Nim - the first high performance language with full support for hot code-reloading at runtime

by Viktor Kirilov

Me, myself and I

my name is Viktor Kirilov - from Bulgaria
creator of doctest - the fastest C++ testing framework
apparently I like text-heavy slides and reading from them...!
- deal with it :|

Talk agenda

some Nim code
the performant programming language landscape
- read: heavily biased C++ rant
Nim compilation model
hot code reloading
- usage & implementation
- ".dll" => assume .so/.dylib (platform-agnostic)
demo
comments & conclusions
a bit on REPLs

Hello

echo "Hello World"

Currencies

type
  # or use {.borrow.} here to inherit everything
  Dollars* = distinct float

proc `+` *(a, b: Dollars): Dollars {.borrow.}

var a = 20.Dollars

a = 25  # Doesn't compile
a = 25.Dollars  # Works fine

a = 20.Dollars * 20.Dollars # Doesn't compile
a = 20.Dollars + 20.Dollars # Works fine

Sets

Iterators

type
  CustomRange = object
    low: int
    high: int

iterator items(range: CustomRange): int =
  var i = range.low
  while i <= range.high:
    yield i
    inc i

iterator pairs(range: CustomRange): tuple[a: int, b: char] =
  for i in range:  # uses CustomRange.items
    yield (i, char(i + ord('a')))

for i, c in CustomRange(low: 1, high: 3):
  echo c

# prints: b, c, d

Variants

# This is an example how an abstract syntax tree could be modelled in Nim
type
  NodeKind = enum  # the different node types
    nkInt,          # a leaf with an integer value
    nkFloat,        # a leaf with a float value
    nkString,       # a leaf with a string value
    nkAdd,          # an addition
    nkSub,          # a subtraction
    nkIf            # an if statement
  Node = ref object
    case kind: NodeKind  # the ``kind`` field is the discriminator
    of nkInt: intVal: int
    of nkFloat: floatVal: float
    of nkString: strVal: string
    of nkAdd, nkSub:
      leftOp, rightOp: Node
    of nkIf:
      condition, thenPart, elsePart: Node

var n = Node(kind: nkFloat, floatVal: 1.0)
# the following statement raises an `FieldError` exception, because
# n.kind's value does not fit:
n.strVal = ""

Multi methods

type
  Thing = ref object of RootObj
  Unit = ref object of Thing
    x: int

method collide(a, b: Thing) {.inline.} =
  quit "to override!"

method collide(a: Thing, b: Unit) {.inline.} =
  echo "1"

method collide(a: Unit, b: Thing) {.inline.} =
  echo "2"

var a, b: Unit
new a
new b
collide(a, b) # output: 2

Meta-programming

what is it
- a program that can read, generate, analyze or transform other programs
why do it
- can optimise code – by compile-time rewrites
  - think expression templates
- can enforce better coding patterns
- can increase code readability and maintainability
  - with great power comes great responsibility
reflection - when the meta language is the actual language

Meta-programming in Nim

works on the Abstract Syntax Tree
respects the type system
levels of complexity:
- normal procs and inline iterators
- generic procs and closure iterators
- templates
- macros

Templates

template withFile(f: untyped, filename: string,
                  mode: FileMode,
                  body: untyped): typed =
  let fn = filename
  var f: File
  if open(f, fn, mode):
    try:
      body
    finally:
      close(f)
  else:
    quit("cannot open: " & fn)

withFile(txt, "ttempl3.txt", fmWrite):
  txt.writeLine("line 1")
  txt.writeLine("line 2")

AST

dumpTree:
  var mt: MyType = MyType(a:123.456, b:"abcdef")

# output:
#   StmtList
#     VarSection
#       IdentDefs
#         Ident "mt"
#         Ident "MyType"
#         ObjConstr
#           Ident "MyType"
#           ExprColonExpr
#             Ident "a"
#             FloatLit 123.456
#           ExprColonExpr
#             Ident "b"
#             StrLit "abcdef"

Macros

import macros

type
  MyType = object
    a: float
    b: string

macro myMacro(arg: untyped): untyped =
  var mt: MyType = MyType(a:123.456, b:"abcdef")
  let mtLit = newLit(mt)
  
  result = quote do:
    echo `arg`
    echo `mtLit`

myMacro("Hallo")

# The call to myMacro will generate the following code:
echo "Hallo"
echo MyType(a: 123.456'f64, b: "abcdef")

More macros

import macros
dumpAstGen:
  proc hello() =
    echo "hi"

nnkStmtList.newTree(
  nnkProcDef.newTree(
    newIdentNode(!"hello"),
    newEmptyNode(),
    newEmptyNode(),
    nnkFormalParams.newTree(
      newEmptyNode()
    ),
    newEmptyNode(),
    newEmptyNode(),
    nnkStmtList.newTree(
      nnkCommand.newTree(
        newIdentNode(!"echo"),
        newLit("hi")
      )
    )
  )
)

HTML DSL

import html_dsl

html page:
  head:
    title("Title")
  body:
    p("Hello")
    p("World")
    dv:
      p "Example"

echo render(page())

HTML DSL result

<!DOCTYPE html>
  <html class='has-navbar-fixed-top' >
  <head>
    <meta charset="utf-8">
    <meta name="viewport" content="width=device-width, initial-scale=1">
    <title>Title</title>
  </head>
  <body class='has-navbar-fixed-top' >
    <p >Hello</p>
    <p >World</p>
    <div>
      <p>Example</p>
    </div>
  </body>
</html>

Simply Nim

statically typed
high performance (compiles to native binaries - comparable to C/C++)
very clean & elegant - no, beauty is NOT subjective!
garbage collected (can do manual memory management too)
expressive - some of the most powerful metaprogramming
- compiler has an interpreter inside
compiles to: C/C++/ObjC/Javascript
- non-idiomatic - not for reading but optimal for execution
suited for: systems programming, applications & web
- all types of software!
backed by Status since 2018 (#65 cryptocurrency by marketshare)
- Status - working on one of the first implementations of Ethereum 2.0
- just like Rust is backed by Mozilla (although with a lot less...)
has a rich stdlib, package manager, docs, some IDE support

Feature rundown

uniform call syntax (extension methods) - obj.method() OR method(obj)
- that's why there are no real "methods" defined in types
function call parens are optional - echo("hello") OR echo "hello"
case-insensitive - also underscore-insensitive but that's another topic :|
generics
templates (meta-programming^2)
macros (meta-programming^3) - evaluated in the compiler by the NimVM
concepts
discriminated unions
strong typedefs (distinct type) - can has $ currency?
coroutines & closures
switch & pattern matching
dynamic dispatch & multi-methods
converters - explicit (for implicit conversions)
effect system (transitive)
extensible pragmas, "defer", exceptions, "discard", named args... good defaults!

My "favourite" aspect of C++

A bit on C++

C++20 is shaping up to be a huge release
- lots of cool stuff, but complexity is through the roof
- Expert-"tolerable" - prestige when you come up with yet more complicated TMP
simple example using ranges from C++20 - blog post
- 3 seconds of compile time for ~20 lines of code, forget about "Debug" builds
Remember the Vasa! - Bjarne Stroustrup
There should come a time for a clean slate
- C++ is a great and valuable ongoing research
- The 2 biggest reasons C++ is so widely used today:
  - legacy and maturity - too much software written already
  - inertia - attachment and lack of interest to learn new languages
- C++ is a HUGE time/money cost on the scale of hundreds of millions
  - developer productivity, bug & safety
  - business should back a better language & push for development + learning

Some quotes & thoughts

Fifty years of programming language research, and we end up with C++?
- Richard A. O’Keefe
There are only two kinds of programming languages: those people always bitch about and those nobody uses.
- Bjarne Stroustrup
Nim is the next iteration of practical language design
- by humble !!! >> me << !!!
Nim: speed of C, elegance of Python, flexibility of Perl
- Peter Munch-Ellingsen
Nim is to C++ as CoffeeScript is to JavaScript
- cjhanks, hackernews Apr 18, 2017

Comparison with others

D, Rust, Jai, Zig
- out of scope for this talk
Go
- not really a *pinnacle* of abstraction and innovation :|
C++
- <optional> - 5k+ LOC for a T and a bool... safe_int - same horror story
- The next big thing: "Design by introspection" - Andrei Alexandrescu
Nim is one of the most logical paths forward
- on-par performance with C/C++ (compiles to them)
- some of the most easy interop with C/C++ ........ (compiles to them)
- uses any C/C++ compiler ........ (compiles to them)
- already quite far in terms of implementation
- meta-programming on steroids

Nim compilation model

nim c -d:release main.nim
- always compile only the main file, follow the imports
- whole program analysis
- a .c file for each .nim file in a "nimcache" (temp) folder (also .obj files)
- only referenced (imported) modules are compiled in the end
entire project is always "compiled" by Nim (currently no "minimal" rebuild)
- ~4-5 sec for the entire source of Nim - 135 files (without the C compiler)
- the C/C++ compiler rebuilds only changed files (takes a bit more time)
- will change when per-module caching is introduced - even faster!

# main.nim

import a

echo a()

# a.nim

import b

proc a*(): string =
  return from_b

# b.nim

let local = "B!"

let from_b* = local
          ^
# means "exported"

Nim to C/C++: nimbase.h

// nimbase.h

#define N_NIMCALL(rettype, name) rettype __fastcall name
#define N_CDECL(rettype, name) rettype __cdecl name
//...
#define N_NIMCALL_PTR(rettype, name) rettype (__fastcall *name)
//...
#define N_LIB_PRIVATE __attribute__((visibility("hidden")))
//...
#define N_LIB_EXPORT  extern __declspec(dllexport)
//...
#define STRING_LITERAL(name, str, length) \
   static const struct {                  \
     TGenericSeq Sup;                     \
     char data[(length) + 1];             \
  } name = {{length, (int) ((unsigned)length | NIM_STRLIT_FLAG)}, str}

included by all .c/.cpp files

handles different platforms - convenience macros

Nim procs to C/C++

proc foo() =
  echo "hello"

foo()

#include <nimbase.h>

// forward declarations / type definitions / constants section
struct TGenericSeq { int len; int reserved; };
struct NimStringDesc : public TGenericSeq { ... };
typedef NimStringDesc* tyArray_nHXaesL0DJZHyVS07ARPRA[1];

STRING_LITERAL(TM_r9bkcJ6PRJ5n7ORNxxJ5ryg_3, "hello", 5); // << string literal
NIM_CONST tyArray_nHXaesL0DJZHyVS07ARPRA TM_r9bkcJ6PRJ5n7ORNxxJ5ryg_2 =
    {((NimStringDesc*) &TM_r9bkcJ6PRJ5n7ORNxxJ5ryg_3)};

N_LIB_PRIVATE N_NIMCALL(void, foo_iineYNh8S9cE6Ry7dr2Tz2A)(void); // << fwd decl

// definition section
N_LIB_PRIVATE N_NIMCALL(void, foo_iineYNh8S9cE6Ry7dr2Tz2A)(void) { // << def
    echoBinSafe(TM_r9bkcJ6PRJ5n7ORNxxJ5ryg_2, 1); // the echo call
}

// code execution section
foo_iineYNh8S9cE6Ry7dr2Tz2A(); // << call

Nim types to C/C++

type
  MyData = object
    answer: int
    ready: bool
proc newData(): MyData = return MyData(answer: 42, ready: true)
echo newData().answer

// forward declarations / type definitions / constants section
struct tyObject_MyData {
    int answer;
    bool ready;
};
// definition section
N_LIB_PRIVATE N_NIMCALL(tyObject_MyData, newData)(void) {
    tyObject_MyData result; // always an implicit "result"
    nimZeroMem((void*)(&result), sizeof(tyObject_MyData));
    result.answer = ((int) 42);
    result.ready = true;
    return result;
}

// code execution section
tyObject_MyData T2_;
T2_ = newData(); // << call
//...

Nim closures to C/C++ (resumable funcs)

iterator closure_iter*(): int {.closure.} = # a resumable function
  var x = 1
  while x < 10:
    yield x
    inc x
for i in closure_iter(): echo i

struct state_type : public RootObj {
    int colonstate_; // state progress - there are some GOTOs using this
    int x1; // the state
};

struct closure_type {
    N_NIMCALL_PTR(int, c_ptr) (void* e_ptr); // function ptr
    void* e_ptr; // environment ptr
};

N_LIB_PRIVATE N_CLOSURE(int, func)(void* e_ptr) { // def omitted for simplicity

state_type st; // the state
closure_type local; // the closure
local.c_ptr = func; // assign the func
local.e_ptr = &st; // assign environment
//...
i = local.c_ptr(local.e_ptr); // the call in the loop

Nim compilation to C/C++: a BIG win

smaller scope for the compiler
all the cutting-edge optimization for C/C++ for free
out-of-the-box support for tons of platforms
easiest C/C++ interop possible
exceptions - reusing those of C++ when using that backend
nim to C/C++ code mapping with #line directives for debuggers
no generated headers for the exported parts of modules
each .c/.cpp file contains everything (and only what) it needs
- forward declarations for external functions
- type definitions
each .c/.cpp file includes nimbase.h and a few C stdlib headers
high level macros & templates => simple structs and functions

Interfacing with C/C++

Foreign Function Interface

proc printf(formatstr: cstring)
    {.header: "<stdio.h>", importc: "printf", varargs.}

{.emit: """
using namespace core;
""".}

{.compile: "logic.c".}

other pragmas - for use in Nim:

We can also call Nim code from C/C++:

# fib.nim
proc fib(a: cint): cint {.exportc.} # do not mangle

nim c --noMain --noLinking --header:fib.h fib.nim

// user.c
#include <fib.h>

Interfacing with C/C++

c2nim tool - generate C/C++ bindings for Nim

type
  StdMap {.importcpp: "std::map", header: "<map>".} [K, V] = object
proc `[]=`[K, V](this: var StdMap[K, V]; key: K; val: V) {.
  importcpp: "#[#] = #", header: "<map>".}

var x: StdMap[cint, cdouble]
x[6] = 91.4

std::map<int, double> x;
x[6] = 91.4;

C++ template constructs

Generated C++

much faster iteration times
- no need to restart the program - can preserve state
less need for a scripting language
- no need for a virtual machine
- no binding layer
- code in one language
can hack something quickly
- introspection, queries
  - debuggers aren't infinitely powerful
- fine-tuning values
interactive (REPL-like): very useful for exploration and teaching

Runtime compilation - WHY

replacing entire functions: using shared libraries OR hot-patching:
- possible for decades - but not widely used
- usually quite intrusive (interfaces, constraints, complicated setup)
- in game engines: Unreal, others...
- hot-patching (with very little setup): Live++, Recode
- Visual Studio "Edit & Continue" - 0 setup, but limited
- https://github.com/crosire/blink
- https://github.com/ddovod/jet-live
- http://bit.ly/runtime-compilation-alternatives << "one link to rule them all"
interactive: REPL-like
- cling - by researchers at CERN - built on top of LLVM
  - inspector, Jupiter
  - hard to integrate in a platform/compiler agnostic way
- RCRL - basically a hack - the inspiration for the Nim implementation

Runtime compilation for C/C++: HOW

Replace "compiling" with "restarting"

Hot code-reloading (HCR) in Nim

inspired by a hacky REPL for C++ (called RCRL - by me)
https://github.com/nim-lang/Nim/issues/8927
- mentored by Zahary
compile with --hotCodeReloading:on
need also 2 .dlls (the HCR runtime + the GC of Nim)

# main.nim

import hotcodereloading # for reload
import other

while true:
  echo readLine(stdin) # pause
  performCodeReload()  # reload
  echo getInt()        # call

# other.nim

import hotcodereloading # for after handler

var glob = 42

proc getInt*(): int = return glob + 1 # exported

afterCodeReload:
  glob = 666

built as an .exe/.dll depending on the project type

built as a reloadable .dll

ends up in the "nimcache"

Effects of HCR

all interaction between .nim modules => through pointers
functions - changes:
- forward declarations become function pointers
- definitions get "_actual" as a suffix
- pointers are assigned the "_actual" on startup
- calls stay the same (pointer has the same name)
globals - changes:
- turned into pointers
- allocated on the heap and initialized on startup
  - state is preserved when reloading
- dereferenced wherever used

Effects of HCR

// fwd decl/globals section
static N_NIMCALL_PTR(int, getInt_omy6T2FkprLEReOy2ITmIQ)(void);
static int* glob_v1zK9aUOu9aNNcsxruuK8NdA;

// definitions
N_LIB_PRIVATE N_NIMCALL(int, getInt_omy6T2FkprLEReOy2ITmIQ_actual)(void) {
	int result;                                       // ^^ the suffix
	result = (*glob_v1zK9aUOu9aNNcsxruuK8NdA);
	return result;
}

// usage
(*glob_v1zK9aUOu9aNNcsxruuK8NdA) = getInt_omy6T2FkprLEReOy2ITmIQ();

// init on startup (naive)
glob_v1zK9aUOu9aNNcsxruuK8NdA = new int(42);
getInt_omy6T2FkprLEReOy2ITmIQ = getInt_omy6T2FkprLEReOy2ITmIQ_actual

Trampolines

Initialization

// naive
glob_v1zK9aUOu9aNNcsxruuK8NdA = new int(42);
getInt_omy6T2FkprLEReOy2ITmIQ = getInt_omy6T2FkprLEReOy2ITmIQ_actual

// reality
getInt_omy6T2FkprLEReOy2ITmIQ = (tyProc_vVu2P82aVLv9c8X0xbI1NJw) hcrRegisterProc(
    "D:\\play\\nimcache/play.cpp.dll",            // "domain" (AKA module)
    "getInt_omy6T2FkprLEReOy2ITmIQ",              // "key"
    (void*)getInt_omy6T2FkprLEReOy2ITmIQ_actual); // the real function

if(hcrRegisterGlobal("D:\\play\\nimcache/play.cpp.dll",        // "domain" (AKA module)
                     "glob_v1zK9aUOu9aNNcsxruuK8NdA",          // "key"
                     sizeof((*glob_v1zK9aUOu9aNNcsxruuK8NdA)), // size for allocation
                     NULL, // for the GC - simple integer is simple, so NULL
                     (void**)&glob_v1zK9aUOu9aNNcsxruuK8NdA))  // address to pointer
{
    // hcrRegisterGlobal returns "true" only if not already inited
    (*glob_v1zK9aUOu9aNNcsxruuK8NdA) = ((int) 42); // init with value (or side effects)
}

Initialization

the HCR.dll runtime holds pointers to all globals/functions
hcrRegisterProc
- allocates executable memory (a few bytes)
- writes a jump instruction (trampoline) to the "_actual"
- returns an address to the trampoline
- this way "_actual" can be changed on reloading
  - changed by calling it again with a different address
  - all pointers to the trampoline stay the same
all symbols are registered per "domain" (.dll)
- no name clashes (even though they are mangled...)
- better management - can remove all symbols for module X

Initialization

# main.nim

import a, b

echo from_a()
echo from_b()

# a.nim

import b

proc from_a*(): string =
  result = "A!"
  result.add from_b()

# b.nim

proc from_b*(): string =
  return "B!"

main.exe loads the hcr.dll (and the Nim GC in rtl.dll)
main.exe calls init() from hcr.dll and passes a list of imports (a, b)
hcr.dll loads a.dll and gets a list of imports (b)
hcr.dll loads b.dll and fully initializes it (it has no imports)
1. registers from_b() and does nothing else
hcr.dll fully initializes a.dll
1. registers from_a() and gets the address for from_b()
hcr.dll skips b.dll (part of the imports of main.exe) since it is already initialized
main.exe is initialized
1. gets the addresses for from_a() and from_b()
2. executes the top-level code (the 2 echo statements)

a DFS traversal with POST visit
when module A imports a symbol from B
- symbol is first registered in B
- symbol is "gotten" in A after B is inited
basically a custom dynamic linker :|
imports are discovered on-the-go
HCR.dll constructs a tree of imports and maintains it
many details omitted
- initialization is broken into multiple passes
  - registration of type infos (for the GC) is a pre-pass
each .dll exports just a few functions which the HCR.dll uses
- getImports(), and the ones for the passes

Initialization

when we call performCodeReload():

HCR.dll will check hasAnyModuleChanged()
- basically scanning if any .dll has been modified (timestamp)
changes shouldn't affect .dll files which are part of the current active callstack when reload() is called! or crash :|
- ==> main module can never be reloaded
execute the "beforeCodeReload" handlers if about to reload
in a DFS traversal, for each modified module:
- same as the init - get its imports, load them (if changed or new), init everything in proper order
  - supports discovery of new imports!
  - also removes no longer referenced modules and their symbols
execute the "afterCodeReload" handlers

Reloading

Reloading - handlers

# main.nim
import a, b, hotcodereloading

beforeCodeReload:
  echo "before main"
afterCodeReload:
  echo "after main"

# a.nim
import b, hotcodereloading

beforeCodeReload:
  echo "before a"
afterCodeReload:
  echo "after a"

# b.nim
import hotcodereloading

beforeCodeReload:
  echo "before b"
afterCodeReload:
  echo "after b"

DFS traversal with POST visit
handlers can be added/removed
can be used to update globals
fine-grained control:
- hasModuleChanged(<module>)

only A changes => all handlers are executed on reload:

before b
before a
before main
after b
after a
after main

Reloading - global scope

top-level code (global scope) is executed only on initial load
- for new top-level code use before/after handlers
changing the initializer of a global doesn't do anything
- use a before/after handler
- or remove the global entirely, reload, and re-add it
  - brand new symbol!
new globals can be added - and will be initialized properly

The initial HCR example revisited

# main.nim

import hotcodereloading # for reload
import other

while true:
  echo readLine(stdin) # pause
  performCodeReload() # reload
  echo getInt() # call

# other.nim

import hotcodereloading # for after handler

var glob = 42

proc getInt*(): int = return glob # exported

afterCodeReload:
  glob = 666

Makes more sense now, doesn't it?

LIVE
DEMO

Encountered problems

processes lock loaded .dll files in the filesystem on Windows
- when reloading we copy x.dll to x_copy.dll and load the copy
changing module X can affect module Y
- such changes shouldn't reach the main module
- mangling of symbols being affected by attributes (purity)
- mangling affected by where "inline" functions get used first
- mangling affected by which module instantiates a generic
C vs C++
- missing forward declarations - fine in C!
- multiple identical forward declarations
  - multiple definitions of global function pointers - fine in C!

Visual Studio debug symbols - PDB drama

.dll/.exe have hardcoded paths to the .pdb (copying the .dll doesn't matter)
the VS Debugger keeps the .pdb files locked for .dlls even after unloaded

someone managed to close the file handles to no longer needed .pdb files (.dll has been unloaded) to the external VS debugger process (live++)
embed the debug info in the actual binaries just like on unix
- /Z7 embeds it in .obj files but not for the final .dll/.exe when linking them
different names for the .pdb using /PDB:<filename> (with the date/time (including milliseconds) as a suffix)
- the "hardcoded" paths to .pdb files are always different
- try to delete all <dll_name>_*.pdb files for a given .dll when linking
  - failure to delete them means the VS debugger still holds them locked
  - links: l1, l2, l3, l4

solutions:

HCR performance

snappy compression algorithm - x2-x4 times slower
- for reference: zlib (c code) to javascript (asm.js) ==> x2 slow down
calls within a translation unit are direct (the "_actual" version gets called)
calls between modules => indirection: pointer to function
- + additional jump from trampoline to actual function
link time optimization (AKA whole program optimization) cannot help
- devirtualization techniques are not applicable either
compactness in memory VS a single binary => instruction cache misses
/hotpatch for MSVC and Live++ (which are faster):
- not going through function pointers
- by default there are no jumps in the function preamble (padding)
slowdown depends a lot on the type/scale of software - x2 to x5...

HCR performance

possible optimizations:

write more "inline" procs
- their body is emitted wherever used => skip indirections
pragmas for excluding files (extension of the first point in this list)
- register the module procs but no indirections between them
relocate all code from loaded binaries close in memory?
PLOT TWIST!
- debug builds are currently affected a lot less (<x2 slowdown)
  - HCR is mainly for development => probably debug builds

HCR TODO

Nim stdlib has trouble compiling with the GC as a separate SO
- "-d:useNimRtl" needs to be enabled for all compiler tests
- currently no real-world project can be built with HCR
detecting type changes
- error when detected
- OR ability for users to handle it (migrate data)
check if "reload" would affect functions from the current call stack
expose state for outside manipulation with interactive speeds
- imagine a slider in the IDE for a variable or a color picker widget
performance & bug fixes

HCR Implementation choice

pros
- any modern (desktop) OS supports dynamic libraries
- works with any C/C++ compiler
- near-native speeds
- final binaries are debuggable
- a REPL is easily built on top of this
- (arguably) less complex than using LLVM / JIT / whatever
- changes are isolated (only the C backend which is a few files)
- program can be changed in (almost) any way
- novel approach - someone had to try it
cons
- not as optimal as the /hotpatch for MSVC or Live++
- (arguably) more complex than using LLVM / JIT / whatever
- not sure how NLVM (Nim on top of LLVM) will support HCR

REPL - Read Eval Print Loop

interpreted languages have it (JavaScript, Python, etc.)
consoles/shells - cmd.exe, bash
can iteratively append/execute code (definitions, side effects, etc.)
education, scientific community, rapid prototyping of any kind

REPL/Nim quote

Nim is the language I have always thought was a brilliant idea that I never get to use. It's a shame.
Nim is to C/C++ as CoffeeScript is to JavaScript. A highly extensible template language atop a portable language with libraries for practically everything.
So why haven't I hopped on the bandwagon? Outside of C++, C, and Fortran - the only way I have ever learned a new language is through using a REPL. How much of Python's and MATLAB's (and maybe even Julia's) success is due to having a brilliant REPL?
I am not complaining, and I do not have any free time to fix it. But man... if Nim just had a killer REPL that allowed me to slowly learn the language properly while not being blocked from my daily work... it would be just killer!

cjhanks on Apr 18, 2017

https://news.ycombinator.com/item?id=14143521

REPL on top of HCR

2 files:

main module
- has the main loop
- handles code submissions
imported file
- gets modified based on submissions
- rebuilt + reloaded

Talk abstract was a lie! didn't get to implementing it in time...

should be well below half a second

REPL on top of HCR

you submit this:

import tables

var a = {1: "one", 2: "two"}.toTable

echo a

and it gets translated to this:

import hotcodereloading # for the before/after handlers

import tables

var a = {1: "one", 2: "two"}.toTable

afterCodeReload:
  echo a

REPL on top of HCR

later you append:

let b = a

echo b

and it gets translated to this:

import hotcodereloading # for the before/after handlers

import tables

var a = {1: "one", 2: "two"}.toTable

let b = a # the new code

# only the new side effects are still present
afterCodeReload:
  echo b

Jupyter kernel

yesterday on ACCU: Interactive C++ : Meet Jupyter / Cling - The data scientist’s geeky younger sibling - by Neil Horlock
A Jupyter Notebook is an interactive document - a collaborative platform for prototyping, experimentation and analysis
Mix and share: code, text, data, computation and visualization
"Notebooks are the most popular tool for working with data at Netflix."
Nim REPL => Nim Jupyter kernel

The road ahead for Nim

version 1.0 - promise of stability
compiler cache for unchanged modules
- because compilation starts always from the main module
- of great benefit for HCR/REPL
more features
better tooling
better docs
taking over the world
get involved - still in early stages - you can have an impact

Q&A

Slides: https://slides.com/onqtam/nim_hot_code_reloading
Blog: http://onqtam.github.io
GitHub: https://github.com/onqtam
Twitter: @KirilovVik
E-Mail: vik.kirilov@gmail.com