help make V a productive tool for programming. It's a shame it seems to have been hyped so much for comparatively so little as a result. The developer has been hyping and selling this language like it's the new sliced bread. It is not. This is a very alpha product. I bet you could use it for productive development as is if you really stuck your head into it, but as it stands I recommend against using it for anything.