Internals of the Nim Compiler

Author: Andreas Rumpf
Version: 0.11.2

"Abstraction is layering ignorance on top of reality." -- unknown

Directory structure

The Nim project's directory structure is:

bingenerated binary files
buildgenerated C code for the installation
compilerthe Nim compiler itself; note that this code has been translated from a bootstrapping version written in Pascal, so the code is not a poster child of good Nim code
configconfiguration files for Nim
distadditional packages for the distribution
docthe documentation; it is a bunch of reStructuredText files
libthe Nim library
webwebsite of Nim; generated by nimweb from the *.txt and *.tmpl files

Bootstrapping the compiler

As of version 0.8.5 the compiler is maintained in Nim. (The first versions have been implemented in Object Pascal.) The Python-based build system has been rewritten in Nim too.

Compiling the compiler is a simple matter of running:

nim c koch.nim
./koch boot

For a release version use:

nim c koch.nim
./koch boot -d:release

And for a debug version compatible with GDB:

nim c koch.nim
./koch boot --debuginfo --linedir:on

The koch program is Nim's maintenance script. It is a replacement for make and shell scripting with the advantage that it is much more portable. More information about its options can be found in the koch documentation.

Coding Guidelines

  • Use CamelCase, not underscored_identifiers.
  • Indent with two spaces.
  • Max line length is 80 characters.
  • Provide spaces around binary operators if that enhances readability.
  • Use a space after a colon, but not before it.
  • Start types with a capital T, unless they are pointers/references which start with P.

See also the API naming design document.

Porting to new platforms

Porting Nim to a new architecture is pretty easy, since C is the most portable programming language (within certain limits) and Nim generates C code, porting the code generator is not necessary.

POSIX-compliant systems on conventional hardware are usually pretty easy to port: Add the platform to platform (if it is not already listed there), check that the OS, System modules work and recompile Nim.

The only case where things aren't as easy is when the garbage collector needs some assembler tweaking to work. The standard version of the GC uses C's setjmp function to store all registers on the hardware stack. It may be necessary that the new platform needs to replace this generic code by some assembler code.

Runtime type information

Runtime type information (RTTI) is needed for several aspects of the Nim programming language:

Garbage collection
The most important reason for RTTI. Generating traversal procedures produces bigger code and is likely to be slower on modern hardware as dynamic procedure binding is hard to predict.
Complex assignments
Sequences and strings are implemented as pointers to resizeable buffers, but Nim requires copying for assignments. Apart from RTTI the compiler could generate copy procedures for any type that needs one. However, this would make the code bigger and the RTTI is likely already there for the GC.

We already know the type information as a graph in the compiler. Thus we need to serialize this graph as RTTI for C code generation. Look at the file lib/system/hti.nim for more information.

Debugging the compiler

You can of course use GDB or Visual Studio to debug the compiler (via --debuginfo --lineDir:on). However, there are also lots of procs that aid in debugging:

# pretty prints the Nim AST
echo renderTree(someNode)
# outputs some JSON representation
# pretty prints some type
echo typeToString(someType)
# pretty prints the Nim ast, but annotates symbol IDs:
echo renderTree(someNode, {renderIds})
if ?? "temp.nim":
  # only output when it comes from "temp.nim"
  echo renderTree(n)
if ?? "temp.nim":
  # why does it process temp.nim here?

To create a new compiler for each run, use koch temp:

./koch temp c /tmp/test.nim

koch temp creates a debug build of the compiler, which is useful to create stacktraces for compiler debugging.

koch temp returns 125 as the exit code in case the compiler compilation fails. This exit code tells git bisect to skip the current commit.:

git bisect start bad-commit good-commit
git bisect ./koch -r c test-source.nim

The compiler's architecture

Nim uses the classic compiler architecture: A lexer/scanner feds tokens to a parser. The parser builds a syntax tree that is used by the code generator. This syntax tree is the interface between the parser and the code generator. It is essential to understand most of the compiler's code.

In order to compile Nim correctly, type-checking has to be separated from parsing. Otherwise generics cannot work.

Short description of Nim's modules

nimmain module: parses the command line and calls main.MainCommand
mainimplements the top-level command dispatching
nimconfimplements the config file reader
syntaxesdispatcher for the different parsers and filters
filter_tmplstandard template filter (#? stdtempl)
lexbasebuffer handling of the lexical analyser
lexerlexical analyser
parserNim's parser
rendererNim code renderer (AST back to its textual form)
optionscontains global and local compiler options
asttype definitions of the abstract syntax tree (AST) and node constructors
astalgoalgorithms for containers of AST nodes; converting the AST to YAML; the symbol table
passesimplement the passes manager for passes over the AST
treessome algorithms for nodes; this module is less important
typesmodule for traversing type graphs; also contain several helpers for dealing with types
sigmatchcontains the matching algorithm that is used for proc calls
semexprscontains the semantic checking phase for expressions
semstmtscontains the semantic checking phase for statements
semtypescontains the semantic checking phase for types
seminstinstantiation of generic procs and types
semfoldcontains code to deal with constant folding
semthreadsdeep program analysis for threads
evalscontains an AST interpreter for compile time evaluation
pragmassemantic checking of pragmas
identsimplements a general mapping from identifiers to an internal representation (PIdent) that is used so that a simple id-comparison suffices to say whether two Nim identifiers are equivalent
ropesimplements long strings represented as trees for lazy evaluation; used mainly by the code generators
transftransformations on the AST that need to be done before code generation
cgenmain file of the C code generator
ccgutilscontains helpers for the C code generator
ccgtypesthe generator for C types
ccgstmtsthe generator for statements
ccgexprsthe generator for expressions
extccompthis module calls the C compiler and linker; interesting if you want to add support for a new C compiler

The syntax tree

The syntax tree consists of nodes which may have an arbitrary number of children. Types and symbols are represented by other nodes, because they may contain cycles. The AST changes its shape after semantic checking. This is needed to make life easier for the code generators. See the "ast" module for the type definitions. The macros module contains many examples how the AST represents each syntactic structure.

How the RTL is compiled

The system module contains the part of the RTL which needs support by compiler magic (and the stuff that needs to be in it because the spec says so). The C code generator generates the C code for it just like any other module. However, calls to some procedures like addInt are inserted by the CCG. Therefore the module magicsys contains a table (compilerprocs) with all symbols that are marked as compilerproc. compilerprocs are needed by the code generator. A magic proc is not the same as a compilerproc: A magic is a proc that needs compiler magic for its semantic checking, a compilerproc is a proc that is used by the code generator.

Compilation cache

The implementation of the compilation cache is tricky: There are lots of issues to be solved for the front- and backend. In the following sections global means shared between modules or property of the whole program.

Frontend issues

Methods and type converters

Nim contains language features that are global. The best example for that are multi methods: Introducing a new method with the same name and some compatible object parameter means that the method's dispatcher needs to take the new method into account. So the dispatching logic is only completely known after the whole program has been translated!

Other features that are implicitly triggered cause problems for modularity too. Type converters fall into this category:

# module A
converter toBool(x: int): bool =
  result = x != 0
# module B
import A

if 1:
  echo "ugly, but should work"

If in the above example module B is re-compiled, but A is not then B needs to be aware of toBool even though toBool is not referenced in B explicitly.

Both the multi method and the type converter problems are solved by storing them in special sections in the ROD file that are loaded unconditionally when the ROD file is read.


If we generate an instance of a generic, we'd like to re-use that instance if possible across module boundaries. However, this is not possible if the compilation cache is enabled. So we give up then and use the caching of generics only per module, not per project. This means that --symbolFiles:on hurts a bit for efficiency. A better solution would be to persist the instantiations in a global cache per project. This might be implemented in later versions.

Backend issues

  • Init procs must not be "forgotten" to be called.
  • Files must not be "forgotten" to be linked.
  • Anything that is contained in nim__dat.c is shared between modules implicitly.
  • Method dispatchers are global.
  • DLL loading via dlsym is global.
  • Emulated thread vars are global.

However the biggest problem is that dead code elimination breaks modularity! To see why, consider this scenario: The module G (for example the huge Gtk2 module...) is compiled with dead code elimination turned on. So none of G's procs is generated at all.

Then module B is compiled that requires G.P1. Ok, no problem, G.P1 is loaded from the symbol file and G.c now contains G.P1.

Then module A (that depends onto B and G) is compiled and B and G are left unchanged. A requires G.P2.

So now G.c MUST contain both P1 and P2, but we haven't even loaded P1 from the symbol file, nor do we want to because we then quickly would restore large parts of the whole program. But we also don't want to store P1 in B.c because that would mean to store every symbol where it is referred from which ultimately means the main module and putting everything in a single C file.

There is however another solution: The old file G.c containing P1 is merged with the new file G.c containing P2. This is the solution that is implemented in the C code generator (have a look at the ccgmerge module). The merging may lead to cruft (aka dead code) in generated C code which can only be removed by recompiling a project with the compilation cache turned off. Nevertheless the merge solution is way superior to the cheap solution "turn off dead code elimination if the compilation cache is turned on".

Debugging Nim's memory management

The following paragraphs are mostly a reminder for myself. Things to keep in mind:

  • If an assertion in Nim's memory manager or GC fails, the stack trace keeps allocating memory! Thus a stack overflow may happen, hiding the real issue.
  • What seem to be C code generation problems is often a bug resulting from not producing prototypes, so that some types default to cint. Testing without the -w option helps!

The Garbage Collector


I use the term cell here to refer to everything that is traced (sequences, refs, strings). This section describes how the new GC works.

The basic algorithm is Deferrent Reference Counting with cycle detection. References on the stack are not counted for better performance and easier C code generation.

Each cell has a header consisting of a RC and a pointer to its type descriptor. However the program does not know about these, so they are placed at negative offsets. In the GC code the type PCell denotes a pointer decremented by the right offset, so that the header can be accessed easily. It is extremely important that pointer is not confused with a PCell as this would lead to a memory corruption.

The CellSet data structure

The GC depends on an extremely efficient datastructure for storing a set of pointers - this is called a TCellSet in the source code. Inserting, deleting and searching are done in constant time. However, modifying a TCellSet during traversation leads to undefined behaviour.

  TCellSet # hidden

proc cellSetInit(s: var TCellSet) # initialize a new set
proc cellSetDeinit(s: var TCellSet) # empty the set and free its memory
proc incl(s: var TCellSet, elem: PCell) # include an element
proc excl(s: var TCellSet, elem: PCell) # exclude an element

proc `in`(elem: PCell, s: TCellSet): bool # tests membership

iterator elements(s: TCellSet): (elem: PCell)

All the operations have to perform efficiently. Because a Cellset can become huge a hash table alone is not suitable for this.

We use a mixture of bitset and hash table for this. The hash table maps pages to a page descriptor. The page descriptor contains a bit for any possible cell address within this page. So including a cell is done as follows:

  • Find the page descriptor for the page the cell belongs to.
  • Set the appropriate bit in the page descriptor indicating that the cell points to the start of a memory block.

Removing a cell is analogous - the bit has to be set to zero. Single page descriptors are never deleted from the hash table. This is not needed as the data structures needs to be rebuilt periodically anyway.

Complete traversal is done in this way:

for each page descriptor d:
  for each bit in d:
    if bit == 1:
      traverse the pointer belonging to this bit

Further complications

In Nim the compiler cannot always know if a reference is stored on the stack or not. This is caused by var parameters. Consider this example:

proc setRef(r: var ref TNode) =

proc usage =
    r: ref TNode
  setRef(r) # here we should not update the reference counts, because
            # r is on the stack
  setRef(r.left) # here we should update the refcounts!

We have to decide at runtime whether the reference is on the stack or not. The generated code looks roughly like this:

void setref(TNode** ref) {
  unsureAsgnRef(ref, newObj(TNode_TI, sizeof(TNode)))
void usage(void) {

Note that for systems with a continuous stack (which most systems have) the check whether the ref is on the stack is very cheap (only two comparisons).

Code generation for closures

Code generation for closures is implemented by lambda lifting.


A closure proc var can call ordinary procs of the default Nim calling convention. But not the other way round! A closure is implemented as a tuple[prc, env]. env can be nil implying a call without a closure. This means that a call through a closure generates an if but the interoperability is worth the cost of the if. Thunk generation would be possible too, but it's slightly more effort to implement.

Tests with GCC on Amd64 showed that it's really beneficical if the 'environment' pointer is passed as the last argument, not as the first argument.

Proper thunk generation is harder because the proc that is to wrap could stem from a complex expression:


A thunk would need to call 'returnsDefaultCC[i]' somehow and that would require an additional closure generation... Ok, not really, but it requires to pass the function to call. So we'd end up with 2 indirect calls instead of one. Another much more severe problem which this solution is that it's not GC-safe to pass a proc pointer around via a generic ref type.

Example code:

proc add(x: int): proc (y: int): int {.closure.} =
  return proc (y: int): int =
    return x + y

var add2 = add(2)
echo add2(5) #OUT 7

This should produce roughly this code:

  PEnv = ref object
    x: int # data

proc anon(y: int, c: PClosure): int =
  return y + c.x

proc add(x: int): tuple[prc, data] =
  var env: PEnv
  new env
  env.x = x
  result = (anon, env)

var add2 = add(2)
let tmp = if == nil: add2.prc(5) else: add2.prc(5,
echo tmp

Beware of nesting:

proc add(x: int): proc (y: int): proc (z: int): int {.closure.} {.closure.} =
  return lamba (y: int): proc (z: int): int {.closure.} =
    return lambda (z: int): int =
      return x + y + z

var add24 = add(2)(4)
echo add24(5) #OUT 11

This should produce roughly this code:

  PEnvX = ref object
    x: int # data
  PEnvY = ref object
    y: int
    ex: PEnvX

proc lambdaZ(z: int, ey: PEnvY): int =
  return ey.ex.x + ey.y + z

proc lambdaY(y: int, ex: PEnvX): tuple[prc, data: PEnvY] =
  var ey: PEnvY
  new ey
  ey.y = y
  ey.ex = ex
  result = (lambdaZ, ey)

proc add(x: int): tuple[prc, data: PEnvX] =
  var ex: PEnvX
  ex.x = x
  result = (labmdaY, ex)

var tmp = add(2)
var tmp2 = tmp.fn(4,
var add24 = tmp2.fn(4,
echo add24(5)

We could get rid of nesting environments by always inlining inner anon procs. More useful is escape analysis and stack allocation of the environment, however.


Process the closure of all inner procs in one pass and accumulate the environments. This is however not always possible.


proc getAccumulator(start: int): proc (): int {.closure} =
  var i = start
  return lambda: int =
    inc i
    return i

proc p =
  var delta = 7
  proc accumulator(start: int): proc(): int =
    var x = start-1
    result = proc (): int =
      x = x + delta
      inc delta
      return x
  var a = accumulator(3)
  var b = accumulator(4)
  echo a() + b()


Lambda lifting is implemented as part of the transf pass. The transf pass generates code to setup the environment and to pass it around. However, this pass does not change the types! So we have some kind of mismatch here; on the one hand the proc expression becomes an explicit tuple, on the other hand the tyProc(ccClosure) type is not changed. For C code generation it's also important the hidden formal param is void* and not something more specialized. However the more specialized env type needs to passed to the backend somehow. We deal with this by modifying s.ast[paramPos] to contain the formal hidden parameter, but not s.typ!