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
26 %\newcommand{\note}[1]{{\em Note: #1}}
27 \newcommand{\note}[1]{{{\bf Note:}\sl #1}}
28 \newcommand{\ToDo}[1]{{{\bf ToDo:}\sl #1}}
29 \newcommand{\Arg}[1]{\mbox{${\tt arg}_{#1}$}}
30 \newcommand{\bottom}{\perp}
32 \newcommand{\secref}[1]{Section~\ref{sec:#1}}
33 \newcommand{\figref}[1]{Figure~\ref{fig:#1}}
34 \newcommand{\Section}[2]{\section{#1}\label{sec:#2}}
35 \newcommand{\Subsection}[2]{\subsection{#1}\label{sec:#2}}
36 \newcommand{\Subsubsection}[2]{\subsubsection{#1}\label{sec:#2}}
47 \setlength{\parskip}{0.15cm}
48 \setlength{\parsep}{0.15cm}
49 \setlength{\topsep}{0cm} % Reduces space before and after verbatim,
50 % which is implemented using trivlist
51 \setlength{\parindent}{0cm}
53 \renewcommand{\textfraction}{0.2}
54 \renewcommand{\floatpagefraction}{0.7}
58 \title{The STG runtime system (revised)}
59 \author{Simon Peyton Jones \\ Microsoft Research Ltd., Cambridge \and
60 Simon Marlow \\ Microsoft Research Ltd., Cambridge \and
61 Alastair Reid \\ Yale University}
69 \Section{Overview}{overview}
71 This document describes the GHC/Hugs run-time system. It serves as
72 a Glasgow/Yale/Nottingham ``contract'' about what the RTS does.
74 \Subsection{New features compared to GHC 3.xx}{new-features}
77 \item The RTS supports mixed compiled/interpreted execution, so
78 that a program can consist of a mixture of GHC-compiled and Hugs-interpreted
81 \item The RTS supports concurrency by default.
82 This has some costs (eg we can't do hardware stack checks) but
83 reduces the number of different configurations we need to support.
85 \item CAFs are only retained if they are
86 reachable. Since they are referred to by implicit references buried
87 in code, this means that the garbage collector must traverse the whole
88 accessible code tree. This feature eliminates a whole class of painful
91 \item A running thread has only one stack, which contains a mixture of
92 pointers and non-pointers. \secref{TSO} describes how we find out
93 which is which. (GHC has used two stacks for some while. Using one
94 stack instead of two reduces register pressure, reduces the size of
95 update frames, and eliminates ``stack-stubbing'' instructions.)
97 \item The ``return in registers'' return convention has been dropped
98 because it was complicated and doesn't work well on register-poor
99 architectures. It has been partly replaced by unboxed tuples
100 (\secref{unboxed-tuples}) which allow the programmer to
101 explicitly state where results should be returned in registers (or on
102 the stack) instead of on the heap.
104 \item Exceptions are supported by the RTS.
106 \item Weak Pointers generalise the previously available Foreign Object
109 \item The garbage collector supports a number of new features,
110 including a dynamically resizable heap and multiple generations with
111 aging within a generation.
115 \Subsection{Wish list}{wish-list}
117 Here's a list of things we'd like to support in the future.
119 \item Interrupts, speculative computation.
122 The SM could tune the size of the allocation arena, the number of
123 generations, etc taking into account residency, GC rate and page fault
127 We could trigger a GC when all threads are blocked waiting for IO if
128 the allocation arena (or some of the generations) are nearly full.
132 \Subsection{Configuration}{configuration}
134 Some of the above features are expensive or less portable, so we
135 envision building a number of different configurations supporting
136 different subsets of the above features.
138 You can make the following choices:
141 Support for parallelism. There are three mutually-exclusive choices.
144 \item[@SEQUENTIAL@] Support for concurrency but not for parallelism.
145 \item[@GRANSIM@] Concurrency support and simulated parallelism.
146 \item[@PARALLEL@] Concurrency support and real parallelism.
149 \item @PROFILING@ adds cost-centre profiling.
151 \item @TICKY@ gathers internal statistics (often known as ``ticky-ticky'' code).
153 \item @DEBUG@ does internal consistency checks.
155 \item Persistence. (well, not yet).
158 Which garbage collector to use. At the moment we
159 only anticipate one, however.
162 \Subsection{Glossary}{glossary}
164 \ToDo{This terminology is not used consistently within the document.
165 If you find something which disagrees with this terminology, fix the
168 In the type system, we have boxed and unboxed types.
172 \item A \emph{pointed} type is one that contains $\bot$. Variables with
173 pointed types are the only things which can be lazily evaluated. In
174 the STG machine, this means that they are the only things that can be
175 \emph{entered} or \emph{updated} and it requires that they be boxed.
177 \item An \emph{unpointed} type is one that does not contain $\bot$.
178 Variables with unpointed types are never delayed --- they are always
179 evaluated when they are constructed. In the STG machine, this means
180 that they cannot be \emph{entered} or \emph{updated}. Unpointed objects
181 may be boxed (like @Array#@) or unboxed (like @Int#@).
185 In the implementation, we have different kinds of objects:
189 \item \emph{boxed} objects are heap objects used by the evaluators
191 \item \emph{unboxed} objects are not heap allocated
193 \item \emph{stack} objects are allocated on the stack
195 \item \emph{closures} are objects which can be \emph{entered}.
196 They are always boxed and always have boxed types.
197 They may be in WHNF or they may be unevaluated.
199 \item A \emph{thunk} is a (representation of) a value of a \emph{pointed}
200 type which is \emph{not} in WHNF.
202 \item A \emph{value} is an object in WHNF. It can be pointed or unpointed.
208 At the hardware level, we have \emph{word}s and \emph{pointer}s.
212 \item A \emph{word} is (at least) 32 bits and can hold either a signed
215 \item A \emph{pointer} is (at least) 32 bits and big enough to hold a
216 function pointer or a data pointer.
220 Occasionally, a field of a data structure must hold either a word or a
221 pointer. In such circumstances, it is \emph{not safe} to assume that
222 words and pointers are the same size. \ToDo{GHC currently makes words
223 the same size as pointers to reduce complexity in the code
224 generator/RTS. It would be useful to relax this restriction, and have
225 eg. 32-bit Ints on a 64-bit machine.}
227 % should define terms like SRT, CAF, PAP, etc. here? --KSW 1999-03
229 \subsection{Subtle Dependencies}
231 Some decisions have very subtle consequences which should be written
232 down in case we want to change our minds.
238 If the garbage collector is allowed to shrink the stack of a thread,
239 we cannot omit the stack check in return continuations
240 (\secref{heap-and-stack-checks}).
244 When we return to the scheduler, the top object on the stack is a closure.
245 The scheduler restarts the thread by entering the closure.
247 \secref{hugs-return-convention} discusses how Hugs returns an
248 unboxed value to GHC and how GHC returns an unboxed value to Hugs.
252 When we return to the scheduler, we need a few empty words on the stack
253 to store a closure to reenter. \secref{heap-and-stack-checks}
254 discusses who does the stack check and how much space they need.
258 Heap objects never contain slop --- this is required if we want to
259 support mostly-copying garbage collection.
261 This is a big problem when updating since the updatee is usually
262 bigger than an indirection object. The fix is to overwrite the end of
263 the updatee with ``slop objects'' (described in
264 \secref{slop-objects}). This is hard to arrange if we do
265 \emph{lazy} blackholing (\secref{lazy-black-holing}) so we
266 currently plan to blackhole an object when we push the update frame.
268 % Idea: have specialised update code for various common sizes of
269 % updatee, the update frame hence encodes the length of the object.
270 % Specialised indirections will also encode the length of the object. A
271 % generic version of the update code will overwrite the slop with a slop
272 % object. We can do the same thing for blackhole objects, or just have
273 % a generic version that is the same size as an indirection and
274 % overwrite the slop with a slop object when blackholing. So: does this
275 % avoid the need to do eager black holing?
279 Info tables for constructors contain enough information to decide which
280 return convention they use. This allows Hugs to use a single piece of
281 entry code for all constructors and insulates Hugs from changes in the
282 choice of return convention.
286 \Section{Source Language}{source-language}
288 \Subsection{Explicit Allocation}{explicit-allocation}
290 As in the original STG machine, (almost) all heap allocation is caused
291 by executing a let(rec). Since we no longer support the return in
292 registers convention for data constructors, constructors now cause heap
293 allocation and so they should be let-bound.
295 For example, we now write
297 > cons = \ x xs -> let r = (:) x xs in r
301 > cons = \ x xs -> (:) x xs
304 \note{For historical reasons, GHC doesn't use this syntax --- but it should.}
306 \Subsection{Unboxed tuples}{unboxed-tuples}
308 Functions can take multiple arguments as easily as they can take one
309 argument: there's no cost for adding another argument. But functions
310 can only return one result: the cost of adding a second ``result'' is
311 that the function must construct a tuple of ``results'' on the heap.
312 The assymetry is rather galling and can make certain programming
313 styles quite expensive. For example, consider a simple state transformer
316 > type S a = State -> (a,State)
317 > bindS m k s0 = case m s0 of { (a,s1) -> k a s1 }
318 > returnS a s = (a,s)
322 Here, every use of @returnS@, @getS@ or @setS@ constructs a new tuple
323 in the heap which is instantly taken apart (and becomes garbage) by
324 the case analysis in @bind@. Even a short state-transformer program
325 will construct a lot of these temporary tuples.
327 Unboxed tuples provide a way for the programmer to indicate that they
328 do not expect a tuple to be shared and that they do not expect it to
329 be allocated in the heap. Syntactically, unboxed tuples are just like
330 single constructor datatypes except for the annotation @unboxed@.
332 > data unboxed AAndState# a = AnS a State
333 > type S a = State -> AAndState# a
334 > bindS m k s0 = case m s0 of { AnS a s1 -> k a s1 }
335 > returnS a s = AnS a s
337 > setS s _ = AnS () s
339 Semantically, unboxed tuples are just unlifted tuples and are subject
340 to the same restrictions as other unpointed types.
342 Operationally, unboxed tuples are never built on the heap. When
343 an unboxed tuple is returned, it is returned in multiple registers
344 or multiple stack slots. At first sight, this seems a little strange
345 but it's no different from passing double precision floats in two
351 Unboxed tuples can only have one constructor and that
352 thunks never have unboxed types --- so we'll never try to update an
353 unboxed constructor. The restriction to a single constructor is
354 largely to avoid garbage collection complications.
357 The core syntax does not allow variables to be bound to
358 unboxed tuples (ie in default case alternatives or as function arguments)
359 and does not allow unboxed tuples to be fields of other constructors.
360 However, there's no harm in allowing it in the source syntax as a
361 convenient, but easily removed, syntactic sugar.
364 The compiler generates a closure of the form
366 > c = \ x y z -> C x y z
368 for every constructor (whether boxed or unboxed).
370 This closure is normally used during desugaring to ensure that
371 constructors are saturated and to apply any strictness annotations.
372 They are also used when returning unboxed constructors to the machine
373 code evaluator from the bytecode evaluator and when a heap check fails
374 in a return continuation for an unboxed-tuple scrutinee.
378 \Subsection{STG Syntax}{stg-syntax}
381 \ToDo{Insert STG syntax with appropriate changes.}
384 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
385 \part{System Overview}
386 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
388 This part is concerned with defining the external interfaces of the
389 major components of the system; the next part is concerned with their
392 The major components of the system are:
397 The evaluators (\secref{sm-overview}) are responsible for
398 evaluating heap objects. The system supports two evaluators: the
399 machine code evaluator; and the bytecode evaluator.
403 The scheduler (\secref{scheduler-overview}) acts as the
404 coordinator for the whole system. It is responsible for switching
405 between evaluators, switching between threads, garbage collection,
406 communication between multiple processors, etc.
410 The storage manager (\secref{evaluators-overview}) is
411 responsible for allocating blocks of contiguous memory and for garbage
416 The loader (\secref{loader-overview}) is responsible for
417 loading machine code and bytecode files from the file system and for
418 resolving references between separately compiled modules.
422 The compilers (\secref{compilers-overview}) generate machine
423 code and bytecode files which can be loaded by the loader.
427 \ToDo{Insert diagram showing all components underneath the scheduler
428 and communicating only with the scheduler}
431 \Section{The Evaluators}{evaluators-overview}
433 There are two evaluators: a machine code evaluator and a bytecode
434 evaluator. The evaluators task is to evaluate code within a thread
435 until one of the following happens:
440 \item it is preempted
441 \item it blocks in one of the concurrency primitives
442 \item it performs a safe ccall
443 \item it needs to switch to the other evaluator.
446 The evaluators expect to find a closure on top of the thread's stack
447 and terminate with a closure on top of the thread's stack.
449 \Subsection{Evaluation Model}{evaluation-model}
451 Whilst the evaluators differ internally, they share a common
452 evaluation model and many object representations.
454 \Subsubsection{Heap objects}{heap-objects-overview}
456 The choice of heap and stack objects used by the evaluators is tightly
457 bound to the evaluation model. This section provides an overview of
458 the most important heap and stack objects; further details are given
461 All heap objects look like this:
464 \begin{tabular}{|l|l|l|l|}\hline
465 \emph{Header} & \emph{Payload} \\ \hline
469 The headers vary between different kinds of object but they all start
470 with a pointer to a pair consisting of an \emph{info table} and some
471 \emph{entry code}. The info table is used both by the evaluators and
472 by the storage manager and contains a @type@ field which identifies
473 which kind of heap object uses it and determines the interpretation of
474 the payload and of the other fields of the info table. The entry code
475 is some machine code used by the machine code evaluator to evaluate
476 closures and raises an error for other kinds of objects.
478 The major kinds of heap object used are as follows. (For simplicity,
479 this description omits certain optimisations and extra fields required
480 by the garbage collector.)
484 \item[Constructors] are used to represent data constructors. Their
485 payload consists of the fields of the constructor; the tag of the
486 constructor is stored in the info table.
489 \begin{tabular}{|l|l|l|l|}\hline
490 @CONSTR@ & \emph{Fields} \\ \hline
494 \item[Primitive objects] are used to represent objects with unlifted
495 types which are too large to fit in a register (or stack slot) or for
496 which sharing must be preserved. Primitive objects include large
497 objects such as multiple precision integers and immutable arrays and
498 mutable objects such as mutable arrays, mutable variables, MVar's,
499 IVar's and foreign object pointers. Since primitive objects are not
500 lifted, they cannot be entered. Their payload varies according to the
503 \item[Function closures] are used to represent functions. Their
504 payload (if any) consists of the free variables of the function.
507 \begin{tabular}{|l|l|l|l|}\hline
508 @FUN@ & \emph{Free Variables} \\ \hline
512 Function closures are only generated by the machine code compiler.
514 \item[Thunks] are used to represent unevaluated expressions which will
515 be updated with their result. Their payload (if any) consists of the
516 free variables of the function. The entry code for a thunk starts by
517 pushing an \emph{update frame} onto the stack. When evaluation of the
518 thunk completes, the update frame will cause the thunk to be
519 overwritten again with an \emph{indirection} to the result of the
520 thunk, which is always a constructor or a partial application.
523 \begin{tabular}{|l|l|l|l|}\hline
524 @THUNK@ & \emph{Free Variables} \\ \hline
528 Thunks are only generated by the machine code evaluator.
530 \item[Byte-code Objects (@BCO@s)] are generated by the bytecode
531 compiler. In conjunction with \emph{updatable applications} and
532 \emph{non-updatable applications} they are used to represent
533 functions, unevaluated expressions and return addresses.
536 \begin{tabular}{|l|l|l|l|}\hline
537 @BCO@ & \emph{Constant Pool} & \emph{Bytecodes} \\ \hline
541 \item[Non-updatable (Partial) Applications] are used to represent the
542 application of a function to an insufficient number of arguments.
543 Their payload consists of the function and the arguments received so far.
546 \begin{tabular}{|l|l|l|l|}\hline
547 @PAP@ & \emph{Function Closure} & \emph{Arguments} \\ \hline
551 @PAP@s are used when a function is applied to too few arguments and by
552 code generated by the lambda-lifting phase of the bytecode compiler.
554 \item[Updatable Applications] are used to represent the application of
555 a function to a sufficient number of arguments. Their payload
556 consists of the function and its arguments.
558 Updateable applications are like thunks: on entering an updateable
559 application, the evaluators push an \emph{update frame} onto the stack
560 and overwrite the application with a \emph{black hole}; when
561 evaluation completes, the evaluators overwrite the application with an
562 \emph{indirection} to the result of the application.
565 \begin{tabular}{|l|l|l|l|}\hline
566 @AP@ & \emph{Function Closure} & \emph{Arguments} \\ \hline
570 @AP@s are only generated by the bytecode compiler.
572 \item[Black holes] are used to mark updateable closures which are
573 currently being evaluated. ``Black holing'' an object cures a
574 potential space leak and detects certain classes of infinite loops.
575 More imporantly, black holes act as synchronisation objects between
576 separate threads: if a second thread tries to enter an updateable
577 closure which is already being evaluated, the second thread is added
578 to a list of blocked threads and the thread is suspended.
580 When evaluation of the black-holed closure completes, the black hole
581 is overwritten with an indirection to the result of the closure and
582 any blocked threads are restored to the runnable queue.
584 Closures are overwritten by black-holes during a ``lazy black-holing''
585 phase which runs on each thread when it returns to the scheduler.
586 \ToDo{section describing lazy black-holing}.
589 \begin{tabular}{|l|l|l|l|}\hline
590 @BLACKHOLE@ & \emph{Blocked threads} \\ \hline
594 \ToDo{In a single threaded system, it's trivial to detect infinite
595 loops: reentering a BLACKHOLE is always an error. How easy is it in a
596 multi-threaded system?}
598 \item[Indirections] are used to update an unevaluated closure with its
599 (usually fully evaluated) result in situations where it isn't possible
600 to perform an update in place. (In the current system, we always
601 update with an indirection to avoid duplicating the result when doing
605 \begin{tabular}{|l|l|l|l|}\hline
606 @IND@ & \emph{Closure} \\ \hline
610 Indirections needn't always point to a closure in WHNF. They can
611 point to a chain of indirections which point to an evaluated closure.
613 \item[Thread State Objects (@TSO@s)] represent Haskell threads. Their
614 payload consists of some per-thread information such as the Thread ID
615 and the status of the thread (runnable, blocked etc.), and the
616 thread's stack. See @TSO.h@ for the full story. @TSO@s may be
617 resized by the scheduler if its stack is too small or too large.
619 The thread stack grows downwards from higher to lower addresses.
622 \begin{tabular}{|l|l|l|l|}\hline
623 @TSO@ & \emph{Thread info} & \emph{Stack} \\ \hline
629 \Subsubsection{Stack objects}{stack-objects-overview}
631 The stack contains a mixture of \emph{pending arguments} and
632 \emph{stack objects}.
634 Pending arguments are arguments to curried functions which have not
635 yet been incorporated into an activation frame. For example, when
636 evaluating @let { g x y = x + y; f x = g{x} } in f{3,4}@, the
637 evaluator pushes both arguments onto the stack and enters @f@. @f@
638 only requires one argument so it leaves the second argument as a
639 \emph{pending argument}. The pending argument remains on the stack
640 until @f@ calls @g@ which requires two arguments: the argument passed
641 to it by @f@ and the pending argument which was passed to @f@.
643 Unboxed pending arguments are always preceeded by a ``tag'' which says
644 how large the argument is. This allows the garbage collector to
645 locate pointers within the stack.
647 There are three kinds of stack object: return addresses, update frames
648 and seq frames. All stack objects look like this
651 \begin{tabular}{|l|l|l|l|}\hline
652 \emph{Header} & \emph{Payload} \\ \hline
656 As with heap objects, the header starts with a pointer to a pair
657 consisting of an \emph{info table} and some \emph{entry code}.
661 \item[Return addresses] are used to cause selection and execution of
662 case alternatives when a constructor is returned. Return addresses
663 generated by the machine code compiler look like this:
666 \begin{tabular}{|l|l|l|l|}\hline
667 @RET_XXX@ & \emph{Free Variables of the case alternatives} \\ \hline
671 The free variables are a mixture of pointers and non-pointers whose
672 layout is described by a bitmask in the info table.
674 There are several kinds of @RET_XXX@ return address - see
675 \secref{activation-records} for the details.
677 Return addresses generated by the bytecode compiler look like this:
679 \begin{tabular}{|l|l|l|l|}\hline
680 @BCO_RET@ & \emph{BCO} & \emph{Free Variables of the case alternatives} \\ \hline
684 There is just one @BCO_RET@ info pointer. We avoid needing different
685 @BCO_RET@s for each stack layout by tagging unboxed free variables as
686 though they were pending arguments.
688 \item[Update frames] are used to trigger updates. When an update
689 frame is entered, it overwrites the updatee with an indirection to the
690 result, restarts any threads blocked on the @BLACKHOLE@ and returns to
691 the stack object underneath the update frame.
694 \begin{tabular}{|l|l|l|l|}\hline
695 @UPDATE_FRAME@ & \emph{Next Update Frame} & \emph{Updatee} \\ \hline
699 \item[Seq frames] are used to implement the polymorphic @seq@
700 primitive. They are a special kind of update frame, and are linked on
701 the update frame list.
704 \begin{tabular}{|l|l|l|l|}\hline
705 @SEQ_FRAME@ & \emph{Next Update Frame} \\ \hline
709 \item[Stop frames] are put on the bottom of each thread's stack, and
710 act as sentinels for the update frame list (i.e. the last update frame
711 points to the stop frame). Returning to a stop frame terminates the
712 thread. Stop frames have no payload:
715 \begin{tabular}{|l|l|l|l|}\hline
716 @SEQ_FRAME@ \\ \hline
722 \Subsubsection{Case expressions}{case-expr-overview}
724 In the STG language, all evaluation is triggered by evaluating a case
725 expression. When evaluating a case expression @case e of alts@, the
726 evaluators pushes a return address onto the stack and evaluate the
727 expression @e@. When @e@ eventually reduces to a constructor, the
728 return address on the stack is entered. The details of how the
729 constructor is passed to the return address and how the appropriate
730 case alternative is selected vary between evaluators.
732 Case expressions for unboxed data types are essentially the same: the
733 case expression pushes a return address onto the stack before
734 evaluating the scrutinee; when a function returns an unboxed value, it
735 enters the return address on top of the stack.
738 \Subsubsection{Function applications}{fun-app-overview}
740 In the STG language, all function calls are tail calls. The arguments
741 are pushed onto the stack and the function closure is entered. If any
742 arguments are unboxed, they must be tagged as unboxed pending
743 arguments. Entering a closure is just a special case of calling a
744 function with no arguments.
747 \Subsubsection{Let expressions}{let-expr-overview}
749 In the STG language, almost all heap allocation is caused by let
750 expressions. Filling in the contents of a set of mutually recursive
751 heap objects is simple enough; the only difficulty is that once the
752 heap space has been allocated, the thread must not return to the
753 scheduler until after the objects are filled in.
756 \Subsubsection{Primitive operations}{primop-overview}
760 Most primops are simple, some aren't.
767 \Section{Scheduler}{scheduler-overview}
769 The Scheduler is the heart of the run-time system. A running program
770 consists of a single running thread, and a list of runnable and
771 blocked threads. A thread is represented by a \emph{Thread Status
772 Object} (TSO), which contains a few words status information and a
773 stack. Except for the running thread, all threads have a closure on
774 top of their stack; the scheduler restarts a thread by entering an
775 evaluator which performs some reduction and returns to the scheduler.
777 \Subsection{The scheduler's main loop}{scheduler-main-loop}
779 The scheduler consists of a loop which chooses a runnable thread and
780 invokes one of the evaluators which performs some reduction and
783 The scheduler also takes care of system-wide issues such as heap
784 overflow or communication with other processors (in the parallel
785 system) and thread-specific problems such as stack overflow.
787 \Subsection{Creating a thread}{create-thread}
795 When the scheduler is first invoked.
799 When a message is received from another processor (I think). (Parallel
804 When a C program calls some Haskell code.
808 By @forkIO@, @takeMVar@ and (maybe) other Concurrent Haskell primitives.
813 \Subsection{Restarting a thread}{thread-restart}
815 When the scheduler decides to run a thread, it has to decide which
816 evaluator to use. It does this by looking at the type of the closure
819 \item @BCO@ $\Rightarrow$ bytecode evaluator
820 \item @FUN@ or @THUNK@ $\Rightarrow$ machine code evaluator
821 \item @CONSTR@ $\Rightarrow$ machine code evaluator
822 \item other $\Rightarrow$ either evaluator.
825 The only surprise in the above is that the scheduler must enter the
826 machine code evaluator if there's a constructor on top of the stack.
827 This allows the bytecode evaluator to return a constructor to a
828 machine code return address by pushing the constructor on top of the
829 stack and returning to the scheduler. If the return address under the
830 constructor is @HUGS_RET@, the entry code for @HUGS_RET@ will
831 rearrange the stack so that the return @BCO@ is on top of the stack
832 and return to the scheduler which will then call the bytecode
833 evaluator. There is little point in trying to shorten this slightly
834 indirect route since it is will happen very rarely if at all.
836 \note{As an optimisation, we could store the choice of evaluator in
837 the TSO status whenever we leave the evaluator. This is required for
838 any thread, no matter what state it is in (blocked, stack overflow,
839 etc). It isn't clear whether this would accomplish anything.}
841 \Subsection{Returning from a thread}{thread-return}
843 The evaluators return to the scheduler when any of the following
847 \item A heap check fails, and a garbage collection is required.
849 \item A stack check fails, and the scheduler must either enlarge the
850 current thread's stack, or flag an out of memory condition.
852 \item A thread enters a closure built by the other evaluator. That
853 is, when the bytecode interpreter enters a closure compiled by GHC or
854 when the machine code evaluator enters a BCO.
856 \item A thread returns to a return continuation built by the other
857 evaluator. That is, when the machine code evaluator returns to a
858 continuation built by Hugs or when the bytecode evaluator returns to a
859 continuation built by GHC.
861 \item The evaluator needs to perform a ``safe'' C call
864 \item The thread becomes blocked. This happens when a thread requires
865 the result of a computation currently being performed by another
866 thread, or it reads a synchronisation variable that is currently empty
869 \item The thread is preempted (the preemption mechanism is described
870 in \secref{thread-preemption}).
872 \item The thread terminates.
875 Except when the thread terminates, the thread always terminates with a
876 closure on the top of the stack. The mechanism used to trigger the
877 world switch and the choice of closure left on top of the stack varies
878 according to which world is being left and what is being returned.
880 \Subsubsection{Leaving the bytecode evaluator}{hugs-to-ghc-switch}
882 \paragraph{Entering a machine code closure}
884 When it enters a closure, the bytecode evaluator performs a switch
885 based on the type of closure (@AP@, @PAP@, @Ind@, etc). On entering a
886 machine code closure, it returns to the scheduler with the closure on
889 \paragraph{Returning a constructor}
891 When it enters a constructor, the bytecode evaluator tests the return
892 continuation on top of the stack. If it is a machine code
893 continuation, it returns to the scheduler with the constructor on top
896 \note{This is why the scheduler must enter the machine code evaluator
897 if it finds a constructor on top of the stack.}
899 \paragraph{Returning an unboxed value}
901 \note{Hugs doesn't support unboxed values in source programs but they
902 are used for a few complex primops.}
904 When it returns an unboxed value, the bytecode evaluator tests the
905 return continuation on top of the stack. If it is a machine code
906 continuation, it returns to the scheduler with the tagged unboxed
907 value and a special closure on top of the stack. When the closure is
908 entered (by the machine code evaluator), it returns the unboxed value
909 on top of the stack to the return continuation under it.
911 The runtime library for GHC provides one of these closures for each unboxed
912 type. Hugs cannot generate them itself since the entry code is really
915 \paragraph{Heap/Stack overflow and preemption}
917 The bytecode evaluator tests for heap/stack overflow and preemption
918 when entering a BCO and simply returns with the BCO on top of the
921 \Subsubsection{Leaving the machine code evaluator}{ghc-to-hugs-switch}
923 \paragraph{Entering a BCO}
925 The entry code for a BCO pushes the BCO onto the stack and returns to
928 \paragraph{Returning a constructor}
930 We avoid the need to test return addresses in the machine code
931 evaluator by pushing a special return address on top of a pointer to
932 the bytecode return continuation. \figref{hugs-return-stack1}
933 shows the state of the stack just before evaluating the scrutinee.
945 %\input{hugs_return1.pstex_t}
947 \caption{Stack layout for evaluating a scrutinee}
948 \label{fig:hugs-return-stack1}
951 This return address rearranges the stack so that the bco pointer is
952 above the constructor on the stack (as shown in
953 \figref{hugs-boxed-return}) and returns to the scheduler.
960 | con |--> Constructor
965 %\input{hugs_return2.pstex_t}
967 \caption{Stack layout for entering a Hugs return address}
968 \label{fig:hugs-boxed-return}
971 \paragraph{Returning an unboxed value}
973 We avoid the need to test return addresses in the machine code
974 evaluator by pushing a special return address on top of a pointer to
975 the bytecode return continuation. This return address rearranges the
976 stack so that the bco pointer is above the tagged unboxed value (as
977 shown in \figref{hugs-entering-unboxed-return}) and returns to the
992 %\input{hugs_return2.pstex_t}
994 \caption{Stack layout for returning an unboxed value}
995 \label{fig:hugs-entering-unboxed-return}
998 \paragraph{Heap/Stack overflow and preemption}
1003 \Subsection{Preempting a thread}{thread-preemption}
1005 Strictly speaking, threads cannot be preempted --- the scheduler
1006 merely sets a preemption request flag which the thread must arrange to
1007 test on a regular basis. When an evaluator finds that the preemption
1008 request flag is set, it pushes an appropriate closure onto the stack
1009 and returns to the scheduler.
1011 In the bytecode interpreter, the flag is tested whenever we enter a
1012 closure. If the preemption flag is set, it leaves the closure on top
1013 of the stack and returns to the scheduler.
1015 In the machine code evaluator, the flag is only tested when a heap or
1016 stack check fails. This is less expensive than testing the flag on
1017 entering every closure but runs the risk that a thread will enter an
1018 infinite loop which does not allocate any space. If the flag is set,
1019 the evaluator returns to the scheduler exactly as if a heap check had
1022 \Subsection{``Safe'' and ``unsafe'' C calls}{c-calls}
1024 There are two ways of calling C:
1028 \item[``Unsafe'' C calls] are used if the programer is certain that
1029 the C function will not do anything dangerous. Unsafe C calls are
1030 faster but must be hand-checked by the programmer.
1032 Dangerous things include:
1038 Call a system function such as @getchar@ which might block
1039 indefinitely. This is dangerous because we don't want the entire
1040 runtime system to block just because one thread blocks.
1044 Call an RTS function which will block on the RTS access semaphore.
1045 This would lead to deadlock.
1049 Call a Haskell function. This is just a special case of calling an
1054 Unsafe C calls are performed by pushing the arguments onto the C stack
1055 and jumping to the C function's entry point. On exit, the result of
1056 the function is in a register which is returned to the Haskell code as
1059 \item[``Safe'' C calls] are used if the programmer suspects that the
1060 thread may do something dangerous. Safe C calls are relatively slow
1061 but are less problematic.
1063 Safe C calls are performed by pushing the arguments onto the Haskell
1064 stack, pushing a return continuation and returning a \emph{C function
1065 descriptor} to the scheduler. The scheduler suspends the Haskell thread,
1066 spawns a new operating system thread which pops the arguments off the
1067 Haskell stack onto the C stack, calls the C function, pushes the
1068 function result onto the Haskell stack and informs the scheduler that
1069 the C function has completed and the Haskell thread is now runnable.
1073 The bytecode evaluator will probably treat all C calls as being safe.
1075 \ToDo{It might be good for the programmer to indicate how the program
1076 is unsafe. For example, if we distinguish between C functions which
1077 might call Haskell functions and those which might block, we could
1078 perform an unsafe call for blocking functions in a single-threaded
1079 system or, perhaps, in a multi-threaded system which only happens to
1080 have a single thread at the moment.}
1084 \Section{The Storage Manager}{sm-overview}
1086 The storage manager is responsible for managing the heap and all
1087 objects stored in it. It provides special support for lazy evaluation
1088 and for foreign function calls.
1090 \Subsection{SM support for lazy evaluation}{sm-lazy-evaluation}
1095 Indirections are shorted out.
1099 Update frames pointing to unreachable objects are squeezed out.
1101 \ToDo{Part IV suggests this doesn't happen.}
1105 Adjacent update frames (for different closures) are compressed to a
1106 single update frame pointing to a single black hole.
1111 \Subsection{SM support for foreign function calls}{sm-foreign-calls}
1117 Stable pointers allow other languages to access Haskell objects.
1121 Weak pointers and foreign objects provide finalisation support for
1122 Haskell references to external objects.
1126 \Subsection{Misc}{sm-misc}
1132 If the stack contains a large amount of free space, the storage
1133 manager may shrink the stack. If it shrinks the stack, it guarantees
1134 never to leave less than @MIN_SIZE_SHRUNKEN_STACK@ empty words on the
1135 stack when it does so.
1139 For efficiency reasons, very large objects (eg large arrays and TSOs)
1140 are not moved if possible.
1145 \Section{The Compilers}{compilers-overview}
1147 Need to describe interface files, format of bytecode files, symbols
1148 defined by machine code files.
1150 \Subsection{Interface Files}{interface-files}
1152 Here's an example - but I don't know the grammar - ADR.
1158 1 main _:_ IOBase.IO PrelBase.();;
1161 \Subsection{Bytecode files}{bytecode-files}
1163 (All that matters here is what the loader sees.)
1165 \Subsection{Machine code files}{asm-files}
1167 (Again, all that matters is what the loader sees.)
1169 \Section{The Loader}{loader-overview}
1171 In a batch mode system, we can statically link all the modules
1172 together. In an interactive system we need a loader which will
1173 explicitly load and unload individual modules (or, perhaps, blocks of
1174 mutually dependent modules) and resolve references between modules.
1176 While many operating systems provide support for dynamic loading and
1177 will automatically resolve cross-module references for us, we generally
1178 cannot rely on being able to load mutually dependent modules.
1180 A portable solution is to perform some of the linking ourselves. Each module
1181 should provide three global symbols:
1184 An initialisation routine. (Might also be used for finalisation.)
1186 A table of symbols it exports.
1187 Entries in this table consist of the symbol name and the address of the
1190 A table of symbols it imports.
1191 Entries in this table consist of the symbol name and a list of references
1195 On loading a group of modules, the loader adds the contents of the
1196 export lists to a symbol table and then fills in all the references in the
1199 References in import lists are of two types:
1201 \item[ References in machine code ]
1203 The most efficient approach is to patch the machine code directly, but
1204 this will be a lot of work, very painful to port and rather fragile.
1206 Alternatively, the loader could store the value of each symbol in the
1207 import table for each module and the compiled code can access all
1208 external objects through the import table. This requires that the
1209 import table be writable but does not require that the machine code or
1210 info tables be writable.
1212 \item[ References in data structures (SRTs and static data constructors) ]
1214 Either we patch the SRTs and constructors directly or we somehow use
1215 indirections through the symbol table. Patching the SRTs requires
1216 that we make them writable and prevents us from making effective use
1217 of virtual memories that use copy-on-write policies (this only makes a
1218 difference if we want to run several copies of the same program
1219 simultaneously). Using an indirection is possible but tricky.
1221 Note: We could avoid patching machine code if all references to
1222 external references went through the SRT --- then we just have one
1223 thing to patch. But the SRT always contains a pointer to the closure
1224 rather than the fast entry point (say), so we'd take a big performance
1229 Using the above scheme, all accesses to ``external'' objects involve a
1230 layer of indirection. To avoid this overhead, the machine code
1231 compiler might provide a way for the programmer to specify which
1232 modules will be statically linked and which will be dynamically linked
1233 --- the idea being that statically linked code and data will be
1237 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1238 \part{Internal details}
1239 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1241 This part is concerned with the internal details of the components
1242 described in the previous part.
1244 The major components of the system are:
1246 \item The scheduler (\secref{scheduler-internals})
1247 \item The storage manager (\secref{storage-manager-internals})
1248 \item The evaluators
1253 \Section{The Scheduler}{scheduler-internals}
1255 \ToDo{Detailed description of scheduler}
1257 Many heap objects contain fields allowing them to be inserted onto lists
1258 during evaluation or during garbage collection. The lists required by
1259 the evaluator and storage manager are as follows.
1263 \item 4 lists of threads: runnable threads, sleeping threads, threads
1264 waiting for timeout and threads waiting for I/O.
1266 \item The \emph{mutables list} is a list of all objects in the old
1267 generation which might contain pointers into the new generation. Most
1268 of the objects on this list are indirections (\secref{IND})
1269 or ``mutable.'' (\secref{mutables}.)
1271 \item The \emph{Foreign Object list} is a list of all foreign objects
1272 which have not yet been deallocated. (\secref{FOREIGN}.)
1274 \item The \emph{Spark pool} is a doubly(?) linked list of Spark objects
1275 maintained by the parallel system. (\secref{SPARK}.)
1277 \item The \emph{Blocked Fetch list} (or
1278 lists?). (\secref{BLOCKED_FETCH}.)
1280 \item For each thread, there is a list of all update frames on the
1281 stack. (\secref{data-updates}.)
1283 \item The Stable Pointer Table is a table of pointers to objects which
1284 are known to the outside world and must be retained by the garbage
1285 collector even if they are not accessible from within the heap.
1289 \ToDo{The links for these fields are usually inserted immediately
1290 after the fixed header except ...}
1294 \Section{The Storage Manager}{storage-manager-internals}
1296 \subsection{Misc Text looking for a home}
1298 A \emph{value} may be:
1300 \item \emph{Boxed}, i.e.~represented indirectly by a pointer to a heap object (e.g.~foreign objects, arrays); or
1301 \item \emph{Unboxed}, i.e.~represented directly by a bit-pattern in one or more registers (e.g.~@Int#@ and @Float#@).
1303 All \emph{pointed} values are \emph{boxed}.
1306 \Subsection{Heap Objects}{heap-objects}
1307 \label{sec:fixed-header}
1313 \ToDo{Fix this picture}
1318 Every \emph{heap object} is a contiguous block of memory, consisting
1319 of a fixed-format \emph{header} followed by zero or more \emph{data
1322 The header consists of the following fields:
1324 \item A one-word \emph{info pointer}, which points to
1325 the object's static \emph{info table}.
1326 \item Zero or more \emph{admin words} that support
1328 \item Profiling (notably a \emph{cost centre} word).
1329 \note{We could possibly omit the cost centre word from some
1330 administrative objects.}
1331 \item Parallelism (e.g. GranSim keeps the object's global address here,
1332 though GUM keeps a separate hash table).
1333 \item Statistics (e.g. a word to track how many times a thunk is entered.).
1335 We add a Ticky word to the fixed-header part of closures. This is
1336 used to indicate if a closure has been updated but not yet entered. It
1337 is set when the closure is updated and cleared when subsequently
1338 entered. \footnote{% NB: It is \emph{not} an ``entry count'', it is
1339 an ``entries-after-update count.'' The commoning up of @CONST@,
1340 @CHARLIKE@ and @INTLIKE@ closures is turned off(?) if this is
1341 required. This has only been done for 2s collection. }
1346 Most of the RTS is completely insensitive to the number of admin
1347 words. The total size of the fixed header is given by
1348 @sizeof(StgHeader)@.
1350 \Subsection{Info Tables}{info-tables}
1352 An \emph{info table} is a contiguous block of memory, laid out as follows:
1355 \begin{tabular}{|r|l|}
1356 \hline Parallelism Info & variable
1357 \\ \hline Profile Info & variable
1358 \\ \hline Debug Info & variable
1359 \\ \hline Static reference table & pointer word (optional)
1360 \\ \hline Storage manager layout info & pointer word
1361 \\ \hline Closure flags & 8 bits
1362 \\ \hline Closure type & 8 bits
1363 \\ \hline Constructor Tag / SRT length & 16 bits
1364 \\ \hline entry code
1369 On a 64-bit machine the tag, type and flags fields will all be doubled
1370 in size, so the info table is a multiple of 64 bits.
1372 An info table has the following contents (working backwards in memory
1377 \item The \emph{entry code} for the closure. This code appears
1378 literally as the (large) last entry in the info table, immediately
1379 preceded by the rest of the info table. An \emph{info pointer} always
1380 points to the first byte of the entry code.
1382 \item A 16-bit constructor tag / SRT length. For a constructor info
1383 table this field contains the tag of the constructor, in the range
1384 $0..n-1$ where $n$ is the number of constructors in the datatype.
1385 Otherwise, it contains the number of entries in this closure's Static
1386 Reference Table (\secref{srt}).
1388 \item An 8-bit {\em closure type field}, which identifies what kind of
1389 closure the object is. The various types of closure are described in
1392 \item an 8-bit flags field, which holds various flags pertaining to
1395 \item A single pointer or word --- the {\em storage manager info
1396 field}, contains auxiliary information describing the closure's
1397 precise layout, for the benefit of the garbage collector and the code
1398 that stuffs graph into packets for transmission over the network.
1399 There are three kinds of layout information:
1402 \item Standard layout information is for closures which place pointers
1403 before non-pointers in instances of the closure (this applies to most
1404 heap-based and static closures, but not activation records). The
1405 layout information for standard closures is
1408 \item Number of pointer fields (16 bits).
1409 \item Number of non-pointer fields (16 bits).
1412 \item Activation records don't have pointers before non-pointers,
1413 since stack-stubbing requires that the record has holes in it. The
1414 layout is therefore represented by a bitmap in which each '1' bit
1415 represents a non-pointer word. This kind of layout info is used for
1416 @RET_SMALL@ and @RET_VEC_SMALL@ closures.
1418 \item If an activation record is longer than 32 words, then the layout
1419 field contains a pointer to a bitmap record, consisting of a length
1420 field followed by two or more bitmap words. This layout information
1421 is used for @RET_BIG@ and @RET_VEC_BIG@ closures.
1423 \item Selector Thunks (\secref{THUNK_SELECTOR}) use the closure
1424 layout field to hold the selector index, since the layout is always
1425 known (the closure contains a single pointer field).
1428 \item A one-word {\em Static Reference Table} field. This field
1429 points to the static reference table for the closure (\secref{srt}),
1430 and is only present for the following closure types:
1438 \ToDo{Expand the following explanation.}
1440 An SRT is basically a vector of pointers to static closures. A
1441 top-level function or thunk will have an SRT (which might be empty),
1442 which points to all the static closures referenced by that function or
1443 thunk. Every non-top-level thunk or function also has an SRT, but
1444 it'll be a sub-sequence of the top-level SRT, so we just store a
1445 pointer and a length in the info table - the pointer points into the
1446 middle of the larger SRT.
1448 At GC time, the garbage collector traverses the transitive closure of
1449 all the SRTs reachable from the roots, and thereby discovers which
1452 \item \emph{Profiling info\/}
1454 \ToDo{The profiling info is completely bogus. I've not deleted it
1455 from the document but I've commented it all out.}
1457 % change to \iftrue to uncomment this section
1460 Closure category records are attached to the info table of the
1461 closure. They are declared with the info table. We put pointers to
1462 these ClCat things in info tables. We need these ClCat things because
1463 they are mutable, whereas info tables are immutable. Hashing will map
1464 similar categories to the same hash value allowing statistics to be
1465 grouped by closure category.
1467 Cost Centres and Closure Categories are hashed to provide indexes
1468 against which arbitrary information can be stored. These indexes are
1469 memoised in the appropriate cost centre or category record and
1470 subsequent hashes avoided by the index routine (it simply returns the
1473 There are different features which can be hashed allowing information
1474 to be stored for different groupings. Cost centres have the cost
1475 centre recorded (using the pointer), module and group. Closure
1476 categories have the closure description and the type
1477 description. Records with the same feature will be hashed to the same
1480 The initialisation routines, @init_index_<feature>@, allocate a hash
1481 table in which the cost centre / category records are stored. The
1482 lower bound for the table size is taken from @max_<feature>_no@. They
1483 return the actual table size used (the next power of 2). Unused
1484 locations in the hash table are indicated by a 0 entry. Successive
1485 @init_index_<feature>@ calls just return the actual table size.
1487 Calls to @index_<feature>@ will insert the cost centre / category
1488 record in the @<feature>@ hash table, if not already inserted. The hash
1489 index is memoised in the record and returned.
1491 CURRENTLY ONLY ONE MEMOISATION SLOT IS AVILABLE IN EACH RECORD SO
1492 HASHING CAN ONLY BE DONE ON ONE FEATURE FOR EACH RECORD. This can be
1493 easily relaxed at the expense of extra memoisation space or continued
1496 The initialisation routines must be called before initialisation of
1497 the stacks and heap as they require to allocate storage. It is also
1498 expected that the caller may want to allocate additional storage in
1499 which to store profiling information based on the return table size
1503 \begin{tabular}{|l|}
1507 \\ \hline Description String
1508 \\ \hline Type String
1514 \item[Hash Index] Memoised copy
1516 Is this category selected (-1 == not memoised, selected? 0 or 1)
1518 One of the following values (defined in CostCentre.lh):
1526 A partial application.
1528 A thunk, or suspension.
1533 \item[@ForeignObj_K@]
1534 A Foreign object (non-Haskell heap resident).
1536 The Stable Pointer table. (There should only be one of these but it
1537 represents a form of weak space leak since it can't shrink to meet
1538 non-demand so it may be worth watching separately? ADR)
1539 \item[@INTERNAL_KIND@]
1540 Something internal to the runtime system.
1544 \item[Description] Source derived string detailing closure description.
1545 \item[Type] Source derived string detailing closure type.
1548 \fi % end of commented out stuff
1550 \item \emph{Parallelism info\/}
1553 \item \emph{Debugging info\/}
1559 %-----------------------------------------------------------------------------
1560 \Subsection{Kinds of Heap Object}{closures}
1562 Heap objects can be classified in several ways, but one useful one is
1566 \emph{Static closures} occupy fixed, statically-allocated memory
1567 locations, with globally known addresses.
1570 \emph{Dynamic closures} are individually allocated in the heap.
1573 \emph{Stack closures} are closures allocated within a thread's stack
1574 (which is itself a heap object). Unlike other closures, there are
1575 never any pointers to stack closures. Stack closures are discussed in
1579 A second useful classification is this:
1582 \item \emph{Executive objects}, such as thunks and data constructors,
1583 participate directly in a program's execution. They can be subdivided
1584 into three kinds of objects according to their type: \begin{itemize}
1586 \item \emph{Pointed objects}, represent values of a \emph{pointed}
1587 type (<.pointed types launchbury.>) --i.e.~a type that includes
1588 $\bottom$ such as @Int@ or @Int# -> Int#@.
1590 \item \emph{Unpointed objects}, represent values of a \emph{unpointed}
1591 type --i.e.~a type that does not include $\bottom$ such as @Int#@ or
1594 \item \emph{Activation frames}, represent ``continuations''. They are
1595 always stored on the stack and are never pointed to by heap objects or
1596 passed as arguments. \note{It's not clear if this will still be true
1597 once we support speculative evaluation.}
1601 \item \emph{Administrative objects}, such as stack objects and thread
1602 state objects, do not represent values in the original program.
1605 Only pointed objects can be entered. If an unpointed object is
1606 entered the program will usually terminate with a fatal error.
1608 This section enumerates all the kinds of heap objects in the system.
1609 Each is identified by a distinct closure type field in its info table.
1611 \begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|}
1614 closure type & Section \\
1620 @CONSTR@ & \ref{sec:CONSTR} \\
1621 @CONSTR_p_n@ & \ref{sec:CONSTR} \\
1622 @CONSTR_STATIC@ & \ref{sec:CONSTR} \\
1623 @CONSTR_NOCAF_STATIC@ & \ref{sec:CONSTR} \\
1625 @FUN@ & \ref{sec:FUN} \\
1626 @FUN_p_n@ & \ref{sec:FUN} \\
1627 @FUN_STATIC@ & \ref{sec:FUN} \\
1629 @THUNK@ & \ref{sec:THUNK} \\
1630 @THUNK_p_n@ & \ref{sec:THUNK} \\
1631 @THUNK_STATIC@ & \ref{sec:THUNK} \\
1632 @THUNK_SELECTOR@ & \ref{sec:THUNK_SELECTOR} \\
1634 @BCO@ & \ref{sec:BCO} \\
1636 @AP_UPD@ & \ref{sec:AP_UPD} \\
1637 @PAP@ & \ref{sec:PAP} \\
1639 @IND@ & \ref{sec:IND} \\
1640 @IND_OLDGEN@ & \ref{sec:IND} \\
1641 @IND_PERM@ & \ref{sec:IND} \\
1642 @IND_OLDGEN_PERM@ & \ref{sec:IND} \\
1643 @IND_STATIC@ & \ref{sec:IND} \\
1645 @CAF_UNENTERED@ & \ref{sec:CAF} \\
1646 @CAF_ENTERED@ & \ref{sec:CAF} \\
1647 @CAF_BLACKHOLE@ & \ref{sec:CAF} \\
1653 @BLACKHOLE@ & \ref{sec:BLACKHOLE} \\
1654 @BLACKHOLE_BQ@ & \ref{sec:BLACKHOLE_BQ} \\
1656 @MVAR@ & \ref{sec:MVAR} \\
1658 @ARR_WORDS@ & \ref{sec:ARR_WORDS} \\
1660 @MUTARR_PTRS@ & \ref{sec:MUT_ARR_PTRS} \\
1661 @MUTARR_PTRS_FROZEN@ & \ref{sec:MUT_ARR_PTRS_FROZEN} \\
1663 @MUT_VAR@ & \ref{sec:MUT_VAR} \\
1665 @WEAK@ & \ref{sec:WEAK} \\
1666 @FOREIGN@ & \ref{sec:FOREIGN} \\
1667 @STABLE_NAME@ & \ref{sec:STABLE_NAME} \\
1671 Activation frames do not live (directly) on the heap --- but they have
1672 a similar organisation.
1674 \begin{tabular}{|l|l|}\hline
1675 closure type & Section \\ \hline
1676 @RET_SMALL@ & \ref{sec:activation-records} \\
1677 @RET_VEC_SMALL@ & \ref{sec:activation-records} \\
1678 @RET_BIG@ & \ref{sec:activation-records} \\
1679 @RET_VEC_BIG@ & \ref{sec:activation-records} \\
1680 @UPDATE_FRAME@ & \ref{sec:activation-records} \\
1681 @CATCH_FRAME@ & \ref{sec:activation-records} \\
1682 @SEQ_FRAME@ & \ref{sec:activation-records} \\
1683 @STOP_FRAME@ & \ref{sec:activation-records} \\
1687 There are also a number of administrative objects. It is an error to
1688 enter one of these objects.
1690 \begin{tabular}{|l|l|}\hline
1691 closure type & Section \\ \hline
1692 @TSO@ & \ref{sec:TSO} \\
1693 @SPARK_OBJECT@ & \ref{sec:SPARK} \\
1694 @BLOCKED_FETCH@ & \ref{sec:BLOCKED_FETCH} \\
1695 @FETCHME@ & \ref{sec:FETCHME} \\
1699 \Subsection{Predicates}{closure-predicates}
1701 The runtime system sometimes needs to be able to distinguish objects
1702 according to their properties: is the object updateable? is it in weak
1703 head normal form? etc. These questions can be answered by examining
1704 the closure type field of the object's info table.
1706 We define the following predicates to detect families of related
1707 info types. They are mutually exclusive and exhaustive.
1710 \item @isCONSTR@ is true for @CONSTR@s.
1711 \item @isFUN@ is true for @FUN@s.
1712 \item @isTHUNK@ is true for @THUNK@s.
1713 \item @isBCO@ is true for @BCO@s.
1714 \item @isAP@ is true for @AP@s.
1715 \item @isPAP@ is true for @PAP@s.
1716 \item @isINDIRECTION@ is true for indirection objects.
1717 \item @isBH@ is true for black holes.
1718 \item @isFOREIGN_OBJECT@ is true for foreign objects.
1719 \item @isARRAY@ is true for array objects.
1720 \item @isMVAR@ is true for @MVAR@s.
1721 \item @isIVAR@ is true for @IVAR@s.
1722 \item @isFETCHME@ is true for @FETCHME@s.
1723 \item @isSLOP@ is true for slop objects.
1724 \item @isRET_ADDR@ is true for return addresses.
1725 \item @isUPD_ADDR@ is true for update frames.
1726 \item @isTSO@ is true for @TSO@s.
1727 \item @isSTABLE_PTR_TABLE@ is true for the stable pointer table.
1728 \item @isSPARK_OBJECT@ is true for spark objects.
1729 \item @isBLOCKED_FETCH@ is true for blocked fetch objects.
1730 \item @isINVALID_INFOTYPE@ is true for all other info types.
1734 The following predicates detect other interesting properties:
1738 \item @isPOINTED@ is true if an object has a pointed type.
1740 If an object is pointed, the following predicates may be true
1741 (otherwise they are false). @isWHNF@ and @isUPDATEABLE@ are
1745 \item @isWHNF@ is true if the object is in Weak Head Normal Form.
1746 Note that unpointed objects are (arbitrarily) not considered to be in WHNF.
1748 @isWHNF@ is true for @PAP@s, @CONSTR@s, @FUN@s and all @BCO@s.
1750 \ToDo{Need to distinguish between whnf BCOs and non-whnf BCOs in their
1753 \item @isUPDATEABLE@ is true if the object may be overwritten with an
1756 @isUPDATEABLE@ is true for @THUNK@s, @AP@s and @BH@s.
1760 It is possible for a pointed object to be neither updatable nor in
1761 WHNF. For example, indirections.
1763 \item @isUNPOINTED@ is true if an object has an unpointed type.
1764 All such objects are boxed since only boxed objects have info pointers.
1766 It is true for @ARR_WORDS@, @ARR_PTRS@, @MUTVAR@, @MUTARR_PTRS@,
1767 @MUTARR_PTRS_FROZEN@, @FOREIGN@ objects, @MVAR@s and @IVAR@s.
1769 \item @isACTIVATION_FRAME@ is true for activation frames of all sorts.
1771 It is true for return addresses and update frames.
1773 \item @isVECTORED_RETADDR@ is true for vectored return addresses.
1774 \item @isDIRECT_RETADDR@ is true for direct return addresses.
1777 \item @isADMINISTRATIVE@ is true for administrative objects:
1778 @TSO@s, the stable pointer table, spark objects and blocked fetches.
1780 \item @hasSRT@ is true if the info table for the object contains an
1783 @hasSRT@ is true for @THUNK@s, @FUN@s, and @RET@s.
1789 \item @isSTATIC@ is true for any statically allocated closure.
1791 \item @isMUTABLE@ is true for objects with mutable pointer fields:
1792 @MUT_ARR@s, @MUTVAR@s, @MVAR@s and @IVAR@s.
1794 \item @isSparkable@ is true if the object can (and should) be sparked.
1795 It is true of updateable objects which are not in WHNF with the
1796 exception of @THUNK_SELECTOR@s and black holes.
1800 As a minor optimisation, we might use the top bits of the @INFO_TYPE@
1801 field to ``cache'' the answers to some of these predicates.
1803 An indirection either points to HNF (post update); or is result of
1804 overwriting a FetchMe, in which case the thing fetched is either under
1805 evaluation (BLACKHOLE), or by now an HNF. Thus, indirections get
1808 \subsection{Closures (aka Pointed Objects)}
1810 An object can be entered iff it is a closure.
1812 \Subsubsection{Function closures}{FUN}
1814 Function closures represent lambda abstractions. For example,
1815 consider the top-level declaration:
1817 f = \x -> let g = \y -> x+y
1820 Both @f@ and @g@ are represented by function closures. The closure
1821 for @f@ is \emph{static} while that for @g@ is \emph{dynamic}.
1823 The layout of a function closure is as follows:
1825 \begin{tabular}{|l|l|l|l|}\hline
1826 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \\ \hline
1830 The data words (pointers and non-pointers) are the free variables of
1831 the function closure. The number of pointers and number of
1832 non-pointers are stored in @info->layout.ptrs@ and
1833 @info->layout.nptrs@ respecively.
1835 There are several different sorts of function closure, distinguished
1836 by their closure type field:
1840 \item @FUN@: a vanilla, dynamically allocated on the heap.
1842 \item $@FUN_@p@_@np$: to speed up garbage collection a number of
1843 specialised forms of @FUN@ are provided, for particular $(p,np)$
1844 pairs, where $p$ is the number of pointers and $np$ the number of
1847 \item @FUN_STATIC@. Top-level, static, function closures (such as @f@
1848 above) have a different layout than dynamic ones:
1851 \begin{tabular}{|l|l|l|}\hline
1852 \emph{Fixed header} & \emph{Static object link} \\ \hline
1856 Static function closures have no free variables. (However they may
1857 refer to other static closures; these references are recorded in the
1858 function closure's SRT.) They have one field that is not present in
1859 dynamic closures, the \emph{static object link} field. This is used
1860 by the garbage collector in the same way that to-space is, to gather
1861 closures that have been determined to be live but that have not yet
1864 \note{Static function closures that have no static references, and
1865 hence a null SRT pointer, don't need the static object link field. We
1866 don't take advantage of this at the moment, but we could. See
1867 @CONSTR\_NOCAF\_STATIC@.}
1870 Each lambda abstraction, $f$, in the STG program has its own private
1871 info table. The following labels are relevant:
1875 \item $f$@_info@ is $f$'s info table.
1877 \item $f$@_entry@ is $f$'s slow entry point (i.e. the entry code of
1878 its info table; so it will label the same byte as $f$@_info@).
1880 \item $f@_fast_@k$ is $f$'s fast entry point. $k$ is the number of
1881 arguments $f$ takes; encoding this number in the fast-entry label
1882 occasionally catches some nasty code-generation errors.
1886 \Subsubsection{Data constructors}{CONSTR}
1888 Data-constructor closures represent values constructed with algebraic
1889 data type constructors. The general layout of data constructors is
1890 the same as that for function closures. That is
1893 \begin{tabular}{|l|l|l|l|}\hline
1894 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \\ \hline
1898 There are several different sorts of constructor:
1902 \item @CONSTR@: a vanilla, dynamically allocated constructor.
1904 \item @CONSTR_@$p$@_@$np$: just like $@FUN_@p@_@np$.
1906 \item @CONSTR_INTLIKE@. A dynamically-allocated heap object that
1907 looks just like an @Int@. The garbage collector checks to see if it
1908 can common it up with one of a fixed set of static int-like closures,
1909 thus getting it out of the dynamic heap altogether.
1911 \item @CONSTR_CHARLIKE@: same deal, but for @Char@.
1913 \item @CONSTR_STATIC@ is similar to @FUN_STATIC@, with the
1914 complication that the layout of the constructor must mimic that of a
1915 dynamic constructor, because a static constructor might be returned to
1916 some code that unpacks it. So its layout is like this:
1919 \begin{tabular}{|l|l|l|l|l|}\hline
1920 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} & \emph{Static object link}\\ \hline
1924 The static object link, at the end of the closure, serves the same purpose
1925 as that for @FUN_STATIC@. The pointers in the static constructor can point
1926 only to other static closures.
1928 The static object link occurs last in the closure so that static
1929 constructors can store their data fields in exactly the same place as
1930 dynamic constructors.
1932 \item @CONSTR_NOCAF_STATIC@. A statically allocated data constructor
1933 that guarantees not to point (directly or indirectly) to any CAF
1934 (\secref{CAF}). This means it does not need a static object
1935 link field. Since we expect that there might be quite a lot of static
1936 constructors this optimisation makes sense. Furthermore, the @NOCAF@
1937 tag allows the compiler to indicate that no CAFs can be reached
1938 anywhere \emph{even indirectly}.
1942 For each data constructor $Con$, two info tables are generated:
1945 \item $Con$@_con_info@ labels $Con$'s dynamic info table,
1946 shared by all dynamic instances of the constructor.
1947 \item $Con$@_static@ labels $Con$'s static info table,
1948 shared by all static instances of the constructor.
1951 Each constructor also has a \emph{constructor function}, which is a
1952 curried function which builds an instance of the constructor. The
1953 constructor function has an info table labelled as @$Con$_info@, and
1954 entry code pointed to by @$Con$_entry@.
1956 Nullary constructors are represented by a single static info table,
1957 which everyone points to. Thus for a nullary constructor we can omit
1958 the dynamic info table and the constructor function.
1960 \subsubsection{Thunks}
1962 \label{sec:THUNK_SELECTOR}
1964 A thunk represents an expression that is not obviously in head normal
1965 form. For example, consider the following top-level definitions:
1967 range = between 1 10
1968 f = \x -> let ys = take x range
1971 Here the right-hand sides of @range@ and @ys@ are both thunks; the former
1972 is static while the latter is dynamic.
1974 The layout of a thunk is the same as that for a function closure.
1975 However, thunks must have a payload of at least @MIN_UPD_SIZE@
1976 words to allow it to be overwritten with a black hole and an
1977 indirection. The compiler may have to add extra non-pointer fields to
1978 satisfy this constraint.
1981 \begin{tabular}{|l|l|l|l|l|}\hline
1982 \emph{Fixed header} & \emph{Pointers} & \emph{Non-pointers} \\ \hline
1986 The layout word in the info table contains the same information as for
1987 function closures; that is, number of pointers and number of
1990 A thunk differs from a function closure in that it can be updated.
1992 There are several forms of thunk:
1996 \item @THUNK@ and $@THUNK_@p@_@np$: vanilla, dynamically allocated
1997 thunks. Dynamic thunks are overwritten with normal indirections
1998 (@IND@), or old generation indirections (@IND_OLDGEN@): see
2001 \item @THUNK_STATIC@. A static thunk is also known as a
2002 \emph{constant applicative form}, or \emph{CAF}. Static thunks are
2003 overwritten with static indirections.
2006 \begin{tabular}{|l|l|}\hline
2007 \emph{Fixed header} & \emph{Static object link}\\ \hline
2011 \item @THUNK_SELECTOR@ is a (dynamically allocated) thunk whose entry
2012 code performs a simple selection operation from a data constructor
2013 drawn from a single-constructor type. For example, the thunk
2015 x = case y of (a,b) -> a
2017 is a selector thunk. A selector thunk is laid out like this:
2020 \begin{tabular}{|l|l|l|l|}\hline
2021 \emph{Fixed header} & \emph{Selectee pointer} \\ \hline
2025 The layout word contains the byte offset of the desired word in the
2026 selectee. Note that this is different from all other thunks.
2028 The garbage collector ``peeks'' at the selectee's tag (in its info
2029 table). If it is evaluated, then it goes ahead and does the
2030 selection, and then behaves just as if the selector thunk was an
2031 indirection to the selected field. If it is not evaluated, it treats
2032 the selector thunk like any other thunk of that shape.
2033 [Implementation notes. Copying: only the evacuate routine needs to be
2034 special. Compacting: only the PRStart (marking) routine needs to be
2037 There is a fixed set of pre-compiled selector thunks built into the
2038 RTS, representing offsets from 0 to @MAX_SPEC_SELECTOR_THUNK@. The
2039 info tables are labelled @__sel_$n$_upd_info@ where $n$ is the offset.
2040 Non-updating versions are also built in, with info tables labelled
2041 @__sel_$n$_noupd_info@.
2045 The only label associated with a thunk is its info table:
2048 \item[$f$@\_info@] is $f$'s info table.
2052 \Subsubsection{Byte-code objects}{BCO}
2054 A Byte-Code Object (BCO) is a container for a a chunk of byte-code,
2055 which can be executed by Hugs. The byte-code represents a
2056 supercombinator in the program: when Hugs compiles a module, it
2057 performs lambda lifting and each resulting supercombinator becomes a
2058 byte-code object in the heap.
2060 BCOs are not updateable; the bytecode compiler represents updatable
2061 thunks using a combination of @AP@s and @BCO@s.
2063 The semantics of BCOs are described in \secref{hugs-heap-objects}. A
2064 BCO has the following structure:
2067 \begin{tabular}{|l|l|l|l|l|l|}
2069 \emph{Fixed Header} & \emph{Layout} & \emph{Offset} & \emph{Size} &
2070 \emph{Literals} & \emph{Byte code} \\
2077 \item The entry code is a static code fragment/info table that returns
2078 to the scheduler to invoke Hugs (\secref{ghc-to-hugs-switch}).
2079 \item \emph{Layout} contains the number of pointer literals in the
2080 \emph{Literals} field.
2081 \item \emph{Offset} is the offset to the byte code from the start of
2083 \item \emph{Size} is the number of words of byte code in the object.
2084 \item \emph{Literals} contains any pointer and non-pointer literals used in
2085 the byte-codes (including jump addresses), pointers first.
2086 \item \emph{Byte code} contains \emph{Size} words of non-pointer byte
2091 \Subsubsection{Partial applications}{PAP}
2093 A partial application (PAP) represents a function applied to too few
2094 arguments. It is only built as a result of updating after an
2095 argument-satisfaction check failure. A PAP has the following shape:
2098 \begin{tabular}{|l|l|l|l|}\hline
2099 \emph{Fixed header} & \emph{No of words of stack} & \emph{Function closure} & \emph{Stack chunk ...} \\ \hline
2103 The ``Stack chunk'' is a copy of the chunk of stack above the update
2104 frame; ``No of words of stack'' tells how many words it consists of.
2105 The function closure is (a pointer to) the closure for the function
2106 whose argument-satisfaction check failed.
2108 In the normal case where a PAP is built as a result of an argument
2109 satisfaction check failure, the stack chunk will just contain
2110 ``pending arguments'', ie. pointers and tagged non-pointers. It may
2111 in fact also contain activation records, but not update frames, seq
2112 frames, or catch frames. The reason is the garbage collector uses the
2113 same code to scavenge a stack as it does to scavenge the payload of a
2114 PAP, but an update frame contains a link to the next update frame in
2115 the chain and this link would need to be relocated during garbage
2116 collection. Revertible black holes and asynchronous exceptions use
2117 the more general form of PAPs (see Section \ref{revertible-bh}).
2119 There is just one standard form of PAP. There is just one info table
2120 too, called @PAP_info@. Its entry code simply copies the arg stack
2121 chunk back on top of the stack and enters the function closure. (It
2122 has to do a stack overflow test first.)
2124 There is just one way to build a PAP: by calling @stg_update_PAP@ with
2125 the function closure in register @R1@ and the pending arguments on the
2126 stack. The @stg_update_PAP@ function will build the PAP, perform the
2127 update, and return to the next activation record on the stack. If
2128 there are \emph{no} pending arguments on the stack, then no PAP need
2129 be built: in this case @stg_update_PAP@ just overwrites the updatee
2130 with an indirection to the function closure.
2132 PAPs are also used to implement Hugs functions (where the arguments
2133 are free variables). PAPs generated by Hugs can be static so we need
2134 both @PAP@ and @PAP_STATIC@.
2136 \Subsubsection{\texttt{AP\_UPD} objects}{AP_UPD}
2138 @AP_UPD@ objects are used to represent thunks built by Hugs. The only
2139 distintion between an @AP_UPD@ and a @PAP@ is that an @AP_UPD@ is
2143 \begin{tabular}{|l|l|l|l|}
2145 \emph{Fixed Header} & \emph{No of stack words} & \emph{Function closure} & \emph{Stack chunk} \\
2150 The entry code pushes an update frame, copies the arg stack chunk on
2151 top of the stack, and enters the function closure. (It has to do a
2152 stack overflow test first.)
2154 The ``stack chunk'' is a block of stack not containing update frames,
2155 seq frames or catch frames (just like a PAP). In the case of Hugs,
2156 the stack chunk will contain the free variables of the thunk, and the
2157 function closure is (a pointer to) the closure for the thunk. The
2158 argument stack may be empty if the thunk has no free variables.
2160 \note{Since @AP\_UPD@s are updateable, the @MIN\_UPD\_SIZE@ constraint applies here too.}
2162 \Subsubsection{Indirections}{IND}
2164 Indirection closures just point to other closures. They are introduced
2165 when a thunk is updated to point to its value. The entry code for all
2166 indirections simply enters the closure it points to.
2168 There are several forms of indirection:
2171 \item[@IND@] is the vanilla, dynamically-allocated indirection.
2172 It is removed by the garbage collector. It has the following
2175 \begin{tabular}{|l|l|l|}\hline
2176 \emph{Fixed header} & \emph{Target closure} \\ \hline
2180 An @IND@ only exists in the youngest generation. In older
2181 generations, we have @IND_OLDGEN@s. The update code
2182 (@Upd_frame_$n$_entry@) checks whether the updatee is in the youngest
2183 generation before deciding which kind of indirection to use.
2185 \item[@IND\_OLDGEN@] is the vanilla, dynamically-allocated indirection.
2186 It is removed by the garbage collector. It has the following
2189 \begin{tabular}{|l|l|l|}\hline
2190 \emph{Fixed header} & \emph{Target closure} & \emph{Mutable link field} \\ \hline
2193 It contains a \emph{mutable link field} that is used to string together
2194 mutable objects in each old generation.
2197 For lexical profiling, it is necessary to maintain cost centre
2198 information in an indirection, so ``permanent indirections'' are
2199 retained forever. Otherwise they are just like vanilla indirections.
2200 \note{If a permanent indirection points to another permanent
2201 indirection or a @CONST@ closure, it is possible to elide the indirection
2202 since it will have no effect on the profiler.}
2204 \note{Do we still need @IND@ in the profiling build, or do we just
2205 need @IND@ but its behaviour changes when profiling is on?}
2207 \item[@IND\_OLDGEN\_PERM@]
2208 Just like an @IND_OLDGEN@, but sticks around like an @IND_PERM@.
2210 \item[@IND\_STATIC@] is used for overwriting CAFs when they have been
2211 evaluated. Static indirections are not removed by the garbage
2212 collector; and are statically allocated outside the heap (and should
2213 stay there). Their static object link field is used just as for
2214 @FUN_STATIC@ closures.
2217 \begin{tabular}{|l|l|l|}
2219 \emph{Fixed header} & \emph{Target closure} & \emph{Static link field} \\
2226 \subsubsection{Black holes and blocking queues}
2227 \label{sec:BLACKHOLE}
2228 \label{sec:BLACKHOLE_BQ}
2230 Black hole closures are used to overwrite closures currently being
2231 evaluated. They inform the garbage collector that there are no live
2232 roots in the closure, thus removing a potential space leak.
2234 Black holes also become synchronization points in the concurrent
2235 world. When a thread attempts to enter a blackhole, it must wait for
2236 the result of the computation, which is presumably in progress in
2239 \note{In a single-threaded system, entering a black hole indicates an
2240 infinite loop. In a concurrent system, entering a black hole
2241 indicates an infinite loop only if the hole is being entered by the
2242 same thread that originally entered the closure. It could also bring
2243 about a deadlock situation where several threads are waiting
2244 circularly on computations in progress.}
2246 There are two types of black hole:
2251 A straightforward blackhole just consists of an info pointer and some
2252 padding to allow updating with an @IND_OLDGEN@ if necessary. This
2253 type of blackhole has no waiting threads.
2256 \begin{tabular}{|l|l|l|}
2258 \emph{Fixed header} & \emph{Padding} & \emph{Padding} \\
2263 If we're doing \emph{eager blackholing} then a thunk's info pointer is
2264 overwritten with @BLACKHOLE_info@ at the time of entry; hence the need
2265 for blackholes to be small, otherwise we'd be overwriting part of the
2268 \item[@BLACKHOLE\_BQ@]
2269 When a thread enters a @BLACKHOLE@, it is turned into a @BLACKHOLE_BQ@
2270 (blocking queue), which contains a linked list of blocked threads in
2271 addition to the info pointer.
2274 \begin{tabular}{|l|l|l|}
2276 \emph{Fixed header} & \emph{Blocked thread link} & \emph{Mutable link field} \\
2281 The \emph{Blocked thread link} points to the TSO of the first thread
2282 waiting for the value of this thunk. All subsequent TSOs in the list
2283 are linked together using their @tso->link@ field, ending in
2284 @END_TSO_QUEUE_closure@.
2286 Because new threads can be added to the \emph{Blocked thread link}, a
2287 blocking queue is \emph{mutable}, so we need a mutable link field in
2288 order to chain it on to a mutable list for the generational garbage
2293 \Subsubsection{FetchMes}{FETCHME}
2295 In the parallel systems, FetchMes are used to represent pointers into
2296 the global heap. When evaluated, the value they point to is read from
2299 \ToDo{Describe layout}
2301 Because there may be offsets into these arrays, a primitive array
2302 cannot be handled as a FetchMe in the parallel system, but must be
2303 shipped in its entirety if its parent closure is shipped.
2307 \Subsection{Unpointed Objects}{unpointed-objects}
2309 A variable of unpointed type is always bound to a \emph{value}, never
2310 to a \emph{thunk}. For this reason, unpointed objects cannot be
2313 \subsubsection{Immutable objects}
2314 \label{sec:ARR_WORDS}
2317 \item[@ARR\_WORDS@] is a variable-sized object consisting solely of
2318 non-pointers. It is used for arrays of all sorts of things (bytes,
2319 words, floats, doubles... it doesn't matter).
2321 Strictly speaking, an @ARR_WORDS@ could be mutable, but because it
2322 only contains non-pointers we don't need to track this fact.
2325 \begin{tabular}{|c|c|c|c|}
2327 \emph{Fixed Hdr} & \emph{No of non-pointers} & \emph{Non-pointers\ldots} \\ \hline
2332 \subsubsection{Mutable objects}
2333 \label{sec:mutables}
2335 \label{sec:MUT_ARR_PTRS}
2336 \label{sec:MUT_ARR_PTRS_FROZEN}
2339 Some of these objects are \emph{mutable}; they represent objects which
2340 are explicitly mutated by Haskell code through the @ST@ or @IO@
2341 monads. They're not used for thunks which are updated precisely once.
2342 Depending on the garbage collector, mutable closures may contain extra
2343 header information which allows a generational collector to implement
2344 the ``write barrier.''
2346 Notice that mutable objects all have the same general layout: there is
2347 a mutable link field as the second word after the header. This is so
2348 that code to process old-generation mutable lists doesn't need to look
2349 at the type of the object to determine where its link field is.
2353 \item[@MUT\_VAR@] is a mutable variable.
2355 \begin{tabular}{|c|c|c|}
2357 \emph{Fixed Hdr} \emph{Pointer} & \emph{Mutable link} & \\ \hline
2361 \item[@MUT\_ARR\_PTRS@] is a mutable array of pointers. Such an array
2362 may be \emph{frozen}, becoming an @MUT_ARR_PTRS_FROZEN@, with a
2363 different info-table.
2366 \begin{tabular}{|c|c|c|c|}
2368 \emph{Fixed Hdr} & \emph{No of ptrs} & \emph{Mutable link} & \emph{Pointers\ldots} \\ \hline
2372 \item[@MUT\_ARR\_PTRS\_FROZEN@] This is the immutable version of
2373 @MUT_ARR_PTRS@. It still has a mutable link field for two reasons: we
2374 need to keep it on the mutable list for an old generation at least
2375 until the next garbage collection, and it may become mutable again via
2379 \begin{tabular}{|c|c|c|c|}
2381 \emph{Fixed Hdr} & \emph{No of ptrs} & \emph{Mutable link} & \emph{Pointers\ldots} \\ \hline
2388 \begin{tabular}{|l|l|l|l|l|}
2390 \emph{Fixed header} & \emph{Head} & \emph{Mutable link} & \emph{Tail}
2401 \Subsubsection{Foreign objects}{FOREIGN}
2403 Here's what a ForeignObj looks like:
2406 \begin{tabular}{|l|l|l|l|}
2408 \emph{Fixed header} & \emph{Data} \\
2413 A foreign object is simple a boxed pointer to an address outside the
2414 Haskell heap, possible to @malloc@ed data. The only reason foreign
2415 objects exist is so that we can track the lifetime of one using weak
2416 pointers (see \secref{WEAK}) and run a finaliser when the foreign
2417 object is unreachable.
2419 \subsubsection{Weak pointers}
2423 \begin{tabular}{|l|l|l|l|l|}
2425 \emph{Fixed header} & \emph{Key} & \emph{Value} & \emph{Finaliser}
2431 \ToDo{Weak poitners}
2433 \subsubsection{Stable names}
2434 \label{sec:STABLE_NAME}
2437 \begin{tabular}{|l|l|l|l|}
2439 \emph{Fixed header} & \emph{Index} \\
2446 The remaining objects types are all administrative --- none of them
2449 \subsection{Other weird objects}
2451 \label{sec:BLOCKED_FETCH}
2454 \item[@BlockedFetch@ heap objects (`closures')] (parallel only)
2456 @BlockedFetch@s are inbound fetch messages blocked on local closures.
2457 They arise as entries in a local blocking queue when a fetch has been
2458 received for a local black hole. When awakened, we look at their
2459 contents to figure out where to send a resume.
2461 A @BlockedFetch@ closure has the form:
2463 \begin{tabular}{|l|l|l|l|l|l|}\hline
2464 \emph{Fixed header} & link & node & gtid & slot & weight \\ \hline
2468 \item[Spark Closures] (parallel only)
2470 Spark closures are used to link together all closures in the spark pool. When
2471 the current processor is idle, it may choose to speculatively evaluate some of
2472 the closures in the pool. It may also choose to delete sparks from the pool.
2474 \begin{tabular}{|l|l|l|l|l|l|}\hline
2475 \emph{Fixed header} & \emph{Spark pool link} & \emph{Sparked closure} \\ \hline
2479 \item[Slop Objects]\label{sec:slop-objects}
2481 Slop objects are used to overwrite the end of an updatee if it is
2482 larger than an indirection. Normal slop objects consist of an info
2483 pointer a size word and a number of slop words.
2486 \begin{tabular}{|l|l|l|l|l|l|}\hline
2487 \emph{Info Pointer} & \emph{Size} & \emph{Slop Words} \\ \hline
2491 This is too large for single word slop objects which consist of a
2494 Note that slop objects only contain an info pointer, not a standard
2495 fixed header. This doesn't cause problems because slop objects are
2496 always unreachable --- they can only be accessed by linearly scanning
2499 \note{Currently we don't use slop objects because the storage manager
2500 isn't reliant on objects being adjacent, but if we move to a ``mostly
2501 copying'' style collector, this will become an issue.}
2505 \Subsection{Thread State Objects (TSOs)}{TSO}
2507 In the multi-threaded system, the state of a suspended thread is
2508 packed up into a Thread State Object (TSO) which contains all the
2509 information needed to restart the thread and for the garbage collector
2510 to find all reachable objects. When a thread is running, it may be
2511 ``unpacked'' into machine registers and various other memory locations
2512 to provide faster access.
2514 Single-threaded systems don't really \emph{need\/} TSOs --- but they do
2515 need some way to tell the storage manager about live roots so it is
2516 convenient to use a single TSO to store the mutator state even in
2517 single-threaded systems.
2519 Rather than manage TSOs' alloc/dealloc, etc., in some \emph{ad hoc}
2520 way, we instead alloc/dealloc/etc them in the heap; then we can use
2521 all the standard garbage-collection/fetching/flushing/etc machinery on
2522 them. So that's why TSOs are ``heap objects,'' albeit very special
2525 \begin{tabular}{|l|l|}
2526 \hline \emph{Fixed header}
2527 \\ \hline \emph{Link field}
2528 \\ \hline \emph{Mutable link field}
2529 \\ \hline \emph{What next}
2530 \\ \hline \emph{State}
2531 \\ \hline \emph{Thread Id}
2532 \\ \hline \emph{Exception Handlers}
2533 \\ \hline \emph{Ticky Info}
2534 \\ \hline \emph{Profiling Info}
2535 \\ \hline \emph{Parallel Info}
2536 \\ \hline \emph{GranSim Info}
2537 \\ \hline \emph{Stack size}
2538 \\ \hline \emph{Max Stack size}
2541 \\ \hline \emph{SpLim}
2549 The contents of a TSO are:
2552 \item[\emph{Link field}] This is a pointer used to maintain a list of
2553 threads with a similar state (e.g.~all runnable, all sleeping, all
2554 blocked on the same black hole, all blocked on the same MVar,
2557 \item[\emph{Mutable link field}] Because the stack is mutable by
2558 definition, the generational collector needs to track TSOs in older
2559 generations that may point into younger ones (which is just about any
2560 TSO for a thread that has run recently). Hence the need for a mutable
2561 link field (see \secref{mutables}).
2563 \item[\emph{What next}]
2564 This field has five values:
2566 \item[@ThreadEnterGHC@] The thread can be started by entering the
2567 closure pointed to by the word on the top of the stack.
2568 \item[@ThreadRunGHC@] The thread can be started by jumping to the
2569 address on the top of the stack.
2570 \item[@ThreadEnterHugs@] The stack has a pointer to a Hugs-built
2571 closure on top of the stack: enter the closure to run the thread.
2572 \item[@ThreadKilled@] The thread has been killed (by @killThread#@).
2573 It is probably still around because it is on some queue somewhere and
2574 hasn't been garbage collected yet.
2575 \item[@ThreadComplete@] The thread has finished. Its @TSO@ hasn't
2576 been garbage collected yet.
2579 \item[\emph{Thread Id}]
2580 This field contains a (not necessarily unique) integer that identifies
2581 the thread. It can be used eg. for hashing.
2583 \item[\emph{Ticky Info}] Optional information for ``Ticky Ticky''
2584 statistics: @TSO_STK_HWM@ is the maximum number of words allocated to
2587 \item[\emph{Profiling Info}] Optional information for profiling:
2588 @TSO_CCC@ is the current cost centre.
2590 \item[\emph{Parallel Info}]
2591 Optional information for parallel execution.
2595 % \item The types of threads (@TSO_TYPE@):
2596 % \begin{description}
2597 % \item[@T_MAIN@] Must be executed locally.
2598 % \item[@T_REQUIRED@] A required thread -- may be exported.
2599 % \item[@T_ADVISORY@] An advisory thread -- may be exported.
2600 % \item[@T_FAIL@] A failure thread -- may be exported.
2603 % \item I've no idea what else
2607 \item[\emph{GranSim Info}]
2608 Optional information for gransim execution.
2610 % \item Optional information for GranSim execution:
2616 % \item basic blocks
2623 % \item global sparks
2624 % \item local sparks
2627 % \item clock (gransim light only)
2631 % Here are the various queues for GrAnSim-type events.
2640 \item[\emph{Stack Info}] Various fields contain information on the
2641 stack: its current size, its maximum size (to avoid infinite loops
2642 overflowing the memory), the current stack pointer (\emph{Sp}), the
2643 current stack update frame pointer (\emph{Su}), and the stack limit
2644 (\emph{SpLim}). The latter three fields are loaded into the relevant
2645 registers when the thread is run.
2647 \item[\emph{Stack}] This is the actual stack for the thread,
2648 \emph{Stack size} words long. It grows downwards from higher
2649 addresses to lower addresses. When the stack overflows, it will
2650 generally be relocated into larger premises unless \emph{Max stack
2655 The garbage collector needs to be able to find all the
2656 pointers in a stack. How does it do this?
2660 \item Within the stack there are return addresses, pushed
2661 by @case@ expressions. Below a return address (i.e. at higher
2662 memory addresses, since the stack grows downwards) is a chunk
2663 of stack that the return address ``knows about'', namely the
2664 activation record of the currently running function.
2666 \item Below each such activation record is a \emph{pending-argument
2667 section}, a chunk of
2668 zero or more words that are the arguments to which the result
2669 of the function should be applied. The return address does not
2671 ``know'' how many pending arguments there are, or their types.
2672 (For example, the function might return a result of type $\alpha$.)
2674 \item Below each pending-argument section is another return address,
2675 and so on. Actually, there might be an update frame instead, but we
2676 can consider update frames as a special case of a return address with
2677 a well-defined activation record.
2681 The game plan is this. The garbage collector walks the stack from the
2682 top, traversing pending-argument sections and activation records
2683 alternately. Next we discuss how it finds the pointers in each of
2684 these two stack regions.
2687 \Subsubsection{Activation records}{activation-records}
2689 An \emph{activation record} is a contiguous chunk of stack,
2690 with a return address as its first word, followed by as many
2691 data words as the return address ``knows about''. The return
2692 address is actually a fully-fledged info pointer. It points
2693 to an info table, replete with:
2696 \item entry code (i.e. the code to return to).
2698 \item closure type is either @RET_SMALL/RET_VEC_SMALL@ or
2699 @RET_BIG/RET_VEC_BIG@, depending on whether the activation record has
2700 more than 32 data words (\note{64 for 8-byte-word architectures}) and
2701 on whether to use a direct or a vectored return.
2703 \item the layout info for @RET_SMALL@ is a bitmap telling the layout
2704 of the activation record, one bit per word. The least-significant bit
2705 describes the first data word of the record (adjacent to the fixed
2706 header) and so on. A ``@1@'' indicates a non-pointer, a ``@0@''
2707 indicates a pointer. We don't need to indicate exactly how many words
2708 there are, because when we get to all zeros we can treat the rest of
2709 the activation record as part of the next pending-argument region.
2711 For @RET_BIG@ the layout field points to a block of bitmap words,
2712 starting with a word that tells how many words are in the block.
2714 \item the info table contains a Static Reference Table pointer for the
2715 return address (\secref{srt}).
2718 The activation record is a fully fledged closure too. As well as an
2719 info pointer, it has all the other attributes of a fixed header
2720 (\secref{fixed-header}) including a saved cost centre which
2721 is reloaded when the return address is entered.
2723 In other words, all the attributes of closures are needed for
2724 activation records, so it's very convenient to make them look alike.
2727 \Subsubsection{Pending arguments}{pending-args}
2729 So that the garbage collector can correctly identify pointers in
2730 pending-argument sections we explicitly tag all non-pointers. Every
2731 non-pointer in a pending-argument section is preceded (at the next
2732 lower memory word) by a one-word byte count that says how many bytes
2733 to skip over (excluding the tag word).
2735 The garbage collector traverses a pending argument section from the
2736 top (i.e. lowest memory address). It looks at each word in turn:
2739 \item If it is less than or equal to a small constant @ARGTAG_MAX@
2740 then it treats it as a tag heralding zero or more words of
2741 non-pointers, so it just skips over them.
2743 \item If it points to the code segment, it must be a return
2744 address, so we have come to the end of the pending-argument section.
2746 \item Otherwise it must be a bona fide heap pointer.
2750 \Subsection{The Stable Pointer Table}{STABLEPTR_TABLE}
2752 A stable pointer is a name for a Haskell object which can be passed to
2753 the external world. It is ``stable'' in the sense that the name does
2754 not change when the Haskell garbage collector runs---in contrast to
2755 the address of the object which may well change.
2757 A stable pointer is represented by an index into the
2758 @StablePointerTable@. The Haskell garbage collector treats the
2759 @StablePointerTable@ as a source of roots for GC.
2761 In order to provide efficient access to stable pointers and to be able
2762 to cope with any number of stable pointers (eg $0 \ldots 100000$), the
2763 table of stable pointers is an array stored on the heap and can grow
2764 when it overflows. (Since we cannot compact the table by moving
2765 stable pointers about, it seems unlikely that a half-empty table can
2766 be reduced in size---this could be fixed if necessary by using a
2767 hash table of some sort.)
2769 In general a stable pointer table closure looks like this:
2772 \begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|}
2774 \emph{Fixed header} & \emph{No of pointers} & \emph{Free} & $SP_0$ & \ldots & $SP_{n-1}$
2782 \item[@NPtrs@:] number of (stable) pointers.
2784 \item[@Free@:] the byte offset (from the first byte of the object) of the first free stable pointer.
2786 \item[$SP_i$:] A stable pointer slot. If this entry is in use, it is
2787 an ``unstable'' pointer to a closure. If this entry is not in use, it
2788 is a byte offset of the next free stable pointer slot.
2792 When a stable pointer table is evacuated
2794 \item the free list entries are all set to @NULL@ so that the evacuation
2795 code knows they're not pointers;
2797 \item The stable pointer slots are scanned linearly: non-@NULL@ slots
2798 are evacuated and @NULL@-values are chained together to form a new free list.
2801 There's no need to link the stable pointer table onto the mutable
2802 list because we always treat it as a root.
2804 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2805 \Subsection{Garbage Collecting CAFs}{CAF}
2806 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2808 % begin{direct quote from current paper}
2809 A CAF (constant applicative form) is a top-level expression with no
2810 arguments. The expression may need a large, even unbounded, amount of
2811 storage when it is fully evaluated.
2813 CAFs are represented by closures in static memory that are updated
2814 with indirections to objects in the heap space once the expression is
2815 evaluated. Previous version of GHC maintained a list of all evaluated
2816 CAFs and traversed them during GC, the result being that the storage
2817 allocated by a CAF would reside in the heap until the program ended.
2818 % end{direct quote from current paper}
2820 % begin{elaboration on why CAFs are very very bad}
2821 Treating CAFs this way has two problems:
2824 It can cause a very large space leak. For example, this program
2825 should run in constant space but, instead, will run out of memory.
2831 > nats = [0..maxInt]
2835 Expressions with no arguments have very different space behaviour
2836 depending on whether or not they occur at the top level. For example,
2837 if we make \verb+nats+ a local definition, the space leak goes away
2838 and the resulting program runs in constant space, as expected.
2844 > nats = [0..maxInt]
2847 This huge change in the operational behaviour of the program
2848 is a problem for optimising compilers and for programmers.
2849 For example, GHC will normally flatten a set of let bindings using
2850 this transformation:
2852 let x1 = let x2 = e2 in e1 ==> let x2 = e2 in let x1 = e1
2854 but it does not do so if this would raise \verb+x2+ to the top level
2855 since that may create a CAF. Many Haskell programmers avoid creating
2856 large CAFs by adding a dummy argument to a CAF or by moving a CAF away
2860 % end{elaboration on why CAFs are very very bad}
2862 Solving the CAF problem requires different treatment in interactive
2863 systems such as Hugs than in batch-mode systems such as GHC
2866 In a batch-mode the program the runtime system is terminated
2867 after every execution of the runtime system. In such systems,
2868 the garbage collector can completely ``destroy'' a CAF when it
2869 is no longer live --- in much the same way as it ``destroys''
2870 normal closures when they are no longer live.
2873 In an interactive system, many expressions are evaluated without
2874 restarting the runtime system between each evaluation. In such
2875 systems, the garbage collector cannot completely ``destroy'' a CAF
2876 when it is no longer live because, whilst it might not be required in
2877 the evaluation of the current expression, it might be required in the
2880 There are two possible behaviours we might want:
2883 When a CAF is no longer required for the current evaluation, the CAF
2884 should be reverted to its original form. This behaviour ensures that
2885 the operational behaviour of the interactive system is a reasonable
2886 predictor of the operational behaviour of the batch-mode system. This
2887 allows us to use Hugs for performance debugging (in particular, trying
2888 to understand and reduce the heap usage of a program) --- an area of
2889 increasing importance as Haskell is used more and more to solve ``real
2890 problems'' in ``real problem domains''.
2893 Even if a CAF is no longer required for the current evaluation, we might
2894 choose to hang onto it by collecting it in the normal way. This keeps
2895 the space leak but might be useful in a teaching environment when
2896 trying to teach the difference between call by name evaluation (which
2897 doesn't share work) and lazy evaluation (which does share work).
2901 It turns out that it is easy to support both styles of use, so the
2902 runtime system provides a switch which lets us turn this on and off
2903 during execution. \ToDo{What is this switch called?} It would also
2904 be easy to provide a function \verb+RevertCAF+ to let the interpreter
2905 revert any CAF it wanted between (but not during) executions, if we so
2906 desired. Running \verb+RevertCAF+ during execution would lose some sharing
2907 but is otherwise harmless.
2911 % % begin{even more pointless observation?}
2912 % The simplest fix would be to remove the special treatment of
2913 % top level variables. This works but is very inefficient.
2915 % (Note: delete this paragraph from final version.)
2916 % % end{even more pointless observation?}
2918 % begin{pointless observation?}
2919 An easy but inefficient fix to the CAF problem would be to make a
2920 complete copy of the heap before every evaluation and discard the copy
2921 after evaluation. This works but is inefficient.
2922 % end{pointless observation?}
2924 An efficient way to achieve a similar effect is to revert all
2925 updatable thunks to their original form as they become unnecessary for
2926 the current evaluation. To do this, we modify the compiler to ensure
2927 that the only updatable thunks generated by the compiler are CAFs and
2928 we modify the garbage collector to revert entered CAFs to unentered
2929 CAFs as their value becomes unnecessary.
2932 \subsubsection{New Heap Objects}
2934 We add three new kinds of heap object: unentered CAF closures, entered
2935 CAF objects and CAF blackholes. We first describe how they are
2936 evaluated and then how they are garbage collected.
2939 Unentered CAF closures contain a pointer to closure representing the
2940 body of the CAF. The ``body closure'' is not updatable.
2942 Unentered CAF closures contain two unused fields to make them the same
2943 size as entered CAF closures --- which allows us to perform an inplace
2944 update. \ToDo{Do we have to add another kind of inplace update operation
2945 to the storage manager interface or do we consider this to be internal
2948 \begin{tabular}{|l|l|l|l|}\hline
2949 \verb+CAF_unentered+ & \emph{body closure} & \emph{unused} & \emph{unused} \\ \hline
2952 When an unentered CAF is entered, we do the following:
2955 allocate a CAF black hole;
2958 push an update frame (to update the CAF black hole) onto the stack;
2961 overwrite the CAF with an entered CAF object (see below) with the same
2962 body and whose value field points to the black hole;
2965 add the CAF to a list of all entered CAFs (called ``the CAF list'');
2969 the closure representing the value of the CAF is entered.
2973 When evaluation of the CAF body returns a value, the update frame
2974 causes the CAF black hole to be updated with the value in the normal
2977 \ToDo{Add a picture}
2980 Entered CAF closures contain two pointers: a pointer to the CAF body
2981 (the same as for unentered CAF closures); a pointer to the CAF value
2982 (this is initialised with a CAF blackhole, as previously described);
2983 and a link to the next CAF in the CAF list
2985 \ToDo{How is the end of the list marked? Null pointer or sentinel value?}.
2988 \begin{tabular}{|l|l|l|l|}\hline
2989 \verb+CAF_entered+ & \emph{body closure} & \emph{value} & \emph{link} \\ \hline
2992 When an entered CAF is entered, it enters its value closure.
2995 CAF blackholes are identical to normal blackholes except that they
2996 have a different infotable. The only reason for having CAF blackholes
2997 is to allow an optimisation of lazy blackholing where we stop scanning
2998 the stack when we see the first {\em normal blackhole} but not
2999 when we see a {\em CAF blackhole.}
3000 \ToDo{The optimisation we want to allow should be described elsewhere
3001 so that all we have to do here is describe the difference.}
3003 Instead of allocating a blackhole to update with the value of the CAF,
3004 it might seem simpler to update the CAF directly. This would require
3005 a new kind of update frame which would update the value field of the
3006 CAF with a pointer to the value and wouldn't catch blackholes caused
3007 by CAFs that depend on themselves so we chose not to do so.
3011 \subsubsection{Garbage Collection}
3013 To avoid the space leak, each run of the garbage collector must revert
3014 the entered CAFs which are not required to complete the current
3015 evaluation (that is all the closures reachable from the set of
3016 runnable threads and the stable pointer table).
3018 It does this by performing garbage collection in three phases:
3021 During the first phase, we ``mark'' all closures reachable from the
3024 How we ``mark'' closures depends on the garbage collector. For
3025 example, in a 2-space collector, closures are ``marked'' by copying
3026 them into ``to-space'', overwriting them with a forwarding node and
3027 ``marking'' all the closures reachable from the copy. The only
3028 requirements are that we can test whether a closure is marked and if a
3029 closure is marked then so are all closures reachable from it.
3031 \ToDo{At present we say that the scheduler state includes any state
3032 that Hugs may have. This is not true anymore.}
3034 Performing this phase first provides us with a cheap test for
3035 execution closures: at this stage in execution, the execution closures
3036 are precisely the marked closures.
3039 During the second phase, we revert all unmarked CAFs on the CAF list
3040 and remove them from the CAF list.
3042 Since the CAF list is exactly the set of all entered CAFs, this reverts
3043 all entered CAFs which are not execution closures.
3046 During the third phase, we mark all top level objects (including CAFs)
3047 by calling \verb+MarkHugsRoots+ which will call \verb+MarkRoot+ for
3048 each top level object known to Hugs.
3052 To implement the second style of interactive behaviour (where we
3053 deliberately keep the CAF-related space leak), we simply omit the
3054 second phase. Omitting the second phase causes the third phase to
3055 mark any unmarked CAF value closures.
3057 So far, we have been describing a pure Hugs system which contains no
3058 machine generated code. The main difference in a hybrid system is
3059 that GHC-generated code is statically allocated in memory instead of
3060 being dynamically allocated on the heap. We split both
3061 \verb+CAF_unentered+ and \verb+CAF_entered+ into two versions: a
3062 static and a dynamic version. The static and dynamic versions of each
3063 CAF differ only in whether they are moved during garbage collection.
3064 When reverting CAFs, we revert dynamic entered CAFs to dynamic
3065 unentered CAFs and static entered CAFs to static unentered CAFs.
3070 \Section{The Bytecode Evaluator}{bytecode-evaluator}
3072 This section describes how the Hugs interpreter interprets code in the
3073 same environment as compiled code executes. Both evaluation models
3074 use a common garbage collector, so they must agree on the form of
3075 objects in the heap.
3077 Hugs interprets code by converting it to byte-code and applying a
3078 byte-code interpreter to it. Wherever possible, we try to ensure that
3079 the byte-code is all that is required to interpret a section of code.
3080 This means not dynamically generating info tables, and hence we can
3081 only have a small number of possible heap objects each with a statically
3082 compiled info table. Similarly for stack objects: in fact we only
3083 have one Hugs stack object, in which all information is tagged for the
3086 There is, however, one exception to this rule. Hugs must generate
3087 info tables for any constructors it is asked to compile, since the
3088 alternative is to force a context-switch each time compiled code
3089 enters a Hugs-built constructor, which would be prohibitively
3092 We achieve this simplicity by forgoing some of the optimisations used
3097 Whereas compiled code has five different ways of entering a closure
3098 (\secref{ghc-fun-call}), interpreted code has only one.
3099 The entry point for interpreted code behaves like slow entry points for
3104 We use just one info table for \emph{all\/} direct returns.
3105 This introduces two problems:
3107 \item How does the interpreter know what code to execute?
3109 Instead of pushing just a return address, we push a return BCO and a
3110 trivial return address which just enters the return BCO.
3112 (In a purely interpreted system, we could avoid pushing the trivial
3115 \item How can the garbage collector follow pointers within the
3118 We could push a third word ---a bitmask describing the location of any
3119 pointers within the record--- but, since we're already tagging unboxed
3120 function arguments on the stack, we use the same mechanism for unboxed
3121 values within the activation record.
3123 \ToDo{Do we have to stub out dead variables in the activation frame?}
3129 We trivially support vectored returns by pushing a return vector whose
3130 entries are all the same.
3134 We avoid the need to build SRTs by putting bytecode objects on the
3135 heap and restricting BCOs to a single basic block.
3139 \Subsection{Hugs Info Tables}{hugs-info-tables}
3141 Hugs requires the following info tables and closures:
3143 \item [@HUGS\_RET@].
3145 Contains both a vectored return table and a direct entry point. All
3146 entry points are the same: they rearrange the stack to match the Hugs
3147 return convention (\secref{hugs-return-convention}) and return to the
3148 scheduler. When the scheduler restarts the thread, it will find a BCO
3149 on top of the stack and will enter the Hugs interpreter.
3153 This is just the standard info table for an update frame.
3155 \item [Constructors].
3157 The entry code for a constructor jumps to a generic entry point in the
3158 runtime system which decides whether to do a vectored or unvectored
3159 return depending on the shape of the constructor/type. This implies that
3160 info tables must have enough info to make that decision.
3162 \item [@AP@ and @PAP@].
3164 \item [Indirections].
3168 Hugs doesn't generate them itself but it ought to recognise them
3170 \item [Complex primops].
3172 Some of the primops are too complex for GHC to generate inline.
3173 Instead, these primops are hand-written and called as normal functions.
3174 Hugs only needs to know their names and types but doesn't care whether
3175 they are generated by GHC or by hand. Two things to watch:
3179 Hugs must be able to enter these primops even if it is working on a
3180 standalone system that does not support genuine GHC generated code.
3182 \item The complex primops often involve unboxed tuple types (which
3183 Hugs does not support at the source level) so we cannot specify their
3184 types in a Haskell source file.
3190 \Subsection{Hugs Heap Objects}{hugs-heap-objects}
3192 \subsubsection{Byte-code objects}
3194 Compiled byte code lives on the global heap, in objects called
3195 Byte-Code Objects (or BCOs). The layout of BCOs is described in
3196 detail in \secref{BCO}, in this section we will describe
3199 Since byte-code lives on the heap, it can be garbage collected just
3200 like any other heap-resident data. Hugs arranges that any BCO's
3201 referred to by the Hugs symbol tables are treated as live objects by
3202 the garbage collector. When a module is unloaded, the pointers to its
3203 BCOs are removed from the symbol table, and the code will be garbage
3204 collected some time later.
3206 A BCO represents a basic block of code --- the (only) entry points is
3207 at the beginning of a BCO, and it is impossible to jump into the
3208 middle of one. A BCO represents not only the code for a function, but
3209 also its closure; a BCO can be entered just like any other closure.
3210 Hugs performs lambda-lifting during compilation to byte-code, and each
3211 top-level combinator becomes a BCO in the heap.
3214 \subsubsection{Thunks and partial applications}
3216 A thunk consists of a code pointer, and values for the free variables
3217 of that code. Since Hugs byte-code is lambda-lifted, free variables
3218 become arguments and are expected to be on the stack by the called
3221 Hugs represents updateable thunks with @AP_UPD@ objects applying a closure
3222 to a list of arguments. (As for @PAP@s, unboxed arguments should be
3223 preceded by a tag.) When it is entered, it pushes an update frame
3224 followed by its payload on the stack, and enters the first word (which
3225 will be a pointer to a BCO). The layout of @AP_UPD@ objects is described
3226 in more detail in \secref{AP_UPD}.
3228 Partial applications are represented by @PAP@ objects, which are
3231 \ToDo{Hugs Constructors}.
3233 \Subsection{Calling conventions}{hugs-calling-conventions}
3235 The calling convention for any byte-code function is straightforward:
3237 \item Push any arguments on the stack.
3238 \item Push a pointer to the BCO.
3239 \item Begin interpreting the byte code.
3242 In a system containing both GHC and Hugs, the bytecode interpreter
3243 only has to be able to enter BCOs: everything else can be handled by
3244 returning to the compiled world (as described in
3245 \secref{hugs-to-ghc-switch}) and entering the closure
3248 This would work but it would obviously be very inefficient if we
3249 entered a @AP@ by switching worlds, entering the @AP@, pushing the
3250 arguments and function onto the stack, and entering the function
3251 which, likely as not, will be a byte-code object which we will enter
3252 by \emph{returning} to the byte-code interpreter. To avoid such
3253 gratuitious world switching, we choose to recognise certain closure
3254 types as being ``standard'' --- and duplicate the entry code for the
3255 ``standard closures'' in the bytecode interpreter.
3257 A closure is said to be ``standard'' if its entry code is entirely
3258 determined by its info table. \emph{Standard Closures} have the
3259 desirable property that the byte-code interpreter can enter the
3260 closure by simply ``interpreting'' the info table instead of switching
3261 to the compiled world. The standard closures include:
3264 \item[Constructor] To enter a constructor, we simply return (see
3265 \secref{hugs-return-convention}).
3268 To enter an indirection, we simply enter the object it points to
3269 after possibly adjusting the current cost centre.
3273 To enter an @AP@, we push an update frame, push the
3274 arguments, push the function and enter the function.
3275 (Not forgetting a stack check at the start.)
3279 To enter a @PAP@, we push the arguments, push the function and enter
3280 the function. (Not forgetting a stack check at the start.)
3284 To enter a selector (\secref{THUNK_SELECTOR}), we test whether the
3285 selectee is a value. If so, we simply select the appropriate
3286 component; if not, it's simplest to treat it as a GHC-built closure
3287 --- though we could interpret it if we wanted.
3291 The most obvious omissions from the above list are @BCO@s (which we
3292 dealt with above) and GHC-built closures (which are covered in
3293 \secref{hugs-to-ghc-switch}).
3296 \Subsection{Return convention}{hugs-return-convention}
3298 When Hugs pushes a return address, it pushes both a pointer to the BCO
3299 to return to, and a pointer to a static code fragment @HUGS_RET@ (this
3300 is described in \secref{ghc-to-hugs-switch}). The
3301 stack layout is shown in \figref{hugs-return-stack}.
3313 %\input{hugs_ret.pstex_t}
3315 \caption{Stack layout for a Hugs return address}
3316 \label{fig:hugs-return-stack}
3317 % this figure apparently duplicates {fig:hugs-return-stack1} earlier.
3328 %\input{hugs_ret2.pstex_t}
3330 \caption{Stack layout on enterings a Hugs return address}
3331 \label{fig:hugs-return2}
3344 %\input{hugs_ret2.pstex_t}
3346 \caption{Stack layout on entering a Hugs return address with an unboxed value}
3347 \label{fig:hugs-return-int1}
3360 %\input{hugs_ret3.pstex_t}
3362 \caption{Stack layout on enterings a GHC return address}
3363 \label{fig:hugs-return3}
3377 | restart |--> id_Int#_closure
3380 %\input{hugs_ret2.pstex_t}
3382 \caption{Stack layout on enterings a GHC return address with an unboxed value}
3383 \label{fig:hugs-return-int}
3386 When a Hugs byte-code sequence enters a closure, it examines the
3387 return address on top of the stack.
3391 \item If the return address is @HUGS_RET@, pop the @HUGS_RET@ and the
3392 bco for the continuation off the stack, push a pointer to the constructor onto
3393 the stack and enter the BCO with the current object pointer set to the BCO
3394 (\figref{hugs-return2}).
3396 \item If the top of the stack is not @HUGS_RET@, we need to do a world
3397 switch as described in \secref{hugs-to-ghc-switch}.
3401 \ToDo{This duplicates what we say about switching worlds
3402 (\secref{switching-worlds}) - kill one or t'other.}
3405 \ToDo{This was in the evaluation model part but it really belongs in
3406 this part which is about the internal details of each of the major
3409 \Subsection{Addressing Modes}{hugs-addressing-modes}
3411 To avoid potential alignment problems and simplify garbage collection,
3412 all literal constants are stored in two tables (one boxed, the other
3413 unboxed) within each BCO and are referred to by offsets into the tables.
3414 Slots in the constant tables are word aligned.
3416 \ToDo{How big can the offsets be? Is the offset specified in the
3417 address field or in the instruction?}
3419 Literals can have the following types: char, int, nat, float, double,
3420 and pointer to boxed object. There is no real difference between
3421 char, int, nat and float since they all occupy 32 bits --- but it
3422 costs almost nothing to distinguish them and may improve portability
3423 and simplify debugging.
3425 \Subsection{Compilation}{hugs-compilation}
3428 \def\is{\mbox{\it is}}
3429 \def\ts{\mbox{\it ts}}
3430 \def\as{\mbox{\it as}}
3431 \def\bs{\mbox{\it bs}}
3432 \def\cs{\mbox{\it cs}}
3433 \def\rs{\mbox{\it rs}}
3434 \def\us{\mbox{\it us}}
3435 \def\vs{\mbox{\it vs}}
3436 \def\ws{\mbox{\it ws}}
3437 \def\xs{\mbox{\it xs}}
3439 \def\e{\mbox{\it e}}
3440 \def\alts{\mbox{\it alts}}
3441 \def\fail{\mbox{\it fail}}
3442 \def\panic{\mbox{\it panic}}
3443 \def\ua{\mbox{\it ua}}
3444 \def\obj{\mbox{\it obj}}
3445 \def\bco{\mbox{\it bco}}
3446 \def\tag{\mbox{\it tag}}
3447 \def\entry{\mbox{\it entry}}
3448 \def\su{\mbox{\it su}}
3450 \def\Ind#1{{\mbox{\it Ind}\ {#1}}}
3451 \def\update#1{{\mbox{\it update}\ {#1}}}
3453 \def\next{$\Longrightarrow$}
3454 \def\append{\mathrel{+\mkern-6mu+}}
3455 \def\reverse{\mbox{\it reverse}}
3456 \def\size#1{{\vert {#1} \vert}}
3457 \def\arity#1{{\mbox{\it arity}{#1}}}
3459 \def\AP{\mbox{\it AP}}
3460 \def\PAP{\mbox{\it PAP}}
3461 \def\GHCRET{\mbox{\it GHCRET}}
3462 \def\GHCOBJ{\mbox{\it GHCOBJ}}
3464 To make sense of the instructions, we need a sense of how they will be
3465 used. Here is a small compiler for the STG language.
3468 > cg (f{a1, ... am}) = do
3469 > pushAtom am; ... pushAtom a1
3473 > cg (let {x1=rhs1; ... xm=rhsm} in e) = do
3474 > ALLOC x1 |rhs1|, ... ALLOC xm |rhsm|
3475 > build x1 rhs1, ... build xm rhsm
3477 > cg (case e of alts) = do
3478 > PUSHALTS (cgAlts alts)
3481 > cgAlts { alt1; ... altm } = cgAlt alt1 $ ... $ cgAlt altm pmFail
3483 > cgAlt (x@C{xs} -> e) fail = do
3485 > HEAPCHECK (heapUse e)
3489 > build x (C{a1, ... am}) = do
3490 > pushUntaggedAtom am; ... pushUntaggedAtom a1
3492 > -- A useful optimisation
3493 > build x ({v1, ... vm} \ {}. f{a1, ... am}) = do
3494 > pushVar am; ... pushVar a1
3497 > build x ({v1, ... vm} \ {}. e) = do
3498 > pushVar vm; ... pushVar v1
3499 > PUSHBCO (cgRhs ({v1, ... vm} \ {}. e))
3501 > build x ({v1, ... vm} \ {x1, ... xm}. e) = do
3502 > pushVar vm; ... pushVar v1
3503 > PUSHBCO (cgRhs ({v1, ... vm} \ {x1, ... xm}. e))
3506 > cgRhs (vs \ xs. e) = do
3507 > ARGCHECK (xs ++ vs) -- can be omitted if xs == {}
3508 > STACKCHECK min(stackUse e,heapOverflowSlop)
3509 > HEAPCHECK (heapUse e)
3512 > pushAtom x = pushVar x
3513 > pushAtom i# = PUSHINT i#
3515 > pushVar x = if isGlobalVar x then PUSHGLOBAL x else PUSHLOCAL x
3517 > pushUntaggedAtom x = pushVar x
3518 > pushUntaggedAtom i# = PUSHUNTAGGEDINT i#
3520 > pushVar x = if isGlobalVar x then PUSHGLOBAL x else PUSHLOCAL x
3523 \ToDo{Is there an easy way to add semi-tagging? Would it be that different?}
3525 \ToDo{Optimise thunks of the form @f{x1,...xm}@ so that we build an AP directly}
3527 \Subsection{Instructions}{hugs-instructions}
3529 We specify the semantics of instructions using transition rules of
3532 \begin{tabular}{|llrrrrr|}
3534 & $\is$ & $s$ & $\su$ & $h$ & $hp$ & $\sigma$ \\
3535 \next & $\is'$ & $s'$ & $\su'$ & $h'$ & $hp'$ & $\sigma$ \\
3539 where $\is$ is an instruction stream, $s$ is the stack, $\su$ is the
3540 update frame pointer and $h$ is the heap.
3543 \Subsection{Stack manipulation}{hugs-stack-manipulation}
3547 \item[ Push a global variable ].
3549 \begin{tabular}{|llrrrrr|}
3551 & PUSHGLOBAL $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3552 \next & $\is$ & $\sigma!o:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3556 \item[ Push a local variable ].
3558 \begin{tabular}{|llrrrrr|}
3560 & PUSHLOCAL $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3561 \next & $\is$ & $s!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3565 \item[ Push an unboxed int ].
3567 \begin{tabular}{|llrrrrr|}
3569 & PUSHINT $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3570 \next & $\is$ & $I\# : \sigma!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3574 The $I\#$ is a tag included for the benefit of the garbage collector.
3575 Similar rules exist for floats, doubles, chars, etc.
3577 \item[ Push an unboxed int ].
3579 \begin{tabular}{|llrrrrr|}
3581 & PUSHUNTAGGEDINT $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3582 \next & $\is$ & $\sigma!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3586 Similar rules exist for floats, doubles, chars, etc.
3588 \item[ Delete environment from stack --- ready for tail call ].
3590 \begin{tabular}{|llrrrrr|}
3592 & SLIDE $m$ $n$ : $\is$ & $\as \append \bs \append \cs$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3593 \next & $\is$ & $\as \append \cs$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3597 where $\size{\as} = m$ and $\size{\bs} = n$.
3600 \item[ Push a return address ].
3602 \begin{tabular}{|llrrrrr|}
3604 & PUSHALTS $o$:$\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3605 \next & $\is$ & $@HUGS_RET@:\sigma!o:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3609 \item[ Push a BCO ].
3611 \begin{tabular}{|llrrrrr|}
3613 & PUSHBCO $o$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3614 \next & $\is$ & $\sigma!o : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3620 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3621 \Subsection{Heap manipulation}{hugs-heap-manipulation}
3622 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3626 \item[ Allocate a heap object ].
3628 \begin{tabular}{|llrrrrr|}
3630 & ALLOC $m$ : $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3631 \next & $\is$ & $hp:s$ & $su$ & $h$ & $hp+m$ & $\sigma$ \\
3635 \item[ Build a constructor ].
3637 \begin{tabular}{|llrrrrr|}
3639 & PACK $o$ $o'$ : $\is$ & $\ws \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3640 \next & $\is$ & $s$ & $su$ & $h[s!o \mapsto Pack C\{\ws\}]$ & $hp$ & $\sigma$ \\
3644 where $C = \sigma!o'$ and $\size{\ws} = \arity{C}$.
3646 \item[ Build an AP or PAP ].
3648 \begin{tabular}{|llrrrrr|}
3650 & MKAP $o$ $m$:$\is$ & $f : \ws \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3651 \next & $\is$ & $s$ & $su$ & $h[s!o \mapsto \AP(f,\ws)]$ & $hp$ & $\sigma$ \\
3655 where $\size{\ws} = m$.
3657 \begin{tabular}{|llrrrrr|}
3659 & MKPAP $o$ $m$:$\is$ & $f : \ws \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3660 \next & $\is$ & $s$ & $su$ & $h[s!o \mapsto \PAP(f,\ws)]$ & $hp$ & $\sigma$ \\
3664 where $\size{\ws} = m$.
3666 \item[ Unpacking a constructor ].
3668 \begin{tabular}{|llrrrrr|}
3670 & UNPACK : $is$ & $a : s$ & $su$ & $h[a \mapsto C\ \ws]$ & $hp$ & $\sigma$ \\
3671 \next & $is'$ & $(\reverse\ \ws) \append a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3675 The $\reverse\ \ws$ looks expensive but, since the stack grows down
3676 and the heap grows up, that's actually the cheap way of copying from
3677 heap to stack. Looking at the compilation rules, you'll see that we
3678 always push the args in reverse order.
3683 \Subsection{Entering a closure}{hugs-entering}
3687 \item[ Enter a BCO ].
3689 \begin{tabular}{|llrrrrr|}
3691 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto BCO\{\is\} ]$ & $hp$ & $\sigma$ \\
3692 \next & $\is$ & $a : s$ & $su$ & $h$ & $hp$ & $a$ \\
3696 \item[ Enter a PAP closure ].
3698 \begin{tabular}{|llrrrrr|}
3700 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \PAP(f,\ws)]$ & $hp$ & $\sigma$ \\
3701 \next & [ENTER] & $f : \ws \append s$ & $su$ & $h$ & $hp$ & $???$ \\
3705 \item[ Entering an AP closure ].
3707 \begin{tabular}{|llrrrrr|}
3709 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \AP(f,ws)]$ & $hp$ & $\sigma$ \\
3710 \next & [ENTER] & $f : \ws \append @UPD_RET@:\su:a:s$ & $su'$ & $h$ & $hp$ & $???$ \\
3716 \item Instead of blindly pushing an update frame for $a$, we can first test whether there's already
3717 an update frame there. If so, overwrite the existing updatee with an indirection to $a$ and
3718 overwrite the updatee field with $a$. (Overwriting $a$ with an indirection to the updatee also
3719 works.) This results in update chains of maximum length 2.
3723 \item[ Returning a constructor ].
3725 \begin{tabular}{|llrrrrr|}
3727 & [ENTER] & $a : @HUGS_RET@ : \alts : s$ & $su$ & $h[a \mapsto C\{\ws\}]$ & $hp$ & $\sigma$ \\
3728 \next & $\alts.\entry$ & $a:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3733 \item[ Entering an indirection node ].
3735 \begin{tabular}{|llrrrrr|}
3737 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \Ind{a'}]$ & $hp$ & $\sigma$ \\
3738 \next & [ENTER] & $a' : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3742 \item[Entering GHC closure].
3744 \begin{tabular}{|llrrrrr|}
3746 & [ENTER] & $a : s$ & $su$ & $h[a \mapsto \GHCOBJ]$ & $hp$ & $\sigma$ \\
3747 \next & [ENTERGHC] & $a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3751 \item[Returning a constructor to GHC].
3753 \begin{tabular}{|llrrrrr|}
3755 & [ENTER] & $a : \GHCRET : s$ & $su$ & $h[a \mapsto C \ws]$ & $hp$ & $\sigma$ \\
3756 \next & [ENTERGHC] & $a : \GHCRET : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3763 \Subsection{Updates}{hugs-updates}
3767 \item[ Updating with a constructor].
3769 \begin{tabular}{|llrrrrr|}
3771 & [ENTER] & $a : @UPD_RET@ : ua : s$ & $su$ & $h[a \mapsto C\{\ws\}]$ & $hp$ & $\sigma$ \\
3772 \next & [ENTER] & $a \append s$ & $su$ & $h[au \mapsto \Ind{a}$ & $hp$ & $\sigma$ \\
3776 \item[ Argument checks].
3778 \begin{tabular}{|llrrrrr|}
3780 & ARGCHECK $m$:$\is$ & $a : \as \append s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3781 \next & $\is$ & $a : \as \append s$ & $su$ & $h'$ & $hp$ & $\sigma$ \\
3785 where $m \ge (su - sp)$
3787 \begin{tabular}{|llrrrrr|}
3789 & ARGCHECK $m$:$\is$ & $a : \as \append @UPD_RET@:su:ua:s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3790 \next & $\is$ & $a : \as \append s$ & $su$ & $h'$ & $hp$ & $\sigma$ \\
3794 where $m < (su - sp)$ and
3795 $h' = h[ua \mapsto \Ind{a'}, a' \mapsto \PAP(a,\reverse\ \as) ]$
3797 Again, we reverse the list of values as we transfer them from the
3798 stack to the heap --- reflecting the fact that the stack and heap grow
3799 in different directions.
3803 \Subsection{Branches}{hugs-branches}
3807 \item[ Testing a constructor ].
3809 \begin{tabular}{|llrrrrr|}
3811 & TEST $tag$ $is'$ : $is$ & $a : s$ & $su$ & $h[a \mapsto C\ \ws]$ & $hp$ & $\sigma$ \\
3812 \next & $is$ & $a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3816 where $C.\tag = tag$
3818 \begin{tabular}{|llrrrrr|}
3820 & TEST $tag$ $is'$ : $is$ & $a : s$ & $su$ & $h[a \mapsto C\ \ws]$ & $hp$ & $\sigma$ \\
3821 \next & $is'$ & $a : s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3825 where $C.\tag \neq tag$
3829 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3830 \Subsection{Heap and stack checks}{hugs-heap-stack-checks}
3831 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3833 \begin{tabular}{|llrrrrr|}
3835 & STACKCHECK $stk$:$\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3836 \next & $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3840 if $s$ has $stk$ free slots.
3842 \begin{tabular}{|llrrrrr|}
3844 & HEAPCHECK $hp$:$\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3845 \next & $\is$ & $s$ & $su$ & $h$ & $hp$ & $\sigma$ \\
3849 if $h$ has $hp$ free slots.
3851 If either check fails, we push the current bco ($\sigma$) onto the
3852 stack and return to the scheduler. When the scheduler has fixed the
3853 problem, it pops the top object off the stack and reenters it.
3858 \item The bytecode CHECK1000 conservatively checks for 1000 words of heap space and 1000 words of stack space.
3859 We use it to reduce code space and instruction decoding time.
3860 \item The bytecode HEAPCHECK1000 conservatively checks for 1000 words of heap space.
3861 It is used in case alternatives.
3865 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3866 \Subsection{Primops}{hugs-primops}
3867 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3869 \ToDo{primops take m words and return n words. The expect boxed arguments on the stack.}
3872 \Section{The Machine Code Evaluator}{asm-evaluator}
3874 This section describes the framework in which compiled code evaluates
3875 expressions. Only at certain points will compiled code need to be
3876 able to talk to the interpreted world; these are discussed in
3877 \secref{switching-worlds}.
3879 \Subsection{Calling conventions}{ghc-calling-conventions}
3881 \Subsubsection{The call/return registers}{ghc-regs}
3883 One of the problems in designing a virtual machine is that we want it
3884 abstract away from tedious machine details but still reveal enough of
3885 the underlying hardware that we can make sensible decisions about code
3886 generation. A major problem area is the use of registers in
3887 call/return conventions. On a machine with lots of registers, it's
3888 cheaper to pass arguments and results in registers than to pass them
3889 on the stack. On a machine with very few registers, it's cheaper to
3890 pass arguments and results on the stack than to use ``virtual
3891 registers'' in memory. We therefore use a hybrid system: the first
3892 $n$ arguments or results are passed in registers; and the remaining
3893 arguments or results are passed on the stack. For register-poor
3894 architectures, it is important that we allow $n=0$.
3896 We'll label the arguments and results \Arg{1} \ldots \Arg{m} --- with
3897 the understanding that \Arg{1} \ldots \Arg{n} are in registers and
3898 \Arg{n+1} \ldots \Arg{m} are on top of the stack.
3900 Note that the mapping of arguments \Arg{1} \ldots \Arg{n} to machine
3901 registers depends on the \emph{kinds} of the arguments. For example,
3902 if the first argument is a Float, we might pass it in a different
3903 register from if it is an Int. In fact, we might find that a given
3904 architecture lets us pass varying numbers of arguments according to
3905 their types. For example, if a CPU has 2 Int registers and 2 Float
3906 registers then we could pass between 2 and 4 arguments in machine
3907 registers --- depending on whether they all have the same kind or they
3908 have different kinds.
3910 \Subsubsection{Entering closures}{entering-closures}
3912 To evaluate a closure we jump to the entry code for the closure
3913 passing a pointer to the closure in \Arg{1} so that the entry code can
3914 access its environment.
3916 \Subsubsection{Function call}{ghc-fun-call}
3918 The function-call mechanism is obviously crucial. There are five different
3922 \item \emph{Known combinator (function with no free variables) and
3925 A fast call can be made: push excess arguments onto stack and jump to
3926 function's \emph{fast entry point} passing arguments in \Arg{1} \ldots
3929 The \emph{fast entry point} is only called with exactly the right
3930 number of arguments (in \Arg{1} \ldots \Arg{m}) so it can instantly
3931 start doing useful work without first testing whether it has enough
3932 registers or having to pop them off the stack first.
3934 \item \emph{Known combinator and insufficient arguments.}
3936 A slow call can be made: push all arguments onto stack and jump to
3937 function's \emph{slow entry point}.
3939 Any unpointed arguments which are pushed on the stack must be tagged.
3940 This means pushing an extra word on the stack below the unpointed
3941 words, containing the number of unpointed words above it.
3943 %Todo: forward ref about tagging?
3946 The \emph{slow entry point} might be called with insufficient arguments
3947 and so it must test whether there are enough arguments on the stack.
3948 This \emph{argument satisfaction check} consists of checking that
3949 @Su-Sp@ is big enough to hold all the arguments (including any tags).
3953 \item If the argument satisfaction check fails, it is because there is
3954 one or more update frames on the stack before the rest of the
3955 arguments that the function needs. In this case, we construct a PAP
3956 (partial application, \secref{PAP}) containing the arguments
3957 which are on the stack. The PAP construction code will return to the
3958 update frame with the address of the PAP in \Arg{1}.
3960 \item If the argument satisfaction check succeeds, we jump to the fast
3961 entry point with the arguments in \Arg{1} \ldots \Arg{arity}.
3963 If the fast entry point expects to receive some of \Arg{i} on the
3964 stack, we can reduce the amount of movement required by making the
3965 stack layout for the fast entry point look like the stack layout for
3966 the slow entry point. Since the slow entry point is entered with the
3967 first argument on the top of the stack and with tags in front of any
3968 unpointed arguments, this means that if \Arg{i} is unpointed, there
3969 should be space below it for a tag and that the highest numbered
3970 argument should be passed on the top of the stack.
3972 We usually arrange that the fast entry point is placed immediately
3973 after the slow entry point --- so we can just ``fall through'' to the
3974 fast entry point without performing a jump.
3979 \item \emph{Known function closure (function with free variables) and
3982 A fast call can be made: push excess arguments onto stack and jump to
3983 function's \emph{fast entry point} passing a pointer to closure in
3984 \Arg{1} and arguments in \Arg{2} \ldots \Arg{m+1}.
3986 Like the fast entry point for a combinator, the fast entry point for a
3987 closure is only called with appropriate values in \Arg{1} \ldots
3988 \Arg{m+1} so we can start work straight away. The pointer to the
3989 closure is used to access the free variables of the closure.
3992 \item \emph{Known function closure and insufficient arguments.}
3994 A slow call can be made: push all arguments onto stack and jump to the
3995 closure's slow entry point passing a pointer to the closure in \Arg{1}.
3997 Again, the slow entry point performs an argument satisfaction check
3998 and either builds a PAP or pops the arguments off the stack into
3999 \Arg{2} \ldots \Arg{m+1} and jumps to the fast entry point.
4002 \item \emph{Unknown function closure, thunk or constructor.}
4004 Sometimes, the function being called is not statically identifiable.
4005 Consider, for example, the @compose@ function:
4007 compose f g x = f (g x)
4009 Since @f@ and @g@ are passed as arguments to @compose@, the latter has
4010 to make a heap call. In a heap call the arguments are pushed onto the
4011 stack, and the closure bound to the function is entered. In the
4012 example, a thunk for @(g x)@ will be allocated, (a pointer to it)
4013 pushed on the stack, and the closure bound to @f@ will be
4014 entered. That is, we will jump to @f@s entry point passing @f@ in
4015 \Arg{1}. If \Arg{1} is passed on the stack, it is pushed on top of
4016 the thunk for @(g x)@.
4018 The \emph{entry code} for an updateable thunk (which must have arity 0)
4019 pushes an update frame on the stack and starts executing the body of
4020 the closure --- using \Arg{1} to access any free variables. This is
4021 described in more detail in \secref{data-updates}.
4023 The \emph{entry code} for a non-updateable closure is just the
4024 closure's slow entry point.
4028 In addition to the above considerations, if there are \emph{too many}
4029 arguments then the extra arguments are simply pushed on the stack with
4032 To summarise, a closure's standard (slow) entry point performs the
4036 \item[Argument satisfaction check.] (function closure only)
4037 \item[Stack overflow check.]
4038 \item[Heap overflow check.]
4039 \item[Copy free variables out of closure.] %Todo: why?
4040 \item[Eager black holing.] (updateable thunk only) %Todo: forward ref.
4041 \item[Push update frame.]
4042 \item[Evaluate body of closure.]
4046 \Subsection{Case expressions and return conventions}{return-conventions}
4048 The \emph{evaluation} of a thunk is always initiated by
4049 a @case@ expression. For example:
4051 case x of (a,b) -> E
4054 The code for a @case@ expression looks like this:
4057 \item Push the free variables of the branches on the stack (fv(@E@) in
4059 \item Push a \emph{return address} on the stack.
4060 \item Evaluate the scrutinee (@x@ in this case).
4063 Once evaluation of the scrutinee is complete, execution resumes at the
4064 return address, which points to the code for the expression @E@.
4066 When execution resumes at the return point, there must be some {\em
4067 return convention} that defines where the components of the pair, @a@
4068 and @b@, can be found. The return convention varies according to the
4069 type of the scrutinee @x@:
4075 (A space for) the return address is left on the top of the stack.
4076 Leaving the return address on the stack ensures that the top of the
4077 stack contains a valid activation record
4078 (\secref{activation-records}) --- should a garbage
4079 collection be required.
4081 \item If @x@ has a boxed type (e.g.~a data constructor or a function),
4082 a pointer to @x@ is returned in \Arg{1}.
4084 \ToDo{Warn that components of E should be extracted as soon as
4085 possible to avoid a space leak.}
4087 \item If @x@ is an unboxed type (e.g.~@Int#@ or @Float#@), @x@ is
4090 \item If @x@ is an unboxed tuple constructor, the components of @x@
4091 are returned in \Arg{1} \ldots \Arg{n} but no object is constructed in
4094 When passing an unboxed tuple to a function, the components are
4095 flattened out and passed in \Arg{1} \ldots \Arg{n} as usual.
4099 \Subsection{Vectored Returns}{vectored-returns}
4101 Many algebraic data types have more than one constructor. For
4102 example, the @Maybe@ type is defined like this:
4104 data Maybe a = Nothing | Just a
4106 How does the return convention encode which of the two constructors is
4107 being returned? A @case@ expression scrutinising a value of @Maybe@
4108 type would look like this:
4114 Rather than pushing a return address before evaluating the scrutinee,
4115 @E@, the @case@ expression pushes (a pointer to) a \emph{return
4116 vector}, a static table consisting of two code pointers: one for the
4117 @Just@ alternative, and one for the @Nothing@ alternative.
4123 The constructor @Nothing@ returns by jumping to the first item in the
4124 return vector with a pointer to a (statically built) Nothing closure
4127 It might seem that we could avoid loading \Arg{1} in this case since the
4128 first item in the return vector will know that @Nothing@ was returned
4129 (and can easily access the Nothing closure in the (unlikely) event
4130 that it needs it. The only reason we load \Arg{1} is in case we have to
4131 perform an update (\secref{data-updates}).
4135 The constructor @Just@ returns by jumping to the second element of the
4136 return vector with a pointer to the closure in \Arg{1}.
4140 In this way no test need be made to see which constructor returns;
4141 instead, execution resumes immediately in the appropriate branch of
4144 \Subsection{Direct Returns}{direct-returns}
4146 When a datatype has a large number of constructors, it may be
4147 inappropriate to use vectored returns. The vector tables may be
4148 large and sparse, and it may be better to identify the constructor
4149 using a test-and-branch sequence on the tag. For this reason, we
4150 provide an alternative return convention, called a \emph{direct
4153 In a direct return, the return address pushed on the stack really is a
4154 code pointer. The returning code loads a pointer to the closure being
4155 returned in \Arg{1} as usual, and also loads the tag into \Arg{2}.
4156 The code at the return address will test the tag and jump to the
4157 appropriate code for the case branch. If \Arg{2} isn't mapped to a
4158 real machine register on this architecture, then we don't load it on a
4159 return, instead using the tag directly from the info table.
4161 The choice of whether to use a vectored return or a direct return is
4162 made on a type-by-type basis --- up to a certain maximum number of
4163 constructors imposed by the update mechanism
4164 (\secref{data-updates}).
4166 Single-constructor data types also use direct returns, although in
4167 that case there is no need to return a tag in \Arg{2}.
4169 \ToDo{for a nullary constructor we needn't return a pointer to the
4170 constructor in \Arg{1}.}
4172 \Subsection{Updates}{data-updates}
4174 The entry code for an updatable thunk (which must be of arity 0):
4177 \item copies the free variables out of the thunk into registers or
4179 \item pushes an \emph{update frame} onto the stack.
4181 An update frame is a small activation record consisting of
4183 \begin{tabular}{|l|l|l|}
4185 \emph{Fixed header} & \emph{Update Frame link} & \emph{Updatee} \\
4190 \note{In the semantics part of the STG paper (section 5.6), an update
4191 frame consists of everything down to the last update frame on the
4192 stack. This would make sense too --- and would fit in nicely with
4193 what we're going to do when we add support for speculative
4195 \ToDo{I think update frames contain cost centres sometimes}
4197 \item If we are doing ``eager blackholing,'' we then overwrite the
4198 thunk with a black hole (\secref{BLACKHOLE}). Otherwise, we leave it
4199 to the garbage collector to black hole the thunk.
4202 Start evaluating the body of the expression.
4206 When the expression finishes evaluation, it will enter the update
4207 frame on the top of the stack. Since the returner doesn't know
4208 whether it is entering a normal return address/vector or an update
4209 frame, we follow exactly the same conventions as return addresses and
4210 return vectors. That is, on entering the update frame:
4213 \item The value of the thunk is in \Arg{1}. (Recall that only thunks
4214 are updateable and that thunks return just one value.)
4216 \item If the data type is a direct-return type rather than a
4217 vectored-return type, then the tag is in \Arg{2}.
4219 \item The update frame is still on the stack.
4222 We can safely share a single statically-compiled update function
4223 between all types. However, the code must be able to handle both
4224 vectored and direct-return datatypes. This is done by arranging that
4225 the update code looks like this:
4233 |---------------| <- update code pointer
4238 Each entry in the return vector (which is large enough to cover the
4239 largest vectored-return type) points to the update code.
4243 \item overwrites the \emph{updatee} with an indirection to \Arg{1};
4244 \item loads @Su@ from the Update Frame link;
4245 \item removes the update frame from the stack; and
4246 \item enters \Arg{1}.
4249 We enter \Arg{1} again, having probably just come from there, because
4250 it knows whether to perform a direct or vectored return. This could
4251 be optimised by compiling special update code for each slot in the
4252 return vector, which performs the correct return.
4254 \Subsection{Semi-tagging}{semi-tagging}
4256 When a @case@ expression evaluates a variable that might be bound
4257 to a thunk it is often the case that the scrutinee is already evaluated.
4258 In this case we have paid the penalty of (a) pushing the return address (or
4259 return vector address) on the stack, (b) jumping through the info pointer
4260 of the scrutinee, and (c) returning by an indirect jump through the
4261 return address on the stack.
4263 If we knew that the scrutinee was already evaluated we could generate
4264 (better) code which simply jumps to the appropriate branch of the
4265 @case@ with a pointer to the scrutinee in \Arg{1}. (For direct
4266 returns to multiconstructor datatypes, we might also load the tag into
4269 An obvious idea, therefore, is to test dynamically whether the heap
4270 closure is a value (using the tag in the info table). If not, we
4271 enter the closure as usual; if so, we jump straight to the appropriate
4272 alternative. Here, for example, is pseudo-code for the expression
4273 @(case x of { (a,_,c) -> E }@:
4275 \Arg{1} = <pointer to x>;
4276 tag = \Arg{1}->entry->tag;
4278 Sp--; \\ insert space for return address
4282 goto \Arg{1}->entry;
4284 <info table for return address goes here>
4285 ret: a = \Arg{1}->data1; \\ suck out a and c to avoid space leak
4289 and here is the code for the expression @(case x of { [] -> E1; x:xs -> E2 }@:
4291 \Arg{1} = <pointer to x>;
4292 tag = \Arg{1}->entry->tag;
4294 Sp--; \\ insert space for return address
4298 goto \Arg{1}->entry;
4302 retvec: \\ reversed return vector
4303 <return info table for case goes here>
4305 panic("Direct return into vectored case");
4309 ret2: x = \Arg{1}->head;
4313 There is an obvious cost in compiled code size (but none in the size
4314 of the bytecodes). There is also a cost in execution time if we enter
4315 more thunks than data constructors.
4317 Both the direct and vectored returns are easily modified to chase chains
4318 of indirections too. In the vectored case, this is most easily done by
4319 making sure that @IND = TAG_1 - 1@, and adding an extra field to every
4320 return vector. In the above example, the indirection code would be
4322 ind: \Arg{1} = \Arg{1}->next;
4325 where @ind_loop@ is the second line of code.
4327 Note that we have to leave space for a return address since the return
4328 address expects to find one. If the body of the expression requires a
4329 heap check, we will actually have to write the return address before
4330 entering the garbage collector.
4333 \Subsection{Heap and Stack Checks}{heap-and-stack-checks}
4335 The storage manager detects that it needs to garbage collect the old
4336 generation when the evaluator requests a garbage collection without
4337 having moved the heap pointer since the last garbage collection. It
4338 is therefore important that the GC routines \emph{not} move the heap
4339 pointer unless the heap check fails. This is different from what
4340 happens in the current STG implementation.
4342 Assuming that the stack can never shrink, we perform a stack check
4343 when we enter a closure but not when we return to a return
4344 continuation. This doesn't work for heap checks because we cannot
4345 predict what will happen to the heap if we call a function.
4347 If we wish to allow the stack to shrink, we need to perform a stack
4348 check whenever we enter a return continuation. Most of these checks
4349 could be eliminated if the storage manager guaranteed that a stack
4350 would always have 1000 words (say) of space after it was shrunk. Then
4351 we can omit stack checks for less than 1000 words in return
4354 When an argument satisfaction check fails, we need to push the closure
4355 (in R1) onto the stack - so we need to perform a stack check. The
4356 problem is that the argument satisfaction check occurs \emph{before}
4357 the stack check. The solution is that the caller of a slow entry
4358 point or closure will guarantee that there is at least one word free
4359 on the stack for the callee to use.
4361 Similarily, if a heap or stack check fails, we need to push the arguments
4362 and closure onto the stack. If we just came from the slow entry point,
4363 there's certainly enough space and it is the responsibility of anyone
4364 using the fast entry point to guarantee that there is enough space.
4366 \ToDo{Be more precise about how much space is required - document it
4367 in the calling convention section.}
4369 \Subsection{Handling interrupts/signals}{signals}
4372 May have to keep C stack pointer in register to placate OS?
4373 May have to revert black holes - ouch!
4378 \section{The Loader}
4379 \section{The Compilers}
4382 \part{Old stuff - needs to be mined for useful info}
4384 \section{The Scheduler}
4386 The Scheduler is the heart of the run-time system. A running program
4387 consists of a single running thread, and a list of runnable and
4388 blocked threads. The running thread returns to the scheduler when any
4389 of the following conditions arises:
4392 \item A heap check fails, and a garbage collection is required
4393 \item Compiled code needs to switch to interpreted code, and vice
4395 \item The thread becomes blocked.
4396 \item The thread is preempted.
4399 A running system has a global state, consisting of
4402 \item @Hp@, the current heap pointer, which points to the next
4403 available address in the Heap.
4404 \item @HpLim@, the heap limit pointer, which points to the end of the
4406 \item The Thread Preemption Flag, which is set whenever the currently
4407 running thread should be preempted at the next opportunity.
4408 \item A list of runnable threads.
4409 \item A list of blocked threads.
4412 Each thread is represented by a Thread State Object (TSO), which is
4413 described in detail in \secref{TSO}.
4415 The following is pseudo-code for the inner loop of the scheduler
4419 while (threads_exist) {
4420 // handle global problems: GC, parallelism, etc
4422 if (external_message) service_message();
4423 // deal with other urgent stuff
4425 pick a runnable thread;
4427 // enter object on top of stack
4428 // if the top object is a BCO, we must enter it
4429 // otherwise appply any heuristic we wish.
4430 if (thread->stack[thread->sp]->info.type == BCO) {
4431 status = runHugs(thread,&smInfo);
4433 status = runGHC(thread,&smInfo);
4435 switch (status) { // handle local problems
4436 case (StackOverflow): enlargeStack; break;
4437 case (Error e) : error(thread,e); break;
4438 case (ExitWith e) : exit(e); break;
4439 case (Yield) : break;
4441 } while (thread_runnable);
4445 \Subsection{Invoking the garbage collector}{ghc-invoking-gc}
4447 \Subsection{Putting the thread to sleep}{ghc-thread-sleeps}
4449 \Subsection{Calling C from Haskell}{ghc-ccall}
4451 We distinguish between "safe calls" where the programmer guarantees
4452 that the C function will not call a Haskell function or, in a
4453 multithreaded system, block for a long period of time and "unsafe
4454 calls" where the programmer cannot make that guarantee.
4456 Safe calls are performed without returning to the scheduler and are
4457 discussed elsewhere (\ToDo{discuss elsewhere}).
4459 Unsafe calls are performed by returning an array (outside the Haskell
4460 heap) of arguments and a C function pointer to the scheduler. The
4461 scheduler allocates a new thread from the operating system
4462 (multithreaded system only), spawns a call to the function and
4463 continues executing another thread. When the ccall completes, the
4464 thread informs the scheduler and the scheduler adds the thread to the
4465 runnable threads list.
4467 \ToDo{Describe this in more detail.}
4470 \Subsection{Calling Haskell from C}{ghc-c-calls-haskell}
4472 When C calls a Haskell closure, it sends a message to the scheduler
4473 thread. On receiving the message, the scheduler creates a new Haskell
4474 thread, pushes the arguments to the C function onto the thread's stack
4475 (with tags for unboxed arguments) pushes the Haskell closure and adds
4476 the thread to the runnable list so that it can be entered in the
4479 When the closure returns, the scheduler sends back a message which
4480 awakens the (C) thread.
4482 \ToDo{Do we need to worry about the garbage collector deallocating the
4483 thread if it gets blocked?}
4485 \Subsection{Switching Worlds}{switching-worlds}
4487 \ToDo{This has all changed: we always leave a closure on top of the
4488 stack if we mean to continue executing it. The scheduler examines the
4489 top of the stack and tries to guess which world we want to be in. If
4490 it finds a @BCO@, it certainly enters Hugs, if it finds a @GHC@
4491 closure, it certainly enters GHC and if it finds a standard closure,
4492 it is free to choose either one but it's probably best to enter GHC
4493 for everything except @BCO@s and perhaps @AP@s.}
4495 Because this is a combined compiled/interpreted system, the
4496 interpreter will sometimes encounter compiled code, and vice-versa.
4498 All world-switches go via the scheduler, ensuring that the world is in
4499 a known state ready to enter either compiled code or the interpreter.
4500 When a thread is run from the scheduler, the @whatNext@ field in the
4501 TSO (\secref{TSO}) is checked to find out how to execute the
4505 \item If @whatNext@ is set to @ReturnGHC@, we load up the required
4506 registers from the TSO and jump to the address at the top of the user
4508 \item If @whatNext@ is set to @EnterGHC@, we load up the required
4509 registers from the TSO and enter the closure pointed to by the top
4511 \item If @whatNext@ is set to @EnterHugs@, we enter the top thing on
4512 the stack, using the interpreter.
4515 There are four cases we need to consider:
4518 \item A GHC thread enters a Hugs-built closure.
4519 \item A GHC thread returns to a Hugs-compiled return address.
4520 \item A Hugs thread enters a GHC-built closure.
4521 \item A Hugs thread returns to a Hugs-compiled return address.
4524 GHC-compiled modules cannot call functions in a Hugs-compiled module
4525 directly, because the compiler has no information about arities in the
4526 external module. Therefore it must assume any top-level objects are
4527 CAFs, and enter their closures.
4529 \ToDo{Hugs-built constructors?}
4531 We now examine the various cases one by one and describe how the
4532 switch happens in each situation.
4534 \subsection{A GHC thread enters a Hugs-built closure}
4535 \label{sec:ghc-to-hugs-switch}
4537 There is three possibilities: GHC has entered a @PAP@, or it has
4538 entered a @AP@, or it has entered the BCO directly (for a top-level
4539 function closure). @AP@s and @PAP@s are ``standard closures'' and
4540 so do not require us to enter the bytecode interpreter.
4542 The entry code for a BCO does the following:
4545 \item Push the address of the object entered on the stack.
4546 \item Save the current state of the thread in its TSO.
4547 \item Return to the scheduler, setting @whatNext@ to @EnterHugs@.
4550 BCO's for thunks and functions have the same entry conventions as
4551 slow entry points: they expect to find their arguments on the stac
4552 with unboxed arguments preceded by appropriate tags.
4554 \subsection{A GHC thread returns to a Hugs-compiled return address}
4555 \label{sec:ghc-to-hugs-switch}
4557 Hugs return addresses are laid out as in \figref{hugs-return-stack}.
4558 If GHC is returning, it will return to the address at the top of the
4559 stack, namely @HUGS_RET@. The code at @HUGS_RET@ performs the
4563 \item pushes \Arg{1} (the return value) on the stack.
4564 \item saves the thread state in the TSO
4565 \item returns to the scheduler with @whatNext@ set to @EnterHugs@.
4568 \noindent When Hugs runs, it will enter the return value, which will
4569 return using the correct Hugs convention
4570 (\secref{hugs-return-convention}) to the return address underneath it
4573 \subsection{A Hugs thread enters a GHC-compiled closure}
4574 \label{sec:hugs-to-ghc-switch}
4576 Hugs can recognise a GHC-built closure as not being one of the
4577 following types of object:
4583 \item An indirection, or
4584 \item A constructor.
4587 When Hugs is called on to enter a GHC closure, it executes the
4588 following sequence of instructions:
4591 \item Push the address of the closure on the stack.
4592 \item Save the current state of the thread in the TSO.
4593 \item Return to the scheduler, with the @whatNext@ field set to
4597 \subsection{A Hugs thread returns to a GHC-compiled return address}
4598 \label{sec:hugs-to-ghc-switch}
4600 When Hugs encounters a return address on the stack that is not
4601 @HUGS_RET@, it knows that a world-switch is required. At this point
4602 the stack contains a pointer to the return value, followed by the GHC
4603 return address. The following sequence is then performed:
4606 \item save the state of the thread in the TSO.
4607 \item return to the scheduler, setting @whatNext@ to @EnterGHC@.
4610 The first thing that GHC will do is enter the object on the top of the
4611 stack, which is a pointer to the return value. This value will then
4612 return itself to the return address using the GHC return convention.
4620 We're nuking the following:
4627 Return in registers.
4628 This lets us remove update code pointers from info tables,
4629 removes the need for phantom info tables, simplifies
4634 Careful analysis suggests that it doesn't buy us very much
4635 and it is hard to work with.
4637 Eliminating threaded GCs eliminates the desire to share SMReps
4638 so they are (once more) part of the Info table.
4642 Doesn't buy us anything on a register-poor architecture and
4643 isn't so important if we have semi-tagging.
4646 - Probably bad on register poor architecture
4647 - Can avoid need to write return address to stack on reg rich arch.
4648 - when a function does a small amount of work, doesn't
4649 enter any other thunks and then returns.
4650 eg entering a known constructor (but semitagging will catch this)
4651 - Adds complications
4657 This lets us drop CONST closures and CHARLIKE closures (assuming we
4658 don't support Unicode). The only point of these closures was to
4659 avoid updating with an indirection.
4661 We also drop @MIN_UPD_SIZE@ --- all we need is space to insert an
4662 indirection or a black hole.
4665 STATIC SMReps are now called CONST
4670 \item The profiling ``kind'' field is now encoded in the @INFO_TYPE@ field.
4671 This identifies the general sort of the closure for profiling purposes.
4673 \item Various papers describe deleting update frames for unreachable objects.
4674 This has never been implemented and we don't plan to anytime soon.