source: doc/theses/andrew_beach_MMath/unwinding.tex@ 21f914e

\chapter{Unwinding in \CFA}

When a function returns, a \emph{single} stack frame is unwound, removing the
function's parameters and local variables, and control continues in the
function caller using the caller's stack frame.  When an exception is raised,
\emph{multiple} stack frames are unwound, removing the function parameters and
local variables for called functions from the exception raise-frame to the
exception catch-frame.

Unwinding multiple levels is simple for a programming languages without object
destructors or block finalizers because a direct transfer is possible from the
current stack frame to a prior stack frame, where control continues at a
location within the prior caller's function. For example, C provides non-local
transfer using $longjmp$, which stores a function's state including its
frame pointer and program counter, and simply reloads this information to
continue at this prior location on the stack.

For programming languages with object destructors or block finalizers it is
necessary to walk the stack frames from raise to catch, checking for code that
must be executed as part of terminating each frame. Walking the stack has a
higher cost, and necessary information must available to detect
destructors/finalizers and call them.

A powerful package to provide stack-walking capabilities is $libunwind$,
which is used in this work to provide exception handling in \CFA. The following
explains how $libunwind$ works and how it is used.

% Stack unwinding is the process of removing things from the stack from outside
% the functions there. In languages that don't provide a way to guaranty that
% code will run when the program leaves a scope or finishes a function, this
% can be relatively trivial. C does this with $longjmp$ by setting the
% stack pointer and a few other registers.

\section{libunwind Usage}

There are two primary functions that \CFA uses in $libunwind$ to create most of
the exceptional control-flow. These functions are
$_Unwind_RaiseException$ and $_Unwind_ForcedUnwind$.

Their operation is divided into two phases: search and clean-up. The search
phase -- phase 1 -- is used to scan the stack but not unwinding it. The
clean-up phase -- phase 2 -- is used to during unwinding.

% Somewhere around here I need to talk about the control structures.
% $_Unwind_Exception$ is used to carry the API's universal data. Some
% of this is internal, other fields are used to communicate between different
% exception handling mechanisms in different runtimes.
% $_Unwind_Context$ is an opaque data structure that is used to pass
% information to helper functions.

The raise-exception function uses both phases. It starts by searching for the
handler, and if found, performs a clean-up phase to unwind the stack to the
hander. If a handler is not found, control returns allowing the
exception-handling policy for unhandled exception to be executed.  During both
phases, the raise exception function searches down the stack, calling each
function's \emph{personality function}.

A personality function performs three tasks, although not all of them have to
be present. The tasks performed are decided by the actions provided.
% Something argument something bitmask.
\begin{itemize}
\item$_UA_SEARCH_PHASE$ is called during the clean-up phase and means search
for handlers. If the hander is found, the personality function should return
$_URC_HANDLER_FOUND$, otherwise it returns $_URC_CONTINUE_UNWIND$.
\item$_UA_CLEANUP_PHASE$ is passed in during the clean-up phase and means part
or all of the stack frame is removed. The personality function should do
whatever clean-up the language defines (such as running destructors/finalizers)
and then generally returns $_URC_CONTINUE_UNWIND$.
\item$_UA_HANDLER_FRAME$ means the personality function must install a
handler. It is also passed in during the clean-up phase and is in addition to
the clean-up action. Libunwind provides several helpers for the personality
function here. Once it is done, the personality function must return
$_URC_INSTALL_CONTEXT$.
\end{itemize}

Forced unwind only performs the clean-up phase. It is similar to the phase 2
section of raise exception with a few changes. A simple one is that it passes
in an extra action to the personality function $_UA_FORCE_UNWIND$, which means
a handler cannot be installed. The most significant is the addition of the $stop$
function, which is passed in as an argument to the forced unwind.

The stop function is a lot like a personality function. It takes an extra
argument, a void pointer passed into force unwind. It may return
$_URC_NO_REASON$ to continue unwinding or it can transfer control out of the
unwind code using its own mechanism.
% Is there a reason that NO_REASON is used instead of CONTINUE_UNWIND?
The stop function is called once on every stack frame and once at the end of
the stack. In a stack frame, it is called before the personality routine with
the same arguments (except for the extra void pointer). At the end of the
stack, the arguments are mostly the same, except the stack pointer stored in
the context is set to null. Because of the significance of that change, both
GCC and Clang add an extra action in this case $_UA_END_OF_STACK$.  The stop
function may not return at the end of the stack.

\section{\CFA Implementation}

To use $libunwind$, \CFA provides several wrappers, its own storage, the
personality functions, and a stop function.

The wrappers perform three tasks: set-up, clean-up and controlling the
unwinding. The set-up allocates a copy of the \CFA exception into a handler
so it can control its lifetime, and stores it in the exception context.
Clean-up -- run when control exits a catch clause and return to normal code --
frees the exception copy.
% It however does not set up the unwind exception so we can't use any inter-
% runtime/language features. Also the exception context is global.

The control code in the middle is run every time a throw or re-throw is
called. It uses raise exception to search for a handler and to run it if one
is found. Otherwise it uses forced unwind to unwind the stack, running all
destructors, before terminating the process.

The stop function is very simple, it checks the end of stack flag to see if it
is finished unwinding. If so, it calls exit to end the process, otherwise it
tells the system to continue unwinding.
% Yeah, this is going to have to change.

The personality routine is much more complicated.  First, because it has to get
some information about the function by scanning the LSDA (Language Specific
Data Area). This allows a single personality function to be used for multiple
functions and it accounts for multiple regions and possible handlers in a
single function.
% Not that we do that yet.

However generating the LSDA is very difficult. It requires a lot of knowledge
about the location of the instruction pointer and stack layout that can be
broken by optimization. So instead for frames where there are only destructors
we use GCC's attribute cleanup and the -fexception flag to handle unwinding
without adding our own functionality.

For functions that do have handlers (defined in try statements) the function
is split into several functions. Everything outside the try statement is the
first function, which has then has only destructors to be run during
unwinding. The clauses of the try block are then converted into GCC inner
functions which can be passed around with function pointers while still having
access to the outer function's scope. $catchResume$ and $finally$
clauses are handled separately and will not be discussed here.

The $try$ clause is converted to a function directly. The $catch$
clauses are combined and then create two functions. The first is the match
function which is used during the search phase to see if any of the handler
here. The second it the catch function, which is a large switch-case block
with the different handlers. All of these function do not interact with
unwinding except for running destructors and so can be handled by GCC.

These three functions are passed into $try_terminate$, an internal
function that represents the try statement. This is the only function with
our personality function as well as assembly statements that create the LSDA.
In normal execution all the function does is call the try block closure and
return once that has finished executing. However using $libunwind$ its
personality function now handles exception matching and catching.

During the search phase the personality function retrieves the match function
from the stack using the saved stack pointer. The function is called, either
returning 0 for no match or the index (a positive integer) of the handler
that will handle it. The if a handler was found the personality function
reports it after saving the index to the exception context.

During the clean-up phase there is nothing for the personality function to
clean-up in $try_terminate$. So if this is not the handler frame
unwinding continues. If this is the handler frame than control is transferred
to the catch function, giving it the exception and the handler index.