Revisiting AVSampleBufferDisplayLayer on iOS 11

When I tried using iOS’s AVSampleBufferDisplayLayer in OGVKit last year, I had a few problems. Most notably a rendering bug on one device, and inability to display frames with 4:4:4 subsampling (higher chroma quality).

Since the rendering path I used instead is OpenGL-based, and OpenGL is being deprecated in iOS 12… figured it might be worth another look rather than converting the shader and rendering code to Metal.

The rendering bug that was striking at 360p on some devices was fixed by Apple in iOS 11 (thanks all!), so that’s a nice improvement!

Fiddling around with the available pixel formats, I found that while 4:4:4 YCbCr still doesn’t work, 4:4:4:4 AYCbCr does work in iOS 11 and iOS 12 beta!

First I swapped out the pixel format without adjusting my data format conversions, which produced a naturally corrupt image:

Then I switched up my SIMD-accelerated sample format conversion to combine the three planes of data and a fourth fixed alpha value, but messed it up and got this:

Turned out to be that I’d only filled out 4 items of a 16×8 vector literal. Durrrr. :) Fixed that and it’s now preeeeetty:

(The sample video is a conversion of a silly GIF, with lots of colored text and sharp edges that degrade visibly at 4:2:0 and 4:2:2 subsampling. I believe the name of the original file was “yo-im-hacking-furiously-dude.gif”)

A remaining downside is that on my old 32-bit devices stuck on iOS 9 and iOS 10, the 4:2:2 and 4:4:4 samples don’t play. So either I need to keep the OpenGL code path for them, or just document it won’t work, or maybe do a runtime check and downsample to 4:2:0.

Mobile video encoding tradeoffs

I spent a little time yesterday and today poking at an old project to encode video into the WebM format we use at Wikipedia on your iPhone or iPad so you could, potentially, take video and upload it directly.

The default encoding settings were mmmuuuucccchhhh tttoooooo ssslllooowww to be practical, so I had to tune it to use a faster configuration. But to maintain quality, you have to bump up the bitrate (and thus file size) significantly.

This same tradeoff is made in the hardware video encoders in the device, too! When you’re making a regular camera recording the bitrate is actually several times higher than it would be on a typical download/stream of the same video from YouTube/Netflix/etc. You just don’t have the luxury of the extra encoding time on a modest mobile chip, especially not if you’re recording live.

Video files and thumbnails in MediaWiki’s TimedMediaHandler

One of my ongoing tasks is modernizing and fixing up issues with TimedMediaHandler, the MediaWiki extension we use to handle audio and video files on Wikipedia and Wikimedia Commons.

One of many “little cuts” has been that the thumbnail frame of a video file defaults to the halfway point in the file, which is sometimes great and sometimes useless (such as a blank transition frame).  You can override the thumbnail to use for a given invocation in a wiki page, but can’t set a default to be used in places where no parameter is added… such as when a WebM or Ogg video got selected as the default poster image via the PageImages extension.

To resolve this, we’re reviving an old request to make it possible to set the default thumbnail frame time for a video.

I’ve got a patch partially working, using a parser function so you can put something like {{#thumbtime:90}} on the file page to set the default to 90 seconds in, or {{#thumbtime:35:17}} for 35 minutes and 17 seconds in, etc. The expanded time in seconds is saved as a page property on the File: page, then extracted and copied into the extended file metadata by a hook in TimedMediaHandler. The metadata gets queried when we the generate HTML harness for the video/thumbnail, and we pass the time down in the URL of the thumb just as we would if you had specifically requested that time.

Currently however it’s incompatible with CommonsMetadata extension, which we kind of need. ;) So I’ve got to figure that out and either get them to play nice, or store the data in a different way.

(Just using the page props locally isn’t enough because we have to export the data from Wikimedia Commons to other sites.)

emscripten fun: libffi on WebAssembly part 2

A couple weeks ago I started porting libffi to work on emscripten’s WebAssembly output mode, in the hopes of eventually getting glib or even gtk+ working in browser-compiled apps.

I got it back set up on my Linux box (moved from my Windows/WSL environment) and spent a little time this morning fixing a few more things.

First, 64-bit integers are now supported in call arguments and return values. This requires jumping through some hoops because JavaScript doesn’t yet have large ints, and so the C<->JS ABI divides them into two 32-bit words. For args, you get two args in a row, first the low then the high word. For return values, the low word is returned and the high word is set in a module-global variable which can be retrieved through Module.getTemp0().

Second, I fixed a bunch of float tests that failed because currently the test suite is being built and run in asm.js mode instead of Wasm. There’s a slight ABI difference again, in that in asm.js floats are always promoted to doubles in args  and return values. Checking Module.usingWasm and using the appropriate “f” or “d” in the signature resolves this.

Note the tests are in asm.js mode because the DejaGnu test runner seems to want to run the executables directly, so I tricked emcc into thinking it’s being run from emconfigure with EMCONFIGURE_JS=2 to force it to build a executable with a Node shebang instead of a native executable (whaaaaa) or bare JS that’s not executable… and currently emcc forces Wasm off in this mode, though I think it should not need to. Will look further into that.

Still have hundreds of failing tests as I haven’t implemented structs or closure generation, but it’s progress!

JavaScript engine internals part 5: conclusions

Thoughts from the last couple weeks of prototyping a JavaScript-like scripting language engine core

Scripting inside WebAssembly

A JavaScript-like garbage-collected dynamically-typed scripting engine runtime can indeed be implemented in WebAssembly, with relatively sensible performance for ahead-of-time compiled code. (C++/emscripten-based prototype showing within 3-4x of native JS in node for a floating-point math test without too much micro-optimization yet.)

  • Garbage collection keepalive requires explicit handling of a GC-scanable stack. There’s no way to scan the native stack, as not everything spills from locals to the emscripten-managed stack in linear memory.
  • Tricks like NaN boxing work for squishing variant values into a 64-bit word, and can perform reasonably well.
  • C++/emscripten prototype engine works well enough, but the build is large using standard library (for hashmaps, vectors, strings, and some string formatting at present).
  • It may be worth investigating a dynamic linking system where the engine and the compiled code are separate .wasm files, but this may have performance implications on calls into the engine.

Sandboxing

  • Memory usage can be strictly limited at compile or instantiation time, but moderate DoS vectors exist in long loops and deep recursion (can mitigate with Worker thread isolation)
  • Sandboxing the whole runtime so you can swap out just the .wasm file with a user-provided engine would be a lot more work (avoiding emscripten’s JS-side wrappers and functions or reducing them to a bare safe interface implemented for all engines).
  • Using separate .wasm modules for the engine and the program would expose engine internals in linear memory to the program’s module. Be aware that this is not a secure boundary unless the compiler is trusted.

Areas for further work

Adding an actual ahead-of-time compiler instead of hand-translating JS to C++ source would be a lot more work.

  • Would have to decide what tools to use to parse JS source
  • What kind of lifecycle and toolchain to use (does this get compiled server-side? client-side?)
  • What to emit (C++ as input to emscripten, or direct to LLVM, or super direct to WebAssembly?)

An in-engine interpreter for a REPL loop would be a different sort of work, and might or might not be able to share code with compiler for parsing etc.

How would compiled scripts be debugged, other than by slipping in console.log() invocations and watching error messages?

  • Is it possible to emit source maps for use in browser developer tools so can step through watching the original source instead of Wasm disassembly? How workable is this with Wasm at this point?

Need to imagine some example plugin APIs that could be used for scenarios such as:

  • SVG diagram with some event handlers and animations attached to it
  • SVG diagram or Canvas implementing Logo-like turtle graphics with simple input commands
  • Text editor plugin that describes a simple search-and-replace dialog and implementation

Immediate plans

I’ll probably keep tinkering with this because it’s fun as heck to research how other scripting engines work and see how I can make it myself. ;) But it’s a looooong way from being usable, and I wouldn’t likely be able to push it all the way there any time soon myself.

Will probably prototype the “other side” of the plugin scenarios above and see if I can get more folks at Wikimedia excited about sandboxed plugins and widgets with secure, narrow, versionable interfaces. I need to think of a clever name for it! ;)

In the meantime …. other tasks beckon! TimedMediaHandler needs its fixes finished…

JavaScript engine internals part 4: NaN-boxing vs heap boxing in Mandelbrot

One of my favorite programming samples is calculating a visualization of the Mandelbrot set, a famous fractal. I’ve added it to my experimental JS-like scripting engine to exercise the floating-point performance:

Using the tagged-pointer scheme in my earlier post for storage of variant local types, there was the disadvantage that double-precision floating point numbers (64 bits wide) would not fit in the 31 bits of tagged-pointer space, so they had to be “boxed” (allocated on the heap as an object). This made for runtimes on the Mandelbrot set demo approximately 200x slower than the original JS running in node. Ouch!

I decided to reimplement NaN boxing (which allows storing full 64-bit doubles or a pointer or an int32, all in a single 64-bit word), this time following the JavaScriptCore method that’s easier to unbox pointers versus the SpiderMonkey method. Now I can flip my code between NaN-boxing and heap-boxing with a compile-time #define and compare!

original js: 9 ms

native w/ NaN box: 48 ms
wasm with NaN box: 71 ms

native w/ box doubles: 2677 ms
wasm with box doubles: 1794 ms

NaN-boxing *screams* versus heap-boxing. In part it’s the indirection, but I think it’s mostly having to allocate a hojillion objects and then clean them up, as every temporary result in calculations gets turned into a live object and then garbage-collected after a while. (Note that the native build on my MacBook Pro is slower than the wasm build with heap boxing, probably due to a different allocator and 32-bit vs 64-bit pointers!)

But do we really have to do all that temporary allocation? Most math operators only return a specific type (-, *, and / always return doubles; + can return a string depending on its input). For now the JS code is hand-translated to C++ using operator overloads, so having my C++ operator overloads return doubles when possible, and accept them when possible, lets temporaries that can only be doubles skip the type-checking and boxing/unboxing overhead.

original js: 9 ms

native w/ NaN box: 17 ms
wasm with NaN box: 28 ms

native w/ box doubles: 467 ms
wasm with box doubles: 248 ms

Boxing doubles on the heap is still really slow, but a lot less slow thanks to sidestepping it on many intermediate values… And with the 64-bit NaN-boxing we’re only about 3x slower than native node when running in wasm in node, a big improvement that I’m really happy with!

This makes me feel that 64-bit NaN boxing is the way to go; it’s only slightly more complex than tagging a pointer-sized word and it makes floating point math a ton faster when it’s needed, which sometimes it just is.

Floating point fun: iterating the number space

As a joke the other day I posted a code snippet on Mastodon:

for (var i = -Infinity; i < Infinity; i++) {
  /* this should cover all possibilities */
}

Such an infinite loop is not possible, as adding 1 to -Infinity just gives you -Infinity again, and you never get anywhere. ;)

But even if you started from the minimum representable individual number in a double-precision floating point you wouldn’t actually iterate anywhere because there’s not enough precision to represent the added 1.

Which got me thinking, how *would* you iterate over the entire set of floating point numbers between -Infinity to +Infinity, knowing that the intervals between the numbers are variable?

Well if you want to get everything, including non-integral values, you can iterate over the available range of exponents and fractional values, and either bit-shift them together manually (ick!) or multiply/add/pow them together in floating point land.

Something like this:

function iterate_all_floats(callback) {
  // exponent is additively biased in the binary rep,
  // so we go from negative to positive vals.
  let min_exp = -(Math.pow(2, 10) - 2); // all binary 1s are reserved
  let max_exp = Math.pow(2, 10) - 1;

  let min_frac = 0;
  let max_frac = Math.pow(2, 52) - 1;
  let frac_divisor = Math.pow(2, 52);

  let min_sub = 0;
  let max_sub = Math.pow(2, 52) - 1;
  let sub_divisor = Math.pow(2, 52);

  callback(-Infinity);

  // Start from the biggest negative numbers and go up towards -0
  for (let exp = max_exp; exp >= min_exp; exp--) {
    for (let frac = max_frac; frac >= min_frac; frac--) {
      callback(-(1.0 + frac / frac_divisor) * Math.pow(2, exp));
    }
  }

  // Negative subnormals and -0
  for (let sub = max_sub; sub >= min_sub; sub--) {
    callback(-(sub / sub_divisor) * Math.pow(2, -1022));
  }

  // +0 and positive subnormals
  for (let sub = min_sub; sub <= max_sub; sub++) {
    callback((sub / sub_divisor) * Math.pow(2, -1022));
  }

  // Start from just past +0 and move up to the biggest numbers
  for (let exp = max_exp; exp <= max_exp; exp++) {
    for (let frac = min_frac; frac <= max_frac; frac++) {
      callback((1.0 + frac / frac_divisor) * Math.pow(2, exp));
    }
  }

  callback(Infinity);
}

iterate_all_floats(function(f) {
  // Convert to binary for display
  let floats = new Float64Array(1);
  let bytes = new Uint8Array(floats.buffer);
  let str = '';
  floats[0] = f;
  for (let i = 3; i >= 0; i--) {
    let b = bytes[i];
    for (let bit = 7; bit >= 0; bit--) {
      let val = (b >> bit) & 1;
      str += val;
    }
  }
  console.log(str, f);
});

Enjoy!

JavaScript engine internals part 3: garbage collection

Another part of implementing a JavaScript-like engine to run inside a WebAssembly sandbox is memory management: JS requires garbage collection because objects may hold references to each other, creating cycles that simpler schemes like reference counting can’t break.

Mark and sweep

Most JS engines implement some form of mark-and-sweep garbage collection, usually enhanced with various fancy features we won’t get into here (moving/compaction, generational sweeps, etc).

The mark phase involves tracing your way through the object graph. Starting with globals and objects that are held in variables, you mark the object, setting a flag unique to that object. For any references it holds to other objects, you repeat — whenever you reach an object that’s already been marked you can skip over to the next one without re-tracing its references.

Any object that didn’t get marked is considered unreachable, and gets deleted in the sweep phase. This requires having some way to iterate through all allocated objects, which in my experimental engine I’ve temporarily implemented as a C++ std::unordered_set (a hash map with no payload) which new objects are always added to. This should eventually be replaced with a custom allocator that knows how to iterate through its own objects, or something. :)

The stack

Traditional C/C++ programs store some of their local variables (those that aren’t optimized into registers) in a region of memory known as the stack: a pointer to the stack top is incremented and decremented as functions run to provide storage space, and the same stack is used to store return addresses for function calls (a cause of many security problems).

Some garbage collection systems like Boehm libgc take advantage of this, and scan the stack looking for object references to keep alive during the mark phase. Clever!

The WebAssembly environment has stricter memory safety: byte-addressable “linear memory” has a limited range and you can only access local variables in the stack frame through compile-time-addressed, pre-verified indexes. There’s no way to access a calling function’s stack frames to check what it had saved in its local variables…

When using the emscripten compiler, there is a second, more C/C++ like stack maintained inside the linear memory for variables and data structures that can’t fit in WebAssembly locals (for instance structure types larger than a single word, or anything that has its address taken and a pointer passed somewhere). Theoretically we could scan through that looking for our pointer values, but there may be false positives and that could be trouble — you wouldn’t want to accidentally try to traverse the object references inside a pointer that was actually a random integer or something!

Stacks, scopes, and smart pointers

For now I’ve implemented my own secondary stack which holds only encoded JS values, and modeled JS local variable bindings as C++ pointers to the value on the stack. When a garbage collection happens, the known globals and anything referenced on the stack are marked and traversed, so objects considered in use stay alive.

The way the stack is managed in the C++ engine code is inspired by V8’s Handle & HandleScope/EscapableHandleScope system. It requires a little bit of explicit discipline, and could probably be further improved in terms of the classes making usage correct, but seems to work correctly.

At the beginning of each function, a Scope or ScopeRetVal object is instantiated:

void someFunc() {
  Scope scope;
  // do stuff
}

This saves the stack position, and restores it to its original state when the function exits — so you don’t have to manually remember to pop. That’ll free up any of your locally allocated values for garbage collection.

Any given variable binding is either a ‘Binding’ (alias to Val*) to an existing, elsewhere-managed Val, or a ‘Local’ “smart pointer” object when compiles down to a pointer, but with a constructor that pushes a slot onto the stack.

Arguments can be passed as Local instances as well, constructed on the parent frame and eventually destructed with it, or if you know they’re GC-rooted already for sure you can pass a direct binding.

The big trick is return values, which I needed inspiration from V8 to figure out. If you try to return a Local, it’ll be cleaned up from your own scope before you return — and if you try to return a Val directly it won’t be GC-rooted and can disappear unexpectedly.

RetVal somefunc() {
  ScopeRetVal scope;
  // do stuff
  return scope.escape(new String("keep it safe!"));
}

Functions return a RetVal which doesn’t push, instead of a Local which does. A variant of the Scope, here ScopeRetVal (V8 uses the awesomer name EscapableHandleScope!) first reserves space on the parent scope for that return value. When you return, you have the scope insert the return val into the saved slot, and then the destructor restores the stack to the saved spot _after_ the return value, safely still alive.

Without this trick, return values could end up de-allocated during code like:

var retval = a() + b();

translated to C++ as

Local retval;
*retval = *(a->call(Null(), {})) + *(b->call(Null(), {}));

where the return value from a() gets de-allocated during an allocation inside the call to b().

 

JavaScript engine internals part 2: tagged pointers

In my last post I explored NaN-boxing in JavaScript engines, showing how Firefox and Webkit manage to squish multiple types, including pointers and floating point, into a single 64-bit word.

In my WebAssembly-based JavaScript-like engine experiment I initially implemented NaN boxing similarly to how Firefox/SpiderMonkey do it. It works nicely, but has some downsides:

  • Extra operations needed to unmask a pointer when dealing with lots of objects
  • In-memory values which are pointers have to be manually masked when peeking in the debugger. Kind of a pain.
  • There’s no equivalent aliasing system available in the proposal for native garbage-collected types for WebAssembly.

Inspired by the GC proposal, I’ve tried switching from the 64-bit NaN boxing to native pointer-size (32-bit for Wasm32) words with a minimal tagging system to store integers, which is equivalent to what is available there.

Tagged pointers

So what’s a pointer anyway? It’s a number that points (ha!) to a location in memory where you store stuff. This has a very important consequence:

Because stuff has to be laid out in memory in a way it can be efficiently accessed, structs and classes containing pointers are themselves always located on a pointer-size-aligned boundary. That is, if pointers are 32 bits wide (4 bytes), the structs have to be at a memory location that’s evenly divisible by 4, or it won’t be efficiently accessible.

This means that every pointer to every garbage-collected object is a number divisible by 4… which means its bottom 2 bits are always 0 in binary representation.

Thus, if your value has the bottom bit set, it’s definitely not a pointer to an object. Knowing that, we can divide the 32-bit number space into 32-bit object pointers on the one hand, and on the other hand either 1 tag bit with a 31-bit payload, or 2 tag bits with a 30-bit payload.

The Wasm GC proposal includes an “int31ref” type which can be stored in an “anyref” slot alongside other GC’d reference types, which is meant to use such a scheme.

Implementing tags

In the Val wrapper class in my engine, I’ve adopted this system now and it’s got some trade-offs, but I kinda like it.

The tag check is now a single bit-mask against 1 rather than a giant constant like 0xffff000000000000, which feels cleaner. Integers can be promoted back to 32 bits by right-shifting one bit, while pointers can be consumed as-is once the test is confirmed (or if you know for sure they’re safe).

The biggest downside is that there’s simply not room in 31 bits for very large integers (abs value > 1 billion-ish) or for floating point numbers (64 bits, ay ay!) so they must be boxed on the heap.

Boxing

“Boxing” is the fun task of taking a nice atomic value type, and wrapping it in a giant heap-allocated object. Horribly inefficient if you use lot of them!

But, it can be automatically done in the accessors, and seems to work so far.

I would prefer to avoid using boxing for floats, but there’s just no room in the int space for them unless I do something crazy like using two tag bits and then storing a 24-bit or 16-bit float when a value can be represented that way… yikes! And you’d still have to box values that are out of range or would lose precision.

For the other JS types — undefined, null, boolean, etc — I’m using boxed types as well, but using a single well-known instance value. So all “undefined” values refer to the same Box<Undefined>* and all “booleans” refer to one of two Box<bool>*s, one for true and one for false. This avoids runtime allocation for those special types, while allowing them to be treated as objects for dynamic dispatch or making the type checks cheap pointer equality comparisons for static dispatch.

JavaScript engine internals: NaN-boxing

In researching code plugin sandboxing with WebAssembly, my thoughts naturally turned to how to let people actually program for that environment without forcing use of low-level C++ or such… so naturally, now I’m writing a small JavaScript-like runtime engine and compiler that targets WebAssembly. ;)

My explorations so far are themselves written in C++, with tiny JS samples hand-translated into C++ code that calls the runtime’s classes. If I continue the project, I may switch to directly emitting Wasm source from the translator, with an eye towards future Wasm improvements like native garbage-collected reference types. (Right now I have to implement garbage collection myself, which will be the subject of another blog post!)

The first step of implementing a JavaScript engine is to implement a representation for values, which is tricky because JS values can be any of several distinct types:

  • undefined
  • null
  • boolean
  • number (float64)
  • object reference (string, Symbol, Object, etc)

Many other dynamic languages like PHP and Python similarly allow a value to hold multiple competing types like this. There are two main ways to implement a value type like this.

Tagged unions

The first is to use a “tagged union” struct, something like this:

enum TypeTag {
  TAG_UNDEFINED,
  TAG_INT32,
  TAG_DOUBLE,
  ...
  TAG_OBJECT
};

struct Value {
  TypeTag tag;
  union {
    int32_t int32_val;
    double double_val;
    Object* object_val;
  };
};

Then add accessor/mutator functions which check the tag and perform appropriate accesses or conversions. This is simple, but has the downside of requiring 16 bytes per value (to maintain 8-byte alignment on the double-precision float or a 64-bit pointer or int64). That means most reads require two loads, writes require two stores, and you can’t pass a value in a register reliably. (WebAssembly’s local variables and function arguments are sort of like CPU registers in how they’re addressed, so would be the equivalent.)

To get around these, a common trick in high-performance dynamic language runtimes like LuaJIT and all major JavaScript engines is known as NaN boxing.

NaN boxing the SpiderMonkey way

Floating point numbers are a complex beast! A 64-bit double-precision float can be broken down into several sets of bits:

  • 1 bit for sign
  • 11 bits for exponent part
  • 52 bits for binary fractional part

If the exponent bits are all 1, then you have a special value:

  • if the fractional bits are all 0, then it represents infinity (positive or negative, depending on the sign bit)
  • if any fractional bit is 1, then it represents “not a number” (NaN), used as a special signal that an operation was invalid or a non-value is being represented.
  • and oh by the way, regular operations only ever produce a canonical NaN value with just the topmost fractional bit set.

The interesting part is that means you have an entire number space within the range of representable NaN values that will never be produced by legit math operations on legit numbers!

  • 1 bit for sign (set to 1 for a signalling NaN that won’t be produced by math ops)
  • 11 bits for exponent (all set to 1)
  • 1 bit to force it to NaN rather than Infinity if rest is 0s
  • 51 bits for tag and payload

Firefox’s SpiderMonkey engine divides this up into 4 bits of tag (16 types) and 47 bits of address, which is enough for x86_64 (where bit 48 will always be 0 in user space and bits 49-64 aren’t needed … yet) but apparently causes problems on AArch64 and Sparc64. (They work around this by deliberately mapping memory in the lower 47 bits of address space with mmap.)

I’ve chosen to experiment with a similar technique, but only using 3 bits of tag (8 types) to leave a clean 48 bits of address. Though when building to WebAssembly, you only need 32 bits of address so it doesn’t make much difference. ;)

When reading a value of as-yet-unknown type, you mask off the top X bits and compare against some constant values:

  • each type other than double-precision its has a particular tag value, so you just == for it
  • actual double-precision values can be detected quickly by forcing the sign bit on and then doing a uint64 compare against a cutoff value: any legit double including canonical NaNs will be <= 0b1111111111111000’0000000000000000’0000000000000000’0000000000000000 once sign/signalling bit is set.
  • if there are multiple type tags for pointer types, you can put them at the high end of the tag range and do a uint64 compare against a cutoff value

One downside is that every object dereference has to be masked from the stored value, and typical JavaScript code does a lot of objects. It’s a bit-op so it’s cheap, though… and in WebAssembly you can actually wrap the 64-bit value down to 32-bits without masking, since pointers are 32 bits. :)

JavaScriptCore style

WebKit’s JavaScriptCore uses a different technique on 64-bit native architecture.

When storing a double value, the bit pattern is treated as a uint64 and has a constant added to it to “rotate” the NaN patterns such that the top 1 bits in a range of the signaling NaNs turn into 0s.

  • object references have 0x0000 in the top 16 bits, meaning a 48-bit pointer can be read directly with a 64-bit read — no bit-masking penalty on dereference!
  • int32 values have 0xffff in the top 16 bits, making common integer operations easy with a mask
  • true/false/undefined/null are stored as special fake pointer values in the object pointer space
  • most double-precision float values sit in the 0x0001-0xfffe space, and have to subtract the constant to get back to their original bit pattern.
  • -Infinity has to be specially stored in the object space because the shift puts it in the wrong place

Sounds clever!

I haven’t checked what Chrome’s V8 and Edge’s ChakraCore do, but I believe they are similar to one or the other of the above.

Security considerations

The downside of all this magic bit-fiddling is that you have to be careful at API boundaries between the JS world and the external world it’s connected to to canonicalize all incoming NaN values in doubles, or it would be possible to create arbitrary object pointer references from within the scripting language — a huge security hole.

So, you know, be careful with that. ;)