2 % (c) The OBFUSCATION-THROUGH-GRATUITOUS-PREPROCESSOR-ABUSE Project,
3 % Glasgow University, 1990-1994
8 % o I (ADR) think it would be worth making the connection with CPS explicit.
9 % Now that we have explicit activation records (on the stack), we can
10 % explain the whole system in terms of CPS and tail calls --- with the
11 % one requirement that we carefuly distinguish stack-allocated objects
12 % from heap-allocated objects.
14 % \documentstyle[preprint]{acmconf}
15 \documentclass[11pt]{article}
16 \oddsidemargin 0.1 in % Note that \oddsidemargin = \evensidemargin
17 \evensidemargin 0.1 in
18 \marginparwidth 0.85in % Narrow margins require narrower marginal notes
24 %\newcommand{\note}[1]{{\em Note: #1}}
25 \newcommand{\note}[1]{{{\bf Note:}\sl #1}}
26 \newcommand{\ToDo}[1]{{{\bf ToDo:}\sl #1}}
27 \newcommand{\Arg}[1]{\mbox{${\tt arg}_{#1}$}}
28 \newcommand{\bottom}{bottom} % foo, can't remember the symbol name
30 \newcommand{\secref}[1]{Section~\ref{sec:#1}}
31 \newcommand{\figref}[1]{Figure~\ref{fig:#1}}
32 \newcommand{\Section}[2]{\section{#1}\label{sec:#2}}
33 \newcommand{\Subsection}[2]{\subsection{#1}\label{sec:#2}}
34 \newcommand{\Subsubsection}[2]{\subsubsection{#1}\label{sec:#2}}
45 \setlength{\parskip}{0.15cm}
46 \setlength{\parsep}{0.15cm}
47 \setlength{\topsep}{0cm} % Reduces space before and after verbatim,
48 % which is implemented using trivlist
49 \setlength{\parindent}{0cm}
51 \renewcommand{\textfraction}{0.2}
52 \renewcommand{\floatpagefraction}{0.7}
56 \title{The STG runtime system (revised)}
57 \author{Simon Peyton Jones \\ Glasgow University and Oregon Graduate Institute \and
58 Simon Marlow \\ Glasgow University \and
59 Alastair Reid \\ Yale University}
67 \Section{Overview}{overview}
69 This document describes the GHC/Hugs run-time system. It serves as
70 a Glasgow/Yale/Nottingham ``contract'' about what the RTS does.
72 \Subsection{New features compared to GHC 2.04}{new-features}
75 \item The RTS supports mixed compiled/interpreted execution, so
76 that a program can consist of a mixture of GHC-compiled and Hugs-interpreted
79 \item The RTS supports concurrency by default.
80 This has some costs (eg we can't do hardware stack checks) but
81 reduces the number of different configurations we need to support.
83 \item CAFs are only retained if they are
84 reachable. Since they are referred to by implicit references buried
85 in code, this means that the garbage collector must traverse the whole
86 accessible code tree. This feature eliminates a whole class of painful
89 \item A running thread has only one stack, which contains a mixture of
90 pointers and non-pointers. \secref{stacks} describes how we find out
91 which is which. (GHC has used two stacks for some while. Using one
92 stack instead of two reduces register pressure, reduces the size of
93 update frames, and eliminates ``stack-stubbing'' instructions.)
95 \item The ``return in registers'' return convention has been dropped
96 because it was complicated and doesn't work well on register-poor
97 architectures. It has been partly replaced by unboxed tuples
98 (\secref{unboxed-tuples}) which allow the programmer to
99 explicitly state where results should be returned in registers (or on
100 the stack) instead of on the heap.
104 Lazy black-holing has been replaced by eager black-holing. The
105 problem with lazy black-holing is that it leaves slop in the heap
106 which conflicts with the use of a mostly-copying collector.
107 \ToDo{Why? I think we can still do lazy black holing without leaving
112 \Subsection{Wish list}{wish-list}
114 Here's a list of things we'd like to support in the future.
116 \item Interrupts, speculative computation.
119 The SM could tune the size of the allocation arena, the number of
120 generations, etc taking into account residency, GC rate and page fault
124 There should be no need to specify the amnount of stack/heap space to
125 allocate when you started a program - let it just take as much or as
126 little as it wants. (It might be useful to be able to specify maximum
127 sizes and to be able to suggest an initial size.)
130 We could trigger a GC when all threads are blocked waiting for IO if
131 the allocation arena (or some of the generations) are nearly full.
135 \Subsection{Configuration}{configuration}
137 Some of the above features are expensive or less portable, so we
138 envision building a number of different configurations supporting
139 different subsets of the above features.
141 You can make the following choices:
144 Support for parallelism. There are three mutually-exclusive choices.
147 \item[@SEQUENTIAL@] Support for concurrency but not for parallelism.
148 \item[@GRANSIM@] Concurrency support and simulated parallelism.
149 \item[@PARALLEL@] Concurrency support and real parallelism.
152 \item @PROFILING@ adds cost-centre profiling.
154 \item @TICKY@ gathers internal statistics (often known as ``ticky-ticky'' code).
156 \item @DEBUG@ does internal consistency checks.
158 \item Persistence. (well, not yet).
161 Which garbage collector to use. At the moment we
162 only anticipate one, however.
165 \Subsection{Glossary}{glossary}
167 \ToDo{This terminology is not used consistently within the document.
168 If you find something which disagrees with this terminology, fix the
171 In the type system, we have boxed and unboxed types.
175 \item A \emph{pointed} type is one that contains $\bot$. Variables with
176 pointed types are the only things which can be lazily evaluated. In
177 the STG machine, this means that they are the only things that can be
178 \emph{entered} or \emph{updated} and it requires that they be boxed.
180 \item An \emph{unpointed} type is one that does not contain $\bot$.
181 Variables with unpointed types are never delayed --- they are always
182 evaluated when they are constructed. In the STG machine, this means
183 that they cannot be \emph{entered} or \emph{updated}. Unpointed objects
184 may be boxed (like @Array#@) or unboxed (like @Int#@).
188 In the implementation, we have different kinds of objects:
192 \item \emph{boxed} objects are heap objects used by the evaluators
194 \item \emph{unboxed} objects are not heap allocated
196 \item \emph{stack} objects are allocated on the stack
198 \item \emph{closures} are objects which can be \emph{entered}.
199 They are always boxed and always have boxed types.
200 They may be in WHNF or they may be unevaluated.
202 \item A \emph{thunk} is a (representation of) a value of a \emph{pointed}
203 type which is \emph{not} in WHNF.
205 \item A \emph{value} is an object in WHNF. It can be pointed or unpointed.
211 At the hardware level, we have \emph{word}s and \emph{pointer}s.
215 \item A \emph{word} is (at least) 32 bits and can hold either a signed
218 \item A \emph{pointer} is (at least) 32 bits and big enough to hold a
219 function pointer or a data pointer.
223 Occasionally, a field of a data structure must hold either a word or a
224 pointer. In such circumstances, it is \emph{not safe} to assume that
225 words and pointers are the same size. \ToDo{GHC currently makes words
226 the same size as pointers to reduce complexity in the code
227 generator/RTS. It would be useful to relax this restriction, and have
228 eg. 32-bit Ints on a 64-bit machine.}
230 \subsection{Subtle Dependencies}
232 Some decisions have very subtle consequences which should be written
233 down in case we want to change our minds.
239 If the garbage collector is allowed to shrink the stack of a thread,
240 we cannot omit the stack check in return continuations
241 (\secref{heap-and-stack-checks}).
245 When we return to the scheduler, the top object on the stack is a closure.
246 The scheduler restarts the thread by entering the closure.
248 \secref{hugs-return-convention} discusses how Hugs returns an
249 unboxed value to GHC and how GHC returns an unboxed value to Hugs.
253 When we return to the scheduler, we need a few empty words on the stack
254 to store a closure to reenter. \secref{heap-and-stack-checks}
255 discusses who does the stack check and how much space they need.
259 Heap objects never contain slop --- this is required if we want to
260 support mostly-copying garbage collection.
262 This is a big problem when updating since the updatee is usually
263 bigger than an indirection object. The fix is to overwrite the end of
264 the updatee with ``slop objects'' (described in
265 \secref{slop-objects}). This is hard to arrange if we do
266 \emph{lazy} blackholing (\secref{lazy-black-holing}) so we
267 currently plan to blackhole an object when we push the update frame.
269 % Idea: have specialised update code for various common sizes of
270 % updatee, the update frame hence encodes the length of the object.
271 % Specialised indirections will also encode the length of the object. A
272 % generic version of the update code will overwrite the slop with a slop
273 % object. We can do the same thing for blackhole objects, or just have
274 % a generic version that is the same size as an indirection and
275 % overwrite the slop with a slop object when blackholing. So: does this
276 % avoid the need to do eager black holing?
280 Info tables for constructors contain enough information to decide which
281 return convention they use. This allows Hugs to use a single piece of
282 entry code for all constructors and insulates Hugs from changes in the
283 choice of return convention.
287 \Section{Source Language}{source-language}
289 \Subsection{Explicit Allocation}{explicit-allocation}
291 As in the original STG machine, (almost) all heap allocation is caused
292 by executing a let(rec). Since we no longer support the return in
293 registers convention for data constructors, constructors now cause heap
294 allocation and so they should be let-bound.
296 For example, we now write
298 > cons = \ x xs -> let r = (:) x xs in r
302 > cons = \ x xs -> (:) x xs
305 \note{For historical reasons, GHC doesn't use this syntax --- but it should.}
307 \Subsection{Unboxed tuples}{unboxed-tuples}
309 Functions can take multiple arguments as easily as they can take one
310 argument: there's no cost for adding another argument. But functions
311 can only return one result: the cost of adding a second ``result'' is
312 that the function must construct a tuple of ``results'' on the heap.
313 The assymetry is rather galling and can make certain programming
314 styles quite expensive. For example, consider a simple state transformer
317 > type S a = State -> (a,State)
318 > bindS m k s0 = case m s0 of { (a,s1) -> k a s1 }
319 > returnS a s = (a,s)
323 Here, every use of @returnS@, @getS@ or @setS@ constructs a new tuple
324 in the heap which is instantly taken apart (and becomes garbage) by
325 the case analysis in @bind@. Even a short state-transformer program
326 will construct a lot of these temporary tuples.
328 Unboxed tuples provide a way for the programmer to indicate that they
329 do not expect a tuple to be shared and that they do not expect it to
330 be allocated in the heap. Syntactically, unboxed tuples are just like
331 single constructor datatypes except for the annotation @unboxed@.
333 > data unboxed AAndState# a = AnS a State
334 > type S a = State -> AAndState# a
335 > bindS m k s0 = case m s0 of { AnS a s1 -> k a s1 }
336 > returnS a s = AnS a s
338 > setS s _ = AnS () s
340 Semantically, unboxed tuples are just unlifted tuples and are subject
341 to the same restrictions as other unpointed types.
343 Operationally, unboxed tuples are never built on the heap. When
344 an unboxed tuple is returned, it is returned in multiple registers
345 or multiple stack slots. At first sight, this seems a little strange
346 but it's no different from passing double precision floats in two
352 Unboxed tuples can only have one constructor and that
353 thunks never have unboxed types --- so we'll never try to update an
354 unboxed constructor. The restriction to a single constructor is
355 largely to avoid garbage collection complications.
358 The core syntax does not allow variables to be bound to
359 unboxed tuples (ie in default case alternatives or as function arguments)
360 and does not allow unboxed tuples to be fields of other constructors.
361 However, there's no harm in allowing it in the source syntax as a
362 convenient, but easily removed, syntactic sugar.
365 The compiler generates a closure of the form
367 > c = \ x y z -> C x y z
369 for every constructor (whether boxed or unboxed).
371 This closure is normally used during desugaring to ensure that
372 constructors are saturated and to apply any strictness annotations.
373 They are also used when returning unboxed constructors to the machine
374 code evaluator from the bytecode evaluator and when a heap check fails
375 in a return continuation for an unboxed-tuple scrutinee.
379 \Subsection{STG Syntax}{stg-syntax}
382 \ToDo{Insert STG syntax with appropriate changes.}
385 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
386 \part{System Overview}
387 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
389 This part is concerned with defining the external interfaces of the
390 major components of the system; the next part is concerned with their
393 The major components of the system are:
398 The evaluators (\secref{sm-overview}) are responsible for
399 evaluating heap objects. The system supports two evaluators: the
400 machine code evaluator; and the bytecode evaluator.
404 The scheduler (\secref{scheduler-overview}) acts as the
405 coordinator for the whole system. It is responsible for switching
406 between evaluators, switching between threads, garbage collection,
407 communication between multiple processors, etc.
411 The storage manager (\secref{evaluators-overview}) is
412 responsible for allocating blocks of contiguous memory and for garbage
417 The loader (\secref{loader-overview}) is responsible for
418 loading machine code and bytecode files from the file system and for
419 resolving references between separately compiled modules.
423 The compilers (\secref{compilers-overview}) generate machine
424 code and bytecode files which can be loaded by the loader.
428 \ToDo{Insert diagram showing all components underneath the scheduler
429 and communicating only with the scheduler}
432 \Section{The Evaluators}{evaluators-overview}
434 There are two evaluators: a machine code evaluator and a bytecode
435 evaluator. The evaluators task is to evaluate code within a thread
436 until one of the following happens:
441 \item it is preempted
442 \item it blocks in one of the concurrency primitives
443 \item it performs a safe ccall
444 \item it needs to switch to the other evaluator.
447 The evaluators expect to find a closure on top of the thread's stack
448 and terminate with a closure on top of the thread's stack.
450 \Subsection{Evaluation Model}{evaluation-model}
452 Whilst the evaluators differ internally, they share a common
453 evaluation model and many object representations.
455 \Subsubsection{Heap objects}{heap-objects-overview}
457 The choice of heap and stack objects used by the evaluators is tightly
458 bound to the evaluation model. This section provides an overview of
459 the most important heap and stack objects; further details are given
462 All heap objects look like this:
465 \begin{tabular}{|l|l|l|l|}\hline
466 \emph{Header} & \emph{Payload} \\ \hline
470 The headers vary between different kinds of object but they all start
471 with a pointer to a pair consisting of an \emph{info table} and some
472 \emph{entry code}. The info table is used both by the evaluators and
473 by the storage manager and contains an @INFO_TYPE@ field which
474 identifies which kind of heap object uses it and determines the
475 interpretation of the payload and of the other fields of the info
476 table. The entry code is some machine code used by the machine code
477 evaluator to evaluate closures and raises an error for other kinds of
480 The major kinds of heap object used are as follows. (For simplicity,
481 this description omits certain optimisations and extra fields required
482 by the garbage collector.)
486 \item[Constructors] are used to represent data constructors. Their
487 payload consists of the fields of the constructor; the tag of the
488 constructor is stored in the info table.
491 \begin{tabular}{|l|l|l|l|}\hline
492 @CONSTR@ & \emph{Fields} \\ \hline
496 \item[Primitive objects] are used to represent objects with unpointed
497 types which are too large to fit in a register (or stack slot) or for
498 which sharing must be preserved. Primitive objects include large
499 objects such as multiple precision integers and immutable arrays and
500 mutable objects such as mutable arrays, mutable variables, MVar's,
501 IVar's and foreign object pointers. Since unpointed objects are not
502 pointed, they cannot be entered. Their payload varies according to
505 \item[Function closures] are used to represent functions. Their
506 payload (if any) consists of the free variables of the function.
509 \begin{tabular}{|l|l|l|l|}\hline
510 @FUN@ & \emph{Free Variables} \\ \hline
514 Function closures are only generated by the machine code compiler.
516 \item[Thunks] are used to represent unevaluated expressions which will
517 be updated with their result. Their payload (if any) consists of the
518 free variables of the function. The entry code for a thunk starts by
519 pushing an \emph{update frame} onto the stack and overwriting the
520 thunk with a \emph{black hole} (see Black Holes, below). When
521 evaluation of the thunk completes, the update frame will cause the
522 thunk to be overwritten again with an \emph{indirection} to the result
523 of the thunk, which is always a constructor or a partial application.
526 \begin{tabular}{|l|l|l|l|}\hline
527 @THUNK@ & \emph{Free Variables} \\ \hline
531 Thunks are only generated by the machine code evaluator.
533 \item[Byte-code Objects (@BCO@s)] are generated by the bytecode
534 compiler. In conjunction with \emph{updatable applications} and
535 \emph{non-updatable applications} they are used to represent
536 functions, unevaluated expressions and return addresses.
539 \begin{tabular}{|l|l|l|l|}\hline
540 @BCO@ & \emph{Constant Pool} & \emph{Bytecodes} \\ \hline
544 \item[Non-updatable (Partial) Applications] are used to represent the
545 application of a function to an insufficient number of arguments.
546 Their payload consists of the function and the arguments received so far.
549 \begin{tabular}{|l|l|l|l|}\hline
550 @PAP@ & \emph{Function Closure} & \emph{Arguments} \\ \hline
554 @PAP@s are used when a function is applied to too few arguments and by
555 code generated by the lambda-lifting phase of the bytecode compiler.
557 \item[Updatable Applications] are used to represent the application of
558 a function to a sufficient number of arguments. Their payload
559 consists of the function and its arguments.
561 Updateable applications are like thunks: on entering an updateable
562 application, the evaluators push an \emph{update frame} onto the stack
563 and overwrite the application with a \emph{black hole}; when
564 evaluation completes, the evaluators overwrite the application with an
565 \emph{indirection} to the result of the application.
568 \begin{tabular}{|l|l|l|l|}\hline
569 @AP@ & \emph{Function Closure} & \emph{Arguments} \\ \hline
573 @AP@s are only generated by the bytecode compiler.
575 \item[Black holes] are used to mark updateable closures which are
576 currently being evaluated. ``Black holing'' an object cures a
577 potential space leak and detects certain classes of infinite loops.
578 More imporantly, black holes act as synchronisation objects between
579 separate threads: if a second thread tries to enter an updateable
580 closure which is already being evaluated, the second thread is added
581 to a list of blocked threads and the thread is suspended.
583 When evaluation of the black-holed closure completes, the black hole
584 is overwritten with an indirection to the result of the closure and
585 any blocked threads are restored to the runnable queue.
588 \begin{tabular}{|l|l|l|l|}\hline
589 @BH@ & \emph{Blocked threads} \\ \hline
593 \ToDo{In a single threaded system, it's trivial to detect infinite
594 loops: reentering a BH is always an error. How easy is it in a
595 multi-threaded system?}
597 \item[Indirections] are used to update an unevaluated closure with its
598 (usually fully evaluated) result in situations where it isn't possible
599 to perform an update in place. (In the current system, we always
600 update with an indirection to avoid duplicating the result when doing
604 \begin{tabular}{|l|l|l|l|}\hline
605 @IND@ & \emph{Closure} \\ \hline
609 Indirections needn't always point to a closure in WHNF. They can
610 point to a chain of indirections which point to an evaluated closure.
611 When revertible black holes are added, they may also point to reverted
614 \item[Thread State Objects (@TSO@s)] represent Haskell threads. Their
615 payload consists of a unique thread id, the status of the thread
616 (runnable, blocked, etc) and the stack. @TSO@s may be resized by the
617 scheduler if its stack is too small or too large.
620 \begin{tabular}{|l|l|l|l|}\hline
621 @TSO@ & \emph{Thread Id} & \emph{Status} & \emph{Stack} \\ \hline
627 \Subsubsection{Stack objects}{stack-objects-overview}
629 The stack contains a mixture of \emph{pending arguments} and
630 \emph{stack objects}.
632 Pending arguments are arguments to curried functions which have not
633 yet been incorporated into an activation frame. For example, when
634 evaluating @let { g x y = x + y; f x = g{x} } in f{3,4}@, the
635 evaluator pushes both arguments onto the stack and enters @f@. @f@
636 only requires one argument so it leaves the second argument as a
637 \emph{pending argument}. The pending argument remains on the stack
638 until @f@ calls @g@ which requires two arguments: the argument passed
639 to it by @f@ and the pending argument which was passed to @f@.
641 Unboxed pending arguments are always preceeded by a ``tag'' which says
642 how large the argument is. This allows the garbage collector to
643 locate pointers within the stack.
645 There are three kinds of stack object: return addresses, update frames
646 and seq frames. All stack objects look like this
649 \begin{tabular}{|l|l|l|l|}\hline
650 \emph{Header} & \emph{Payload} \\ \hline
654 As with heap objects, the header starts with a pointer to a pair
655 consisting of an \emph{info table} and some \emph{entry code}.
659 \item[Return addresses] are used to cause selection and execution of
660 case alternatives when a constructor is returned. Return addresses
661 generated by the machine code compiler look like this:
664 \begin{tabular}{|l|l|l|l|}\hline
665 \emph{@RET_ADDR@} & \emph{Free Variables of the case alternatives} \\ \hline
669 The free variables are a mixture of pointers and non-pointers whose
670 layout is described by the info table.
672 Return addresses generated by the bytecode compiler look like this:
674 \begin{tabular}{|l|l|l|l|}\hline
675 \emph{@BCO_RET@} & \emph{BCO} & \emph{Free Variables of the case alternatives} \\ \hline
679 There is just one @BCO_RET@ info pointer. We avoid needing different
680 @BCO_RET@s for each stack layout by tagging unboxed free variables as
681 though they were pending arguments.
683 \item[Update frames] are used to trigger updates. When an update
684 frame is entered, it overwrites the updatee with an indirection to the
685 result, restarts any threads blocked on the @BH@ and returns to the
686 stack object underneath the update frame.
689 \begin{tabular}{|l|l|l|l|}\hline
690 \emph{@UPDATE@} & \emph{Next Update Frame} & \emph{Updatee} \\ \hline
694 \item[Seq frames] are used to implement the polymorphic @seq@ primitive.
695 They are a special kind of update frame.
697 \ToDo{Describe them properly}
702 \ToDo{We also need a stop frame which goes on the bottom of the stack
703 when the thread terminates.}
706 \Subsubsection{Case expressions}{case-expr-overview}
708 In the STG language, all evaluation is triggered by evaluating a case
709 expression. When evaluating a case expression @case e of alts@, the
710 evaluators pushes a return address onto the stack and evaluate the
711 expression @e@. When @e@ eventually reduces to a constructor, the
712 return address on the stack is entered. The details of how the
713 constructor is passed to the return address and how the appropriate
714 case alternative is selected vary between evaluators.
716 Case expressions for unboxed data types are essentially the same: the
717 case expression pushes a return address onto the stack before
718 evaluating the scrutinee; when a function returns an unboxed value, it
719 enters the return address on top of the stack.
722 \Subsubsection{Function applications}{fun-app-overview}
724 In the STG language, all function calls are tail calls. The arguments
725 are pushed onto the stack and the function closure is entered. If any
726 arguments are unboxed, they must be tagged as unboxed pending
727 arguments. Entering a closure is just a special case of calling a
728 function with no arguments.
731 \Subsubsection{Let expressions}{let-expr-overview}
733 In the STG language, almost all heap allocation is caused by let
734 expressions. Filling in the contents of a set of mutually recursive
735 heap objects is simple enough; the only difficulty is that once the
736 heap space has been allocated, the thread must not return to the
737 scheduler until after the objects are filled in.
740 \Subsubsection{Primitive operations}{primop-overview}
744 Most primops are simple, some aren't.
751 \Section{Scheduler}{scheduler-overview}
753 The Scheduler is the heart of the run-time system. A running program
754 consists of a single running thread, and a list of runnable and
755 blocked threads. A thread is represented by a \emph{Thread Status
756 Object} (TSO), which contains a few words status information and a
757 stack. Except for the running thread, all threads have a closure on
758 top of their stack; the scheduler restarts a thread by entering an
759 evaluator which performs some reduction and returns to the scheduler.
761 \Subsection{The scheduler's main loop}{scheduler-main-loop}
763 The scheduler consists of a loop which chooses a runnable thread and
764 invokes one of the evaluators which performs some reduction and
767 The scheduler also takes care of system-wide issues such as heap
768 overflow or communication with other processors (in the parallel
769 system) and thread-specific problems such as stack overflow.
771 \Subsection{Creating a thread}{create-thread}
779 When the scheduler is first invoked.
783 When a message is received from another processor (I think). (Parallel
788 When a C program calls some Haskell code.
792 By @forkIO@, @takeMVar@ and (maybe) other Concurrent Haskell primitives.
797 \Subsection{Restarting a thread}{thread-restart}
799 When the scheduler decides to run a thread, it has to decide which
800 evaluator to use. It does this by looking at the type of the closure
803 \item @BCO@ $\Rightarrow$ bytecode evaluator
804 \item @FUN@ or @THUNK@ $\Rightarrow$ machine code evaluator
805 \item @CONSTR@ $\Rightarrow$ machine code evaluator
806 \item other $\Rightarrow$ either evaluator.
809 The only surprise in the above is that the scheduler must enter the
810 machine code evaluator if there's a constructor on top of the stack.
811 This allows the bytecode evaluator to return a constructor to a
812 machine code return address by pushing the constructor on top of the
813 stack and returning to the scheduler. If the return address under the
814 constructor is @HUGS_RET@, the entry code for @HUGS_RET@ will
815 rearrange the stack so that the return @BCO@ is on top of the stack
816 and return to the scheduler which will then call the bytecode
817 evaluator. There is little point in trying to shorten this slightly
818 indirect route since it is will happen very rarely if at all.
820 \note{As an optimisation, we could store the choice of evaluator in
821 the TSO status whenever we leave the evaluator. This is required for
822 any thread, no matter what state it is in (blocked, stack overflow,
823 etc). It isn't clear whether this would accomplish anything.}
825 \Subsection{Returning from a thread}{thread-return}
827 The evaluators return to the scheduler when any of the following
831 \item A heap check fails, and a garbage collection is required.
833 \item A stack check fails, and the scheduler must either enlarge the
834 current thread's stack, or flag an out of memory condition.
836 \item A thread enters a closure built by the other evaluator. That
837 is, when the bytecode interpreter enters a closure compiled by GHC or
838 when the machine code evaluator enters a BCO.
840 \item A thread returns to a return continuation built by the other
841 evaluator. That is, when the machine code evaluator returns to a
842 continuation built by Hugs or when the bytecode evaluator returns to a
843 continuation built by GHC.
845 \item The evaluator needs to perform a ``safe'' C call
848 \item The thread becomes blocked. This happens when a thread requires
849 the result of a computation currently being performed by another
850 thread, or it reads a synchronisation variable that is currently empty
853 \item The thread is preempted (the preemption mechanism is described
854 in \secref{thread-preemption}).
856 \item The thread terminates.
859 Except when the thread terminates, the thread always terminates with a
860 closure on the top of the stack. The mechanism used to trigger the
861 world switch and the choice of closure left on top of the stack varies
862 according to which world is being left and what is being returned.
864 \Subsubsection{Leaving the bytecode evaluator}{hugs-to-ghc-switch}
866 \paragraph{Entering a machine code closure}
868 When it enters a closure, the bytecode evaluator performs a switch
869 based on the type of closure (@AP@, @PAP@, @Ind@, etc). On entering a
870 machine code closure, it returns to the scheduler with the closure on
873 \paragraph{Returning a constructor}
875 When it enters a constructor, the bytecode evaluator tests the return
876 continuation on top of the stack. If it is a machine code
877 continuation, it returns to the scheduler with the constructor on top
880 \note{This is why the scheduler must enter the machine code evaluator
881 if it finds a constructor on top of the stack.}
883 \paragraph{Returning an unboxed value}
885 \note{Hugs doesn't support unboxed values in source programs but they
886 are used for a few complex primops.}
888 When it returns an unboxed value, the bytecode evaluator tests the
889 return continuation on top of the stack. If it is a machine code
890 continuation, it returns to the scheduler with the tagged unboxed
891 value and a special closure on top of the stack. When the closure is
892 entered (by the machine code evaluator), it returns the unboxed value
893 on top of the stack to the return continuation under it.
895 The runtime library for GHC provides one of these closures for each unboxed
896 type. Hugs cannot generate them itself since the entry code is really
899 \paragraph{Heap/Stack overflow and preemption}
901 The bytecode evaluator tests for heap/stack overflow and preemption
902 when entering a BCO and simply returns with the BCO on top of the
905 \Subsubsection{Leaving the machine code evaluator}{ghc-to-hugs-switch}
907 \paragraph{Entering a BCO}
909 The entry code for a BCO pushes the BCO onto the stack and returns to
912 \paragraph{Returning a constructor}
914 We avoid the need to test return addresses in the machine code
915 evaluator by pushing a special return address on top of a pointer to
916 the bytecode return continuation. \figref{hugs-return-stack}
917 shows the state of the stack just before evaluating the scrutinee.
929 %\input{hugs_return1.pstex_t}
931 \caption{Stack layout for evaluating a scrutinee}
932 \label{fig:hugs-return-stack}
935 This return address rearranges the stack so that the bco pointer is
936 above the constructor on the stack (as shown in
937 \figref{hugs-boxed-return}) and returns to the scheduler.
944 | con |--> Constructor
949 %\input{hugs_return2.pstex_t}
951 \caption{Stack layout for entering a Hugs return address}
952 \label{fig:hugs-boxed-return}
955 \paragraph{Returning an unboxed value}
957 We avoid the need to test return addresses in the machine code
958 evaluator by pushing a special return address on top of a pointer to
959 the bytecode return continuation. This return address rearranges the
960 stack so that the bco pointer is above the tagged unboxed value (as
961 shown in \figref{hugs-entering-unboxed-return}) and returns to the
976 %\input{hugs_return2.pstex_t}
978 \caption{Stack layout for returning an unboxed value}
979 \label{fig:hugs-entering-unboxed-return}
982 \paragraph{Heap/Stack overflow and preemption}
987 \Subsection{Preempting a thread}{thread-preemption}
989 Strictly speaking, threads cannot be preempted --- the scheduler
990 merely sets a preemption request flag which the thread must arrange to
991 test on a regular basis. When an evaluator finds that the preemption
992 request flag is set, it pushes an appropriate closure onto the stack
993 and returns to the scheduler.
995 In the bytecode interpreter, the flag is tested whenever we enter a
996 closure. If the preemption flag is set, it leaves the closure on top
997 of the stack and returns to the scheduler.
999 In the machine code evaluator, the flag is only tested when a heap or
1000 stack check fails. This is less expensive than testing the flag on
1001 entering every closure but runs the risk that a thread will enter an
1002 infinite loop which does not allocate any space. If the flag is set,
1003 the evaluator returns to the scheduler exactly as if a heap check had
1006 \Subsection{``Safe'' and ``unsafe'' C calls}{c-calls}
1008 There are two ways of calling C:
1012 \item[``Unsafe'' C calls] are used if the programer is certain that
1013 the C function will not do anything dangerous. Unsafe C calls are
1014 faster but must be hand-checked by the programmer.
1016 Dangerous things include:
1022 Call a system function such as @getchar@ which might block
1023 indefinitely. This is dangerous because we don't want the entire
1024 runtime system to block just because one thread blocks.
1028 Call an RTS function which will block on the RTS access semaphore.
1029 This would lead to deadlock.
1033 Call a Haskell function. This is just a special case of calling an
1038 Unsafe C calls are performed by pushing the arguments onto the C stack
1039 and jumping to the C function's entry point. On exit, the result of
1040 the function is in a register which is returned to the Haskell code as
1043 \item[``Safe'' C calls] are used if the programmer suspects that the
1044 thread may do something dangerous. Safe C calls are relatively slow
1045 but are less problematic.
1047 Safe C calls are performed by pushing the arguments onto the Haskell
1048 stack, pushing a return continuation and returning a \emph{C function
1049 descriptor} to the scheduler. The scheduler suspends the Haskell thread,
1050 spawns a new operating system thread which pops the arguments off the
1051 Haskell stack onto the C stack, calls the C function, pushes the
1052 function result onto the Haskell stack and informs the scheduler that
1053 the C function has completed and the Haskell thread is now runnable.
1057 The bytecode evaluator will probably treat all C calls as being safe.
1059 \ToDo{It might be good for the programmer to indicate how the program
1060 is unsafe. For example, if we distinguish between C functions which
1061 might call Haskell functions and those which might block, we could
1062 perform an unsafe call for blocking functions in a single-threaded
1063 system or, perhaps, in a multi-threaded system which only happens to
1064 have a single thread at the moment.}
1068 \Section{The Storage Manager}{sm-overview}
1070 The storage manager is responsible for managing the heap and all
1071 objects stored in it. It provides special support for lazy evaluation
1072 and for foreign function calls.
1074 \Subsection{SM support for lazy evaluation}{sm-lazy-evaluation}
1079 Indirections are shorted out.
1083 Update frames pointing to unreachable objects are squeezed out.
1087 Adjacent update frames (for different closures) are compressed to a
1088 single update frame pointing to a single black hole.
1093 \Subsection{SM support for foreign function calls}{sm-foreign-calls}
1099 Stable pointers allow other languages to access Haskell objects.
1103 Foreign Objects are a form of weak pointer which lets Haskell access
1108 \Subsection{Misc}{sm-misc}
1114 If the stack contains a large amount of free space, the storage
1115 manager may shrink the stack. If it shrinks the stack, it guarantees
1116 never to leave less than @MIN_SIZE_SHRUNKEN_STACK@ empty words on the
1117 stack when it does so.
1119 \ToDo{Would it be useful for the storage manager to enlarge the stack?}
1123 For efficiency reasons, very large objects (eg large arrays and TSOs)
1124 are not moved if possible.
1129 \Section{The Compilers}{compilers-overview}
1131 Need to describe interface files, format of bytecode files, symbols
1132 defined by machine code files.
1134 \Subsection{Interface Files}{interface-files}
1136 Here's an example - but I don't know the grammar - ADR.
1142 1 main _:_ IOBase.IO PrelBase.();;
1145 \Subsection{Bytecode files}{bytecode-files}
1147 (All that matters here is what the loader sees.)
1149 \Subsection{Machine code files}{asm-files}
1151 (Again, all that matters is what the loader sees.)
1153 \Section{The Loader}{loader-overview}
1155 In a batch mode system, we can statically link all the modules
1156 together. In an interactive system we need a loader which will
1157 explicitly load and unload individual modules (or, perhaps, blocks of
1158 mutually dependent modules) and resolve references between modules.
1160 While many operating systems provide support for dynamic loading and
1161 will automatically resolve cross-module references for us, we generally
1162 cannot rely on being able to load mutually dependent modules.
1164 A portable solution is to perform some of the linking ourselves. Each module
1165 should provide three global symbols:
1168 An initialisation routine. (Might also be used for finalisation.)
1170 A table of symbols it exports.
1171 Entries in this table consist of the symbol name and the address of the
1174 A table of symbols it imports.
1175 Entries in this table consist of the symbol name and a list of references
1179 On loading a group of modules, the loader adds the contents of the
1180 export lists to a symbol table and then fills in all the references in the
1183 References in import lists are of two types:
1185 \item[ References in machine code ]
1187 The most efficient approach is to patch the machine code directly, but
1188 this will be a lot of work, very painful to port and rather fragile.
1190 Alternatively, the loader could store the value of each symbol in the
1191 import table for each module and the compiled code can access all
1192 external objects through the import table. This requires that the
1193 import table be writable but does not require that the machine code or
1194 info tables be writable.
1196 \item[ References in data structures (SRTs and static data constructors) ]
1198 Either we patch the SRTs and constructors directly or we somehow use
1199 indirections through the symbol table. Patching the SRTs requires
1200 that we make them writable and prevents us from making effective use
1201 of virtual memories that use copy-on-write policies (this only makes a
1202 difference if we want to run several copies of the same program
1203 simultaneously). Using an indirection is possible but tricky.
1205 Note: We could avoid patching machine code if all references to
1206 external references went through the SRT --- then we just have one
1207 thing to patch. But the SRT always contains a pointer to the closure
1208 rather than the fast entry point (say), so we'd take a big performance
1213 Using the above scheme, all accesses to ``external'' objects involve a
1214 layer of indirection. To avoid this overhead, the machine code
1215 compiler might provide a way for the programmer to specify which
1216 modules will be statically linked and which will be dynamically linked
1217 --- the idea being that statically linked code and data will be
1221 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1222 \part{Internal details}
1223 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1225 This part is concerned with the internal details of the components
1226 described in the previous part.
1228 The major components of the system are:
1230 \item The scheduler (\secref{storage-manager-internals})
1231 \item The storage manager (\secref{storage-manager-internals})
1232 \item The evaluators
1237 \Section{The Scheduler}{scheduler-internals}
1239 \ToDo{Detailed description of scheduler}
1241 Many heap objects contain fields allowing them to be inserted onto lists
1242 during evaluation or during garbage collection. The lists required by
1243 the evaluator and storage manager are as follows.
1247 \item 4 lists of threads: runnable threads, sleeping threads, threads
1248 waiting for timeout and threads waiting for I/O.
1250 \item The \emph{mutables list} is a list of all objects in the old
1251 generation which might contain pointers into the new generation. Most
1252 of the objects on this list are indirections (\secref{IND})
1253 or ``mutable.'' (\secref{mutables}.)
1255 \item The \emph{Foreign Object list} is a list of all foreign objects
1256 which have not yet been deallocated. (\secref{FOREIGN}.)
1258 \item The \emph{Spark pool} is a doubly(?) linked list of Spark objects
1259 maintained by the parallel system. (\secref{SPARK}.)
1261 \item The \emph{Blocked Fetch list} (or
1262 lists?). (\secref{BLOCKED_FETCH}.)
1264 \item For each thread, there is a list of all update frames on the
1265 stack. (\secref{data-updates}.)
1267 \item The Stable Pointer Table is a table of pointers to objects which
1268 are known to the outside world and must be retained by the garbage
1269 collector even if they are not accessible from within the heap.
1273 \ToDo{The links for these fields are usually inserted immediately
1274 after the fixed header except ...}
1278 \Section{The Storage Manager}{storage-manager-internals}
1280 \subsection{Misc Text looking for a home}
1282 A \emph{value} may be:
1284 \item \emph{Boxed}, i.e.~represented indirectly by a pointer to a heap object (e.g.~foreign objects, arrays); or
1285 \item \emph{Unboxed}, i.e.~represented directly by a bit-pattern in one or more registers (e.g.~@Int#@ and @Float#@).
1287 All \emph{pointed} values are \emph{boxed}.
1290 \Subsection{Heap Objects}{heap-objects}
1296 \ToDo{Fix this picture}
1301 Every \emph{heap object} is a contiguous block of memory, consisting
1302 of a fixed-format \emph{header} followed by zero or more \emph{data
1305 The header consists of the following fields:
1307 \item A one-word \emph{info pointer}, which points to
1308 the object's static \emph{info table}.
1309 \item Zero or more \emph{admin words} that support
1311 \item Profiling (notably a \emph{cost centre} word).
1312 \note{We could possibly omit the cost centre word from some
1313 administrative objects.}
1314 \item Parallelism (e.g. GranSim keeps the object's global address here,
1315 though GUM keeps a separate hash table).
1316 \item Statistics (e.g. a word to track how many times a thunk is entered.).
1318 We add a Ticky word to the fixed-header part of closures. This is
1319 used to indicate if a closure has been updated but not yet entered. It
1320 is set when the closure is updated and cleared when subsequently
1321 entered. \footnote{% NB: It is \emph{not} an ``entry count'', it is
1322 an ``entries-after-update count.'' The commoning up of @CONST@,
1323 @CHARLIKE@ and @INTLIKE@ closures is turned off(?) if this is
1324 required. This has only been done for 2s collection. }
1329 Most of the RTS is completely insensitive to the number of admin
1330 words. The total size of the fixed header is @FIXED_HS@.
1332 \Subsection{Info Tables}{info-tables}
1334 An \emph{info table} is a contiguous block of memory, laid out as follows:
1337 \begin{tabular}{|r|l|}
1338 \hline Parallelism Info & variable
1339 \\ \hline Profile Info & variable
1340 \\ \hline Debug Info & variable
1341 \\ \hline Static reference table & 32 bits (optional)
1342 \\ \hline Storage manager layout info & 32 bits
1343 \\ \hline Closure type & 16 bits
1344 \\ \hline Constructor Tag & 16 bits
1345 \\ \hline entry code
1350 An info table has the following contents (working backwards in memory
1355 \item The \emph{entry code} for the closure. This code appears
1356 literally as the (large) last entry in the info table, immediately
1357 preceded by the rest of the info table. An \emph{info pointer} always
1358 points to the first byte of the entry code.
1360 \item A 16-bit constructor tag. This field is used for constructor
1361 info-tables only (\secref{CONSTR}), and contains an integer
1362 representing the tag value of the constructor, in the range $0..n-1$
1363 where $n$ is the number of constructors in the datatype.
1365 \item An 16-bit {\em closure type field}, which identifies what kind of
1366 closure the object is. The various types of closure are described in
1369 \item A single pointer or word --- the {\em storage manager info
1370 field}, contains auxiliary information describing the closure's
1371 precise layout, for the benefit of the garbage collector and the code
1372 that stuffs graph into packets for transmission over the network.
1373 There are three kinds of layout information:
1376 \item Standard layout information is for closures which place pointers
1377 before non-pointers in instances of the closure (this applies to most
1378 heap-based and static closures, but not activation records). The
1379 layout information for standard closures is
1382 \item Number of pointer fields (16 bits).
1383 \item Number of non-pointer fields (16 bits).
1386 \item Activation records don't have pointers before non-pointers,
1387 since stack-stubbing requires that the record has holes in it. The
1388 layout is therefore represented by a bitmap in which each '1' bit
1389 represents a non-pointer word. This kind of layout info is used for
1390 @RET_SMALL@ and @RET_VEC_SMALL@ closures.
1392 \item If an activation record is longer than 32 words, then the layout
1393 field contains a pointer to a bitmap record, consisting of a length
1394 field followed by two or more bitmap words. This layout information
1395 is used for @RET_BIG@ and @RET_VEC_BIG@ closures.
1397 \item Selector Thunks (\secref{THUNK_SEL}) use the closure
1398 layout field to hold the selector index, since the layout is always
1399 known (the closure contains a single pointer field).
1402 \item A one-word {\em Static Reference Table} field. This field
1403 points to the static reference table for the closure (\secref{srt}),
1404 and is only present for the following closure types:
1412 \item \emph{Profiling info\/}
1414 \ToDo{The profiling info is completely bogus. I've not deleted it
1415 from the document but I've commented it all out.}
1417 % change to \iftrue to uncomment this section
1420 Closure category records are attached to the info table of the
1421 closure. They are declared with the info table. We put pointers to
1422 these ClCat things in info tables. We need these ClCat things because
1423 they are mutable, whereas info tables are immutable. Hashing will map
1424 similar categories to the same hash value allowing statistics to be
1425 grouped by closure category.
1427 Cost Centres and Closure Categories are hashed to provide indexes
1428 against which arbitrary information can be stored. These indexes are
1429 memoised in the appropriate cost centre or category record and
1430 subsequent hashes avoided by the index routine (it simply returns the
1433 There are different features which can be hashed allowing information
1434 to be stored for different groupings. Cost centres have the cost
1435 centre recorded (using the pointer), module and group. Closure
1436 categories have the closure description and the type
1437 description. Records with the same feature will be hashed to the same
1440 The initialisation routines, @init_index_<feature>@, allocate a hash
1441 table in which the cost centre / category records are stored. The
1442 lower bound for the table size is taken from @max_<feature>_no@. They
1443 return the actual table size used (the next power of 2). Unused
1444 locations in the hash table are indicated by a 0 entry. Successive
1445 @init_index_<feature>@ calls just return the actual table size.
1447 Calls to @index_<feature>@ will insert the cost centre / category
1448 record in the @<feature>@ hash table, if not already inserted. The hash
1449 index is memoised in the record and returned.
1451 CURRENTLY ONLY ONE MEMOISATION SLOT IS AVILABLE IN EACH RECORD SO
1452 HASHING CAN ONLY BE DONE ON ONE FEATURE FOR EACH RECORD. This can be
1453 easily relaxed at the expense of extra memoisation space or continued
1456 The initialisation routines must be called before initialisation of
1457 the stacks and heap as they require to allocate storage. It is also
1458 expected that the caller may want to allocate additional storage in
1459 which to store profiling information based on the return table size
1463 \begin{tabular}{|l|}
1467 \\ \hline Description String
1468 \\ \hline Type String
1474 \item[Hash Index] Memoised copy
1476 Is this category selected (-1 == not memoised, selected? 0 or 1)
1478 One of the following values (defined in CostCentre.lh):
1486 A partial application.
1488 A thunk, or suspension.
1493 \item[@ForeignObj_K@]
1494 A Foreign object (non-Haskell heap resident).
1496 The Stable Pointer table. (There should only be one of these but it
1497 represents a form of weak space leak since it can't shrink to meet
1498 non-demand so it may be worth watching separately? ADR)
1499 \item[@INTERNAL_KIND@]
1500 Something internal to the runtime system.
1504 \item[Description] Source derived string detailing closure description.
1505 \item[Type] Source derived string detailing closure type.
1508 \fi % end of commented out stuff
1510 \item \emph{Parallelism info\/}
1513 \item \emph{Debugging info\/}
1519 %-----------------------------------------------------------------------------
1520 \Subsection{Kinds of Heap Object}{closures}
1522 Heap objects can be classified in several ways, but one useful one is
1526 \emph{Static closures} occupy fixed, statically-allocated memory
1527 locations, with globally known addresses.
1530 \emph{Dynamic closures} are individually allocated in the heap.
1533 \emph{Stack closures} are closures allocated within a thread's stack
1534 (which is itself a heap object). Unlike other closures, there are
1535 never any pointers to stack closures. Stack closures are discussed in
1539 A second useful classification is this:
1542 \item \emph{Executive objects}, such as thunks and data constructors,
1543 participate directly in a program's execution. They can be subdivided
1544 into three kinds of objects according to their type: \begin{itemize}
1546 \item \emph{Pointed objects}, represent values of a \emph{pointed}
1547 type (<.pointed types launchbury.>) --i.e.~a type that includes
1548 $\bottom$ such as @Int@ or @Int# -> Int#@.
1550 \item \emph{Unpointed objects}, represent values of a \emph{unpointed}
1551 type --i.e.~a type that does not include $\bottom$ such as @Int#@ or
1554 \item \emph{Activation frames}, represent ``continuations''. They are
1555 always stored on the stack and are never pointed to by heap objects or
1556 passed as arguments. \note{It's not clear if this will still be true
1557 once we support speculative evaluation.}
1561 \item \emph{Administrative objects}, such as stack objects and thread
1562 state objects, do not represent values in the original program.
1565 Only pointed objects can be entered. All pointed objects share a
1566 common header format: the ``pointed header''; while all unpointed
1567 objects share a common header format: the ``unpointed header''.
1568 \ToDo{Describe the difference and update the diagrams to mention an
1569 appropriate header type.}
1571 This section enumerates all the kinds of heap objects in the system.
1572 Each is identified by a distinct closure type field in its info table.
1574 \begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|}
1577 closure type & Section \\
1583 @CONSTR@ & \ref{sec:CONSTR} \\
1584 @CONSTR_STATIC@ & \ref{sec:CONSTR} \\
1585 @CONSTR_STATIC_NOCAF@ & \ref{sec:CONSTR} \\
1587 @FUN@ & \ref{sec:FUN} \\
1588 @FUN_STATIC@ & \ref{sec:FUN} \\
1590 @THUNK@ & \ref{sec:THUNK} \\
1591 @THUNK_STATIC@ & \ref{sec:THUNK} \\
1592 @THUNK_SELECTOR@ & \ref{sec:THUNK_SEL} \\
1594 @BCO@ & \ref{sec:BCO} \\
1595 @BCO_CAF@ & \ref{sec:BCO} \\
1597 @AP@ & \ref{sec:AP} \\
1598 @PAP@ & \ref{sec:PAP} \\
1600 @IND@ & \ref{sec:IND} \\
1601 @IND_OLDGEN@ & \ref{sec:IND} \\
1602 @IND_PERM@ & \ref{sec:IND} \\
1603 @IND_OLDGEN_PERM@ & \ref{sec:IND} \\
1604 @IND_STATIC@ & \ref{sec:IND} \\
1610 @ARR_WORDS@ & \ref{sec:ARR_WORDS1},\ref{sec:ARR_WORDS2} \\
1611 @ARR_PTRS@ & \ref{sec:ARR_PTRS} \\
1612 @MUTVAR@ & \ref{sec:MUTVAR} \\
1613 @MUTARR_PTRS@ & \ref{sec:MUTARR_PTRS} \\
1614 @MUTARR_PTRS_FROZEN@ & \ref{sec:MUTARR_PTRS_FROZEN} \\
1616 @FOREIGN@ & \ref{sec:FOREIGN} \\
1618 @BH@ & \ref{sec:BH} \\
1619 @MVAR@ & \ref{sec:MVAR} \\
1620 @IVAR@ & \ref{sec:IVAR} \\
1621 @FETCHME@ & \ref{sec:FETCHME} \\
1625 Activation frames do not live (directly) on the heap --- but they have
1626 a similar organisation.
1628 \begin{tabular}{|l|l|}\hline
1629 closure type & Section \\ \hline
1630 @RET_SMALL@ & \ref{sec:activation-records} \\
1631 @RET_VEC_SMALL@ & \ref{sec:activation-records} \\
1632 @RET_BIG@ & \ref{sec:activation-records} \\
1633 @RET_VEC_BIG@ & \ref{sec:activation-records} \\
1634 @UPDATE_FRAME@ & \ref{sec:activation-records} \\
1638 There are also a number of administrative objects.
1640 \begin{tabular}{|l|l|}\hline
1641 closure type & Section \\ \hline
1642 @TSO@ & \ref{sec:TSO} \\
1643 @STABLEPTR_TABLE@ & \ref{sec:STABLEPTR_TABLE} \\
1644 @SPARK_OBJECT@ & \ref{sec:SPARK} \\
1645 @BLOCKED_FETCH@ & \ref{sec:BLOCKED_FETCH} \\
1649 \ToDo{I guess the parallel system has something like a stable ptr
1650 table. Is there any opportunity for sharing code/data structures
1654 \Subsection{Predicates}{closure-predicates}
1656 \ToDo{The following is a first attempt at defining a useful set of
1657 predicates. Some (such as @isWHNF@ and @isSparkable@) may need their
1658 definitions tweaked a little.}
1660 The runtime system sometimes needs to be able to distinguish objects
1661 according to their properties: is the object updateable? is it in weak
1662 head normal form? etc. These questions can be answered by examining
1663 the closure type field of the object's info table.
1665 We define the following predicates to detect families of related
1666 info types. They are mutually exclusive and exhaustive.
1669 \item @isCONSTR@ is true for @CONSTR@s.
1670 \item @isFUN@ is true for @FUN@s.
1671 \item @isTHUNK@ is true for @THUNK@s.
1672 \item @isBCO@ is true for @BCO@s.
1673 \item @isAP@ is true for @AP@s.
1674 \item @isPAP@ is true for @PAP@s.
1675 \item @isINDIRECTION@ is true for indirection objects.
1676 \item @isBH@ is true for black holes.
1677 \item @isFOREIGN_OBJECT@ is true for foreign objects.
1678 \item @isARRAY@ is true for array objects.
1679 \item @isMVAR@ is true for @MVAR@s.
1680 \item @isIVAR@ is true for @IVAR@s.
1681 \item @isFETCHME@ is true for @FETCHME@s.
1682 \item @isSLOP@ is true for slop objects.
1683 \item @isRET_ADDR@ is true for return addresses.
1684 \item @isUPD_ADDR@ is true for update frames.
1685 \item @isTSO@ is true for @TSO@s.
1686 \item @isSTABLE_PTR_TABLE@ is true for the stable pointer table.
1687 \item @isSPARK_OBJECT@ is true for spark objects.
1688 \item @isBLOCKED_FETCH@ is true for blocked fetch objects.
1689 \item @isINVALID_INFOTYPE@ is true for all other info types.
1693 The following predicates detect other interesting properties:
1697 \item @isPOINTED@ is true if an object has a pointed type.
1699 If an object is pointed, the following predicates may be true
1700 (otherwise they are false). @isWHNF@ and @isUPDATEABLE@ are
1704 \item @isWHNF@ is true if the object is in Weak Head Normal Form.
1705 Note that unpointed objects are (arbitrarily) not considered to be in WHNF.
1707 @isWHNF@ is true for @PAP@s, @CONSTR@s, @FUN@s and all @BCO@s.
1709 \ToDo{Need to distinguish between whnf BCOs and non-whnf BCOs in their
1712 \item @isUPDATEABLE@ is true if the object may be overwritten with an
1715 @isUPDATEABLE@ is true for @THUNK@s, @AP@s and @BH@s.
1719 It is possible for a pointed object to be neither updatable nor in
1720 WHNF. For example, indirections.
1722 \item @isUNPOINTED@ is true if an object has an unpointed type.
1723 All such objects are boxed since only boxed objects have info pointers.
1725 It is true for @ARR_WORDS@, @ARR_PTRS@, @MUTVAR@, @MUTARR_PTRS@,
1726 @MUTARR_PTRS_FROZEN@, @FOREIGN@ objects, @MVAR@s and @IVAR@s.
1728 \item @isACTIVATION_FRAME@ is true for activation frames of all sorts.
1730 It is true for return addresses and update frames.
1732 \item @isVECTORED_RETADDR@ is true for vectored return addresses.
1733 \item @isDIRECT_RETADDR@ is true for direct return addresses.
1736 \item @isADMINISTRATIVE@ is true for administrative objects:
1737 @TSO@s, the stable pointer table, spark objects and blocked fetches.
1739 \item @hasSRT@ is true if the info table for the object contains an
1742 @hasSRT@ is true for @THUNK@s, @FUN@s, and @RET@s.
1748 \item @isSTATIC@ is true for any statically allocated closure.
1750 \item @isMUTABLE@ is true for objects with mutable pointer fields:
1751 @MUT_ARR@s, @MUTVAR@s, @MVAR@s and @IVAR@s.
1753 \item @isSparkable@ is true if the object can (and should) be sparked.
1754 It is true of updateable objects which are not in WHNF with the
1755 exception of @THUNK_SELECTOR@s and black holes.
1759 As a minor optimisation, we might use the top bits of the @INFO_TYPE@
1760 field to ``cache'' the answers to some of these predicates.
1762 An indirection either points to HNF (post update); or is result of
1763 overwriting a FetchMe, in which case the thing fetched is either
1764 under evaluation (BH), or by now an HNF. Thus, indirections get NoSpark flag.
1769 #define _NF 0x0001 /* Normal form */
1770 #define _NS 0x0002 /* Don't spark */
1771 #define _ST 0x0004 /* Is static */
1772 #define _MU 0x0008 /* Is mutable */
1773 #define _UP 0x0010 /* Is updatable (but not mutable) */
1774 #define _BM 0x0020 /* Is a "rimitive" array */
1775 #define _BH 0x0040 /* Is a black hole */
1776 #define _IN 0x0080 /* Is an indirection */
1777 #define _TH 0x0100 /* Is a thunk */
1782 SPEC_N SPEC | _NF | _NS
1784 SPEC_U SPEC | _UP | _TH
1787 GEN_N GEN | _NF | _NS
1789 GEN_U GEN | _UP | _TH
1792 TUPLE _NF | _NS | _BM
1793 DATA _NF | _NS | _BM
1794 MUTUPLE _NF | _NS | _MU | _BM
1795 IMMUTUPLE _NF | _NS | _BM
1807 CAF _NF | _NS | _ST | _IN
1816 STKO_DYNAMIC STKO | _MU
1817 STKO_STATIC STKO | _ST
1819 SPEC_RBH _NS | _MU | _BH
1820 GEN_RBH _NS | _MU | _BH
1828 \subsection{Closures (aka Pointed Objects)}
1830 An object can be entered iff it is a closure.
1832 \Subsubsection{Function closures}{FUN}
1834 Function closures represent lambda abstractions. For example,
1835 consider the top-level declaration:
1837 f = \x -> let g = \y -> x+y
1840 Both @f@ and @g@ are represented by function closures. The closure
1841 for @f@ is \emph{static} while that for @g@ is \emph{dynamic}.
1843 The layout of a function closure is as follows:
1845 \begin{tabular}{|l|l|l|l|}\hline
1846 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \\ \hline
1849 The data words (pointers and non-pointers) are the free variables of
1850 the function closure.
1851 The number of pointers
1852 and number of non-pointers are stored in the @INFO_SM@ word, in the least significant
1853 and most significant half-word respectively.
1855 There are several different sorts of function closure, distinguished
1856 by their closure type field:
1860 \item @FUN@: a vanilla, dynamically allocated on the heap.
1862 \item $@FUN_@p@_@np$: to speed up garbage collection a number of
1863 specialised forms of @FUN@ are provided, for particular $(p,np)$
1864 pairs, where $p$ is the number of pointers and $np$ the number of
1867 \item @FUN_STATIC@. Top-level, static, function closures (such as @f@
1868 above) have a different layout than dynamic ones:
1871 \begin{tabular}{|l|l|l|}\hline
1872 \emph{Fixed header} & \emph{Static object link} \\ \hline
1876 Static function closures have no free variables. (However they may
1877 refer to other static closures; these references are recorded in the
1878 function closure's SRT.) They have one field that is not present in
1879 dynamic closures, the \emph{static object link} field. This is used
1880 by the garbage collector in the same way that to-space is, to gather
1881 closures that have been determined to be live but that have not yet
1884 \note{Static function closures that have no static references, and
1885 hence a null SRT pointer, don't need the static object link field. Is
1886 it worth taking advantage of this? See @CONSTR_STATIC_NOCAF@.}
1889 Each lambda abstraction, $f$, in the STG program has its own private
1890 info table. The following labels are relevant:
1894 \item $f$@_info@ is $f$'s info table.
1896 \item $f$@_entry@ is $f$'s slow entry point (i.e. the entry code of
1897 its info table; so it will label the same byte as $f$@_info@).
1899 \item $f@_fast_@k$ is $f$'s fast entry point. $k$ is the number of
1900 arguments $f$ takes; encoding this number in the fast-entry label
1901 occasionally catches some nasty code-generation errors.
1905 \Subsubsection{Data constructors}{CONSTR}
1907 Data-constructor closures represent values constructed with algebraic
1908 data type constructors. The general layout of data constructors is
1909 the same as that for function closures. That is
1912 \begin{tabular}{|l|l|l|l|}\hline
1913 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \\ \hline
1917 There are several different sorts of constructor:
1921 \item @CONSTR@: a vanilla, dynamically allocated constructor.
1923 \item @CONSTR_@$p$@_@$np$: just like $@FUN_@p@_@np$.
1925 \item @CONSTR_INTLIKE@. A dynamically-allocated heap object that
1926 looks just like an @Int@. The garbage collector checks to see if it
1927 can common it up with one of a fixed set of static int-like closures,
1928 thus getting it out of the dynamic heap altogether.
1930 \item @CONSTR_CHARLIKE@: same deal, but for @Char@.
1932 \item @CONSTR_STATIC@ is similar to @FUN_STATIC@, with the
1933 complication that the layout of the constructor must mimic that of a
1934 dynamic constructor, because a static constructor might be returned to
1935 some code that unpacks it. So its layout is like this:
1938 \begin{tabular}{|l|l|l|l|l|}\hline
1939 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} & \emph{Static object link}\\ \hline
1943 The static object link, at the end of the closure, serves the same purpose
1944 as that for @FUN_STATIC@. The pointers in the static constructor can point
1945 only to other static closures.
1947 The static object link occurs last in the closure so that static
1948 constructors can store their data fields in exactly the same place as
1949 dynamic constructors.
1951 \item @CONSTR_STATIC_NOCAF@. A statically allocated data constructor
1952 that guarantees not to point (directly or indirectly) to any CAF
1953 (\secref{CAF}). This means it does not need a static object
1954 link field. Since we expect that there might be quite a lot of static
1955 constructors this optimisation makes sense. Furthermore, the @NOCAF@
1956 tag allows the compiler to indicate that no CAFs can be reached
1957 anywhere \emph{even indirectly}.
1961 For each data constructor $Con$, two info tables are generated:
1964 \item $Con$@_con_info@ labels $Con$'s dynamic info table,
1965 shared by all dynamic instances of the constructor.
1966 \item $Con$@_static@ labels $Con$'s static info table,
1967 shared by all static instances of the constructor.
1970 Each constructor also has a \emph{constructor function}, which is a
1971 curried function which builds an instance of the constructor. The
1972 constructor function has an info table labelled as @$Con$_info@.
1974 Nullary constructors are represented by a single static info table,
1975 which everyone points to. Thus for a nullary constructor we can omit
1976 the dynamic info table and the constructor function.
1978 \subsubsection{Thunks}
1980 \label{sec:THUNK_SEL}
1982 A thunk represents an expression that is not obviously in head normal
1983 form. For example, consider the following top-level definitions:
1985 range = between 1 10
1986 f = \x -> let ys = take x range
1989 Here the right-hand sides of @range@ and @ys@ are both thunks; the former
1990 is static while the latter is dynamic.
1992 The layout of a thunk is the same as that for a function closure.
1993 However, thunks must have a payload of at least @MIN_UPD_PAYLOAD@
1994 words to allow it to be overwritten with a black hole and an
1995 indirection. The compiler may have to add extra non-pointer fields to
1996 satisfy this constraint.
1999 \begin{tabular}{|l|l|l|l|l|}\hline
2000 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \\ \hline
2004 The layout word in the info table contains the same information as for
2005 function closures; that is, number of pointers and number of
2008 A thunk differs from a function closure in that it can be updated.
2010 There are several forms of thunk:
2014 \item @THUNK@ and $@THUNK_@p@_@np$: vanilla, dynamically allocated
2015 thunks. Dynamic thunks are overwritten with normal indirections.
2017 \item @THUNK_STATIC@. A static thunk is also known as a
2018 \emph{constant applicative form}, or \emph{CAF}. Static thunks are
2019 overwritten with static indirections.
2022 \begin{tabular}{|l|l|l|l|l|}\hline
2023 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \emph{Static object link}\\ \hline
2027 \item @THUNK_SELECTOR@ is a (dynamically allocated) thunk whose entry
2028 code performs a simple selection operation from a data constructor
2029 drawn from a single-constructor type. For example, the thunk
2031 x = case y of (a,b) -> a
2033 is a selector thunk. A selector thunk is laid out like this:
2036 \begin{tabular}{|l|l|l|l|}\hline
2037 \emph{Fixed header} & \emph{Selectee pointer} \\ \hline
2041 The layout word contains the byte offset of the desired word in the
2042 selectee. Note that this is different from all other thunks.
2044 The garbage collector ``peeks'' at the selectee's tag (in its info
2045 table). If it is evaluated, then it goes ahead and does the
2046 selection, and then behaves just as if the selector thunk was an
2047 indirection to the selected field. If it is not evaluated, it treats
2048 the selector thunk like any other thunk of that shape.
2049 [Implementation notes. Copying: only the evacuate routine needs to be
2050 special. Compacting: only the PRStart (marking) routine needs to be
2053 There is a fixed set of pre-compiled selector thunks built into the
2054 RTS, representing offsets from 0 to @MAX_SPEC_SELECTOR_THUNK@. The
2055 info tables are labelled @sel_info_$n$@ where $n$ is the offset.
2059 The only label associated with a thunk is its info table:
2061 \item[$f$@_info@] is $f$'s info table.
2065 \Subsubsection{Byte-code objects}{BCO}
2067 A Byte-Code Object (BCO) is a container for a a chunk of byte-code,
2068 which can be executed by Hugs. The byte-code represents a
2069 supercombinator in the program: when Hugs compiles a module, it
2070 performs lambda lifting and each resulting supercombinator becomes a
2071 byte-code object in the heap.
2073 BCOs are not updateable; the bytecode compiler represents updatable
2074 thunks using a combination of @AP@s and @BCO@s.
2076 The semantics of BCOs are described in \secref{hugs-heap-objects}. A
2077 BCO has the following structure:
2080 \begin{tabular}{|l|l|l|l|l|l|}
2082 \emph{Fixed Header} & \emph{Layout} & \emph{Offset} & \emph{Size} &
2083 \emph{Literals} & \emph{Byte code} \\
2090 \item The entry code is a static code fragment/info table that returns
2091 to the scheduler to invoke Hugs (\secref{ghc-to-hugs-switch}).
2092 \item \emph{Layout} contains the number of pointer literals in the
2093 \emph{Literals} field.
2094 \item \emph{Offset} is the offset to the byte code from the start of
2096 \item \emph{Size} is the number of words of byte code in the object.
2097 \item \emph{Literals} contains any pointer and non-pointer literals used in
2098 the byte-codes (including jump addresses), pointers first.
2099 \item \emph{Byte code} contains \emph{Size} words of non-pointer byte
2104 \Subsubsection{Partial applications}{PAP}
2106 \ToDo{PAPs don't contains update frames or activation frames. When we
2107 add revertible black holes, we'll introduce a new kind of object which
2108 can contain activation frames.}
2110 A partial application (PAP) represents a function applied to too few arguments.
2111 It is only built as a result of updating after an argument-satisfaction
2112 check failure. A PAP has the following shape:
2114 \begin{tabular}{|l|l|l|l|}\hline
2115 \emph{Fixed header} & \emph{No of arg words} & \emph{Function closure} & \emph{Arg stack} \\ \hline
2118 The ``arg stack'' is a copy of the chunk of stack above the update
2119 frame; ``no of arg words'' tells how many words it consists of. The
2120 function closure is (a pointer to) the closure for the function whose
2121 argument-satisfaction check failed.
2123 There is just one standard form of PAP. There is just one info table
2124 too, called @PAP_info@. Its entry code simply copies the arg stack
2125 chunk back on top of the stack and enters the function closure. (It
2126 has to do a stack overflow test first.)
2128 PAPs are also used to implement Hugs functions (where the arguments
2129 are free variables). PAPs generated by Hugs can be static so we need
2130 both @PAP@ and @PAP_STATIC@.
2132 \Subsubsection{@AP@ objects}{AP}
2134 @AP@ objects are used to represent thunks built by Hugs. The only
2135 distintion between an @AP@ and a @PAP@ is that an @AP@ is updateable.
2138 \begin{tabular}{|l|l|l|l|}
2140 \emph{Fixed Header} & \emph{No of arg words} & \emph{Function closure} & \emph{Arg stack} \\
2145 The entry code pushes an update frame, copies the arg stack chunk on
2146 top of the stack, and enters the function closure. (It has to do a
2147 stack overflow test first.)
2149 The ``arg stack'' is a block of (tagged) arguments representing the
2150 free variables of the thunk; ``no of arg words'' tells how many words
2151 it consists of. The function closure is (a pointer to) the closure
2152 for the thunk. The argument stack may be empty if the thunk has no
2155 \note{Since @AP@s are updateable, the @MIN_UPD_PAYLOAD@ constraint
2158 \Subsubsection{Indirections}{IND}
2160 Indirection closures just point to other closures. They are introduced
2161 when a thunk is updated to point to its value. The entry code for all
2162 indirections simply enters the closure it points to.
2164 There are several forms of indirection:
2166 \item[@IND@] is the vanilla, dynamically-allocated indirection.
2167 It is removed by the garbage collector. It has the following
2170 \begin{tabular}{|l|l|l|}\hline
2171 \emph{Fixed header} & \emph{Mutable link field} & \emph{Target closure} \\ \hline
2174 It contains a \emph{mutable link field} that is used to string together
2175 indirections in each generation.
2178 \item[@IND_PERMANENT@]
2179 for lexical profiling, it is necessary to maintain cost centre
2180 information in an indirection, so ``permanent indirections'' are
2181 retained forever. Otherwise they are just like vanilla indirections.
2182 \note{If a permanent indirection points to another permanent
2183 indirection or a @CONST@ closure, it is possible to elide the indirection
2184 since it will have no effect on the profiler.}
2186 \note{Do we still need @IND@ in the profiling build, or do we just
2187 need @IND@ but its behaviour changes when profiling is on?}
2189 \item[@IND_STATIC@] is used for overwriting CAFs when they have been
2190 evaluated. Static indirections are not removed by the garbage
2191 collector; and are statically allocated outside the heap (and should
2192 stay there). Their static object link field is used just as for
2193 @FUN_STATIC@ closures.
2196 \begin{tabular}{|l|l|l|}
2198 \emph{Fixed header} & \emph{Target closure} & \emph{Static object link} \\
2205 \Subsubsection{Black holes and blocking queues}{BH}
2207 Black hole closures are used to overwrite closures currently being
2208 evaluated. They inform the garbage collector that there are no live
2209 roots in the closure, thus removing a potential space leak.
2211 Black holes also become synchronization points in the threaded world.
2212 They contain a pointer to a list of blocked threads to be awakened
2213 when the black hole is updated (or @NULL@ if the list is empty).
2216 \begin{tabular}{|l|l|l|}
2218 \emph{Fixed header} & \emph{Mutable link} & \emph{Blocked thread link} \\
2223 The \emph{Blocked thread link} points to the TSO of the first thread
2224 waiting for the value of this thunk. All subsequent TSOs in the list
2225 are linked together using their @TSO_LINK@ field.
2227 When the blocking queue is non-@NULL@ and the @BH@ is in the old
2228 generation, the black hole must be added to the mutables list since
2229 the TSOs on the list may contain pointers into the new generation.
2230 There is no need to clutter up the mutables list with black holes with
2231 empty blocking queues.
2233 \note{In a single-threaded system, entering a black hole indicates an
2234 infinite loop. In a concurrent system, entering a black hole
2235 indicates an infinite loop only if the hole is being entered by the
2236 same thread that originally entered the closure.}
2239 \Subsubsection{FetchMes}{FETCHME}
2241 In the parallel systems, FetchMes are used to represent pointers into
2242 the global heap. When evaluated, the value they point to is read from
2245 \ToDo{Describe layout}
2247 Because there may be offsets into these arrays, a primitive array
2248 cannot be handled as a FetchMe in the parallel system, but must be
2249 shipped in its entirety if its parent closure is shipped.
2253 \Subsection{Unpointed Objects}{unpointed-objects}
2255 A variable of unpointed type is always bound to a \emph{value}, never
2256 to a \emph{thunk}. For this reason, unpointed objects cannot be
2259 \subsubsection{Immutable objects}
2260 \label{sec:ARR_WORDS1}
2261 \label{sec:ARR_PTRS}
2264 \item[@ARR_WORDS@] is a variable-sized object consisting solely of
2265 non-pointers. It is used for arrays of all sorts of things (bytes,
2266 words, floats, doubles... it doesn't matter).
2268 \begin{tabular}{|c|c|c|c|}
2270 \emph{Fixed Hdr} & \emph{No of non-pointers} & \emph{Non-pointers\ldots} \\ \hline
2274 \item[@ARR_PTRS@] is an immutable, variable sized array of pointers.
2276 \begin{tabular}{|c|c|c|c|}
2278 \emph{Fixed Hdr} & \emph{Mutable link} & \emph{No of pointers} & \emph{Pointers\ldots} \\ \hline
2281 The mutable link is present so that we can easily freeze and thaw an
2282 array (by changing the header and adding/removing the array to the
2287 \subsubsection{Mutable objects}
2288 \label{sec:mutables}
2289 \label{sec:ARR_WORDS2}
2291 \label{sec:MUTARR_PTRS}
2292 \label{sec:MUTARR_PTRS_FROZEN}
2294 Some of these objects are \emph{mutable}; they represent objects which
2295 are explicitly mutated by Haskell code through the @ST@ monad.
2296 They're not used for thunks which are updated precisely once.
2297 Depending on the garbage collector, mutable closures may contain extra
2298 header information which allows a generational collector to implement
2299 the ``write barrier.''
2303 \item[@ARR_WORDS@] is also used to represent \emph{mutable} arrays of
2304 bytes, words, floats, doubles, etc. It's possible to use the same
2305 object type because even generational collectors don't need to
2308 \item[@MUTVAR@] is a mutable variable.
2310 \begin{tabular}{|c|c|c|}
2312 \emph{Fixed Hdr} & \emph{Mutable link} & \emph{Pointer} \\ \hline
2316 \item[@MUTARR_PTRS@] is a mutable array of pointers.
2317 Such an array may be \emph{frozen}, becoming an @ARR_PTRS@, with a
2318 different info-table.
2320 \begin{tabular}{|c|c|c|c|}
2322 \emph{Fixed Hdr} & \emph{Mutable link} & \emph{No of ptrs} & \emph{Pointers\ldots} \\ \hline
2329 \Subsubsection{Foreign objects}{FOREIGN}
2331 Here's what a ForeignObj looks like:
2334 \begin{tabular}{|l|l|l|l|}
2336 \emph{Fixed header} & \emph{Data} & \emph{Free Routine} & \emph{Foreign object link} \\
2341 The @FreeRoutine@ is a reference to the finalisation routine to call
2342 when the @ForeignObj@ becomes garbage. If @freeForeignObject@ is
2343 called on a Foreign Object, the @FreeRoutine@ is set to zero and the
2344 garbage collector will not attempt to call @FreeRoutine@ when the
2345 object becomes garbage.
2347 The Foreign object link is a link to the next foreign object in the
2348 list. This list is traversed at the end of garbage collection: if an
2349 object is about to be deallocated (e.g.~it was not marked or
2350 evacuated), the free routine is called and the object is deleted from
2353 \subsubsection{MVars and IVars}
2357 \ToDo{MVars and IVars}
2361 The remaining objects types are all administrative --- none of them
2364 \subsection{Other weird objects}
2366 \label{sec:BLOCKED_FETCH}
2369 \item[@BlockedFetch@ heap objects (`closures')] (parallel only)
2371 @BlockedFetch@s are inbound fetch messages blocked on local closures.
2372 They arise as entries in a local blocking queue when a fetch has been
2373 received for a local black hole. When awakened, we look at their
2374 contents to figure out where to send a resume.
2376 A @BlockedFetch@ closure has the form:
2378 \begin{tabular}{|l|l|l|l|l|l|}\hline
2379 \emph{Fixed header} & link & node & gtid & slot & weight \\ \hline
2383 \item[Spark Closures] (parallel only)
2385 Spark closures are used to link together all closures in the spark pool. When
2386 the current processor is idle, it may choose to speculatively evaluate some of
2387 the closures in the pool. It may also choose to delete sparks from the pool.
2389 \begin{tabular}{|l|l|l|l|l|l|}\hline
2390 \emph{Fixed header} & \emph{Spark pool link} & \emph{Sparked closure} \\ \hline
2394 \item[Slop Objects]\label{sec:slop-objects}
2396 Slop objects are used to overwrite the end of an updatee if it is
2397 larger than an indirection. Normal slop objects consist of an info
2398 pointer a size word and a number of slop words.
2401 \begin{tabular}{|l|l|l|l|l|l|}\hline
2402 \emph{Info Pointer} & \emph{Size} & \emph{Slop Words} \\ \hline
2406 This is too large for single word slop objects which consist of a
2409 Note that slop objects only contain an info pointer, not a standard
2410 fixed header. This doesn't cause problems because slop objects are
2411 always unreachable --- they can only be accessed by linearly scanning
2416 \Subsection{Thread State Objects (TSOs)}{TSO}
2418 \ToDo{This is very out of date. We now embed a single stack object
2419 within the TSO. TSOs include an ID number which can be used to
2420 generate a hash value. The gransim, profiling and ticky info is
2423 In the multi-threaded system, the state of a suspended thread is
2424 packed up into a Thread State Object (TSO) which contains all the
2425 information needed to restart the thread and for the garbage collector
2426 to find all reachable objects. When a thread is running, it may be
2427 ``unpacked'' into machine registers and various other memory locations
2428 to provide faster access.
2430 Single-threaded systems don't really \emph{need\/} TSOs --- but they do
2431 need some way to tell the storage manager about live roots so it is
2432 convenient to use a single TSO to store the mutator state even in
2433 single-threaded systems.
2435 Rather than manage TSOs' alloc/dealloc, etc., in some \emph{ad hoc}
2436 way, we instead alloc/dealloc/etc them in the heap; then we can use
2437 all the standard garbage-collection/fetching/flushing/etc machinery on
2438 them. So that's why TSOs are ``heap objects,'' albeit very special
2441 \begin{tabular}{|l|l|}
2442 \hline \emph{Fixed header}
2443 \\ \hline @TSO_LINK@
2444 \\ \hline @TSO_STATE@
2445 \\ \hline \emph{Exception Handlers}
2446 \\ \hline \emph{Ticky Info}
2447 \\ \hline \emph{Profiling Info}
2448 \\ \hline \emph{Parallel Info}
2449 \\ \hline \emph{GranSim Info}
2457 The contents of a TSO are:
2460 \item A pointer (@TSO_LINK@) used to maintain a list of threads with a similar
2461 state (e.g.~all runnable, all sleeping, all blocked on the same black
2462 hole, all blocked on the same MVar, etc.)
2464 \item A word (@TSO_STATE@) which records the current state of a thread: running, runnable, blocked, etc.
2466 \item Optional information for ``Ticky Ticky'' statistics: @TSO_STK_HWM@ is
2467 the maximum number of words allocated to this thread.
2469 \item Optional information for profiling:
2470 @TSO_CCC@ is the current cost centre.
2472 \item Optional information for parallel execution:
2476 % \item The types of threads (@TSO_TYPE@):
2477 % \begin{description}
2478 % \item[@T_MAIN@] Must be executed locally.
2479 % \item[@T_REQUIRED@] A required thread -- may be exported.
2480 % \item[@T_ADVISORY@] An advisory thread -- may be exported.
2481 % \item[@T_FAIL@] A failure thread -- may be exported.
2484 % \item I've no idea what else
2488 % \item Optional information for GranSim execution:
2494 % \item basic blocks
2501 % \item global sparks
2502 % \item local sparks
2505 % \item clock (gransim light only)
2509 % Here are the various queues for GrAnSim-type events.
2520 \subsection{Stack Objects}
2521 \label{sec:STACK_OBJECT}
2524 \ToDo{Merge this in with the section on TSOs}
2526 These are ``stack objects,'' which are used in the threaded world as
2527 the stack for each thread is allocated from the heap in smallish
2528 chunks. (The stack in the sequential world is allocated outside of
2531 Each reduction thread has to have its own stack space. As there may
2532 be many such threads, and as any given one may need quite a big stack,
2533 a naive give-'em-a-big-stack-and-let-'em-run approach will cost a {\em
2536 Our approach is to give a thread a small stack space, and then link
2537 on/off extra ``chunks'' as the need arises. Again, this is a
2538 storage-management problem, and, yet again, we choose to graft the
2539 whole business onto the existing heap-management machinery. So stack
2540 objects will live in the heap, be garbage collected, etc., etc..
2542 A stack object is laid out like this:
2545 \begin{tabular}{|l|}
2549 \emph{Link to next stack object (0 for last)}
2551 \emph{N, the payload size in words}
2553 \emph{@Sp@ (byte offset from head of object)}
2555 \emph{@Su@ (byte offset from head of object)}
2557 \emph{Payload (N words)}
2562 The stack grows downwards, towards decreasing
2563 addresses. This makes it easier to print out the stack
2564 when debugging, and it means that a return address is
2565 at the lowest address of the chunk of stack it ``knows about''
2566 just like an info pointer on a closure.
2568 The garbage collector needs to be able to find all the
2569 pointers in a stack. How does it do this?
2573 \item Within the stack there are return addresses, pushed
2574 by @case@ expressions. Below a return address (i.e. at higher
2575 memory addresses, since the stack grows downwards) is a chunk
2576 of stack that the return address ``knows about'', namely the
2577 activation record of the currently running function.
2579 \item Below each such activation record is a \emph{pending-argument
2580 section}, a chunk of
2581 zero or more words that are the arguments to which the result
2582 of the function should be applied. The return address does not
2584 ``know'' how many pending arguments there are, or their types.
2585 (For example, the function might return a result of type $\alpha$.)
2587 \item Below each pending-argument section is another return address,
2588 and so on. Actually, there might be an update frame instead, but we
2589 can consider update frames as a special case of a return address with
2590 a well-defined activation record.
2594 The game plan is this. The garbage collector
2595 walks the stack from the top, traversing pending-argument sections and
2596 activation records alternately. Next we discuss how it finds
2597 the pointers in each of these two stack regions.
2600 \Subsubsection{Activation records}{activation-records}
2603 An \emph{activation record} is a contiguous chunk of stack,
2604 with a return address as its first word, followed by as many
2605 data words as the return address ``knows about''. The return
2606 address is actually a fully-fledged info pointer. It points
2607 to an info table, replete with:
2610 \item entry code (i.e. the code to return to).
2612 \item closure type is either @RET_SMALL/RET_VEC_SMALL@ or
2613 @RET_BIG/RET_VEC_BIG@, depending on whether the activation record has
2614 more than 32 data words (\note{64 for 8-byte-word architectures}) and
2615 on whether to use a direct or a vectored return.
2617 \item the layout info for @RET_SMALL@ is a bitmap telling the layout
2618 of the activation record, one bit per word. The least-significant bit
2619 describes the first data word of the record (adjacent to the fixed
2620 header) and so on. A ``@1@'' indicates a non-pointer, a ``@0@''
2621 indicates a pointer. We don't need to indicate exactly how many words
2622 there are, because when we get to all zeros we can treat the rest of
2623 the activation record as part of the next pending-argument region.
2625 For @RET_BIG@ the layout field points to a block of bitmap words,
2626 starting with a word that tells how many words are in the block.
2628 \item the info table contains a Static Reference Table pointer for the
2629 return address (\secref{srt}).
2632 The activation record is a fully fledged closure too. As well as an
2633 info pointer, it has all the other attributes of a fixed header
2634 (\secref{fixed-header}) including a saved cost centre which
2635 is reloaded when the return address is entered.
2637 In other words, all the attributes of closures are needed for
2638 activation records, so it's very convenient to make them look alike.
2641 \Subsubsection{Pending arguments}{pending-args}
2643 So that the garbage collector can correctly identify pointers in
2644 pending-argument sections we explicitly tag all non-pointers. Every
2645 non-pointer in a pending-argument section is preceded (at the next
2646 lower memory word) by a one-word byte count that says how many bytes
2647 to skip over (excluding the tag word).
2649 The garbage collector traverses a pending argument section from the
2650 top (i.e. lowest memory address). It looks at each word in turn:
2653 \item If it is less than or equal to a small constant @MAX_STACK_TAG@
2654 then it treats it as a tag heralding zero or more words of
2655 non-pointers, so it just skips over them.
2657 \item If it points to the code segment, it must be a return
2658 address, so we have come to the end of the pending-argument section.
2660 \item Otherwise it must be a bona fide heap pointer.
2664 \Subsection{The Stable Pointer Table}{STABLEPTR_TABLE}
2666 A stable pointer is a name for a Haskell object which can be passed to
2667 the external world. It is ``stable'' in the sense that the name does
2668 not change when the Haskell garbage collector runs---in contrast to
2669 the address of the object which may well change.
2671 A stable pointer is represented by an index into the
2672 @StablePointerTable@. The Haskell garbage collector treats the
2673 @StablePointerTable@ as a source of roots for GC.
2675 In order to provide efficient access to stable pointers and to be able
2676 to cope with any number of stable pointers (eg $0 \ldots 100000$), the
2677 table of stable pointers is an array stored on the heap and can grow
2678 when it overflows. (Since we cannot compact the table by moving
2679 stable pointers about, it seems unlikely that a half-empty table can
2680 be reduced in size---this could be fixed if necessary by using a
2681 hash table of some sort.)
2683 In general a stable pointer table closure looks like this:
2686 \begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|}
2688 \emph{Fixed header} & \emph{No of pointers} & \emph{Free} & $SP_0$ & \ldots & $SP_{n-1}$
2696 \item[@NPtrs@:] number of (stable) pointers.
2698 \item[@Free@:] the byte offset (from the first byte of the object) of the first free stable pointer.
2700 \item[$SP_i$:] A stable pointer slot. If this entry is in use, it is
2701 an ``unstable'' pointer to a closure. If this entry is not in use, it
2702 is a byte offset of the next free stable pointer slot.
2706 When a stable pointer table is evacuated
2708 \item the free list entries are all set to @NULL@ so that the evacuation
2709 code knows they're not pointers;
2711 \item The stable pointer slots are scanned linearly: non-@NULL@ slots
2712 are evacuated and @NULL@-values are chained together to form a new free list.
2715 There's no need to link the stable pointer table onto the mutable
2716 list because we always treat it as a root.
2718 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2719 \Subsection{Garbage Collecting CAFs}{CAF}
2720 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2722 % begin{direct quote from current paper}
2723 A CAF (constant applicative form) is a top-level expression with no
2724 arguments. The expression may need a large, even unbounded, amount of
2725 storage when it is fully evaluated.
2727 CAFs are represented by closures in static memory that are updated
2728 with indirections to objects in the heap space once the expression is
2729 evaluated. Previous version of GHC maintained a list of all evaluated
2730 CAFs and traversed them during GC, the result being that the storage
2731 allocated by a CAF would reside in the heap until the program ended.
2732 % end{direct quote from current paper}
2734 % begin{elaboration on why CAFs are very very bad}
2735 Treating CAFs this way has two problems:
2738 It can cause a very large space leak. For example, this program
2739 should run in constant space but, instead, will run out of memory.
2745 > nats = [0..maxInt]
2749 Expressions with no arguments have very different space behaviour
2750 depending on whether or not they occur at the top level. For example,
2751 if we make \verb+nats+ a local definition, the space leak goes away
2752 and the resulting program runs in constant space, as expected.
2758 > nats = [0..maxInt]
2761 This huge change in the operational behaviour of the program
2762 is a problem for optimising compilers and for programmers.
2763 For example, GHC will normally flatten a set of let bindings using
2764 this transformation:
2766 let x1 = let x2 = e2 in e1 ==> let x2 = e2 in let x1 = e1
2768 but it does not do so if this would raise \verb+x2+ to the top level
2769 since that may create a CAF. Many Haskell programmers avoid creating
2770 large CAFs by adding a dummy argument to a CAF or by moving a CAF away
2774 % end{elaboration on why CAFs are very very bad}
2776 Solving the CAF problem requires different treatment in interactive
2777 systems such as Hugs than in batch-mode systems such as GHC
2780 In a batch-mode the program the runtime system is terminated
2781 after every execution of the runtime system. In such systems,
2782 the garbage collector can completely ``destroy'' a CAF when it
2783 is no longer live --- in much the same way as it ``destroys''
2784 normal closures when they are no longer live.
2787 In an interactive system, many expressions are evaluated without
2788 restarting the runtime system between each evaluation. In such
2789 systems, the garbage collector cannot completely ``destroy'' a CAF
2790 when it is no longer live because, whilst it might not be required in
2791 the evaluation of the current expression, it might be required in the
2794 There are two possible behaviours we migth want:
2797 When a CAF is no longer required for the current evaluation, the CAF
2798 should be reverted to its original form. This behaviour ensures that
2799 the operational behaviour of the interactive system is a reasonable
2800 predictor of the operational behaviour of the batch-mode system. This
2801 allows us to use Hugs for performance debugging (in particular, trying
2802 to understand and reduce the heap usage of a program) --- an area of
2803 increasing importance as Haskell is used more and more to solve ``real
2804 problems'' in ``real problem domains''.
2807 Even if a CAF is no longer required for the current evaluation, we might
2808 choose to hang onto it by collecting it in the normal way. This keeps
2809 the space leak but might be useful in a teaching environment when
2810 trying to teach the difference between call by name evaluation (which
2811 doesn't share work) and lazy evaluation (which does share work).
2815 It turns out that it is easy to support both styles of use, so the
2816 runtime system provides a switch which lets us turn this on and off
2817 during execution. \ToDo{What is this switch called?} It would also
2818 be easy to provide a function \verb+RevertCAF+ to let the interpreter
2819 revert any CAF it wanted between (but not during) executions, if we so
2820 desired. Running \verb+RevertCAF+ during execution would lose some sharing
2821 but is otherwise harmless.
2825 % % begin{even more pointless observation?}
2826 % The simplest fix would be to remove the special treatment of
2827 % top level variables. This works but is very inefficient.
2829 % (Note: delete this paragraph from final version.)
2830 % % end{even more pointless observation?}
2832 % begin{pointless observation?}
2833 An easy but inefficient fix to the CAF problem would be to make a
2834 complete copy of the heap before every evaluation and discard the copy
2835 after evaluation. This works but is inefficient.
2836 % end{pointless observation?}
2838 An efficient way to achieve a similar effect is to revert all
2839 updatable thunks to their original form as they become unnecessary for
2840 the current evaluation. To do this, we modify the compiler to ensure
2841 that the only updatable thunks generated by the compiler are CAFs and
2842 we modify the garbage collector to revert entered CAFs to unentered
2843 CAFs as their value becomes unnecessary.
2846 \subsubsection{New Heap Objects}
2848 We add three new kinds of heap object: unentered CAF closures, entered
2849 CAF objects and CAF blackholes. We first describe how they are
2850 evaluated and then how they are garbage collected.
2853 Unentered CAF closures contain a pointer to closure representing the
2854 body of the CAF. The ``body closure'' is not updatable.
2856 Unentered CAF closures contain two unused fields to make them the same
2857 size as entered CAF closures --- which allows us to perform an inplace
2858 update. \ToDo{Do we have to add another kind of inplace update operation
2859 to the storage manager interface or do we consider this to be internal
2862 \begin{tabular}{|l|l|l|l|}\hline
2863 \verb+CAF_unentered+ & \emph{body closure} & \emph{unused} & \emph{unused} \\ \hline
2866 When an unentered CAF is entered, we do the following:
2869 allocate a CAF black hole;
2872 push an update frame (to update the CAF black hole) onto the stack;
2875 overwrite the CAF with an entered CAF object (see below) with the same
2876 body and whose value field points to the black hole;
2879 add the CAF to a list of all entered CAFs (called ``the CAF list'');
2883 the closure representing the value of the CAF is entered.
2887 When evaluation of the CAF body returns a value, the update frame
2888 causes the CAF black hole to be updated with the value in the normal
2891 \ToDo{Add a picture}
2894 Entered CAF closures contain two pointers: a pointer to the CAF body
2895 (the same as for unentered CAF closures); a pointer to the CAF value
2896 (this is initialised with a CAF blackhole, as previously described);
2897 and a link to the next CAF in the CAF list
2899 \ToDo{How is the end of the list marked? Null pointer or sentinel value?}.
2902 \begin{tabular}{|l|l|l|l|}\hline
2903 \verb+CAF_entered+ & \emph{body closure} & \emph{value} & \emph{link} \\ \hline
2906 When an entered CAF is entered, it enters its value closure.
2909 CAF blackholes are identical to normal blackholes except that they
2910 have a different infotable. The only reason for having CAF blackholes
2911 is to allow an optimisation of lazy blackholing where we stop scanning
2912 the stack when we see the first {\em normal blackhole} but not
2913 when we see a {\em CAF blackhole.}
2914 \ToDo{The optimisation we want to allow should be described elsewhere
2915 so that all we have to do here is describe the difference.}
2917 Instead of allocating a blackhole to update with the value of the CAF,
2918 it might seem simpler to update the CAF directly. This would require
2919 a new kind of update frame which would update the value field of the
2920 CAF with a pointer to the value and wouldn't catch blackholes caused
2921 by CAFs that depend on themselves so we chose not to do so.
2925 \subsubsection{Garbage Collection}
2927 To avoid the space leak, each run of the garbage collector must revert
2928 the entered CAFs which are not required to complete the current
2929 evaluation (that is all the closures reachable from the set of
2930 runnable threads and the stable pointer table).
2932 It does this by performing garbage collection in three phases:
2935 During the first phase, we ``mark'' all closures reachable from the
2938 How we ``mark'' closures depends on the garbage collector. For
2939 example, in a 2-space collector, closures are ``marked'' by copying
2940 them into ``to-space'', overwriting them with a forwarding node and
2941 ``marking'' all the closures reachable from the copy. The only
2942 requirements are that we can test whether a closure is marked and if a
2943 closure is marked then so are all closures reachable from it.
2945 \ToDo{At present we say that the scheduler state includes any state
2946 that Hugs may have. This is not true anymore.}
2948 Performing this phase first provides us with a cheap test for
2949 execution closures: at this stage in execution, the execution closures
2950 are precisely the marked closures.
2953 During the second phase, we revert all unmarked CAFs on the CAF list
2954 and remove them from the CAF list.
2956 Since the CAF list is exactly the set of all entered CAFs, this reverts
2957 all entered CAFs which are not execution closures.
2960 During the third phase, we mark all top level objects (including CAFs)
2961 by calling \verb+MarkHugsRoots+ which will call \verb+MarkRoot+ for
2962 each top level object known to Hugs.
2966 To implement the second style of interactive behaviour (where we
2967 deliberately keep the CAF-related space leak), we simply omit the
2968 second phase. Omitting the second phase causes the third phase to
2969 mark any unmarked CAF value closures.
2971 So far, we have been describing a pure Hugs system which contains no
2972 machine generated code. The main difference in a hybrid system is
2973 that GHC-generated code is statically allocated in memory instead of
2974 being dynamically allocated on the heap. We split both
2975 \verb+CAF_unentered+ and \verb+CAF_entered+ into two versions: a
2976 static and a dynamic version. The static and dynamic versions of each
2977 CAF differ only in whether they are moved during garbage collection.
2978 When reverting CAFs, we revert dynamic entered CAFs to dynamic
2979 unentered CAFs and static entered CAFs to static unentered CAFs.
2984 \Section{The Bytecode Evaluator}{bytecode-evaluator}
2986 This section describes how the Hugs interpreter interprets code in the
2987 same environment as compiled code executes. Both evaluation models
2988 use a common garbage collector, so they must agree on the form of
2989 objects in the heap.
2991 Hugs interprets code by converting it to byte-code and applying a
2992 byte-code interpreter to it. Wherever possible, we try to ensure that
2993 the byte-code is all that is required to interpret a section of code.
2994 This means not dynamically generating info tables, and hence we can
2995 only have a small number of possible heap objects each with a statically
2996 compiled info table. Similarly for stack objects: in fact we only
2997 have one Hugs stack object, in which all information is tagged for the
3000 There is, however, one exception to this rule. Hugs must generate
3001 info tables for any constructors it is asked to compile, since the
3002 alternative is to force a context-switch each time compiled code
3003 enters a Hugs-built constructor, which would be prohibitively
3006 We achieve this simplicity by forgoing some of the optimisations used
3011 Whereas compiled code has five different ways of entering a closure
3012 (\secref{entering-closures}), interpreted code has only one.
3013 The entry point for interpreted code behaves like slow entry points for
3018 We use just one info table for \emph{all\/} direct returns.
3019 This introduces two problems:
3021 \item How does the interpreter know what code to execute?
3023 Instead of pushing just a return address, we push a return BCO and a
3024 trivial return address which just enters the return BCO.
3026 (In a purely interpreted system, we could avoid pushing the trivial
3029 \item How can the garbage collector follow pointers within the
3032 We could push a third word ---a bitmask describing the location of any
3033 pointers within the record--- but, since we're already tagging unboxed
3034 function arguments on the stack, we use the same mechanism for unboxed
3035 values within the activation record.
3037 \ToDo{Do we have to stub out dead variables in the activation frame?}
3043 We trivially support vectored returns by pushing a return vector whose
3044 entries are all the same.
3048 We avoid the need to build SRTs by putting bytecode objects on the
3049 heap and restricting BCOs to a single basic block.
3053 \Subsection{Hugs Info Tables}{hugs-info-tables}
3055 Hugs requires the following info tables and closures:
3059 Contains both a vectored return table and a direct entry point. All
3060 entry points are the same: they rearrange the stack to match the Hugs
3061 return convention (\secref{hugs-return-convention}) and return to the
3062 scheduler. When the scheduler restarts the thread, it will find a BCO
3063 on top of the stack and will enter the Hugs interpreter.
3067 This is just the standard info table for an update frame.
3069 \item [Constructors].
3071 The entry code for a constructor jumps to a generic entry point in the
3072 runtime system which decides whether to do a vectored or unvectored
3073 return depending on the shape of the constructor/type. This implies that
3074 info tables must have enough info to make that decision.
3076 \item [@AP@ and @PAP@].
3078 \item [Indirections].
3082 Hugs doesn't generate them itself but it ought to recognise them
3084 \item [Complex primops].
3086 Some of the primops are too complex for GHC to generate inline.
3087 Instead, these primops are hand-written and called as normal functions.
3088 Hugs only needs to know their names and types but doesn't care whether
3089 they are generated by GHC or by hand. Two things to watch:
3093 Hugs must be able to enter these primops even if it is working on a
3094 standalone system that does not support genuine GHC generated code.
3096 \item The complex primops often involve unboxed tuple types (which
3097 Hugs does not support at the source level) so we cannot specify their
3098 types in a Haskell source file.
3104 \Subsection{Hugs Heap Objects}{hugs-heap-objects}
3106 \subsubsection{Byte-code objects}
3108 Compiled byte code lives on the global heap, in objects called
3109 Byte-Code Objects (or BCOs). The layout of BCOs is described in
3110 detail in \secref{BCO}, in this section we will describe
3113 Since byte-code lives on the heap, it can be garbage collected just
3114 like any other heap-resident data. Hugs arranges that any BCO's
3115 referred to by the Hugs symbol tables are treated as live objects by
3116 the garbage collector. When a module is unloaded, the pointers to its
3117 BCOs are removed from the symbol table, and the code will be garbage
3118 collected some time later.
3120 A BCO represents a basic block of code --- the (only) entry points is
3121 at the beginning of a BCO, and it is impossible to jump into the
3122 middle of one. A BCO represents not only the code for a function, but
3123 also its closure; a BCO can be entered just like any other closure.
3124 Hugs performs lambda-lifting during compilation to byte-code, and each
3125 top-level combinator becomes a BCO in the heap.
3128 \subsubsection{Thunks and partial applications}
3130 A thunk consists of a code pointer, and values for the free variables
3131 of that code. Since Hugs byte-code is lambda-lifted, free variables
3132 become arguments and are expected to be on the stack by the called
3135 Hugs represents updateable thunks with @AP@ objects applying a closure
3136 to a list of arguments. (As for @PAP@s, unboxed arguments should be
3137 preceded by a tag.) When it is entered, it pushes an update frame
3138 followed by its payload on the stack, and enters the first word (which
3139 will be a pointer to a BCO). The layout of @AP@ objects is described
3140 in more detail in \secref{AP}.
3142 Partial applications are represented by @PAP@ objects, which are
3145 \ToDo{Hugs Constructors}.
3147 \Subsection{Calling conventions}{hugs-calling-conventions}
3149 The calling convention for any byte-code function is straightforward:
3151 \item Push any arguments on the stack.
3152 \item Push a pointer to the BCO.
3153 \item Begin interpreting the byte code.
3156 In a system containing both GHC and Hugs, the bytecode interpreter
3157 only has to be able to enter BCOs: everything else can be handled by
3158 returning to the compiled world (as described in
3159 \secref{hugs-to-ghc-switch}) and entering the closure
3162 This would work but it would obviously be very inefficient if we
3163 entered a @AP@ by switching worlds, entering the @AP@, pushing the
3164 arguments and function onto the stack, and entering the function
3165 which, likely as not, will be a byte-code object which we will enter
3166 by \emph{returning} to the byte-code interpreter. To avoid such
3167 gratuitious world switching, we choose to recognise certain closure
3168 types as being ``standard'' --- and duplicate the entry code for the
3169 ``standard closures'' in the bytecode interpreter.
3171 A closure is said to be ``standard'' if its entry code is entirely
3172 determined by its info table. \emph{Standard Closures} have the
3173 desirable property that the byte-code interpreter can enter the
3174 closure by simply ``interpreting'' the info table instead of switching
3175 to the compiled world. The standard closures include:
3178 \item[Constructor] To enter a constructor, we simply return (see
3179 \secref{hugs-return-convention}).
3182 To enter an indirection, we simply enter the object it points to
3183 after possibly adjusting the current cost centre.
3187 To enter an @AP@, we push an update frame, push the
3188 arguments, push the function and enter the function.
3189 (Not forgetting a stack check at the start.)
3193 To enter a @PAP@, we push the arguments, push the function and enter
3194 the function. (Not forgetting a stack check at the start.)
3198 To enter a selector (\secref{THUNK_SEL}), we test whether the
3199 selectee is a value. If so, we simply select the appropriate
3200 component; if not, it's simplest to treat it as a GHC-built closure
3201 --- though we could interpret it if we wanted.
3205 The most obvious omissions from the above list are @BCO@s (which we
3206 dealt with above) and GHC-built closures (which are covered in
3207 \secref{hugs-to-ghc-switch}).
3210 \Subsection{Return convention}{hugs-return-convention}
3212 When Hugs pushes a return address, it pushes both a pointer to the BCO
3213 to return to, and a pointer to a static code fragment @HUGS_RET@ (this
3214 is described in \secref{ghc-to-hugs-switch}). The
3215 stack layout is shown in \figref{hugs-return-stack}.
3227 %\input{hugs_ret.pstex_t}
3229 \caption{Stack layout for a Hugs return address}
3230 \label{fig:hugs-return-stack}
3241 %\input{hugs_ret2.pstex_t}
3243 \caption{Stack layout on enterings a Hugs return address}
3244 \label{fig:hugs-return2}
3257 %\input{hugs_ret2.pstex_t}
3259 \caption{Stack layout on entering a Hugs return address with an unboxed value}
3260 \label{fig:hugs-return-int}
3273 %\input{hugs_ret3.pstex_t}
3275 \caption{Stack layout on enterings a GHC return address}
3276 \label{fig:hugs-return3}
3290 | restart |--> id_Int#_closure
3293 %\input{hugs_ret2.pstex_t}
3295 \caption{Stack layout on enterings a GHC return address with an unboxed value}
3296 \label{fig:hugs-return-int}
3299 When a Hugs byte-code sequence enters a closure, it examines the
3300 return address on top of the stack.
3304 \item If the return address is @HUGS_RET@, pop the @HUGS_RET@ and the
3305 bco for the continuation off the stack, push a pointer to the constructor onto
3306 the stack and enter the BCO with the current object pointer set to the BCO
3307 (\figref{hugs-return2}).
3309 \item If the top of the stack is not @HUGS_RET@, we need to do a world
3310 switch as described in \secref{hugs-to-ghc-switch}.
3314 \ToDo{This duplicates what we say about switching worlds
3315 (\secref{switching-worlds}) - kill one or t'other.}
3318 \ToDo{This was in the evaluation model part but it really belongs in
3319 this part which is about the internal details of each of the major
3322 \Subsection{Addressing Modes}{hugs-addressing-modes}
3324 To avoid potential alignment problems and simplify garbage collection,
3325 all literal constants are stored in two tables (one boxed, the other
3326 unboxed) within each BCO and are referred to by offsets into the tables.
3327 Slots in the constant tables are word aligned.
3329 \ToDo{How big can the offsets be? Is the offset specified in the
3330 address field or in the instruction?}
3332 Literals can have the following types: char, int, nat, float, double,
3333 and pointer to boxed object. There is no real difference between
3334 char, int, nat and float since they all occupy 32 bits --- but it
3335 costs almost nothing to distinguish them and may improve portability
3336 and simplify debugging.
3338 \Subsection{Compilation}{hugs-compilation}
3341 \def\is{\mbox{\it is}}
3342 \def\ts{\mbox{\it ts}}
3343 \def\as{\mbox{\it as}}
3344 \def\bs{\mbox{\it bs}}
3345 \def\cs{\mbox{\it cs}}
3346 \def\rs{\mbox{\it rs}}
3347 \def\us{\mbox{\it us}}
3348 \def\vs{\mbox{\it vs}}
3349 \def\ws{\mbox{\it ws}}
3350 \def\xs{\mbox{\it xs}}
3352 \def\e{\mbox{\it e}}
3353 \def\alts{\mbox{\it alts}}
3354 \def\fail{\mbox{\it fail}}
3355 \def\panic{\mbox{\it panic}}
3356 \def\ua{\mbox{\it ua}}
3357 \def\obj{\mbox{\it obj}}
3358 \def\bco{\mbox{\it bco}}
3359 \def\tag{\mbox{\it tag}}
3360 \def\entry{\mbox{\it entry}}
3361 \def\su{\mbox{\it su}}
3363 \def\Ind#1{{\mbox{\it Ind}\ {#1}}}
3364 \def\update#1{{\mbox{\it update}\ {#1}}}
3366 \def\next{$\Longrightarrow$}
3367 \def\append{\mathrel{+\mkern-6mu+}}
3368 \def\reverse{\mbox{\it reverse}}
3369 \def\size#1{{\vert {#1} \vert}}
3370 \def\arity#1{{\mbox{\it arity}{#1}}}
3372 \def\AP{\mbox{\it AP}}
3373 \def\PAP{\mbox{\it PAP}}
3374 \def\GHCRET{\mbox{\it GHCRET}}
3375 \def\GHCOBJ{\mbox{\it GHCOBJ}}
3377 To make sense of the instructions, we need a sense of how they will be
3378 used. Here is a small compiler for the STG language.
3381 > cg (f{a1, ... am}) = do
3382 > pushAtom am; ... pushAtom a1
3386 > cg (let {x1=rhs1; ... xm=rhsm} in e) = do
3387 > ALLOC x1 |rhs1|, ... ALLOC xm |rhsm|
3388 > build x1 rhs1, ... build xm rhsm
3390 > cg (case e of alts) = do
3391 > PUSHALTS (cgAlts alts)
3394 > cgAlts { alt1; ... altm } = cgAlt alt1 $ ... $ cgAlt altm pmFail
3396 > cgAlt (x@C{xs} -> e) fail = do
3398 > HEAPCHECK (heapUse e)
3402 > build x (C{a1, ... am}) = do
3403 > pushUntaggedAtom am; ... pushUntaggedAtom a1
3405 > -- A useful optimisation
3406 > build x ({v1, ... vm} \ {}. f{a1, ... am}) = do
3407 > pushVar am; ... pushVar a1
3410 > build x ({v1, ... vm} \ {}. e) = do
3411 > pushVar vm; ... pushVar v1
3412 > PUSHBCO (cgRhs ({v1, ... vm} \ {}. e))
3414 > build x ({v1, ... vm} \ {x1, ... xm}. e) = do
3415 > pushVar vm; ... pushVar v1
3416 > PUSHBCO (cgRhs ({v1, ... vm} \ {x1, ... xm}. e))
3419 > cgRhs (vs \ xs. e) = do
3420 > ARGCHECK (xs ++ vs) -- can be omitted if xs == {}
3421 > STACKCHECK min(stackUse e,heapOverflowSlop)
3422 > HEAPCHECK (heapUse e)
3425 > pushAtom x = pushVar x
3426 > pushAtom i# = PUSHINT i#
3428 > pushVar x = if isGlobalVar x then PUSHGLOBAL x else PUSHLOCAL x
3430 > pushUntaggedAtom x = pushVar x
3431 > pushUntaggedAtom i# = PUSHUNTAGGEDINT i#
3433 > pushVar x = if isGlobalVar x then PUSHGLOBAL x else PUSHLOCAL x
3436 \ToDo{Is there an easy way to add semi-tagging? Would it be that different?}
3438 \ToDo{Optimise thunks of the form @f{x1,...xm}@ so that we build an AP directly}
3440 \Subsection{Instructions}{hugs-instructions}
3442 We specify the semantics of instructions using transition rules of
3445 \begin{tabular}{|llrrrrr|}
3447 & $\is$ & $s$ & $\su$ & $h$ & $hp$ & $\sigma$ \\
3448 \next & $\is'$ & $s'$ & $\su'$ & $h'$ & $hp'$ & $\sigma$ \\
3452 where $\is$ is an instruction stream, $s$ is the stack, $\su$ is the
3453 update frame pointer and $h$ is the heap.
3456 \Subsection{Stack manipulation}{hugs-stack-manipulation}
3460 \item[ Push a global variable ].
3462 \begin{tabular}{|llrrrrr|}
3464 & PUSHGLOBAL $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3465 \next & $\is$ & $\sigma!o:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3469 \item[ Push a local variable ].
3471 \begin{tabular}{|llrrrrr|}
3473 & PUSHLOCAL $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3474 \next & $\is$ & $s!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3478 \item[ Push an unboxed int ].
3480 \begin{tabular}{|llrrrrr|}
3482 & PUSHINT $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3483 \next & $\is$ & $I\# : \sigma!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3487 The $I\#$ is a tag included for the benefit of the garbage collector.
3488 Similar rules exist for floats, doubles, chars, etc.
3490 \item[ Push an unboxed int ].
3492 \begin{tabular}{|llrrrrr|}
3494 & PUSHUNTAGGEDINT $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3495 \next & $\is$ & $\sigma!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3499 Similar rules exist for floats, doubles, chars, etc.
3501 \item[ Delete environment from stack --- ready for tail call ].
3503 \begin{tabular}{|llrrrrr|}
3505 & SLIDE $m$ $n$ : $\is$ & $\as \append \bs \append \cs$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3506 \next & $\is$ & $\as \append \cs$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3510 where $\size{\as} = m$ and $\size{\bs} = n$.
3513 \item[ Push a return address ].
3515 \begin{tabular}{|llrrrrr|}
3517 & PUSHALTS $o$:$\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3518 \next & $\is$ & $@HUGS_RET@:\sigma!o:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3522 \item[ Push a BCO ].
3524 \begin{tabular}{|llrrrrr|}
3526 & PUSHBCO $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3527 \next & $\is$ & $\sigma!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3533 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3534 \Subsection{Heap manipulation}{hugs-heap-manipulation}
3535 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3539 \item[ Allocate a heap object ].
3541 \begin{tabular}{|llrrrrr|}
3543 & ALLOC $m$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3544 \next & $\is$ & $hp:s$ & $su$ & $h$ & $hp+m$ & $\sigma$ \\
3548 \item[ Build a constructor ].
3550 \begin{tabular}{|llrrrrr|}
3552 & PACK $o$ $o'$ : $\is$ & $\ws \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3553 \next & $\is$ & $s$ & $su$ & $h[s!o \mapsto Pack C\{\ws\}]$ & $hp$ & $\sigma$ \\
3557 where $C = \sigma!o'$ and $\size{\ws} = \arity{C}$.
3559 \item[ Build an AP or PAP ].
3561 \begin{tabular}{|llrrrrr|}
3563 & MKAP $o$ $m$:$\is$ & $f : \ws \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3564 \next & $\is$ & $s$ & $su$ & $h[s!o \mapsto \AP(f,\ws)]$ & $hp$ & $\sigma$ \\
3568 where $\size{\ws} = m$.
3570 \begin{tabular}{|llrrrrr|}
3572 & MKPAP $o$ $m$:$\is$ & $f : \ws \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3573 \next & $\is$ & $s$ & $su$ & $h[s!o \mapsto \PAP(f,\ws)]$ & $hp$ & $\sigma$ \\
3577 where $\size{\ws} = m$.
3579 \item[ Unpacking a constructor ].
3581 \begin{tabular}{|llrrrrr|}
3583 & UNPACK : $is$ & $a : s$ & $su$ & $h[a \mapsto C\ \ws]$ & $hp$ & $\sigma$ \\
3584 \next & $is'$ & $(\reverse\ \ws) \append a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3588 The $\reverse\ \ws$ looks expensive but, since the stack grows down
3589 and the heap grows up, that's actually the cheap way of copying from
3590 heap to stack. Looking at the compilation rules, you'll see that we
3591 always push the args in reverse order.
3596 \Subsection{Entering a closure}{hugs-entering}
3600 \item[ Enter a BCO ].
3602 \begin{tabular}{|llrrrrr|}
3604 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto BCO\{\is\} ]$ & $hp$ & $\sigma$ \\
3605 \next & $\is$ & $a : s$ & $su$ & $h$ & $hp$ & $a$ \\
3609 \item[ Enter a PAP closure ].
3611 \begin{tabular}{|llrrrrr|}
3613 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \PAP(f,\ws)]$ & $hp$ & $\sigma$ \\
3614 \next & [ENTER] & $f : \ws \append s$ & $su$ & $h$ & $hp$ & $???$ \\
3618 \item[ Entering an AP closure ].
3620 \begin{tabular}{|llrrrrr|}
3622 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \AP(f,ws)]$ & $hp$ & $\sigma$ \\
3623 \next & [ENTER] & $f : \ws \append @UPD_RET@:\su:a:s$ & $su'$ & $h$ & $hp$ & $???$ \\
3629 \item Instead of blindly pushing an update frame for $a$, we can first test whether there's already
3630 an update frame there. If so, overwrite the existing updatee with an indirection to $a$ and
3631 overwrite the updatee field with $a$. (Overwriting $a$ with an indirection to the updatee also
3632 works.) This results in update chains of maximum length 2.
3636 \item[ Returning a constructor ].
3638 \begin{tabular}{|llrrrrr|}
3640 & [ENTER] & $a : @HUGS_RET@ : \alts : s$ & $su$ & $h[a \mapsto C\{\ws\}]$ & $hp$ & $\sigma$ \\
3641 \next & $\alts.\entry$ & $a:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3646 \item[ Entering an indirection node ].
3648 \begin{tabular}{|llrrrrr|}
3650 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \Ind{a'}]$ & $hp$ & $\sigma$ \\
3651 \next & [ENTER] & $a' : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3655 \item[Entering GHC closure].
3657 \begin{tabular}{|llrrrrr|}
3659 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \GHCOBJ]$ & $hp$ & $\sigma$ \\
3660 \next & [ENTERGHC] & $a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3664 \item[Returning a constructor to GHC].
3666 \begin{tabular}{|llrrrrr|}
3668 & [ENTER] & $a : \GHCRET : s$ & $su$ & $h[a \mapsto C \ws]$ & $hp$ & $\sigma$ \\
3669 \next & [ENTERGHC] & $a : \GHCRET : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3676 \Subsection{Updates}{hugs-updates}
3680 \item[ Updating with a constructor].
3682 \begin{tabular}{|llrrrrr|}
3684 & [ENTER] & $a : @UPD_RET@ : ua : s$ & $su$ & $h[a \mapsto C\{\ws\}]$ & $hp$ & $\sigma$ \\
3685 \next & [ENTER] & $a \append s$ & $su$ & $h[au \mapsto \Ind{a}$ & $hp$ & $\sigma$ \\
3689 \item[ Argument checks].
3691 \begin{tabular}{|llrrrrr|}
3693 & ARGCHECK $m$:$\is$ & $a : \as \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3694 \next & $\is$ & $a : \as \append s$ & $su$ & $h'$ & $hp$ & $\sigma$ \\
3698 where $m \ge (su - sp)$
3700 \begin{tabular}{|llrrrrr|}
3702 & ARGCHECK $m$:$\is$ & $a : \as \append @UPD_RET@:su:ua:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3703 \next & $\is$ & $a : \as \append s$ & $su$ & $h'$ & $hp$ & $\sigma$ \\
3707 where $m < (su - sp)$ and
3708 $h' = h[ua \mapsto \Ind{a'}, a' \mapsto \PAP(a,\reverse\ \as) ]$
3710 Again, we reverse the list of values as we transfer them from the
3711 stack to the heap --- reflecting the fact that the stack and heap grow
3712 in different directions.
3716 \Subsection{Branches}{hugs-branches}
3720 \item[ Testing a constructor ].
3722 \begin{tabular}{|llrrrrr|}
3724 & TEST $tag$ $is'$ : $is$ & $a : s$ & $su$ & $h[a \mapsto C\ \ws]$ & $hp$ & $\sigma$ \\
3725 \next & $is$ & $a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3729 where $C.\tag = tag$
3731 \begin{tabular}{|llrrrrr|}
3733 & TEST $tag$ $is'$ : $is$ & $a : s$ & $su$ & $h[a \mapsto C\ \ws]$ & $hp$ & $\sigma$ \\
3734 \next & $is'$ & $a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3738 where $C.\tag \neq tag$
3742 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3743 \Subsection{Heap and stack checks}{hugs-heap-stack-checks}
3744 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3746 \begin{tabular}{|llrrrrr|}
3748 & STACKCHECK $stk$:$\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3749 \next & $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3753 if $s$ has $stk$ free slots.
3755 \begin{tabular}{|llrrrrr|}
3757 & HEAPCHECK $hp$:$\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3758 \next & $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3762 if $h$ has $hp$ free slots.
3764 If either check fails, we push the current bco ($\sigma$) onto the
3765 stack and return to the scheduler. When the scheduler has fixed the
3766 problem, it pops the top object off the stack and reenters it.
3771 \item The bytecode CHECK1000 conservatively checks for 1000 words of heap space and 1000 words of stack space.
3772 We use it to reduce code space and instruction decoding time.
3773 \item The bytecode HEAPCHECK1000 conservatively checks for 1000 words of heap space.
3774 It is used in case alternatives.
3778 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3779 \Subsection{Primops}{hugs-primops}
3780 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3782 \ToDo{primops take m words and return n words. The expect boxed arguments on the stack.}
3785 \Section{The Machine Code Evaluator}{asm-evaluator}
3787 This section describes the framework in which compiled code evaluates
3788 expressions. Only at certain points will compiled code need to be
3789 able to talk to the interpreted world; these are discussed in
3790 \secref{switching-worlds}.
3792 \Subsection{Calling conventions}{ghc-calling-conventions}
3794 \Subsubsection{The call/return registers}{ghc-regs}
3796 One of the problems in designing a virtual machine is that we want it
3797 abstract away from tedious machine details but still reveal enough of
3798 the underlying hardware that we can make sensible decisions about code
3799 generation. A major problem area is the use of registers in
3800 call/return conventions. On a machine with lots of registers, it's
3801 cheaper to pass arguments and results in registers than to pass them
3802 on the stack. On a machine with very few registers, it's cheaper to
3803 pass arguments and results on the stack than to use ``virtual
3804 registers'' in memory. We therefore use a hybrid system: the first
3805 $n$ arguments or results are passed in registers; and the remaining
3806 arguments or results are passed on the stack. For register-poor
3807 architectures, it is important that we allow $n=0$.
3809 We'll label the arguments and results \Arg{1} \ldots \Arg{m} --- with
3810 the understanding that \Arg{1} \ldots \Arg{n} are in registers and
3811 \Arg{n+1} \ldots \Arg{m} are on top of the stack.
3813 Note that the mapping of arguments \Arg{1} \ldots \Arg{n} to machine
3814 registers depends on the \emph{kinds} of the arguments. For example,
3815 if the first argument is a Float, we might pass it in a different
3816 register from if it is an Int. In fact, we might find that a given
3817 architecture lets us pass varying numbers of arguments according to
3818 their types. For example, if a CPU has 2 Int registers and 2 Float
3819 registers then we could pass between 2 and 4 arguments in machine
3820 registers --- depending on whether they all have the same kind or they
3821 have different kinds.
3823 \Subsubsection{Entering closures}{entering-closures}
3825 To evaluate a closure we jump to the entry code for the closure
3826 passing a pointer to the closure in \Arg{1} so that the entry code can
3827 access its environment.
3829 \Subsubsection{Function call}{ghc-fun-call}
3831 The function-call mechanism is obviously crucial. There are five different
3835 \item \emph{Known combinator (function with no free variables) and
3838 A fast call can be made: push excess arguments onto stack and jump to
3839 function's \emph{fast entry point} passing arguments in \Arg{1} \ldots
3842 The \emph{fast entry point} is only called with exactly the right
3843 number of arguments (in \Arg{1} \ldots \Arg{m}) so it can instantly
3844 start doing useful work without first testing whether it has enough
3845 registers or having to pop them off the stack first.
3847 \item \emph{Known combinator and insufficient arguments.}
3849 A slow call can be made: push all arguments onto stack and jump to
3850 function's \emph{slow entry point}.
3852 Any unpointed arguments which are pushed on the stack must be tagged.
3853 This means pushing an extra word on the stack below the unpointed
3854 words, containing the number of unpointed words above it.
3856 %Todo: forward ref about tagging?
3859 The \emph{slow entry point} might be called with insufficient arguments
3860 and so it must test whether there are enough arguments on the stack.
3861 This \emph{argument satisfaction check} consists of checking that
3862 @Su-Sp@ is big enough to hold all the arguments (including any tags).
3866 \item If the argument satisfaction check fails, it is because there is
3867 one or more update frames on the stack before the rest of the
3868 arguments that the function needs. In this case, we construct a PAP
3869 (partial application, \secref{PAP}) containing the arguments
3870 which are on the stack. The PAP construction code will return to the
3871 update frame with the address of the PAP in \Arg{1}.
3873 \item If the argument satisfaction check succeeds, we jump to the fast
3874 entry point with the arguments in \Arg{1} \ldots \Arg{arity}.
3876 If the fast entry point expects to receive some of \Arg{i} on the
3877 stack, we can reduce the amount of movement required by making the
3878 stack layout for the fast entry point look like the stack layout for
3879 the slow entry point. Since the slow entry point is entered with the
3880 first argument on the top of the stack and with tags in front of any
3881 unpointed arguments, this means that if \Arg{i} is unpointed, there
3882 should be space below it for a tag and that the highest numbered
3883 argument should be passed on the top of the stack.
3885 We usually arrange that the fast entry point is placed immediately
3886 after the slow entry point --- so we can just ``fall through'' to the
3887 fast entry point without performing a jump.
3892 \item \emph{Known function closure (function with free variables) and
3895 A fast call can be made: push excess arguments onto stack and jump to
3896 function's \emph{fast entry point} passing a pointer to closure in
3897 \Arg{1} and arguments in \Arg{2} \ldots \Arg{m+1}.
3899 Like the fast entry point for a combinator, the fast entry point for a
3900 closure is only called with appropriate values in \Arg{1} \ldots
3901 \Arg{m+1} so we can start work straight away. The pointer to the
3902 closure is used to access the free variables of the closure.
3905 \item \emph{Known function closure and insufficient arguments.}
3907 A slow call can be made: push all arguments onto stack and jump to the
3908 closure's slow entry point passing a pointer to the closure in \Arg{1}.
3910 Again, the slow entry point performs an argument satisfaction check
3911 and either builds a PAP or pops the arguments off the stack into
3912 \Arg{2} \ldots \Arg{m+1} and jumps to the fast entry point.
3915 \item \emph{Unknown function closure, thunk or constructor.}
3917 Sometimes, the function being called is not statically identifiable.
3918 Consider, for example, the @compose@ function:
3920 compose f g x = f (g x)
3922 Since @f@ and @g@ are passed as arguments to @compose@, the latter has
3923 to make a heap call. In a heap call the arguments are pushed onto the
3924 stack, and the closure bound to the function is entered. In the
3925 example, a thunk for @(g x)@ will be allocated, (a pointer to it)
3926 pushed on the stack, and the closure bound to @f@ will be
3927 entered. That is, we will jump to @f@s entry point passing @f@ in
3928 \Arg{1}. If \Arg{1} is passed on the stack, it is pushed on top of
3929 the thunk for @(g x)@.
3931 The \emph{entry code} for an updateable thunk (which must have arity 0)
3932 pushes an update frame on the stack and starts executing the body of
3933 the closure --- using \Arg{1} to access any free variables. This is
3934 described in more detail in \secref{data-updates}.
3936 The \emph{entry code} for a non-updateable closure is just the
3937 closure's slow entry point.
3941 In addition to the above considerations, if there are \emph{too many}
3942 arguments then the extra arguments are simply pushed on the stack with
3945 To summarise, a closure's standard (slow) entry point performs the
3949 \item[Argument satisfaction check.] (function closure only)
3950 \item[Stack overflow check.]
3951 \item[Heap overflow check.]
3952 \item[Copy free variables out of closure.] %Todo: why?
3953 \item[Eager black holing.] (updateable thunk only) %Todo: forward ref.
3954 \item[Push update frame.]
3955 \item[Evaluate body of closure.]
3959 \Subsection{Case expressions and return conventions}{return-conventions}
3961 The \emph{evaluation} of a thunk is always initiated by
3962 a @case@ expression. For example:
3964 case x of (a,b) -> E
3967 The code for a @case@ expression looks like this:
3970 \item Push the free variables of the branches on the stack (fv(@E@) in
3972 \item Push a \emph{return address} on the stack.
3973 \item Evaluate the scrutinee (@x@ in this case).
3976 Once evaluation of the scrutinee is complete, execution resumes at the
3977 return address, which points to the code for the expression @E@.
3979 When execution resumes at the return point, there must be some {\em
3980 return convention} that defines where the components of the pair, @a@
3981 and @b@, can be found. The return convention varies according to the
3982 type of the scrutinee @x@:
3988 (A space for) the return address is left on the top of the stack.
3989 Leaving the return address on the stack ensures that the top of the
3990 stack contains a valid activation record
3991 (\secref{activation-records}) --- should a garbage
3992 collection be required.
3994 \item If @x@ has a boxed type (e.g.~a data constructor or a function),
3995 a pointer to @x@ is returned in \Arg{1}.
3997 \ToDo{Warn that components of E should be extracted as soon as
3998 possible to avoid a space leak.}
4000 \item If @x@ is an unboxed type (e.g.~@Int#@ or @Float#@), @x@ is
4003 \item If @x@ is an unboxed tuple constructor, the components of @x@
4004 are returned in \Arg{1} \ldots \Arg{n} but no object is constructed in
4007 When passing an unboxed tuple to a function, the components are
4008 flattened out and passed in \Arg{1} \ldots \Arg{n} as usual.
4012 \Subsection{Vectored Returns}{vectored-returns}
4014 Many algebraic data types have more than one constructor. For
4015 example, the @Maybe@ type is defined like this:
4017 data Maybe a = Nothing | Just a
4019 How does the return convention encode which of the two constructors is
4020 being returned? A @case@ expression scrutinising a value of @Maybe@
4021 type would look like this:
4027 Rather than pushing a return address before evaluating the scrutinee,
4028 @E@, the @case@ expression pushes (a pointer to) a \emph{return
4029 vector}, a static table consisting of two code pointers: one for the
4030 @Just@ alternative, and one for the @Nothing@ alternative.
4036 The constructor @Nothing@ returns by jumping to the first item in the
4037 return vector with a pointer to a (statically built) Nothing closure
4040 It might seem that we could avoid loading \Arg{1} in this case since the
4041 first item in the return vector will know that @Nothing@ was returned
4042 (and can easily access the Nothing closure in the (unlikely) event
4043 that it needs it. The only reason we load \Arg{1} is in case we have to
4044 perform an update (\secref{data-updates}).
4048 The constructor @Just@ returns by jumping to the second element of the
4049 return vector with a pointer to the closure in \Arg{1}.
4053 In this way no test need be made to see which constructor returns;
4054 instead, execution resumes immediately in the appropriate branch of
4057 \Subsection{Direct Returns}{direct-returns}
4059 When a datatype has a large number of constructors, it may be
4060 inappropriate to use vectored returns. The vector tables may be
4061 large and sparse, and it may be better to identify the constructor
4062 using a test-and-branch sequence on the tag. For this reason, we
4063 provide an alternative return convention, called a \emph{direct
4066 In a direct return, the return address pushed on the stack really is a
4067 code pointer. The returning code loads a pointer to the closure being
4068 returned in \Arg{1} as usual, and also loads the tag into \Arg{2}.
4069 The code at the return address will test the tag and jump to the
4070 appropriate code for the case branch. If \Arg{2} isn't mapped to a
4071 real machine register on this architecture, then we don't load it on a
4072 return, instead using the tag directly from the info table.
4074 The choice of whether to use a vectored return or a direct return is
4075 made on a type-by-type basis --- up to a certain maximum number of
4076 constructors imposed by the update mechanism
4077 (\secref{data-updates}).
4079 Single-constructor data types also use direct returns, although in
4080 that case there is no need to return a tag in \Arg{2}.
4082 \ToDo{for a nullary constructor we needn't return a pointer to the
4083 constructor in \Arg{1}.}
4085 \Subsection{Updates}{data-updates}
4087 The entry code for an updatable thunk (which must be of arity 0):
4090 \item copies the free variables out of the thunk into registers or
4092 \item pushes an \emph{update frame} onto the stack.
4094 An update frame is a small activation record consisting of
4096 \begin{tabular}{|l|l|l|}
4098 \emph{Fixed header} & \emph{Update Frame link} & \emph{Updatee} \\
4103 \note{In the semantics part of the STG paper (section 5.6), an update
4104 frame consists of everything down to the last update frame on the
4105 stack. This would make sense too --- and would fit in nicely with
4106 what we're going to do when we add support for speculative
4108 \ToDo{I think update frames contain cost centres sometimes}
4111 If we are doing ``eager blackholing,'' we then overwrite the thunk
4112 with a black hole. Otherwise, we leave it to the garbage collector to
4113 black hole the thunk.
4116 Start evaluating the body of the expression.
4120 When the expression finishes evaluation, it will enter the update
4121 frame on the top of the stack. Since the returner doesn't know
4122 whether it is entering a normal return address/vector or an update
4123 frame, we follow exactly the same conventions as return addresses and
4124 return vectors. That is, on entering the update frame:
4127 \item The value of the thunk is in \Arg{1}. (Recall that only thunks
4128 are updateable and that thunks return just one value.)
4130 \item If the data type is a direct-return type rather than a
4131 vectored-return type, then the tag is in \Arg{2}.
4133 \item The update frame is still on the stack.
4136 We can safely share a single statically-compiled update function
4137 between all types. However, the code must be able to handle both
4138 vectored and direct-return datatypes. This is done by arranging that
4139 the update code looks like this:
4147 |---------------| <- update code pointer
4152 Each entry in the return vector (which is large enough to cover the
4153 largest vectored-return type) points to the update code.
4157 \item overwrites the \emph{updatee} with an indirection to \Arg{1};
4158 \item loads @Su@ from the Update Frame link;
4159 \item removes the update frame from the stack; and
4160 \item enters \Arg{1}.
4163 We enter \Arg{1} again, having probably just come from there, because
4164 it knows whether to perform a direct or vectored return. This could
4165 be optimised by compiling special update code for each slot in the
4166 return vector, which performs the correct return.
4168 \Subsection{Semi-tagging}{semi-tagging}
4170 When a @case@ expression evaluates a variable that might be bound
4171 to a thunk it is often the case that the scrutinee is already evaluated.
4172 In this case we have paid the penalty of (a) pushing the return address (or
4173 return vector address) on the stack, (b) jumping through the info pointer
4174 of the scrutinee, and (c) returning by an indirect jump through the
4175 return address on the stack.
4177 If we knew that the scrutinee was already evaluated we could generate
4178 (better) code which simply jumps to the appropriate branch of the
4179 @case@ with a pointer to the scrutinee in \Arg{1}. (For direct
4180 returns to multiconstructor datatypes, we might also load the tag into
4183 An obvious idea, therefore, is to test dynamically whether the heap
4184 closure is a value (using the tag in the info table). If not, we
4185 enter the closure as usual; if so, we jump straight to the appropriate
4186 alternative. Here, for example, is pseudo-code for the expression
4187 @(case x of { (a,_,c) -> E }@:
4189 \Arg{1} = <pointer to x>;
4190 tag = \Arg{1}->entry->tag;
4192 Sp--; \\ insert space for return address
4196 goto \Arg{1}->entry;
4198 <info table for return address goes here>
4199 ret: a = \Arg{1}->data1; \\ suck out a and c to avoid space leak
4203 and here is the code for the expression @(case x of { [] -> E1; x:xs -> E2 }@:
4205 \Arg{1} = <pointer to x>;
4206 tag = \Arg{1}->entry->tag;
4208 Sp--; \\ insert space for return address
4212 goto \Arg{1}->entry;
4216 retvec: \\ reversed return vector
4217 <return info table for case goes here>
4219 panic("Direct return into vectored case");
4223 ret2: x = \Arg{1}->head;
4227 There is an obvious cost in compiled code size (but none in the size
4228 of the bytecodes). There is also a cost in execution time if we enter
4229 more thunks than data constructors.
4231 Both the direct and vectored returns are easily modified to chase chains
4232 of indirections too. In the vectored case, this is most easily done by
4233 making sure that @IND = TAG_1 - 1@, and adding an extra field to every
4234 return vector. In the above example, the indirection code would be
4236 ind: \Arg{1} = \Arg{1}->next;
4239 where @ind_loop@ is the second line of code.
4241 Note that we have to leave space for a return address since the return
4242 address expects to find one. If the body of the expression requires a
4243 heap check, we will actually have to write the return address before
4244 entering the garbage collector.
4247 \Subsection{Heap and Stack Checks}{heap-and-stack-checks}
4249 The storage manager detects that it needs to garbage collect the old
4250 generation when the evaluator requests a garbage collection without
4251 having moved the heap pointer since the last garbage collection. It
4252 is therefore important that the GC routines \emph{not} move the heap
4253 pointer unless the heap check fails. This is different from what
4254 happens in the current STG implementation.
4256 Assuming that the stack can never shrink, we perform a stack check
4257 when we enter a closure but not when we return to a return
4258 continuation. This doesn't work for heap checks because we cannot
4259 predict what will happen to the heap if we call a function.
4261 If we wish to allow the stack to shrink, we need to perform a stack
4262 check whenever we enter a return continuation. Most of these checks
4263 could be eliminated if the storage manager guaranteed that a stack
4264 would always have 1000 words (say) of space after it was shrunk. Then
4265 we can omit stack checks for less than 1000 words in return
4268 When an argument satisfaction check fails, we need to push the closure
4269 (in R1) onto the stack - so we need to perform a stack check. The
4270 problem is that the argument satisfaction check occurs \emph{before}
4271 the stack check. The solution is that the caller of a slow entry
4272 point or closure will guarantee that there is at least one word free
4273 on the stack for the callee to use.
4275 Similarily, if a heap or stack check fails, we need to push the arguments
4276 and closure onto the stack. If we just came from the slow entry point,
4277 there's certainly enough space and it is the responsibility of anyone
4278 using the fast entry point to guarantee that there is enough space.
4280 \ToDo{Be more precise about how much space is required - document it
4281 in the calling convention section.}
4283 \Subsection{Handling interrupts/signals}{signals}
4286 May have to keep C stack pointer in register to placate OS?
4287 May have to revert black holes - ouch!
4292 \section{The Loader}
4293 \section{The Compilers}
4296 \part{Old stuff - needs to be mined for useful info}
4298 \section{The Scheduler}
4300 The Scheduler is the heart of the run-time system. A running program
4301 consists of a single running thread, and a list of runnable and
4302 blocked threads. The running thread returns to the scheduler when any
4303 of the following conditions arises:
4306 \item A heap check fails, and a garbage collection is required
4307 \item Compiled code needs to switch to interpreted code, and vice
4309 \item The thread becomes blocked.
4310 \item The thread is preempted.
4313 A running system has a global state, consisting of
4316 \item @Hp@, the current heap pointer, which points to the next
4317 available address in the Heap.
4318 \item @HpLim@, the heap limit pointer, which points to the end of the
4320 \item The Thread Preemption Flag, which is set whenever the currently
4321 running thread should be preempted at the next opportunity.
4322 \item A list of runnable threads.
4323 \item A list of blocked threads.
4326 Each thread is represented by a Thread State Object (TSO), which is
4327 described in detail in \secref{TSO}.
4329 The following is pseudo-code for the inner loop of the scheduler
4333 while (threads_exist) {
4334 // handle global problems: GC, parallelism, etc
4336 if (external_message) service_message();
4337 // deal with other urgent stuff
4339 pick a runnable thread;
4341 // enter object on top of stack
4342 // if the top object is a BCO, we must enter it
4343 // otherwise appply any heuristic we wish.
4344 if (thread->stack[thread->sp]->info.type == BCO) {
4345 status = runHugs(thread,&smInfo);
4347 status = runGHC(thread,&smInfo);
4349 switch (status) { // handle local problems
4350 case (StackOverflow): enlargeStack; break;
4351 case (Error e) : error(thread,e); break;
4352 case (ExitWith e) : exit(e); break;
4353 case (Yield) : break;
4355 } while (thread_runnable);
4359 \Subsection{Invoking the garbage collector}{ghc-invoking-gc}
4361 \Subsection{Putting the thread to sleep}{ghc-thread-sleeps}
4363 \Subsection{Calling C from Haskell}{ghc-ccall}
4365 We distinguish between "safe calls" where the programmer guarantees
4366 that the C function will not call a Haskell function or, in a
4367 multithreaded system, block for a long period of time and "unsafe
4368 calls" where the programmer cannot make that guarantee.
4370 Safe calls are performed without returning to the scheduler and are
4371 discussed elsewhere (\ToDo{discuss elsewhere}).
4373 Unsafe calls are performed by returning an array (outside the Haskell
4374 heap) of arguments and a C function pointer to the scheduler. The
4375 scheduler allocates a new thread from the operating system
4376 (multithreaded system only), spawns a call to the function and
4377 continues executing another thread. When the ccall completes, the
4378 thread informs the scheduler and the scheduler adds the thread to the
4379 runnable threads list.
4381 \ToDo{Describe this in more detail.}
4384 \Subsection{Calling Haskell from C}{ghc-c-calls-haskell}
4386 When C calls a Haskell closure, it sends a message to the scheduler
4387 thread. On receiving the message, the scheduler creates a new Haskell
4388 thread, pushes the arguments to the C function onto the thread's stack
4389 (with tags for unboxed arguments) pushes the Haskell closure and adds
4390 the thread to the runnable list so that it can be entered in the
4393 When the closure returns, the scheduler sends back a message which
4394 awakens the (C) thread.
4396 \ToDo{Do we need to worry about the garbage collector deallocating the
4397 thread if it gets blocked?}
4399 \Subsection{Switching Worlds}{switching-worlds}
4401 \ToDo{This has all changed: we always leave a closure on top of the
4402 stack if we mean to continue executing it. The scheduler examines the
4403 top of the stack and tries to guess which world we want to be in. If
4404 it finds a @BCO@, it certainly enters Hugs, if it finds a @GHC@
4405 closure, it certainly enters GHC and if it finds a standard closure,
4406 it is free to choose either one but it's probably best to enter GHC
4407 for everything except @BCO@s and perhaps @AP@s.}
4409 Because this is a combined compiled/interpreted system, the
4410 interpreter will sometimes encounter compiled code, and vice-versa.
4412 All world-switches go via the scheduler, ensuring that the world is in
4413 a known state ready to enter either compiled code or the interpreter.
4414 When a thread is run from the scheduler, the @whatNext@ field in the
4415 TSO (\secref{TSO}) is checked to find out how to execute the
4419 \item If @whatNext@ is set to @ReturnGHC@, we load up the required
4420 registers from the TSO and jump to the address at the top of the user
4422 \item If @whatNext@ is set to @EnterGHC@, we load up the required
4423 registers from the TSO and enter the closure pointed to by the top
4425 \item If @whatNext@ is set to @EnterHugs@, we enter the top thing on
4426 the stack, using the interpreter.
4429 There are four cases we need to consider:
4432 \item A GHC thread enters a Hugs-built closure.
4433 \item A GHC thread returns to a Hugs-compiled return address.
4434 \item A Hugs thread enters a GHC-built closure.
4435 \item A Hugs thread returns to a Hugs-compiled return address.
4438 GHC-compiled modules cannot call functions in a Hugs-compiled module
4439 directly, because the compiler has no information about arities in the
4440 external module. Therefore it must assume any top-level objects are
4441 CAFs, and enter their closures.
4443 \ToDo{Hugs-built constructors?}
4445 We now examine the various cases one by one and describe how the
4446 switch happens in each situation.
4448 \subsection{A GHC thread enters a Hugs-built closure}
4449 \label{sec:ghc-to-hugs-switch}
4451 There is three possibilities: GHC has entered a @PAP@, or it has
4452 entered a @AP@, or it has entered the BCO directly (for a top-level
4453 function closure). @AP@s and @PAP@s are ``standard closures'' and
4454 so do not require us to enter the bytecode interpreter.
4456 The entry code for a BCO does the following:
4459 \item Push the address of the object entered on the stack.
4460 \item Save the current state of the thread in its TSO.
4461 \item Return to the scheduler, setting @whatNext@ to @EnterHugs@.
4464 BCO's for thunks and functions have the same entry conventions as
4465 slow entry points: they expect to find their arguments on the stac
4466 with unboxed arguments preceded by appropriate tags.
4468 \subsection{A GHC thread returns to a Hugs-compiled return address}
4469 \label{sec:ghc-to-hugs-switch}
4471 Hugs return addresses are laid out as in \figref{hugs-return-stack}.
4472 If GHC is returning, it will return to the address at the top of the
4473 stack, namely @HUGS_RET@. The code at @HUGS_RET@ performs the
4477 \item pushes \Arg{1} (the return value) on the stack.
4478 \item saves the thread state in the TSO
4479 \item returns to the scheduler with @whatNext@ set to @EnterHugs@.
4482 \noindent When Hugs runs, it will enter the return value, which will
4483 return using the correct Hugs convention
4484 (\secref{hugs-return-convention}) to the return address underneath it
4487 \subsection{A Hugs thread enters a GHC-compiled closure}
4488 \label{sec:hugs-to-ghc-switch}
4490 Hugs can recognise a GHC-built closure as not being one of the
4491 following types of object:
4497 \item An indirection, or
4498 \item A constructor.
4501 When Hugs is called on to enter a GHC closure, it executes the
4502 following sequence of instructions:
4505 \item Push the address of the closure on the stack.
4506 \item Save the current state of the thread in the TSO.
4507 \item Return to the scheduler, with the @whatNext@ field set to
4511 \subsection{A Hugs thread returns to a GHC-compiled return address}
4512 \label{sec:hugs-to-ghc-switch}
4514 When Hugs encounters a return address on the stack that is not
4515 @HUGS_RET@, it knows that a world-switch is required. At this point
4516 the stack contains a pointer to the return value, followed by the GHC
4517 return address. The following sequence is then performed:
4520 \item save the state of the thread in the TSO.
4521 \item return to the scheduler, setting @whatNext@ to @EnterGHC@.
4524 The first thing that GHC will do is enter the object on the top of the
4525 stack, which is a pointer to the return value. This value will then
4526 return itself to the return address using the GHC return convention.
4534 We're nuking the following:
4541 Return in registers.
4542 This lets us remove update code pointers from info tables,
4543 removes the need for phantom info tables, simplifies
4548 Careful analysis suggests that it doesn't buy us very much
4549 and it is hard to work with.
4551 Eliminating threaded GCs eliminates the desire to share SMReps
4552 so they are (once more) part of the Info table.
4556 Doesn't buy us anything on a register-poor architecture and
4557 isn't so important if we have semi-tagging.
4560 - Probably bad on register poor architecture
4561 - Can avoid need to write return address to stack on reg rich arch.
4562 - when a function does a small amount of work, doesn't
4563 enter any other thunks and then returns.
4564 eg entering a known constructor (but semitagging will catch this)
4565 - Adds complications
4571 This lets us drop CONST closures and CHARLIKE closures (assuming we
4572 don't support Unicode). The only point of these closures was to
4573 avoid updating with an indirection.
4575 We also drop @MIN_UPD_SIZE@ --- all we need is space to insert an
4576 indirection or a black hole.
4579 STATIC SMReps are now called CONST
4584 \item The profiling ``kind'' field is now encoded in the @INFO_TYPE@ field.
4585 This identifies the general sort of the closure for profiling purposes.
4587 \item Various papers describe deleting update frames for unreachable objects.
4588 This has never been implemented and we don't plan to anytime soon.