The Apple Compact Unwinding Format: Documented and Explained

Aria Beingessner

June 9th, 2021

A few years ago, Apple introduced a new kind of unwinding info (I’ll explain what that is in the next section) – the “compact unwinding format”, which can be found in __unwind_info sections of binaries.

Clang also started preferentially emitting that format over DWARF CFI on Apple’s platforms. So to generate good backtraces on Apple platforms, you need to be able to parse and interpret compact unwinding tables.

Unfortunately, this format is defined only by its implementation in llvm. Notably these files include lots of useful details, but you need to do a lot of studying of the actual implementation to figure everything out:

Firefox’s crash reporter needs good backtraces on Apple platforms, so learning this format and implementing a parser/interpreter for it became my problem. As I often do, I proceeded to write complete documentation for the format, to the best of my ability.

This article is an expanded version of the enormous doc-comment I wrote for my implementation which is currently hosted in the docs for symbolic-debuginfo 8.2.0 (although it may eventually be moved into goblin).

This implementation is currently deployed in Firefox’s build system as part of the rust dump_syms tool. dump_syms takes native debuginfo and converts it into breakpad’s text-based symbol file format, which simplifies our server-side minidump analyzing.

This implementation and article are both based on llvm commit d480f968ad8b56d3ee4a6b6df5532d485b0ad01e and observations of MacOS around April-June 2021. They omit details on the “personality” and “LSDA” parts of the format, as those are only needed for things like running destructors which aren’t needed for simple backtraces.

Also: A BIG thank you to the various folks who helped me debug/test/review/investigate this format. While I wrote all the code and documentation, their input was invaluable in getting the work done properly.

§1 Background: Unwinding and Debug Info

(This section can be skipped entirely if you feel like you understand unwinding.)

So you have a running program. Functions call other functions, so you need a mechanism to save the caller’s state while the callee runs, and then restore that state when the callee completes. Some of this is defined by ABIs, but even with a defined ABI there’s a lot of wiggle room for the callee to mess around.

Say for instance that your ABI mandates that the caller’s rbx register must be preserved by the callee. At one extreme, the callee can just not use rbx but that’s usually pretty impractical. At the other extreme, the callee can save (push) rbx to the stack and restore (pop) it when it returns. Then the callee can use rbx as much as it pleases, at the cost of always needing to push and pop it.

In between these two extremes you find many Clever Tricks:

This is generally all Good Fun for an optimizer since these details are private to the implementation of the callee – it can do whatever it wants as long as it figures out how to restore the caller’s registers for every return in the function.

Except backtraces and unwinding exist.

§1.1 Unwinding

Whenever you decide to unwind (panic/throw) or generate a backtrace (crash, gdb, profiler trace), you are faced with the following problem:

You know the current address in the binary your program is executing (rip/pc) and the current stack pointer (rsp/sp). Now figure out:

  1. (optional) What function you’re executing (and ideally what line)
  2. What the return address is (roughly the address the caller was last executing)
  3. How to restore the caller’s registers (most importantly the stack pointer)

And then repeat this over and over until you have walked over the whole stack.

§1.2 Frame Pointer Unwinding (Standard Prologues)

Unwinding Steps 2 and 3 can be done very easily if all functions uses a standard “prologue” that maintains the frame pointer. On x64, the standard prologue is as follows:

  1. caller performs a CALL (implicitly PUSHes the return address (rip) to the stack)
  2. callee PUSHes rbp (the caller’s frame pointer – where its stack starts)
  3. callee sets rbp := rsp (initializes its own frame pointer)

If you know a function uses this calling convention, you can minimally complete Unwinding Steps 2 and 3 (only restoring rip and rsp) with the following:

rsp := rbp + 16
rip := *(rbp + 8)
rbp := *(rbp)

(Note that because that stack grows down, the stack pointer always points to the last thing that was pushed!)

or natively:

mov %rsp %rbp
pop %rbp

This convention kicks ass – it gives you perfect super fast backtraces at the cost of:

If you want to do Unwinding Step 3 properly and restore all the callee-save registers, you can also push all the callee-save registers in the prologue as well.

Good news: Apple Mandates Frame Pointers on ARM64.

Bad news: No one else does, and optimizers gonna optimize. Especially on x86 where register pressure is really high and ebp is so juicy and tempting.

§1.3 Debug Info

At this point you need to know what debug info is to do Unwinding Step 1, and do harder versions of Unwinding Steps 2 and 3.

Debug info is kind of a blanket term for “extra metadata about a binary”. There are many kinds of debug info and container formats. The 3 most notable binary container formats are:

Given the subject of this article, I will only really be focusing on Mach-O, but a lot of this will also apply to ELF, as they both heavily use DWARF.

These container formats are generally broken up into sections – things like constants (.rodata), instructions to execute (.text), and well, debug info (.debug_info). Apple likes to prefix these with two underscores, so I may interchangeably refer to .__eh_frame or .eh_frame – they’re the same.

The most common debug info format you will encounter is DWARF, which is actually more of a family of formats for the various kinds of debug info sections. If you’re dealing with any kind of Unix/Linux/BSD variant you’re gonna get a lot of DWARF, and Apple’s platforms are no exception.

For our purposes, the most notable DWARF sections are:

To properly use these sections you need to know one more detail: an executable is actually many disparate relocatable pieces of code that have been stitched together by both a static linker (at compile time) and a dynamic linker (at runtime). To handle this, the executable is sliced up into “modules” which generally refer to a particular static or dynamic library (e.g. libsystem_pthread.dylib).

So given a particular instruction address, you need to lookup what module has been mapped to that address, and then within that module you need to lookup its .debug_info/.eh_frame sections. All addresses those sections refer to will be relative to the start of that module, so you will need to add/subtract that base address in various places to map back to the actual running executable.

I’ve only ever worked at the point where the module lookup has already been resolved, so I’ll leave that as an exercise to the reader.

With all that said, Unwinding Step 1 is simple:

Lookup the instruction address in the .debug_info section. Boom you have function names and lines. That’s it. That format isn’t the focus of this. There’s plenty of DWARF parsers/interpreters, and the official specification is quite good.

Of more interest to this article is .debug_frame and .eh_frame. These sections contain DWARF CFI (unwinding) tables. I’m also not going to give a complete description of these – again the DWARF spec is great, read it – but I will sketch out the basic idea in the next section.

§1.4 Unwinding Tables (DWARF CFI)

If you don’t have nice function prologues with frame pointers – or if you want to know how to restore the other callee-saved registers, you need something more. You need unwinding tables. There are various formats of these tables but they all tell you how to do the same basic things:

And indeed if we look at DWARF CFI, it’s a giant table mapping every single address in a module to those rules.

The rules to recover the return address/registers are basically a complete virtual machine, because DWARF really wanted to let optimizers and language designers go completely wild and have the most incredibly convoluted register recovery methods possible.

Naively these unwinding tables would be enormous (far bigger than the executable itself, as it would have multiple unwinding instructions for every instruction in the executable), so a compression scheme is needed.

DWARF CFI achieves compression by allowing entries in the table to be a diff on a base entry that covers a range of instructions. The simplest and most efficient diff is just not having a new entry. Unless a new rule is introduced at an address, the previous rule applies.

To simplify these diffs, DWARF CFI defines a “CFA” (canonical frame address) register. This is conceptually the frame pointer, but doesn’t necessarily have to be. Since most rules end up being something to the effect of “rbx is saved at offset 24 from the start of the frame”, most of our rules end up looking like %rbx := *(%cfa - 24).

When we can do this, most of our entries will just be instructions on how to update the CFA. Here’s an example of what the CFI for a function which saves rbp and rbx would look like:

initial rule (covering 0x08 to 0xA8):   (start of function)
    %rip := *(%cfa);
    %cfa := %rsp + 8

diff rule (starting at 0x10):           (function pushes rbp)
    %rbp := *(%cfa - 8)
    %cfa := %rsp + 16

diff rule (starting at 0x18):           (function pushes rbx)
    %rbx := *(%cfa - 16)
    %cfa := %rsp + 24

diff rule (starting at 0x20):           (function reserves remaining stack space)
    %cfa := %rsp + 120

Unfortunately this format isn’t all roses and sunshine. Compilers can easily make mistakes generating it, and it’s apparently quite slow to evaluate. I quite liked the discussion of these issues in the paper Reliable and Fast DWARF-Based Stack Unwinding (Bastian, Kell, Nardelli – Proc. ACM Program. Lang., Vol. 3, No. OOPSLA, Article 146. Publication date: October 2019).

§1.5 Stack Scanning

For completeness sake, here’s the last strategy for unwinding. If you don’t have unwinding tables or frame pointers (or those seem to produce obviously-wrong results), you can resort to stack scanning.

Stack scanning is incredibly simple: just start walking down the stack and reading values out of it. If any of them look like a return address, assume that’s the return address (and the caller’s stack pointer should be restored to that location).

It’s a surprisingly decent fallback.

The main question to answer is “what does a return address look like?”. If you have other kinds of debug info or can lookup modules, then you can check if the address maps to a module/function. Otherwise you can try eliminating things like non-canonical addresses.

Unfortunately if you want to support an application that contains a JIT (Firefox, Chrome, and all electron apps…), things get muddier as executable pages may not actually map to any “module”. Ideally for those you have frame pointers.

§2 The Compact Unwinding Format

Ok, now we can talk about Apple’s Compact Unwinding Format, which is found in a binary’s __unwind_info section. As a reminder, this is based on my implementation of the format.

I personally found llvm’s terminology for compact unwinding very confusing, so this document will use the following terminology changes:

Like all unwinding info formats, the goal of the compact unwinding format is to create a mapping from addresses in the binary to opcodes describing how to unwind from that location.

These opcodes describe:

A user of the compact unwinding format would:

  1. Get the current instruction pointer (e.g. %rip).
  2. Lookup the corresponding opcode in the compact unwinding structure.
  3. Follow the instructions of that opcode to recover the current frame.
  4. Optionally perform runtime unwinding tasks for the current frame (destructors).
  5. Use that information to recover the instruction pointer of the previous frame.
  6. Repeat until unwinding is complete.

The compact unwinding format can be understood as two separate pieces:

Unlike DWARF CFI, compact unwinding doesn’t have facilities for incrementally updating how to recover certain registers as the function progresses.

Empirical analysis suggests that there tends to only be one opcode for an entire function (which explains why llvm refers to instruction addresses as “function offsets”), although nothing in the format seems to require this to be the case.

One consequence of only having one opcode for a whole function is that functions will generally have incorrect instructions for the function’s prologue (where callee-saved registers are individually PUSHed onto the stack before the rest of the stack space is allocated), and epilogue (where callee-saved registers are individually POPed back into registers).

Presumably this isn’t a very big deal, since there’s very few situations where unwinding would involve a function still executing its prologue/epilogue. This might matter when handling a stack overflow that occurred while saving the registers, or when processing a non-crashing thread in a minidump that happened to be in its prologue/epilogue.

Similarly, the way ranges of instructions are mapped means that Compact Unwinding will generally incorrectly map the padding bytes between functions (attributing them to the previous function), while DWARF CFI tends to more carefully exclude those addresses. Presumably also not a big deal.

Both of these things mean that if DWARF CFI and Compact Unwinding are available for a function, the DWARF CFI is expected to be more precise.

It’s possible that LSDA entries have addresses decoupled from the primary opcode so that instructions on how to run destructors can vary more granularly, but LSDA support is still TODO as it’s not needed for backtraces.

§3 Page Tables

This section describes the architecture-agnostic layout of the compact unwinding format. The layout of the format is a two-level page-table with one root first-level node pointing to arbitrarily many second-level nodes, which in turn can hold several hundred opcode entries.

There are two high-level concepts in this format that enable significant compression of the tables:

  1. Eliding duplicate instruction addresses
  2. Palettizing the opcodes

Trick 1 is standard for unwinders: the table of mappings is sorted by address, and any entries that would have the same opcode as the previous one are elided. So for instance the following:

address: 1, opcode: 1
address: 2, opcode: 1
address: 3, opcode: 2

Is just encoded like this:

address: 1, opcode: 1
address: 3, opcode: 2

We have found a few places with “zero-length” entries, where the same address gets repeated, such as the following in libsystem_kernel.dylib:

address: 0x000121c3, opcode: 0x00000000
address: 0x000121c3, opcode: 0x04000680

In this case you can just discard the zero-length one (the first one).

Trick 2 is more novel: At the first level a global palette of up to 127 opcodes is defined. Each second-level “compressed” (leaf) page can also define up to 128 local opcodes. Then the entries mapping instruction addresses to opcodes can use 8-bit indices into those palettes instead of entire 32-bit opcodes. If an index is smaller than the number of global opcodes, it’s global, otherwise it’s local (subtract the global count to get the local index).

Unclear detail: If the global palette is smaller than 127, can the local palette be larger than 128?

To compress these entries into a single 32-bit value, the address is truncated to 24 bits and packed with the index. The addresses stored in these entries are also relative to a base address that each second-level page defines. (This will be made more clear below).

There are also non-palletized “regular” second-level pages with absolute 32-bit addresses, but those are fairly rare. llvm seems to only want to emit them in the final page.

The root page also stores the first address mapped by each second-level page, allowing for more efficient binary search for a particular function offset entry. (This is the base address the compressed pages use.)

The root page always has a final sentinel entry which has a null pointer to its second-level page while still specifying a first address. This makes it easy to lookup the maximum mapped address (the sentinel will store that value +1), and just generally makes everything Work Nicer.

§3.1 Layout of the Page Table

The page table starts at the very beginning of the __unwind_info section with the root page:

struct RootPage {
    /// Only version 1 is currently defined
    version: u32 = 1,

    /// The array of u32 global opcodes (offset relative to start of root page).
    /// These may be indexed by "compressed" second-level pages.
    global_opcodes_offset: u32,
    global_opcodes_len: u32,

    /// The array of u32 global personality codes
    /// (offset relative to start of root page).
    /// Personalities define the style of unwinding that an unwinder should
    /// use, and how to interpret the LSDA entries for a function (see below).
    personalities_offset: u32,
    personalities_len: u32,

    /// The array of FirstLevelPageEntry's describing the second-level pages
    /// (offset relative to start of root page).
    pages_offset: u32,
    pages_len: u32,

    // After this point there are several dynamically-sized arrays whose
    // precise order and positioning don't matter, because they are all
    // accessed using offsets like the ones above. The arrays are:

    global_opcodes: [u32; global_opcodes_len],
    personalities: [u32; personalities_len],
    pages: [FirstLevelPageEntry; pages_len],

    /// An array of LSDA pointers (Language Specific Data Areas).
    /// LSDAs are tables that an unwinder's personality function will use to
    /// find what destructors should be run and whether unwinding should
    /// be caught and normal execution resumed. We can treat them opaquely.
    /// Second-level pages have addresses into this array so that it can
    /// can be indexed, the root page doesn't need to know about them.
    lsdas: [LsdaEntry; unknown_len],

struct FirstLevelPageEntry {
    /// The first address mapped by this page.
    /// This is useful for binary-searching for the page that can map
    /// a specific address in the binary (the primary kind of lookup
    /// performed by an unwinder).
    first_address: u32,

    /// Offset to the second-level page (offset relative to start of root page).
    /// This may point to a RegularSecondLevelPage or a CompressedSecondLevelPage.
    /// Which it is can be determined by the 32-bit "kind" value that is at
    /// the start of both layouts.
    second_level_page_offset: u32,

    /// Base offset into the lsdas array that entries in this page will be
    /// relative to (offset relative to start of root page).
    lsda_index_offset: u32,

struct RegularSecondLevelPage {
    /// Always 2 (use to distinguish from CompressedSecondLevelPage).
    kind: u32 = 2,

    /// The Array of RegularEntry's (offset relative to **start of this page**).
    entries_offset: u16,
    entries_len: u16,

struct RegularEntry {
    /// The address in the binary for this entry (absolute).
    instruction_address: u32,
    /// The opcode for this address.
    opcode: u32,

struct CompressedSecondLevelPage {
    /// Always 3 (use to distinguish from RegularSecondLevelPage).
    kind: u32 = 3,

    /// The array of compressed u32 entries
    /// (offset relative to **start of this page**).
    /// Entries are a u32 that contains two packed values (from high to low):
    /// * 8 bits: opcode index
    ///   * 0..global_opcodes_len => index into global palette
    ///   * global_opcodes_len..255 => index into local palette
    ///     (subtract global_opcodes_len to get the real local index)
    /// * 24 bits: instruction address
    ///   * address is relative to this page's first_address!
    entries_offset: u16,
    entries_len: u16,

    /// The array of u32 local opcodes for this page
    /// (offset relative to **start of this page**).
    local_opcodes_offset: u16,
    local_opcodes_len: u16,

// TODO: why do these have instruction_addresses? Are they not in sync
// with the second-level entries?
struct LsdaEntry {
    instruction_address: u32,
    lsda_address: u32,

§4 Opcode Format

There are 3 architecture-specific opcode formats: x86, x64, and ARM64.

All 3 formats have a “null opcode” (0x0000_0000) which indicates that there is no unwinding information for this range of addresses. This happens with things like hand-written assembly subroutines. This implementation will yield it as a valid opcode that converts into CompactUnwindOp::None.

All 3 formats share a common header in the top 8 bits (from high to low):

/// Whether this instruction is the start of a function.
is_start: u1,

/// Whether there is an lsda entry for this instruction.
has_lsda: u1,

/// An index into the global personalities array
/// (TODO: ignore if has_lsda == false?)
personality_index: u2,

/// The architecture-specific kind of opcode this is, specifying how to
/// interpret the remaining 24 bits of the opcode.
opcode_kind: u4,

§4.1 x86 and x64 Opcodes

x86 and x64 use the same opcode layout, differing only in the registers being restored. Registers are numbered 0-6, with the following mappings:



Note also that encoded sizes/offsets are generally divided by the pointer size (since all values we are interested in are pointer-aligned), which of course differs between x86 and x64.

There are 4 kinds of x86/x64 opcodes (specified by opcode_kind):

(One of the llvm headers refers to a 5th “0=old” opcode. Apparently this was used for initial development of the format, and is basically just reserved to prevent the testing data from ever getting mixed with real data. Nothing should produce or handle it. It does incidentally match the “null opcode”, but it’s fine to regard that as an unknown opcode and do nothing.)

§4.1.1 x86/x64 Opcode 1: Frame-Based

The function has the standard frame pointer (bp) prelude which:

bp has been preserved, and any callee-saved registers that need to be restored are saved on the stack at a known offset from bp. The return address is stored just before the caller’s bp. The caller’s stack pointer should point before where the return address is saved.

So to unwind you just need to do:

%sp := %bp + 2*POINTER_SIZE
%ip := *(%bp + POINTER_SIZE)
%bp := *(%bp)

(and restore all the other callee-saved registers as described below)

Registers are stored in increasing order (so reg1 comes before reg2).

If a register has the “no register” value, continue iterating the offset forward. This lets the registers be stored slightly-non-contiguously on the stack.

The remaining 24 bits of the opcode are interpreted as follows (from high to low):

/// The offset from bp that the registers to restore are saved at,
/// divided by pointer size.
stack_offset: u8,

_unused: u1,

/// Registers to restore (see register mapping in previous section)
reg1: u3,
reg2: u3,
reg3: u3,
reg4: u3,
reg5: u3,

§4.1.2 x86/x64 Opcode 2: Frameless (Stack-Immediate)

The callee’s stack frame has a known size, so we can find the start of the frame by offsetting from sp (the stack pointer). The return address is saved immediately after that location. Any callee-saved registers that need to be restored are saved immediately after that.

So to unwind you just need to do:

%sp := %sp + stack_size * POINTER_SIZE
%ip := *(%sp - POINTER_SIZE)

(and restore all the other callee-saved registers as described below)

Registers are stored in reverse order on the stack from the order the decoding algorithm outputs (so reg[1] comes before reg[0]).

If a register has the “no register” value, do not continue iterating the offset forward – registers are strictly contiguous (it’s possible “no register” can only be trailing due to the encoding, but I haven’t verified this).

The remaining 24 bits of the opcode are interpreted as follows (from high to low):

/// How big the stack frame is, divided by pointer size.
stack_size: u8,

_unused: u3,

/// The number of registers that are saved on the stack.
register_count: u3,

/// The permutation encoding of the registers that are saved
/// on the stack (see below).
register_permutations: u10,

The register permutation encoding is a Lehmer code sequence encoded into a single variable-base number so we can encode the ordering of up to six registers in a 10-bit space.

This can’t really be described well with anything but code, so just read this implementation or llvm’s implementation for how to encode/decode this.

§4.1.3 x86/x64 Opcode 3: Frameless (Stack-Indirect)

(Currently Unimplemented)

Stack-Indirect is exactly the same situation as Stack-Immediate, but the stack-frame size is too large for Stack-Immediate to encode. However, the function prereserved the size of the frame in its prologue, so we can extract the the size of the frame from a sub instruction at a known offset from the start of the function (subl $nnnnnnnn,ESP in x86, subq $nnnnnnnn,RSP in x64).

This requires being able to find the first instruction of the function (TODO: presumably the first is_start entry <= this one?).

TODO: describe how to extract the value from the sub instruction.

/// Offset from the start of the function where the `sub` instruction
/// we need is stored. (NOTE: not divided by anything!)
instruction_offset: u8,

/// An offset to add to the loaded stack size, divided by pointer size.
/// This allows the stack size to differ slightly from the `sub`, to
/// compensate for any function prologue that pushes a bunch of
/// pointer-sized registers.
stack_adjust: u3,

/// The number of registers that are saved on the stack.
register_count: u3,

/// The permutation encoding of the registers that are saved on the stack
/// (see Stack-Immediate for a description of this format).
register_permutations: u10,

Note: apparently binaries generated by the clang in Xcode 6 generated corrupted versions of this opcode, but this was fixed in Xcode 7 (released in September 2015), so presumably this isn’t something we’re likely to encounter. But if you encounter messed up opcodes this might be why.

§4.1.4 x86/x64 Opcode 4: Dwarf

There is no compact unwind info here, and you should instead use the DWARF CFI in .eh_frame for this line. The remaining 24 bits of the opcode are an offset into the .eh_frame section that should hold the DWARF FDE for this instruction address.

§4.2 ARM64 Opcodes

ARM64 (AKA AArch64) is a lot more strict about the ABI of functions, and as such it has fairly simple opcodes. There are 3 kinds of ARM64 opcode:

(Yes there’s no Opcode 1, I don’t know why.)

§4.2.1 ARM64 Opcode 2: Frameless

This is a “frameless” leaf function. The caller is responsible for saving/restoring all of its general purpose registers. The frame pointer is still the caller’s frame pointer and doesn’t need to be touched. The return address is stored in the link register (x30).

So to unwind you just need to do:

%sp := %sp + stack_size * 16
%pc := %x30

(no other registers to restore)

The remaining 24 bits of the opcode are interpreted as follows (from high to low):

/// How big the stack frame is, divided by 16.
stack_size: u12,

_unused: u12,

§4.2.2 ARM64 Opcode 3: Dwarf

There is no compact unwind info here, and you should instead use the DWARF CFI in .eh_frame for this line. The remaining 24 bits of the opcode are an offset into the .eh_frame section that should hold the DWARF FDE for this instruction address.

§4.2.3 ARM64 Opcode 4: Frame-Based

This is a function with the standard prologue. The return address (pc) and the frame pointer (x29) were pushed onto the stack in a pair and in that order (ARM64 registers are saved/restored in pairs), and then the frame pointer was updated to the current stack pointer.

So to unwind you just need to do:

%sp := %x29 + 16
%pc := *(%x29 + 8)
%x29 := *(%x29)

(and restore all the other callee-saved registers as described below)

Any callee-saved registers that need to be restored were then pushed onto the stack in pairs in the following order (if they were pushed at all, see below):

  1. x19, x20
  2. x21, x22
  3. x23, x24
  4. x25, x26
  5. x27, x28
  6. d8, d9
  7. d10, d11
  8. d12, d13
  9. d14, d15

The remaining 24 bits of the opcode are interpreted as follows (from high to low):

_unused: u15,

// Whether each register pair was pushed
d14_and_d15_saved: u1,
d12_and_d13_saved: u1,
d10_and_d11_saved: u1,
d8_and_d9_saved: u1,

x27_and_x28_saved: u1,
x25_and_x26_saved: u1,
x23_and_x24_saved: u1,
x21_and_x22_saved: u1,
x19_and_x20_saved: u1,

§5 Notable Corners

Here’s some notable corner cases and esoterica of the format. Behaviour in these situations is not strictly guaranteed (as in we may decide to make the implemenation more strict or liberal if it is deemed necessary or desirable). But current behaviour is documented here for the sake of maintenance/debugging. Hopefully it also helps highlight all the ways things can go wrong for anyone using this documentation to write their own tooling.

For all these cases, if an Error is reported during iteration/search, the CompactUnwindInfoIter will be in an unspecified state for future queries. It will never violate memory safety but it may start yielding chaotic values.

If this implementation ever panics, that should be regarded as an implementation bug.

Things we allow:

Things we produce errors for:

Things that cause chaos:

§6 Errata

An earlier version of this article had an incorrect layout for the x86/x64 Frame-Based Opcode. This is because the comment describing the layout in llvm was incorrect (the actual masks were correct, but I didn’t compare the two for consistency).