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)

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

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
• Integrated C/C++ compiler
• Integrated linker (lld)
• Integrated cross compiler, including wasm32 support
• Just-in-time execution and REPL (nlvm r) using the LLVM ORCv2 JIT

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

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

Binaries are available from the github releases page.

To do what I do, you will need:

• A C/C++ compiler (for building LLVM and the Nim compiler bootstrap)
  • gcc or clang 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:

# 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

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/lib/nim/examples
../../nlvm/nlvm c fizzbuzz

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

cd $SRC/nlvm/lib/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 lib/nlvm.

nlvm includes an integrated clang that can compile C/C++ code directly.

Use the cc subcommand to pass arguments through to clang (preview, this may change in the future):

# Compile and link a C file
nlvm cc test.c -o test

# Compile to object file only
nlvm cc -c test.c -o test.o

# Cross-compile to windows With optimization and warnings
nlvm --os:windows cc -O3 -Wall test.c -o test.exe

# Compile C++
nlvm cc test.cpp -o test

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.

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)

Put the emitted code in a C file and {.compile.} it.

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

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

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.

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

For cross-compilation, an 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.
  • Place the env in ../<triple> relative to the nlvm binary - the mingw environment for example should be copied to <nlvm-path>/../x86_64-w64-mingw32.
• The compiler runtime for that target (compiler-rt or libgcc)
  • Resides in clang/<ver>/lib

The integrated clang compiler supports cross-compiling out of the box - an external compiler can also be used provided it has been set up.

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

A helper script exists to set up llvm-mingw:

./dl-llvm-mingw.sh

# Set up $PATH use the locally compiler `clang` compiler
. 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.target=x86_64-w64-mingw32 --passl:-static lib/nim/examples/fizzbuzz.nim

# Run the compiled program via wine
wine test.exe

WASM programs run in a runtime provided either standalone or inside a browser.

Standalone runtimes typically ship with a WASI runtime while browsers do not.

• Exceptions are not supported - consequently, --panics:on is used by default
• In WASI mode, what works is limited by what is provided by wasi-libc / WASI Preview 1
• Threads are currently not implemented - --threads:off is used by default
• Flags and linking options are in still in flux and might change between releases

The execution mode is derived from the application type:

# By default, command modules are created
nlvm command.nim

# Use --app:lib to create a reactor module instead
nlvm --app:lib ...

By default, if you select --cpu:wasm32, a standalone WASM module will be created - such modules are suitable for use in runtimes that do not provide WASI - commonly browsers and JavaScript engines.

The standalone WASM mode corresponds to wasm32-unknown-unknown.

# Using Nim flags
nlvm c --cpu:wasm32

# Using a target
nlvm c --nlvm.target=wasm32-unknown-unknown

You will need to provide a panicoverride.nim - here's a starting point to customize using functions imported from the environment:

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

In standalone mode, nlvm will pass --no-entry --allow-undefined to the linker.

With WASI, nlvm can compile many Nim applications out of the box using wasi-libc which is included in the distribution.

# Using Nim flags
nlvm c --cpu:wasm32 --os:wasip1

# Using a target
nlvm c --nlvm.target=wasm32-wasip1

Imports allow your code to call functions from the runtime environment, using importc:

# myfile.nim
proc message*(v: cstring) {.importc.}

Exports are used to call Nim code from the environment - for example to call WASM functions from JavaScript in the browser:

proc add(a, b: int): int {.exportc, dynlib.} =
  a + b

This is the equivalent of __attribute__((export_name("add"))) in clang.

WASM features can be enabled or disabled by with regular --passc flags:

nlvm c ... --passc:-mcompact-imports

See the clang reference for details.

BPF targets are selected with --cpu:bpf|bpfel|bpfeb or with a matching --nlvm.target triple. BPF builds use standalone, no-GC, no-startup and no-linking settings because programs run in the kernel. They default to --opt:size when no --opt flag is given.

BPF map definitions are exported globals whose names start with bpf_map_; they are emitted in the .maps ELF section. An optional license is emitted in the license section, and debug information can be used to generate BTF. For the BPF ABI, use fixed-width integer types and keep global initializers as compile-time constants.

Here is a minimal XDP program with an array map:

# program.nim
type
  BpfMapDef = object
    mapType: uint32
    keySize: uint32
    valueSize: uint32
    maxEntries: uint32

# BPF_MAP_TYPE_ARRAY = 2
var bpf_map_packets {.exportc: "bpf_map_packets".}: BpfMapDef =
  BpfMapDef(
    mapType: 2,
    keySize: 4,
    valueSize: 8,
    maxEntries: 16
  )

# XDP_PASS = 2
proc xdpPass(ctx: pointer): uint32 {.exportc: "xdp_pass".} =
  discard ctx
  2'u32

Compile it to a BPF ELF object with:

nlvm c \
    --nlvm.target:bpfel-unknown-none \
    --nlvm.bpf.entry:xdp_pass \
    --nlvm.bpf.section:xdp \
    --nlvm.bpf.license:GPL \
    program.nim

The entrypoint and section options are required. The entrypoint must match the exported name (xdp_pass above), and the section name must differ from it. Pass --debuginfo to include debug information for BTF-enabled tooling. Dynamic allocation, exception handling, and runtime-dependent global initialization are not currently supported by eBPF programs.

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 lib/nim/lib/system.nim(1595, 6)]
...
>>> import math
.....
>>> log2(100.0)
6.643856189774724: float64

• 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
• Some parts of the Nim standard library are unavailable due to the reliance on C headers to provide accurate ABI descriptions - nlvm will provide a hint for the API that need further investigation.
• 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!