source: doc/theses/mike_brooks_MMath/string.tex @ 1697c40

\chapter{String}

\section{String}

\subsection{Logical overlap}

\input{sharing-demo.tex}

Consider two strings @s1@ and @s1a@ that are in an aliasing relationship, and a third, @s2@, made by a simple copy from @s1@.
\par\noindent
\begin{tabular}{llll}
                                & @s1@  & @s1a@ & @s2@  \\
%\input{sharing-demo1.tex}
\end{tabular}
\par\noindent


\subsection{RAII limitations}

Earlier work on \CFA [to cite Schluntz] implemented the feature of constructors and destructors.  A constructor is a user-defined function that runs implicitly, when control passes an object's declaration, while a destructor runs at the exit of the declaration's lexical scope.  The feature allows programmers to assume that, whenever a runtime object of a certain type is accessible, the system called one of the programmer's constuctor functions on that object, and a matching destructor call will happen in the future.  The feature helps programmers know that their programs' invariants obtain.

The purposes of such invariants go beyond ensuring authentic values for the bits inside the object.   These invariants can track occurrences of the managed objects in other data structures.  Reference counting is a typical application of the latter invariant type.  With a reference-counting smart pointer, the consturctor and destructor \emph{of the pointer type} track the lifecycles of occurrences of these pointers, by incrementing and decrementing a counter (ususally) on the referent object, that is, they maintain a that is state separate from the objects to whose lifecycles they are attached.  Both the \CC and \CFA RAII systems ares powerful enough to achive such reference counting.

The \CC RAII system supports a more advanced application.  A lifecycle function has access to the object under managamanet, by location; constructors and destuctors receive a @this@ parameter providing its memory address.  A lifecycle-function implementation can then add its objects to a collection upon creation, and remove them at destruction.  A modulue that provides such objects, by using and encapsulating such a collection, can traverse the collection at relevant times, to keep the objects ``good.''  Then, if you are the user of such an module, declaring an object of its type means not only receiving an authentically ``good'' value at initialization, but receiving a subscription to a service that will keep the value ``good'' until you are done with it.

In many cases, the relationship between memory location and lifecycle is simple.  But with stack-allocated objects being used as parameters and returns, there is a sender version in one stack frame and a receiver version in another.  \CC is able to treat those versions as distinct objects and guarantee a copy-constructor call for communicating the value from one to the other.  This ability has implications on the language's calling convention.  Consider an ordinary function @void f( Vehicle x )@, which receives an aggregate by value.  If the type @Vehicle@ has custom lifecycle functions, then a call to a user-provided copy constructor occurs, after the caller evaluates its argument expression, after the callee's stack frame exists, with room for its variable @x@ (which is the location that the copy-constructor must target), but before the user-provided body of @f@ begins executing.  \CC achieves this ordering by changing the function signature, in the compiled form, to pass-by-reference and having the callee invoke the copy constructor in its preamble.  On the other hand, if @Vehicle@ is a simple structure then the C calling convention is applied as the code originally appeared, that is, the callsite implementation code performs a bitwise copy from the caller's expression result, into the callee's x.

TODO: learn correction to fix inconcsistency: this discussion says the callee invokes the copy constructor, but only the caller knows which copy constructor to use!

TODO: discuss the return-value piece of this pattern

The \CFA RAII system has limited support for using lifecycle functions to provide a ``stay good'' service.  It works in restricted settings, including on dynamically allocated objects.  It does not work for communicating arguments and returns by value because the system does not produce a constructor call that tracks the implied move from a sender frame to a reciver frame.  This limitation does not prevent a typical reference-counting design from using call-with-value/return-of-value, because the constructor--destructor calls are correctly balanced.  But it impedes a ``stay-good'' service from supporting call-with-value/return-of-value, because the lifecycles presented to the constructor/destor calls do not keep stable locations.  A ``stay-good'' service is acheivable so long as call-with-value/return-of-value do not occur.  The original presentation [to cite Schluntz section] acknoweledges this limitiation; the present discussion makes its consequences more apparent.

The \CFA team sees this limitation as part of a tactical interem state that should some day be improved.  The \CFA compiler is currently a source-to-source translator that targets relativly portable C.  Several details of its features are provisionally awkward or under-performant until finer control of its code generation is feasible.  In the present state, all calls that appear in \CFA source code as call-with-value/return-of-value are emitted this way to the underlying C calling convention.  SO WHAT?

The present string-API contribution has both the ``stay good'' promise and call-with-value/return-of-value being essential.  The main string API uses a work-around to acheive the full feature set, at a runtime performance penalty.  An alternative API sacrifices call-with-value/return-of-value functionality to recover full runtime performance.  These APIs are layered, with the slower, friendlier High Level API (HL) wrapping the faster, more primitive Low Level API (LL).  They present the same features, up to lifecycle management, with call-with-value/return-of-value being disabled in LL and implemented with the workaround in HL.  The intention is for most future code to target HL.  In a more distant future state, where \CFA has an RAII system that can handle the problematic quadrant, the HL layer can be abolished, the LL can be renamed to match today's HL, and LL can have its call-with-value/return-of-value permission reenabled.  Then, programs written originally against HL will simply run faster.  In the meantime, two use cases of LL exist.  Performance-critical sections of applications have LL as an option.  Within [Xref perf experiments], though HL-v-LL penalties are measured, typcial comparisons of the contributed string libary vs similar systems are made using LL.  This measurement gives a fair estimate of the goal state for \CFA while it is an evloving work in progress.



\subsection{Memory management}

A centrepriece of the string module is its memory manager.  The managment scheme defines a large shared buffer for strings' text.  Allocation in this buffer is always bump-pointer; the buffer is compacted and/or relocated with growth when it fills.  A string is a smart pointer into this buffer.

This cycle of frequent cheap allocations, interspersed with infrequent expensive compactions, has obvious similarities to a general-purpose memory manager based on garbage collection (GC).  A few differences are noteworthy.  First, in a general purpose manager, the objects of allocation contain pointers to other such objects, making the transitive reachability of these objects be a critical property.  Here, the allocations are of buffers of text, never pointers, so one allocation never keeps another one alive.  Second, in a general purpose manager, where the handle that keeps an allocation alive is the same as the program's general-purpose inter-object reference, an extremely lean representation of this reference is required.  Here, a fatter representation is acceptable because [why??].


Figure [memmgr-basix.vsdx] shows the representation.  A heap header, with its text buffer, defines a sharing context.  Often, one global sharing context is appropriate for an entire program; exceptions are discussed in [xref TBD].  Strings are handles into the buffer.  They are members of a linked list whose order matches the order of their buffer fragments (exactly, where there is no overlapping, and approximately, where there is).  The header maintains a next-allocation pointer (alloc, in the figure) after the last live allocation of the buffer.  No external references into the buffer are allowed and the management procedure relocates the text allocations as needed.  The string handles contain explicit length fields, null termination characters are not used and all string text is kept in contiguous storage.  When strings (the inter-linked hanldes) are allocated on the program's call stack, a sustained period with no use of the program's dynamic memory allocator can ensue, during which the program nonetheless creates strings, destroys them, and runs length-increasing modifications on existing ones.  

Compaction happens when the heap fills.  It is the first of two uses of the linked list.  The list allows discovering all live string handles, and from them, the ranges of the character buffer that are in use.  With these ranges known, their total character count gives the amount of space in use.  When that amount is small, compared with the current buffer size, an in-place compaction occurs, which enatils copying the in-use ranges, to be adjacent, at the font of the buffer.  When the in-use amount is large, a larger buffer is allocated (using the program's general-purpose dynamic allcator), the in-use strings are copied to be adjacent at the front of it, and the original buffer is freed back to the program's general allocator.  Either way, navigating the links between the handles provides the pointers into the buffer, first read, to find the source fragment, then written with the location of the resultant fragment.  This linkage across the structure is unaffected during a compaction; only the pointers from the handles to the buffer are modified.  This modification, along with the grooming/provisioning of the text storage resouce that it represents, is an example, in the language of [xref RAII limitations] of the string module providing a ``stay good'' service.

Object lifecycle events are the subscription-management triggers in such a service.  There are two fundamental string-creation routines:  importing from external text like a C-string, and initialization from an existing \CFA string.  When importing, a fresh allocation at the free end fo the buffer occurs, into which the text is copied.  The resultant handle is therefore inserted into the list at the position after the incumbent last handle, a position given by the heap manager's ``last handle'' pointer.  When initializing from text already on the \CFA heap, the resultant handle is a second reference onto the original run of characters.  In this case, the resultant handle's linked-list position is beside the original handle.  Both string initialization styles preserve the string module's internal invriant that the linked-list order match the buffer order.  For string destruction, the list being doubly linked provides for easy removal of the disappearing handle.

While a string handle is live, it accepts modification operations, some of which make it reference a different portion of the underlying buffer, and accordingly, move the handle to a different position in the inter-handle list.   While special cases have more optimal handling, the general case requires a fresh buffer run.  In this case, the new run is allocated at the bump-pointer end and filled with the required value.  Then, handles that originally referenced the old location and need to see the new value are pointed at the new buffer location, unlinked from their original positions in the handles' list, and linked in at the end of the list.  An optimal case, when the target is not a substring of something larger, and the source is text from elsewhere in the managed buffer, allows the target to be re-pointed at the source characters, and accordingly, move list position to be beside the source.  Cases where in-place editing happens, addressed further in [xref: TBD], leave affected handles in their original list positions.  In analogy to the two cases of string initialization, the two cases of realizing assignment by moving either to a fresh buffer run, or to overlap references with the source, maintain the invariant of linked list order matching buffer order.


To explain: GCing allocator doing bump-pointer with compaction



At the level of the memory manager, these modifications can aways be explained as assignments; for example, an append is an assignemnt into the empty substring at the end. 

While favourable conditions allow for in-place editing, the general case requires a fresh buffer run.  For example, if the new value does not fit in the old place, or if other handles are still using the old value, then the new value will use a fresh buffer run.

where there is room for the resulting value in the original buffer location, and where all handles referring to the original buffer location should see the new value, 


always boiled down to assignment and appendment.  Assignment has special cases that happen in-place, but in the general case, it is implemented as a sequence of appends onto a fresh allocation at the end of the buffer.  (The sequence has multiple steps when the assignment target is a substring: old before, new middle, old after.)  Similarly, an append request can be serviced in-place when there is room, or as a pair of appends 



\subsection{Sharing implementation}

The \CFA string module has two manners in which serveral string handles can share an unerlying run of characters.  

The first type of sharing is user-requested, following the [xref Logical Overlap].  Here, the user requests, explicitly, that both handles be views of the same logical, modifiable string.  This state is typically prodecd by the substring operation.  In a typical substring call, the source string-handle is referencing an entire string, and the resluting, newly made, string handle is referencing a portion of the orignal.  In this state, a subsequent modification made by either is visible in both.

The second type of sharing happens when the system implicitly delays the physical execution of a logical \emph{copy} operation, as part of its copy-on-write optimization.  This state is typically produced by constructing a new string, using an original string as its intialization source.  In this state, a subsequent modification done on one handle triggers the deferred copy action, leaving the handles referencing different runs within the buffer, holding distinct values.

A further abstraction, in the string module's implementation, helps distinguish the two senses of sharing.  A share-edit set (SES) is an equivalence class over string handles, being the reflexive, symmetric and transitive closure of the relationship of one being constructed from the other, with the ``share edits'' opt-in given.  It is represented by a second linked list among the handles.  A string that shares edits with no other is in a SES by itself.  Inside a SES, a logical modification of one substring portion may change the logical value in another, depending on whether the two actually overlap.  Conversely, no logical value change can flow outside of a SES.  Even if a modification on one string handle does not reveal itself \emph{logically} to anther handle in the same SES (because they don't overlap), if the modification is length-changing, completing the modification requires visiting the second handle to adjust its location in the sliding text.


\subsection{Avoiding implicit sharing}

There are tradeoffs associated with the copy-on-write mechanism.  Several quatitative matters are detailed in the [xref: Performance Assessment] section and the qualitiative issue of multi-threaded support is introduced here.  The \CFA sting library provides a switch to disable the sharing mechanism for situtations where it is inappropriate.

Because of the inter-linked string handles, any participant managing one string is also managing, directly, the neighbouring strings, and from there, a data structure of the ``set of all strings.''  This data structure is intended for sequential access.  A negative consequence of this decision is that multiple threads using strings need to be set up so that they avoid attempting to modify (concurrently) an instance of this structure.  A positive consequence is that a single-threaded program, or a program with several independent threads, can use the sharing context without an overhead from locking.

The \CFA sting library provides the @string_sharectx@ type to control an ambient sharing context for the current thread.  It allows two adjustments: to opt out of sharing entirely, or to begin sharing within a private context.  Either way, the chosen mode applies to the current thread, for the duration of the lifetime of the created  @string_sharectx@ object, up to being suspended by child liftimes of different contexts.  The indended use is with stack-managed lifetimes, in which the established context lasts until the current function returns, and affects all functions called that don't create their own contexts.
\lstinputlisting[language=CFA, firstline=20, lastline=34]{sharectx-demo.cfa}
In this example, the single-letter functions are called in alphabetic order.  The functions @a@ and @d@ share string character ranges within themselves, but not with each other.  The functions @b@, @c@ and @e@ never share anything.

[ TODO: true up with ``is thread local'' (implement that and expand this discussion to give a concurrent example, or adjust this wording) ]

When the string library is running with sharing disabled, it runs without implicit thread-safety challenges (which same as the STL) and with performance goals similar to the STL's.  This thread-safety quality means concurrent users of one string object must still bring their own mutual exlusion, but the string libary will not add any cross thread uses that were not apparent in the user's code.

Running with sharing disabled can be thought of as STL-emulation mode.



\subsection{Future work}

To discuss: Unicode

To discuss: Small-string optimization


\subsection{Performance assessment}

I assessed the CFA string library's speed and memory usage.  I present these results ineven quivalent cases, due to either micro-optimizations foregone, or fundamental costs of the added functionality.  They also show the benefits and tradeoffs, as >100\% effects, of switching to CFA, with the tradeoff points quantified.  The final test shows the overall win of the CFA text-sharing mechanism.  It exercises several operations together, showing CFA enabling clean user code to achieve performance that STL requires less-clean user code to achieve.

To discuss: general goal of ... while STL makes you think about memory management, all the time, and if you do your performance can be great ... CFA sacrifices this advantage modestly in exchange for big wins when you're not thinking about memory mamangement.  [Does this position cover all of it?]

To discuss: revisit HL v LL APIs

To discuss: revisit nosharing as STL emulation modes

These tests use randomly generated text fragments of varying lengths.  A collection of such fragments is a \emph{corpus}.  The mean length of a fragment from corpus is a typical explanatory variable.  Such a length is used in one of three modes:
\begin{description}
    \item [Fixed-size] means all string fragments are of the stated size
    \item [Varying from 1] means string lengths are drawn from a geometric distribution with the stated mean, and all lengths occur
    \item [Varying from 16] means string lengths are drawn from a geometric distribution with the stated mean, but only lengths 16 and obove occur; thus, the stated mean will be above 16.
\end{description}
The geometric distribution implies that lengths much longer than the mean occur frequently.  The special treatment of length 16 deals with comparison to STL, given that STL has short-string optimization (see [todo: write and cross-ref future-work SSO]), currently not implmented in \CFA.  When success notwithstanding SSO is illustrated, a fixed-size or from-16 distribution ensures that extra-optimized cases are not part of the mix on the STL side.  In all experiments that use a corpus, its text is generated and loaded into the SUT before the timed phase begins.

To discuss: vocabulary for reused case variables

To discuss: common approach to iteration and quoted rates

To discuss: hardware and such

To discuss: memory allocator


\subsubsection{Test: Append}

This test measures the speed of appending fragments of text onto a growing string.  Its subcases include both CFA being similar to STL, and their designs offering a tradeoff.

One experimental variable is the user's operation being @a = a + b@ vs. @a += b@.  While experienced programmers expect the latter to be ``what you obviously should do,'' controling the penatly of the former both helps the API be accessible to beginners and also helps offer confidence that when a user tries to compose operations, the forms that are most natural to the user's composition are viable.

Another experimental variable is whether the user's logical allocation is fresh vs reused.  Here, \emph{reusing a logical allocation}, means that the prgram variable, into which the user is concatenating, previously held a long string:\\
\begin{tabular}{ll}
    Logical allocation fresh                   & Logical allocation reused                  \\
                                               & @ string x; @                              \\
    @ for( ... ) { @                           & @ for( ... ) { @                           \\
    @     string x; @                          & @     x = ""; @                            \\
    @     for( ... ) @                         & @     for( ... ) @                         \\
    @        x += ... @                        & @        x += ... @                        \\
    @ } @                                      & @ } @
\end{tabular}\\
These benchmark drivers have an outer loop for ``until a sample-worthy amount of execution has happened'' and an inner loop for ``build up the desired-length string.''  It is sensible to doubt that a user should have to care about this difference, yet the STL performs differently in these cases.  Concretly, both cases incur the cost of copying characters into the target string, but only the allocation-fresh case incurs a further reallocation cost, which is generally paid at points of doubling the length.  For the STL, this cost includes obtaining a fresh buffer from the memory allocator and copying older characters into the new buffer, while CFA-sharing hides such a cost entirely.  The reuse-vs-fresh distinction is only relevant in the currrent \emph{append} tests.

The \emph{append} tests use the varying-from-1 corpus construction; that is they do not assume away the STL's advantage from small-string opitimization.

To discuss: any other case variables intruduced in the performance intro

\begin{figure}
    \includegraphics[width=\textwidth]{string-graph-peq-cppemu.png}
    \caption{Average time per iteration with one \lstinline{x += y} invocation, comparing \CFA with STL implementations (given \CFA running in STL emulation mode), and comparing the ``fresh'' with ``reused'' reset styles, at various string sizes.}
    \label{fig:string-graph-peq-cppemu}
\end{figure}

Figure \ref{fig:string-graph-peq-cppemu} shows this behaviour, by the STL and by \CFA in STL emulation mode.  \CFA reproduces STL's performance, up to a 15\% penalty averaged over the cases shown, diminishing with larger strings, and 50\% in the worst case.  This penatly characterizes the amount of implementation fine tuning done with STL and not done with \CFA in present state.  The larger inherent penalty, for a user mismanaging reuse, is 40\% averaged over the cases shown, is minimally 24\%, shows up consistently between the STL and \CFA implementations, and increases with larger strings.

\begin{figure}
    \includegraphics[width=\textwidth]{string-graph-peq-sharing.png}
    \caption{Average time per iteration with one \lstinline{x += y} invocation, comparing \CFA (having implicit sharing activated) with STL, and comparing the ``fresh'' with ``reused'' reset styles, at various string sizes.}
    \label{fig:string-graph-peq-sharing}
\end{figure}

In sharing mode, \CFA makes the fresh/reuse difference disappear, as shown in Figure \ref{fig:string-graph-peq-sharing}.  At append lengths 5 and above, CFA not only splits the two baseline STL cases, but its slowdown of 16\% over (STL with user-managed reuse) is close to the \CFA-v-STL implementation difference seen with \CFA in STL-emulation mode.

\begin{figure}
    \includegraphics[width=\textwidth]{string-graph-pta-sharing.png}
    \caption{Average time per iteration with one \lstinline{x = x + y} invocation (new, purple bands), comparing \CFA (having implicit sharing activated) with STL.  For context, the results from Figure \ref{fig:string-graph-peq-sharing} are repeated as the bottom bands.  While not a design goal, and not graphed out, \CFA in STL-emulation mode outperformed STL in this case; user-managed allocation reuse did not affect any of the implementations in this case.}
    \label{fig:string-graph-pta-sharing}
\end{figure}

When the user takes a further step beyond the STL's optimal zone, by running @x = x + y@, as in Figure \ref{fig:string-graph-pta-sharing}, the STL's penalty is above $15 \times$ while CFA's (with sharing) is under $2 \times$, averaged across the cases shown here.  Moreover, the STL's gap increases with string size, while \CFA's converges.

\subsubsection{Test: Pass argument}

To have introduced:  STL string library forces users to think about memory management when communicating values across a function call

STL charges a prohibitive penalty for passing a string by value.  With implicit sharing active, \CFA treats this operation as normal and supported.  This test illustrates a main adjantage of the \CFA sharing algorithm.  It also has a case in which STL's small-string optimization provides a successful mitigation.

\begin{figure}
    \includegraphics[width=\textwidth]{string-graph-pbv.png}
    \caption{Average time per iteration with one call to a function that takes a by-value string argument, comparing \CFA (having implicit sharing activated) with STL.  (a) With \emph{Varying-from-1} corpus construction, in which the STL-only benefit of small-string optimization occurs, in varying degrees, at all string sizes.  (b) With \emph{Fixed-size} corpus construction, in which this benefit applies exactly to strings with length below 16.  [TODO: show version (b)]}
    \label{fig:string-graph-pbv}
\end{figure}

Figure \ref{fig:string-graph-pbv} shows the costs for calling a function that receives a string argument by value.  STL's performance worsens as string length increases, while \CFA has the same performance at all sizes.

The \CFA cost to pass is nontrivial.  The contributor is adding and removing the callee's string handle from the global list.  This cost is $1.5 \times$ to $2 \times$ over STL's when small-string optimization applies, though this cost should be avoidable in the same case, given a \CFA realization of this optimization.  At the larger sizes, when STL has to manage storage for the string, STL runs more than $3 \times$ slower, mainly due to time spent in the general-purpose memory allocator.


\subsubsection{Test: Allocate}

This test directly compares the allocation schemes of the \CFA string with sharing, compared with the STL string.  It treats the \CFA scheme as a form of garbage collection, and the STL scheme as an application of malloc-free.  The test shows that \CFA enables faster speed at a cost in memory usage.

A garbage collector, afforded the freedom of managed memory, often runs faster than malloc-free (in an ammortized analysis, even though it must occasionally stop to collect) because it is able to use its collection time to move objects.  (In the case of the mini-allocator powering the \CFA string library, objects are runs of text.)  Moving objects lets fresh allocations consume from a large contiguous store of available memory; the ``bump pointer'' book-keeping for such a scheme is very light.  A malloc-free implementation without the freedom to move objects must, in the general case, allocate in the spaces between existing objects; doing so entails the heavier book-keeping to navigate and maintain a linked structure.

A garbage collector keeps allocations around for longer than the using program can reach them.  By contrast, a program using malloc-free (correctly) releases allocations exactly when they are no longer reachable.  Therefore, the same harness will use more memory while running under garbage collection.  A garbage collector can minimize the memory overhead by searching for these dead allocations aggressively, that is, by collecting more often.  Tuned in this way, it spends a lot of time collecting, easily so much as to overwhelm its speed advantage from bump-pointer allocation.  If it is tuned to collect rarely, then it leaves a lot of garbage allocated (waiting to be collected) but gains the advantage of little time spent doing collection.

[TODO: find citations for the above knowledge]

The speed for memory tradeoff is, therefore, standard for comparisons like \CFA--STL string allocations.  The test verifies that it is so and quantifies the returns available.

These tests manipulate a tuning knob that controls how much extra space to use.  Specific values of this knob are not user-visible and are not presented in the results here.  Instead, its two effects (amount of space used and time per operation) are shown.  The independent variable is the liveness target, which is the fraction of the text buffer that is in use at the end of a collection.  The allocator will expand its text buffer during a collection if the actual fraction live exceeds this target.

This experiment's driver allocates strings by constructing a string handle as a local variable then looping over recursive calls.  The time measurement is of nanoseconds per such allocating call.  The arrangement of recursive calls and their fan-out (iterations per recursion level) makes some of the strings long-lived and some of them short-lived.  String lifetime (measured in number of subsequent string allocations) is ?? distributed, because each node in the call tree survives as long as its descendent calls.  The run presented in this section used a call depth of 1000 and a fan-out of 1.006, which means that approximately one call in 167 makes two recursive calls, while the rest make one.  This sizing was chosen to keep the amounts of consumed memory within the machine's last-level cache.

\begin{figure}
    \includegraphics[width=\textwidth]{string-graph-allocn.png}
    \caption{Space and time performance, under varying fraction-live targets, for the five string lengths shown, at (emph{Fixed-size} corpus construction.  [MISSING] The identified clusters are for the default fraction-live target, which is 30\%.  MISSING: STL results, typically just below the 0.5--0.9 CFA segment.  All runs keep an average of 836 strings live, and the median string lifetime is ?? allocations.}
    \label{fig:string-graph-allocn}
\end{figure}

Figure \ref{fig:string-graph-allocn} shows the results of this experiemnt.  At all string sizes, varying the liveness threshold gives offers speed-for-space tradeoffs relative to STL.  At the default liveness threshold, all measured string sizes see a ??\%--??\% speedup for a ??\%--??\% increase in memory footprint.



\subsubsection{Test: Normalize}

This test is more applied than the earlier ones.  It combines the effects of several operations.  It also demonstrates a case of the CFA API allowing user code to perform well, while being written without overt memory management, while achieving similar performance in STL requires adding memory-management complexity.

To motivate: edits being rare

The program is doing a specialized find-replace operation on a large body of text.  In the program under test, the replacement is just to erase a magic character.  But in the larger software problem represented, the rewrite logic belongs to a module that was originally intended to operate on simple, modest-length strings.  The challenge is to apply this packaged function across chunks taken from the large body.  Using the \CFA string library, the most natural way to write the helper module's function also works well in the adapted context.  Using the STL string, the most natural ways to write the helper module's function, given its requirements in isolation, slow down when it is driven in the adapted context.

\begin{lstlisting}
void processItem( string & item ) {
    // find issues in item and fix them
}
\end{lstlisting}

\section{String I/O}