Introduction

nlvm (the nim-level virtual machine?) is an LLVM-based compiler for the Nim programming language.

From Nim's point of view, it's a backend just like C or JavaScript - from LLVM's point of view, it's a language frontend that emits IR.

Questions, patches, improvement suggestions and reviews welcome. When you find bugs, feel free to fix them as well :)

Fork and enjoy!

Jacek Sieka (arnetheduck on gmail point com)

Table of contents

• Features
• Installation
  • Binaries
  • Compiling nlvm
• Compiling your code
  • Pipeline
  • Porting guide
    • dynlib
    • {.header.}
    • {.emit.}
    • {.asm.}
• wasm32
• Cross compiler
  • Prerequisites
  • Compiling
• REPL / running your code
• Random notes

Features

nlvm works as a drop-in replacement for nim with the following notable differences:

• Fast compile times - no intermediate C compiler step
• DWARF ("zero-cost") exception handling
• High-quality gdb/lldb debug information with source stepping, type information etc
• Smart code generation and optimisation
  • LTO and whole-program optimisation out-of-the-box
  • compiler-intrinsic guided optimisation for overflow checking, memory operations, exception handling
  • heap allocation elision
  • native constant initialization
• Native cross compiler, including wasm32 support with no extra tooling
• Native integrated fast linker (lld)
• Just-in-time execution and REPL (nlvm r) using the LLVM ORCv2 JIT
• Built-in cross-compiler

Most things from nim work just fine (see the porting guide below!):

• the same standard library is used
• similar command line options are supported (just change nim to nlvm!)
• C header files are not used - the declaration in the .nim file needs to be accurate
• If your program has {.compile.} dependencies, these work as usual but require a corresponding compiler to be installed (ie clang, gcc)

Test coverage is not too bad either:

• bootstrapping and compiling itself
• ~95% of all upstream tests - most failures can be traced to the standard library and compiler relying on C implementation details - see skipped-tests.txt for an updated list of issues
• compiling most applications
• platforms with tests:
  • Linux/x86_64
  • Windows/x86_64
• majority of the nim standard library (the rest can be fixed easily - requires upstream changes however)

How you could contribute:

• work on making skipped-tests.txt smaller
• improve platform support (osx should be easy, arm would be nice)
• help nlvm generate better IR - optimizations, builtins, exception handling..
• help upstream make std library smaller and more nlvm-compatible
• send me success stories :)
• leave the computer for a bit and do something real for your fellow earthlings

nlvm does not:

• understand C - as a consequence, header, emit and similar pragmas are ignored - neither will the fancy importcpp/C++ features - see the porting guide below!
• support all nim compiler flags and features - do file bugs for anything useful that's missing

Installation

Binaries

Binaries are available from the github releases page.

Source code

To do what I do, you will need:

• A C/C++ compiler
  • gcc on Linux
  • clang on Windows
• A cup of tea and a good book
  • Compiling llvm takes about an hour the first time (then it's cached)

Start with a clone:

cd $SRC
git clone https://github.com/arnetheduck/nlvm.git --recurse-submodules
cd nlvm

We will need a few development libraries installed, mainly due to how nlvm processes library dependencies (see dynlib section below):

# Fedora
sudo dnf install pcre-devel openssl-devel sqlite-devel ninja-build cmake clang libzstd-devel

# Debian, ubuntu etc
sudo apt-get install libpcre3-dev libssl-dev libsqlite3-dev ninja-build cmake clang libzstd-dev

# MSYS2 CLANG64 (note that we need the gcc shim for clang)
pacboy -S toolchain cmake ninja gcc

Compile nlvm (if needed, this will also build nim and llvm):

make

Compile with itself and compare:

make compare

Run test suite:

make test
make stats

You can link statically to LLVM to create a stand-alone binary - this will use a more optimized version of LLVM as well, but takes longer to build:

make STATIC_LLVM=1

If you want a faster nlvm, you can also try the release build - it will be called nlvmr:

make STATIC_LLVM=1 nlvmr

When you update nlvm from git, don't forget the submodule:

git pull && git submodule update

To build a docker image, use:

make docker

To run built nlvm docker image use:

docker run -v $(pwd):/code/ nlvm c -r /code/test.nim

Compiling your code

On the command line, nlvm is mostly compatible with nim.

For compiling pure Nimm code, you do not need a C compiler - nlvm will compile and link the code using the built-in lld linker:

cd $SRC/nlvm/Nim/examples
../../nlvm/nlvm c fizzbuzz

If you want to see the generated LLVM IR, use the -c option:

cd $SRC/nlvm/Nim/examples
../../nlvm/nlvm c -c fizzbuzz
less fizzbuzz.ll

You can then run the LLVM optimizer on it:

opt -Os fizzbuzz.ll | llvm-dis

... or compile it to assembly (.s):

llc fizzbuzz.ll
less fizzbuzz.s

Apart from the code of your .nim files, the compiler will also mix in the compiler runtime library in nlvm-lib/.

Pipeline

Generally, the nim compiler pipeline looks something like this:

nim --> c files --> IR --> object files --> linker --> executable

In nlvm, we remove one step and bunch all the code together:

nim --> single IR file --> built-in LTO linker --> executable

Going straight to the IR means it's possible to express nim constructs more clearly, allowing llvm to understand the code better and thus do a better job at optimization. It also helps keep compile times down, because the c-to-IR step can be avoided.

The practical effect of generating a single object file is similar to clang -fwhole-program -flto - it is a bit more expensive in terms of memory, but results in slightly smaller and faster binaries. Notably, the IR-to-machine-code step, including any optimizations, is repeated in full for each recompile.

Porting guide

dynlib

nim uses a runtime dynamic library loading scheme to gain access to shared libraries. When compiling, no linking is done - instead, when running your application, nim will try to open anything the user has installed.

nlvm does not support the {.dynlib.} pragma - instead you can use {.passL.} using normal system linking.

# works with `nim`
proc f() {. importc, dynlib: "mylib" .}

# works with both `nim` and `nlvm`
{.passL: "-lmylib".}
proc f() {. importc .}

{.header.}

When nim compiles code, it will generate c code which may include other c code, from headers or directly via emit statements. This means nim has direct access to symbols declared in the c file, which can be both a feature and a problem.

In nlvm, {.header.} directives are ignored - nlvm looks strictly at the signature of the declaration, meaning the declaration must exactly match the c header file or subtly ABI issues and crashes ensue!

# When `nim` encounters this, it will emit `jmp_buf` in the `c` code without
# knowing the true size of the type, letting the `c` compiler determine it
# instead.
type C_JmpBuf {.importc: "jmp_buf", header: "<setjmp.h>".} = object

# nlvm instead ignores the `header` directive completely and will use the
# declaration as written. Failure to correctly declare the type will result
# in crashes and subtle bugs - memory will be overwritten or fields will be
# read from the wrong offsets.
#
# The following works with both `nim` and `nlvm`, but requires you to be
# careful to match the binary size and layout exactly (note how `bycopy`
# sometimes help to further nail down the ABI):

when defined(linux) and defined(amd64):
  type
    C_JmpBuf {.importc: "jmp_buf", bycopy.} = object
      abi: array[200 div sizeof(clong), clong]

# In `nim`, `C` constant defines are often imported using the following trick,
# which makes `nim` emit the right `C` code that the value from the header
# can be read (no writing of course, even though it's a `var`!)
#
# assuming a c header with: `#define RTLD_NOW 2`
# works for nim:
var RTLD_NOW* {.importc: "RTLD_NOW", header: "<dlfcn.h>".}: cint

# both nlvm and nim (note how these values often can be platform-specific):
when defined(linux) and defined(amd64):
  const RTLD_NOW* = cint(2)

{.emit.}

To deal with emit, the recommendation is to put the emitted code in a C file and {.compile.} it.

proc myEmittedFunction() {.importc.}
{.compile: "myemits.c".}

void myEmittedFunction() {
  /* ... */
}

{.asm.}

Similar to {.emit.}, {.asm.} functions must be moved to a separate file and included in the compilation with {.compile.} - this works both with .S and .c files.

Cross compiler

nlvm can be used to cross-compile code for a different platform, for example to create Windows executables on a Linux machine.

Prerequisites

For cross-compilation, a clang-based environment for that platform must first be set up - the environment consists of:

• a sysroot - the basic libraries needed to create executables for the platform
  • for Windows, this is llvm-mingw
  • It can be obtained from the environment you're targeting.
• The compiler runtime for that target (compiler-rt or libgcc)
  • provided by llvm-mingw
• clang - for finding libraries in the sysroot and dealing with {.compile.}
  • a cross-compilation version of gcc may also work, though this hasn't been tested

Compiling

Once the sysroot is set up, compiling is as easy as selecting an alternative OS / CPU using the standard Nim flags, --os: and --cpu:.

Windows

A helper script exists to set up llvm-mingw:

./dl-llvm-mingw.sh

# Set up $PATH to include the `clang` compiler that comes with llvm-mingw
. env.sh

# Assuming we're on Linux, compile for Windows and link dependencies statically
# to create a (mostly) stand-alone binary:
nlvm c --os:windows --passl:-static test.nim

# You can also use a target triple
nlvm c --nlvm.triple=x86_64-w64-mingw32 --passl:-static Nim/examples/fizzbuzz.nim

# Run the compiled program via wine
wine test.exe

wasm32

Use --cpu:wasm32 --os:standalone --gc:none to compile Nim to (barebones) WASM.

You will need to provide a runtime (ie WASI) and use manual memory allocation as the garbage collector hasn't yet been ported to WASM and the Nim standard library lacks WASM / WASI support.

Apart from WASI, an implementation of panicoverride.nim also needs to be provided - here's one that discards all panics:

# panicoverride.nim
proc rawoutput(s: string) = discard
proc panic(s: string) {.noreturn.} = discard

After placing the above code in your project folder, you can compile .nim code to wasm32:

# myfile.nim
proc adder*(v: int): int {.exportc.} =
  v + 4

nlvm c --cpu:wasm32 --os:standalone --gc:none --passl:-Wl,--no-entry myfile.nim
wasm2wat -l myfile.wasm

Most WASM-compile code ends up needing WASM extensions - in particular, the bulk memory extension is needed to process data.

Extensions are enabled by passing --passc:-mattr=+feature,+feature2, for example:

nlvm c --cpu:wasm32 --os:standalone --gc:none --passl:-Wl,--no-entry --passc:-mattr=+bulk-memory

Passing --passc:-mattr=help will print available features (only works while compiling, for now!)

To use functions from the environment (with importc), compile with --passl:-Wl,--allow-undefined.

REPL / running your code

nlvm supports directly running Nim code using just-in-time compilation:

# Compile and run `myfile.nim` without creating a binary first
nlvm r myfile.nim

This mode can also be used to run code directly from the standard input:

$ nlvm r
.......................................................
>>> log2(100.0)
stdin(1, 1) Error: undeclared identifier: 'log2'
candidates (edit distance, scope distance); see '--spellSuggest':
 (2, 2): 'low' [proc declared in /home/arnetheduck/src/nlvm/Nim/lib/system.nim(1595, 6)]
...
>>> import math
.....
>>> log2(100.0)
6.643856189774724: float64

Random notes

• Upstream is pinned using a submodule - nlvm relies heavily on internals that keep changing - it's unlikely that it works with any other versions, patches welcome to update it
• The nim standard library likes to import C headers directly which works because the upstream nim compiler uses a C compiler underneath - ergo, large parts of the standard library don't work with nlvm.
• Happy to take patches for anything, including better platform support!
• For development, it's convenient to build LLVM with assertions turned on - the API is pretty unforgiving
• When I started on this little project, I knew neither llvm nor Nim. Therefore, I'd specially like to thank the friendly folks at the #nim channel that never seemed to tire of my nooby questions. Also, thanks to all tutorial writers out there, on llvm, programming and other topics for providing such fine sources of copy-pa... er, inspiration!