source: doc/theses/andrew_beach_MMath/implement.tex@ 02a43ff

\chapter{Implementation}
% Goes over how all the features are implemented.

The implementation work for this thesis covers two components: the virtual
system and exceptions. Each component is discussed in detail.

\section{Virtual System}
\label{s:VirtualSystem}
% Virtual table rules. Virtual tables, the pointer to them and the cast.
While the \CFA virtual system currently has only one public feature, virtual
cast (see the virtual cast feature \vpageref{p:VirtualCast}),
substantial structure is required to support it,
and provide features for exception handling and the standard library.

\subsection{Virtual Type}
Virtual types only have one change to their structure: the addition of a
pointer to the virtual table, which is called the \emph{virtual-table pointer}.
Internally, the field is called @virtual_table@.
The field is fixed after construction. It is always the first field in the
structure so that its location is always known.
\todo{Talk about constructors for virtual types (after they are working).}

This is what binds an instance of a virtual type to a virtual table. This
pointer can be used as an identity check. It can also be used to access the
virtual table and the virtual members there.

\subsection{Type Id}
Every virtual type has a unique id.
Type ids can be compared for equality (the types reperented are the same)
or used to access the type's type information.
The type information currently is only the parent's type id or, if the
type has no parent, zero.

The id's are implemented as pointers to the type's type information instance.
Derefencing the pointer gets the type information.
By going back-and-forth between the type id and
the type info one can find every ancestor of a virtual type.
It also pushes the issue of creating a unique value (for
the type id) to the problem of creating a unique instance (for type
information) which the linker can solve.

Advanced linker support is required because there is no place that appears
only once to attach the type information to. There should be one structure
definition but it is included in multiple translation units. Each virtual
table definition should be unique but there are an arbitrary number of thoses.
So the special section prefix \texttt{.gnu.linkonce} is used.
With a unique suffix (making the entire section name unique) the linker will
remove multiple definition making sure only one version exists after linking.
Then it is just a matter of making sure there is a unique name for each type.

This is done in three phases.
The first phase is to generate a new structure definition to store the type
information. The layout is the same in each case, just the parent's type id,
but the types are changed.
The structure's name is change, it is based off the virtual type's name, and
the type of the parent's type id.
If the virtual type is polymorphic then the type information structure is
polymorphic as well, with the same polymorphic arguments.

The second phase is to generate an instance of the type information with a
almost unique name, generated by mangling the virtual type name.

The third phase is implicit with \CFA's overloading scheme. \CFA mangles
names with type information so that all of the symbols exported to the linker
are unique even if in \CFA code they are the same. Having two declarations
with the same name and same type is forbidden because it is impossible for
overload resolution to pick between them. This is why a unique type is
generated for each virtual type.
Polymorphic information is included in this mangling so polymorphic
types will have seperate instances for each set of polymorphic arguments.

\begin{cfa}
struct /* type name */ {
        /* parent type name */ const * parent;
};

__attribute__((section(".gnu.linkonce./* instance name */")))
/* type name */ const /* instance name */ = {
        &/* parent instance name */,
};
\end{cfa}

\subsection{Virtual Table}
Each virtual type has a virtual table type that stores its type id and
virtual members.
Each virtual type instance is bound to a table instance that is filled with
the values of virtual members.
Both the layout of the fields and their value are decided by the rules given
below.

The layout always comes in three parts.
The first section is just the type id at the head of the table. It is always
there to ensure that
The second section are all the virtual members of the parent, in the same
order as they appear in the parent's virtual table. Note that the type may
change slightly as references to the ``this" will change. This is limited to
inside pointers/references and via function pointers so that the size (and
hence the offsets) are the same.
The third section is similar to the second except that it is the new virtual
members introduced at this level in the hierarchy.

\begin{figure}
\begin{cfa}
type_id
parent_field0
...
parent_fieldN
child_field0
...
child_fieldN
\end{cfa}
\caption{Virtual Table Layout}
\label{f:VirtualTableLayout}
\todo*{Improve the Virtual Table Layout diagram.}
\end{figure}

The first and second sections together mean that every virtual table has a
prefix that has the same layout and types as its parent virtual table.
This, combined with the fixed offset to the virtual table pointer, means that
for any virtual type it doesn't matter if we have it or any of its
descendants, it is still always safe to access the virtual table through
the virtual table pointer.
From there it is safe to check the type id to identify the exact type of the
underlying object, access any of the virtual members and pass the object to
any of the method-like virtual members.
\todo{Introduce method-like virtual members.}

When a virtual table is declared the user decides where to declare it and its
name. The initialization of the virtual table is entirely automatic based on
the context of the declaration.

The type id is always fixed, each virtual table type will always have one
exactly one possible type id.
The virtual members are usually filled in by resolution. The best match for
a given name and type at the declaration site is filled in.
There are two exceptions to that rule: the @size@ field is the type's size
and is set to the result of a @sizeof@ expression, the @align@ field is the
type's alignment and similarly uses an @alignof@ expression.

\subsubsection{Concurrency Integration}
Coroutines and threads need instances of @CoroutineCancelled@ and
@ThreadCancelled@ respectively to use all of their functionality. When a new
data type is declared with @coroutine@ or @thread@ the forward declaration for
the instance is created as well. The definition of the virtual table is created
at the definition of the main function.
\todo{Add an example with code snipits.}

\subsection{Virtual Cast}
Virtual casts are implemented as a function call that does the subtype check
and a C coercion-cast to do the type conversion.
% The C-cast is just to make sure the generated code is correct so the rest of
% the section is about that function.
The function is implemented in the standard library and has the following
signature:
\begin{cfa}
void * __cfa__virtual_cast(
        struct __cfa__parent_vtable const * parent,
        struct __cfa__parent_vtable const * const * child );
\end{cfa}
\todo{Get rid of \_\_cfa\_\_parent\_vtable in the standard library and then
the document.}
The type id of target type of the virtual cast is passed in as @parent@ and
the cast target is passed in as @child@.

For C generation both arguments and the result are wrapped with type casts.
There is also an internal store inside the compiler to make sure that the
target type is a virtual type.
% It also checks for conflicting definitions.

The virtual cast either returns the original pointer as a new type or null.
So the function just does the parent check and returns the approprate value.
The parent check is a simple linear search of child's ancestors using the
type information.

\section{Exceptions}
% Anything about exception construction.

\section{Unwinding}
% Adapt the unwind chapter, just describe the sections of libunwind used.
% Mention that termination and cancellation use it. Maybe go into why
% resumption doesn't as well.

% Many modern languages work with an interal stack that function push and pop
% their local data to. Stack unwinding removes large sections of the stack,
% often across functions.

Stack unwinding is the process of removing stack frames (activations) from the
stack. On function entry and return, unwinding is handled directly by the
call/return code embedded in the function.
In many cases the position of the instruction pointer (relative to parameter
and local declarations) is enough to know the current size of the stack
frame.

Usually, the stack-frame size is known statically based on parameter and
local variable declarations. Even with dynamic stack-size the information
to determain how much of the stack has to be removed is still contained
within the function.
Allocating/deallocating stack space is usually an $O(1)$ operation achieved by
bumping the hardware stack-pointer up or down as needed.
Constructing/destructing values on the stack takes longer put in terms of
figuring out what needs to be done is of similar complexity.

Unwinding across multiple stack frames is more complex because that
information is no longer contained within the current function.
With seperate compilation a function has no way of knowing what its callers
are so it can't know how large those frames are.
Without altering the main code path it is also hard to pass that work off
to the caller.

The traditional unwinding mechanism for C is implemented by saving a snap-shot
of a function's state with @setjmp@ and restoring that snap-shot with
@longjmp@. This approach bypasses the need to know stack details by simply
reseting to a snap-shot of an arbitrary but existing function frame on the
stack. It is up to the programmer to ensure the snap-shot is valid when it is
reset and that all required clean-up from the unwound stacks is preformed.
This approach is fragile and forces a work onto the surounding code.

With respect to that work forced onto the surounding code,
many languages define clean-up actions that must be taken when certain
sections of the stack are removed. Such as when the storage for a variable
is removed from the stack or when a try statement with a finally clause is
(conceptually) popped from the stack.
None of these should be handled by the user, that would contradict the
intention of these features, so they need to be handled automatically.

To safely remove sections of the stack the language must be able to find and
run these clean-up actions even when removing multiple functions unknown at
the beginning of the unwinding.

One of the most popular tools for stack management is libunwind, a low-level
library that provides tools for stack walking, handler execution, and
unwinding. What follows is an overview of all the relevant features of
libunwind needed for this work, and how \CFA uses them to implement exception
handling.

\subsection{libunwind Usage}
Libunwind, accessed through @unwind.h@ on most platforms, is a C library that
provides \Cpp-style stack-unwinding. Its operation is divided into two phases:
search and cleanup. The dynamic target search -- phase 1 -- is used to scan the
stack and decide where unwinding should stop (but no unwinding occurs). The
cleanup -- phase 2 -- does the unwinding and also runs any cleanup code.

To use libunwind, each function must have a personality function and a Language
Specific Data Area (LSDA). The LSDA has the unique information for each
function to tell the personality function where a function is executing, its
current stack frame, and what handlers should be checked. Theoretically, the
LSDA can contain any information but conventionally it is a table with entries
representing regions of the function and what has to be done there during
unwinding. These regions are bracketed by instruction addresses. If the
instruction pointer is within a region's start/end, then execution is currently
executing in that region. Regions are used to mark out the scopes of objects
with destructors and try blocks.

% Libunwind actually does very little, it simply moves down the stack from
% function to function. Most of the actions are implemented by the personality
% function which libunwind calls on every function. Since this is shared across
% many functions or even every function in a language it will need a bit more
% information.

The GCC compilation flag @-fexceptions@ causes the generation of an LSDA and
attaches a personality function to each function.
In plain C (which \CFA currently compiles down to) this
flag only handles the cleanup attribute:
\begin{cfa}
void clean_up( int * var ) { ... }
int avar __attribute__(( cleanup(clean_up) ));
\end{cfa}
The attribue is used on a variable and specifies a function,
in this case @clean_up@, run when the variable goes out of scope.
This is enough to mimic destructors, but not try statements which can effect
the unwinding.

To get full unwinding support all of this has to be done directly with
assembly and assembler directives. Partiularly the cfi directives
\texttt{.cfi\_lsda} and \texttt{.cfi\_personality}.

\subsection{Personality Functions}
Personality functions have a complex interface specified by libunwind. This
section covers some of the important parts of the interface.

A personality function can preform different actions depending on how it is
called.
\begin{lstlisting}[language=C,{moredelim=**[is][\color{red}]{@}{@}}]
typedef _Unwind_Reason_Code (*@_Unwind_Personality_Fn@) (
        _Unwind_Action @action@,
        _Unwind_Exception_Class @exception_class@,
        _Unwind_Exception * @exception@,
        struct _Unwind_Context * @context@
);
\end{lstlisting}
The @action@ argument is a bitmask of possible actions:
\begin{enumerate}[topsep=5pt]
\item
@_UA_SEARCH_PHASE@ specifies a search phase and tells the personality function
to check for handlers. If there is a handler in a stack frame, as defined by
the language, the personality function returns @_URC_HANDLER_FOUND@; otherwise
it return @_URC_CONTINUE_UNWIND@.

\item
@_UA_CLEANUP_PHASE@ specifies a cleanup phase, where the entire frame is
unwound and all cleanup code is run. The personality function does whatever
cleanup the language defines (such as running destructors/finalizers) and then
generally returns @_URC_CONTINUE_UNWIND@.

\item
\begin{sloppypar}
@_UA_HANDLER_FRAME@ specifies a cleanup phase on a function frame that found a
handler. The personality function must prepare to return to normal code
execution and return @_URC_INSTALL_CONTEXT@.
\end{sloppypar}

\item
@_UA_FORCE_UNWIND@ specifies a forced unwind call. Forced unwind only performs
the cleanup phase and uses a different means to decide when to stop
(see \vref{s:ForcedUnwind}).
\end{enumerate}

The @exception_class@ argument is a copy of the
\code{C}{exception}'s @exception_class@ field.
This a number that identifies the exception handling mechanism that created
the

The \code{C}{exception} argument is a pointer to the user
provided storage object. It has two public fields: the @exception_class@,
which is described above, and the @exception_cleanup@ function.
The clean-up function is used by the EHM to clean-up the exception if it
should need to be freed at an unusual time, it takes an argument that says
why it had to be cleaned up.

The @context@ argument is a pointer to an opaque type passed to helper
functions called inside the personality function.

The return value, @_Unwind_Reason_Code@, is an enumeration of possible messages
that can be passed several places in libunwind. It includes a number of
messages for special cases (some of which should never be used by the
personality function) and error codes. However, unless otherwise noted, the
personality function should always return @_URC_CONTINUE_UNWIND@.

\subsection{Raise Exception}
Raising an exception is the central function of libunwind and it performs a
two-staged unwinding.
\begin{cfa}
_Unwind_Reason_Code _Unwind_RaiseException(_Unwind_Exception *);
\end{cfa}
First, the function begins the search phase, calling the personality function
of the most recent stack frame. It continues to call personality functions
traversing the stack from newest to oldest until a function finds a handler or
the end of the stack is reached. In the latter case, raise exception returns
@_URC_END_OF_STACK@.

Second, when a handler is matched, raise exception moves to the clean-up
phase and walks the stack a second time.
Once again, it calls the personality functions of each stack frame from newest
to oldest. This pass stops at the stack frame containing the matching handler.
If that personality function has not install a handler, it is an error.

If an error is encountered, raise exception returns either
@_URC_FATAL_PHASE1_ERROR@ or @_URC_FATAL_PHASE2_ERROR@ depending on when the
error occurred.

\subsection{Forced Unwind}
\label{s:ForcedUnwind}
Forced Unwind is the other central function in libunwind.
\begin{cfa}
_Unwind_Reason_Code _Unwind_ForcedUnwind(_Unwind_Exception *,
        _Unwind_Stop_Fn, void *);
\end{cfa}
It also unwinds the stack but it does not use the search phase. Instead another
function, the stop function, is used to stop searching. The exception is the
same as the one passed to raise exception. The extra arguments are the stop
function and the stop parameter. The stop function has a similar interface as a
personality function, except it is also passed the stop parameter.
\begin{lstlisting}[language=C,{moredelim=**[is][\color{red}]{@}{@}}]
typedef _Unwind_Reason_Code (*@_Unwind_Stop_Fn@)(
        _Unwind_Action @action@,
        _Unwind_Exception_Class @exception_class@,
        _Unwind_Exception * @exception@,
        struct _Unwind_Context * @context@,
        void * @stop_parameter@);
\end{lstlisting}

The stop function is called at every stack frame before the personality
function is called and then once more after all frames of the stack are
unwound.

Each time it is called, the stop function should return @_URC_NO_REASON@ or
transfer control directly to other code outside of libunwind. The framework
does not provide any assistance here.

\begin{sloppypar}
Its arguments are the same as the paired personality function. The actions
@_UA_CLEANUP_PHASE@ and @_UA_FORCE_UNWIND@ are always set when it is
called. Beyond the libunwind standard, both GCC and Clang add an extra action
on the last call at the end of the stack: @_UA_END_OF_STACK@.
\end{sloppypar}

\section{Exception Context}
% Should I have another independent section?
% There are only two things in it, top_resume and current_exception. How it is
% stored changes depending on whether or not the thread-library is linked.

The exception context is global storage used to maintain data across different
exception operations and to communicate among different components.

Each stack must have its own exception context. In a sequential \CFA program,
there is only one stack with a single global exception-context. However, when
the library @libcfathread@ is linked, there are multiple stacks and each
needs its own exception context.

The exception context should be retrieved by calling the function
@this_exception_context@. For sequential execution, this function is defined as
a weak symbol in the \CFA system-library, @libcfa@. When a \CFA program is
concurrent, it links with @libcfathread@, where this function is defined with a
strong symbol replacing the sequential version.

The sequential @this_exception_context@ returns a hard-coded pointer to the
global exception context.
The concurrent version adds the exception context to the data stored at the
base of each stack. When @this_exception_context@ is called, it retrieves the
active stack and returns the address of the context saved there.

\section{Termination}
% Memory management & extra information, the custom function used to implement
% catches. Talk about GCC nested functions.

\CFA termination exceptions use libunwind heavily because they match \Cpp
\Cpp exceptions closely. The main complication for \CFA is that the
compiler generates C code, making it very difficult to generate the assembly to
form the LSDA for try blocks or destructors.

\subsection{Memory Management}
The first step of a termination raise is to copy the exception into memory
managed by the exception system. Currently, the system uses @malloc@, rather
than reserved memory or the stack top. The exception handling mechanism manages
memory for the exception as well as memory for libunwind and the system's own
per-exception storage.

\begin{figure}
\begin{verbatim}
Fixed Header  | _Unwind_Exception   <- pointer target
              |
              | Cforall storage
              |
Variable Body | the exception       <- fixed offset
              V ...
\end{verbatim}
\caption{Exception Layout}
\label{f:ExceptionLayout}
\end{figure}
\todo*{Convert the exception layout to an actual diagram.}

Exceptions are stored in variable-sized blocks (see \vref{f:ExceptionLayout}).
The first component is a fixed-sized data structure that contains the
information for libunwind and the exception system. The second component is an
area of memory big enough to store the exception. Macros with pointer arthritic
and type cast are used to move between the components or go from the embedded
@_Unwind_Exception@ to the entire node.

Multipe exceptions can exist at the same time because exceptions can be
raised inside handlers, destructors and finally blocks.
Figure~\vref{f:MultipleExceptions} shows a program that has multiple
exceptions active at one time.
Each time an exception is thrown and caught the stack unwinds and the finally
clause runs. This will throw another exception (until @num_exceptions@ gets
high enough) which must be allocated. The previous exceptions may not be
freed because the handler/catch clause has not been run.
So the EHM must keep them alive while it allocates exceptions for new throws.

\begin{figure}
\centering
% Andrew: Figure out what these do and give them better names.
\newsavebox{\myboxA}
\newsavebox{\myboxB}
\begin{lrbox}{\myboxA}
\begin{lstlisting}[language=CFA,{moredelim=**[is][\color{red}]{@}{@}}]
unsigned num_exceptions = 0;
void throws() {
    try {
        try {
            ++num_exceptions;
            throw (Example){table};
        } finally {
            if (num_exceptions < 3) {
                throws();
            }
        }
    } catch (exception_t *) {
        --num_exceptions;
    }
}
int main() {
    throws();
}
\end{lstlisting}
\end{lrbox}

\begin{lrbox}{\myboxB}
\begin{lstlisting}
\end{lstlisting}
\end{lrbox}

{\usebox\myboxA}
\hspace{25pt}
{\usebox\myboxB}

\caption{Multiple Exceptions}
\label{f:MultipleExceptions}
\end{figure}
\todo*{Work on multiple exceptions code sample.}

All exceptions are stored in nodes which are then linked together in lists,
one list per stack, with the
list head stored in the exception context. Within each linked list, the most
recently thrown exception is at the head followed by older thrown
exceptions. This format allows exceptions to be thrown, while a different
exception is being handled. The exception at the head of the list is currently
being handled, while other exceptions wait for the exceptions before them to be
removed.

The virtual members in the exception's virtual table provide the size of the
exception, the copy function, and the free function, so they are specific to an
exception type. The size and copy function are used immediately to copy an
exception into managed memory. After the exception is handled, the free
function is used to clean up the exception and then the entire node is
passed to free so the memory can be given back to the heap.

\subsection{Try Statements and Catch Clauses}
The try statement with termination handlers is complex because it must
compensate for the lack of assembly-code generated from \CFA. Libunwind
requires an LSDA and personality function for control to unwind across a
function. The LSDA in particular is hard to mimic in generated C code.

The workaround is a function called @__cfaehm_try_terminate@ in the standard
library. The contents of a try block and the termination handlers are converted
into functions. These are then passed to the try terminate function and it
calls them.
Because this function is known and fixed (and not an arbitrary function that
happens to contain a try statement), the LSDA can be generated ahead
of time.

Both the LSDA and the personality function are set ahead of time using
embedded assembly. This assembly code is handcrafted using C @asm@ statements
and contains
enough information for the single try statement the function repersents.

The three functions passed to try terminate are:
\begin{description}
\item[try function:] This function is the try block, all the code inside the
try block is placed inside the try function. It takes no parameters and has no
return value. This function is called during regular execution to run the try
block.

\item[match function:] This function is called during the search phase and
decides if a catch clause matches the termination exception. It is constructed
from the conditional part of each handler and runs each check, top to bottom,
in turn, first checking to see if the exception type matches and then if the
condition is true. It takes a pointer to the exception and returns 0 if the
exception is not handled here. Otherwise the return value is the id of the
handler that matches the exception.

\item[handler function:] This function handles the exception. It takes a
pointer to the exception and the handler's id and returns nothing. It is called
after the cleanup phase. It is constructed by stitching together the bodies of
each handler and dispatches to the selected handler.
\end{description}
All three functions are created with GCC nested functions. GCC nested functions
can be used to create closures, functions that can refer to the state of other
functions on the stack. This approach allows the functions to refer to all the
variables in scope for the function containing the @try@ statement. These
nested functions and all other functions besides @__cfaehm_try_terminate@ in
\CFA use the GCC personality function and the @-fexceptions@ flag to generate
the LSDA.
Using this pattern, \CFA implements destructors with the cleanup attribute.
\todo{Add an example of the conversion from try statement to functions.}

\section{Resumption}
% The stack-local data, the linked list of nodes.

Resumption simpler to implement than termination
because there is no stack unwinding.
Instead of storing the data in a special area using assembly,
there is just a linked list of possible handlers for each stack,
with each node on the list reperenting a try statement on the stack.

The head of the list is stored in the exception context.
The nodes are stored in order, with the more recent try statements closer
to the head of the list.
Instead of traversing the stack resumption handling traverses the list.
At each node the EHM checks to see if the try statement the node repersents
can handle the exception. If it can, then the exception is handled and
the operation finishes, otherwise the search continues to the next node.
If the search reaches the end of the list without finding a try statement
that can handle the exception the default handler is executed and the
operation finishes.

In each node is a handler function which does most of the work there.
The handler function is passed the raised the exception and returns true
if the exception is handled and false if it cannot be handled here.

For each @catchResume@ clause the handler function will:
check to see if the raised exception is a descendant type of the declared
exception type, if it is and there is a conditional expression then it will
run the test, if both checks pass the handling code for the clause is run
and the function returns true, otherwise it moves onto the next clause.
If this is the last @catchResume@ clause then instead of moving onto
the next clause the function returns false as no handler could be found.

\todo{Diagram showing a try statement being converted into resumption handlers.}

% Recursive Resumption Stuff:
Search skipping (see \vpageref{s:ResumptionMarking}), which ignores parts of
the stack
already examined, is accomplished by updating the front of the list as the
search continues. Before the handler at a node is called, the head of the list
is updated to the next node of the current node. After the search is complete,
successful or not, the head of the list is reset.

This mechanism means the current handler and every handler that has already
been checked are not on the list while a handler is run. If a resumption is
thrown during the handling of another resumption the active handlers and all
the other handler checked up to this point are not checked again.

This structure also supports new handler added while the resumption is being
handled. These are added to the front of the list, pointing back along the
stack -- the first one points over all the checked handlers -- and the ordering
is maintained.
\todo{Add a diagram for resumption marking.}

\label{p:zero-cost}
Note, the resumption implementation has a cost for entering/exiting a @try@
statement with @catchResume@ clauses, whereas a @try@ statement with @catch@
clauses has zero-cost entry/exit. While resumption does not need the stack
unwinding and cleanup provided by libunwind, it could use the search phase to
providing zero-cost enter/exit using the LSDA. Unfortunately, there is no way
to return from a libunwind search without installing a handler or raising an
error. Although workarounds might be possible, they are beyond the scope of
this thesis. The current resumption implementation has simplicity in its
favour.
% Seriously, just compare the size of the two chapters and then consider
% that unwind is required knowledge for that chapter.

\section{Finally}
% Uses destructors and GCC nested functions.
A finally clause is placed into a GCC nested-function with a unique name,
and no arguments or return values.
This nested function is then set as the cleanup
function of an empty object that is declared at the beginning of a block placed
around the context of the associated @try@ statement.

The rest is handled by GCC. The try block and all handlers are inside this
block. At completion, control exits the block and the empty object is cleaned
up, which runs the function that contains the finally code.

\section{Cancellation}
% Stack selections, the three internal unwind functions.

Cancellation also uses libunwind to do its stack traversal and unwinding,
however it uses a different primary function: @_Unwind_ForcedUnwind@. Details
of its interface can be found in the Section~\vref{s:ForcedUnwind}.

The first step of cancellation is to find the cancelled stack and its type:
coroutine or thread. Fortunately, the thread library stores the main thread
pointer and the current thread pointer, and every thread stores a pointer to
its main coroutine and the coroutine it is currently executing.
\todo*{Consider adding a description of how threads are coroutines.}

If a the current thread's main and current coroutines are the same then the
current stack is a thread stack. Furthermore it is easy to compare the
current thread to the main thread to see if they are the same. And if this
is not a thread stack then it must be a coroutine stack.

However, if the threading library is not linked, the sequential execution is on
the main stack. Hence, the entire check is skipped because the weak-symbol
function is loaded. Therefore, a main thread cancellation is unconditionally
performed.

Regardless of how the stack is chosen, the stop function and parameter are
passed to the forced-unwind function. The general pattern of all three stop
functions is the same: they continue unwinding until the end of stack and
then preform their transfer.

For main stack cancellation, the transfer is just a program abort.

For coroutine cancellation, the exception is stored on the coroutine's stack,
and the coroutine context switches to its last resumer. The rest is handled on
the backside of the resume, which check if the resumed coroutine is
cancelled. If cancelled, the exception is retrieved from the resumed coroutine,
and a @CoroutineCancelled@ exception is constructed and loaded with the
cancelled exception. It is then resumed as a regular exception with the default
handler coming from the context of the resumption call.

For thread cancellation, the exception is stored on the thread's main stack and
then context switched to the scheduler. The rest is handled by the thread
joiner. When the join is complete, the joiner checks if the joined thread is
cancelled. If cancelled, the exception is retrieved and the joined thread, and
a @ThreadCancelled@ exception is constructed and loaded with the cancelled
exception. The default handler is passed in as a function pointer. If it is
null (as it is for the auto-generated joins on destructor call), the default is
used, which is a program abort.
%; which gives the required handling on implicate join.