Language/OS - Multiplatform Resource Library/LANGUAGE OS.iso/pcl/sptmbr16.lha/lap.text

Text File | 1990-01-22 | 26KB | 656 lines

-*- Mode: Text -*-

Copyright (c) 1985, 1986, 1987, 1988, 1989 Xerox Corporation.
All rights reserved.

Use and copying of this document is permitted.  Any distribution of this
document must comply with all applicable United States export control
laws.

Last updated: 6/3/89 by Gregor
              10/26/89 by Gregor -- added :RETURN, removed :ISHIFT

This file contains documentation of the PCL abstract LAP code.  Any port
of PCL is required to implement the abstract LAP code interface.  There
is a portable, relatively good performance implementation in the file
lap.lisp, port-specific implementations are in that file as well.

The PCL abstract LAP code mechanism exists to provide PCL with a way to
create high-performance method lookup functions.  Using this mechanism,
PCL can produce "LAP closures" which do the method lookup.  By allowing
PCL to specify these closures using abstract LAP code rather that Lisp
code we hope to achieve the following:

  * Better runtime performance.  By using abstract LAP code, we
    will get better machine instruction sequences than we would
    from compiling Lisp code.

  * Better load and update time performance.  Because it should
    be possible to "assemble" the LAP code more quickly than
    compiling Lisp code, PCL will spend less time building the
    method lookup code.

  * Ability to use PCL without a compiler.  The LAP assembler will
    still be required but this should be much smaller than the full
    lisp compiler.

Of course, not all implementations of the LAP code mechanism will
satisfy all of these goals.  The first is the most important.

In particular, many PCL ports will use the portable LAP implementation.
KCL will use the portable implementation in all of its ports.  Other
Lisps may have custom LAP implementations for some ports and use the
portable implementation for other ports.

Some Lisps will have a custom LAP implementation but will nonetheless
require the compiler to be loaded to generate LAP closure constructors.

An important point is why we have chosen to take this route rather than
have each implementation implement the method lookup codes itself.  This
was done because we are, at PARC, just beginning to study cache behavior
for CLOS programs.  As we learn more about this we will want to modify
the caching strategy PCL uses.  This architecture, because it leaves
PCL to implement caching behavior makes it possible to do this.  Once
this study is complete, implementations may want to do their own, ultra
high performance implementations of caching strategies.



Production of LAP closures is a two step process.  In the first step, a
port-specific function is called to take abstract LAP code and produce a
a "lap closure generator".  Lap closure generators are functions which
are called with a set of closure variable values and return a LAP
closure.

The intermediary of the lap closure generators provides an important
optimization.  Because it is assumed that producing the LAP closure
generator can take much longer than producing a LAP closure from the
generator, PCL attempts to make only one closure generator for each
sequence of LAP code.  Because of the way PCL generates the LAP code
sequences, this is quite easy for it to do.

The rest of this document is divided into six parts.

  * the metatypes std-instance and fsc-instance

  * an abstraction for simple vector indices

  * important optimizations

  * the port specific function for making lap closure generators

  * the actual abstract LAP code

  * examples

*** The metatypes STD-INSTANCE and FSC-INSTANCE ***

In PCL, instances with metaclass STANDARD-CLASS are represented using
the metatype STD-INSTANCE.  (Note that in Cinco de Mayo PCL, this
metatype is called IWMC-CLASS.)  Each port must implement this metatype.
The metatype could be implemented by the following DEFSTRUCT.

   (defstruct (std-instance
                (:predicate std-instance-p)
                (:conc-name %std-instance-)
                (:constructor %allocate-std-instance (wrapper slots))
                (:constructor %allocate-std-instance-1 ())
                (:print-function print-std-instance))
     (wrapper nil)
     (slots nil))

 PCL itself will guarantee correct access to this structure and the
 accessors and constructors.  With this in mind, the following are
 important.

 * Being able to type test this structure quickly is critical. See
   the :STD-INSTANCE-P opcode.

 * The allocation functions should compile inline, do no argument
   checking and be as fast as possible.

 * The accessor functions should compile inline, do no checking of their
   arguments and be as fast as possible.  SETF of the accessors should
   do the same.

The port is also required to implement the metatype FSC-INSTANCE (called
FUNCALLABLE-INSTANCE, or FIN for short, in Cinco de Mayo PCL).  Objects
with this metatype are used, among other things, to implement generic
functions.  These objects have field structure associated with them and
are also functions that can be applied to arguments.  The fields are the
same as those for STD-INSTANCE, the FSC-INSTANCE metatype has
predicates, print-functions, constructors and accessors as follows:

  fsc-instance-p
  print-fsc-instance
  %fsc-instance-wrapper
  %fsc-instance-slots
  %allocate-fsc-instance (wrapper slots)
  %allocate-fsc-instance-1 ()

In addition, objects of metatype FSC-INSTANCE have a property called the
funcallable instance function.  When an FSC-INSTANCE is applied to
arguments, the funcallable instance function is what is actually called.
The funcallable instance function of an FSC-INSTANCE can be changed
using the function SET-FUNCALLABLE-INSTANCE-FUNCTION.  There is no
mechanism for obtaining the funcallable instance function of an
FSC-INSTANCE.

It is possible to implement the FSC-INSTANCE metatype in pure Common
Lisp. A simple implementation which uses lexical closures as the
instances and a hash table to record that the lexical closures are of
metatype FSC-INSTANCE is easy to write.  Unfortunately, this
implementation adds significant overhead:

   to generic-function-invocation (1 function call)
   to slot-access (1 function call or one hash table lookup)
   to class-of a generic-function (1 hash-table lookup)

In addition, it would prevent the FSC-INSTANCEs from being garbage
collected.  In short, the pure Common Lisp implementation really isn't
practical.

Note that previous implementations of FINS were always based on the
lexical closure metatype.  In some ports, that provides poor
performance.  Those ports may want to consider reimplementing to use the
compiled code metatype.  In that implementation strategy, LAP closure
variables would become constants of the compiled code object.

The following note from JonL is of interest when working on a FIN
implementation:

  Date: Tue, 16 May 89 05:45:56 PDT
  From: Jon L White <[e-mail]>

  This isn't a bug in Lucid's compiler -- it's a lurking bug in PCL
  that will "bite" most implementations where different settings of
  the compiler optimization switches will produce morphologically
  different (but of course functionally equivalent) function objects.

  The difficulty is in how discriminator codes service cache misses.
  They  "call out" to (potentially) random functions that will in some
  cases "smash" the function object that was actually running as the
  discriminator code.  This is all right providing you don't return to
  that function frame, but alas ...

  I know this is a more extensive problem because the code in the
  port-independent function 'notice-methods-change' goes out of
  its way to do a tail-recursive call to the function that is going
  to smash the possibly-executing discriminator code.  Here is the
  commentary from that code (sic):

  ;; In order to prevent this we take a simple measure:  we just
  ;; make sure that it doesn't try to reference our its own closure
  ;; variables after it makes the dcode change.  This is done by
  ;; having notice-methods-change-2 do the work of making the change
  ;; AND calling the actual generic function (a closure variable)
  ;; over.  This means that at the time the dcode change is made,
  ;; there is a pointer to the generic function on the stack where
  ;; it won't be affected by the change to the closure variables.


  A similar thing should be done in the construction of standard-accessor,
  checking, and  caching dcodes.  In an experimental version here at Lucid,
  I rewrote  dcode.lisp to do that, and there is no problem with it.
  Although that code is somewhat Lucid-specific, it could be of help to
  someone who wanted to rewrite the generic dcode.lisp (no pun intended).
  Contact me privately if you are interested.

  Doing a tail-recursive call out of dcodes when there is a cache miss
  is a good thing, regardless of other problems.  I think one might as
  well do it.  However, I should point out that in the presence of
  multiprocessing, there is another more serious problem that cannot be
  solved so simply.  Think about what happens when one process decides
  to update a dcode while another process is still using it; no such
  stack-maintenance discipline will fix this case.  A tail-recursive
  exit from the dcode will *immensely* reduce the likelihood that
  another process can sneak in during the interval in which the dcode
  requires consistency in its function; but it can't reduce that
  likelihood to zero.

  The more desirable thing to do is to put the whole "dcode" down one
  more level of indirection through the symbol-function cell of the
  generic function.  This is effectively what PCL's 'make-trampoline'
  function does, but unfortunately that is not a very efficient approach
  when you consider how most compilers will compile it.  Something akin
  to the "mattress-pads" in Steve Haflich's code (in the fin.lisp file)
  could probably be done for many other implementations as well.


*** Index Operations ***

Indexes are an abstraction for indexes into a simple vector.  This
abstraction is used to make it possible to generate more efficient
code to access simple vectors.  The idea being that this may make it
possible to use alternate addressing modes to address these.

The "index value" of an index is defined to be the fixnum of which that
index is an alternate form.  So, using the Lisp function SVREF with the
index value of an index accesses the same element as using the index
with the appropriate access function or operand.

The format of an index is unspecified, but is assumed to be something
like a fixnum with certain bits ignored.  Accessing a vector using an
index must be done using the appropriate special accessor function or
opcode.

Conversion from index values to indices and vice-versa can be done with
the following functions:

INDEX-VALUE->INDEX (index-value)
INDEX->INDEX-VALUE (index)


The following constant indicates the maximum index value an index can
have in a given port.  This must be at least 2^16.

INDEX-VALUE-LIMIT  - a fixnum, must be at least 2^16.


MAKE-INDEX-MASK (<cache-size> <line-size>)

This function is used to make index masks.  Because I am lazy, I show an
implementation of it in the common case where indexes are just fixnums:

  (defun make-index-mask (cache-size line-size)
    (let ((cache-size-in-bits (floor (log cache-size 2)))
          (line-size-in-bits (floor (log line-size 2)))
          (mask 0))
      (dotimes (i cache-size-in-bits) (setq mask (dpb 1 (byte 1 i) mask)))
      (dotimes (i line-size-in-bits)  (setq mask (dpb 0 (byte 1 i) mask)))
      mask))

*** Optimizations ***

This section discusses two important optimizations related to LAP
closures.  The first relates to calling LAP closures themselves, the
second relates to calling other functions from LAP closures.

The important point about calling LAP closures is that almost all of the
time, LAP closures will be used as the funcallable-instance-function of
funcallable instances.  It is required that LAP closures be funcallable
themselves, but usually they will be stored in a FIN and the fin will
then be funcalled.  This brings up several optimizations, including ones
having to do with access to the closure variables of a LAP closure.

When a LAP closure is used to do method lookup, the function the LAP
closure ends up calling has the same number of required arguments as the
LAP closure itself.  Since the LAP closure must check its required
arguments to do the lookup, it is redundant for the function called to
do so as well.  Since LAP closures do all calls in a tail recursive way,
it should even be possible to optimize out certain parts of the normal
stack frame initialization.

A similar situation occurs between effective method functions and the
individual method functions; the difference is that in effective method
functions, the calls are not necessarily tail recursive.

Consequently, it would be nice to have a way to call certain functions
and inhibit the checking of required arguments.  This is made possible
by use of the PCL-FAST-APPLY and PCL-FAST-FUNCALL macros together with
the PCL-FAST-CALL compiler declaration.

The PCL-FAST-CALL compiler declaration declares that a function may be
fast called.  Not all callers of the function will necessarily fast call
it, but most probably will.  The :JMP opcode can only be used to call a
function compiled with the PCL-FAST-CALL declaration.

The PCL-FAST-APPLY and PCL-FAST-FUNCALL macros are used to fast call a
function.  The function argument must be a compiled function that has
the PCL-FAST-CALL compiler declaration in its lambda declarations.

The basic idea is that the PCL-FAST-CALL compiler declaration causes the
compiler to set up an additional entrypoint to the function.  This
entrypoint comes after checking of required arguments but before
processing of other arguments.

Note:  When FAST-APPLY is used, the required arguments will be given as
separate arguments and all other arguments will appear as a single
spread argument.  For example:

(let ((fn (compile () '(lambda (a b &optional (c 'z))
                         (declare (pcl-fast-call))
                         (list a b c)))))

  (pcl-fast-apply fn 'x 'y ())          ;legal
  (pcl-fast-apply fn 'x 'y '(foo))      ;legal
  (pcl-fast-apply fn '(a b c))          ;illegal
  )

*** Producing LAP Closure Generators ***

Each implementation of the LAP code mechanism must provide a port
specific function making lap closure generators.  In the portable
implementation, this function is called PLAP-CLOSURE-GENERATOR.  In
ExCL it should be called EXCL-LAP-CLOSURE-GENERATOR etc.

At any time, the value of the variable *make-lap-closure-generator* is a
symbol which names the function currently being used to make lap closure
generators.

The port specific function must accept arguments as follows:

PLAP-CLOSURE-GENERATOR (<closure-vars>
                        <args>
                        <index-registers>
                        <simple-vector-registers>
                        <lap-code>)

This returns a lap-closure generator.  A lap-closure generator is a
function which is called with a number of arguments equal to the length
of <closure-vars>.  These arguments are the values of the closure
variables for the lap closure.  These values cannot be changed once the
LAP closure is created.   PCL takes care of keeping track of
lap-closure-generators it already has on hand and reusing them.  The
function RESET-LAP-CLOSURE-GENERATORS can be called to force PCL to
forget all the lap closure generators it has remembered.

  <args>

A list of symbols.  This provides a way to name particular arguments to
the LAP closure. Arguments which will not be referenced by name are
given as NIL. All required arguments to the LAP closure are explicitly
included (perhaps as NIL).  If &REST appears at the end of arguments it
means that non-required arguments are allowed, these will be processed
by the methods.  If &REST does not appear at the end of arguments, the
lap closure should signal an error if more than the indicated number of
arguments are supplied.

Examples:

 -  (obj-0 obj-1)

    Specifies a two argument lap closure.  If more or less than
    two arguments are supplied an error is signalled.  Within
    the actual lap code, both arguments can be referenced by
    name (see the :ARG operand).

 -  (obj-0 nil &rest)

    Specifies a two or more argument lap closure.  If less than
    two arguments are supplied an error is signalled.  Within
    the actual lap code, the first argument can be referenced by
    name (see the :ARG operand).


  <closure-vars>

A list of symbols.  The closure will have these as closure variables.
Within the lap code these can be accessed using the :CVAR operand.  The
lap code cannot change these values.  SET-FUNCALLABLE-INSTANCE-FUNCTION
is permitted to have the special knowledge that there are at most ?? of
these and to be prepared to do something special when the funcallable
instance function of a funcallable instance is set to a lap closure.

  <index-registers>

A list of register numbers.  These registers will be used only to hold
indexes.  Other registers may be used to hold indexes as well, but the
only values put into these registers will be indexes.

  <simple-vector-registers>

A list of register numbers.  These registers will be used only to hold
simple-vectors.  Other registers may be used to hold simple-vectors as
well, but the only values put into these registers will be
simple-vectors.


  <lap-code>

The actual lap code for this closure.  This is a list of LAP code
opcodes.  See the section "Abstract LAP Code" for more details.

Each implementation must also supply a function named PRE-MAKE-xxx where
xxx is the same as the name of its make-lap-closure-generator function.
The macro doesn't evaluate its arguments, and when it appears in a file
it should try to do some of the work at load time.  It might appear in a
file like this:

(eval-when (load)
  (setq 1-arg-std-lap
        (pre-make-plap-closure-generator ...)))

*** Abstract LAP Code ***

Each lap code operand has the form: (opcode operand1 ... operandn).

In some cases, the distinction between an operand and an opcode is
somewhat arbitrary.  In general, opcodes have a significant "action"
component to their behavior.  Operands select a piece of data to operate
on.  Some operands select their data in a more complex way, but they are
operands anyways.

All data must be in a register before it can be operated on.   This
requirement means that the only place a non-register operand can appear
is as the first argument to the :move opcode.  (Actually, there is one
other exception, a :iref operand can be the target of a move as well.)
Moreover, only register operands can appear as the second argument to
the :move opcode and this register must not appear in the <from>
operand.

>> The operands are:

 (:reg <n>)

A pseudo register.  <n> is an integer in the range [0 , 31].

A particular implementation can map this to a real register, a memory
location or the stack.  The abstract LAP code itself does not include
the notion of a stack.

PCL will attempt to optimize register use in two ways.  PCL itself will
attempt to re-use registers whenever possible.  That is, the port should
not have to worry with doing live register analysis for the registers.
In addition, PCL will consider lower numbered registers to be "faster"
than higher numbered ones.


 (:cvar <name>)

A closure variable of the lap-closure.  <name> is a symbol.


 (:arg <name>)

An argument to the LAP closure.  <name> is a symbol.

 (:std-wrapper <from>)
 (:fsc-wrapper <from>)
 (:built-in-wrapper <from>)
 (:structure-wrapper <from>)
 (:other-wrapper <from>)

Get the class wrapper of <from>.  For std-instances and fsc-instances
this just fetches the wrapper field.  The specific port is required to
implement fast access to the wrappers of built-in, structure and other
metatypes.  A callback mechanism allows the port to ask PCL to generate
a class and wrapper for objects for which no class and wrapper exists
yet.  This mechanism is <<to be spec'd out>>.


 (:std-slots <operand>)
 (:fsc-slots <operand>)

Fetch the slots field of a std-instance or a fsc-instance.

 (:constant <constant>)

This just allows inline constants. <constant> can be any Lisp object.

The following operands operate on indexes.  Each is patterned after a
Lisp function which would have a corresponding effect on the index value
of the index.

 (:i1+ <index>)
 (:i+ <index-1> <index-2>)
 (:i- <index-1> <index-2>)
 (:ilogand <index-1> <index-2>)
 (:ilogxor <index-1> <index-2>)

Like the corresponding Lisp functions.


 (:iref <vector> <index>)

Like the SVREF function.  <vector> must be a simple vector.

 (:cref <vector> <constant>)

The :cref operand is for constant vector references.  <constant> must be
a fixnum.

>> The opcodes are:

 (:move <from> <to>)

A full word move operation.


 (:eq <from1> <from2> <label>)
 (:neq <from1> <from2> <label>)

EQ and NOT EQ conditional branches.  If the value contained in <from1>
is EQ (or not) to the value contained in <from2>, jump to <label>.
Otherwise continue execution at the next opcode.  <label> is a symbol.

 (:fix= <from1> <from2> <label>)

A fixnum = conditional branch.

 (:izerop <from> <label>)

Jump to label if and only if the index <from> is 0.

 (:std-instance-p <from> <destination-label>)
 (:fsc-instance-p <from> <destination-label>)
 (:built-in-instance-p <from> <destination-label>)
 (:structure-instance-p <from> <destination-label>)

Test the metatype of <from> and dispatch.  If the metatype of from is of
the specified type execution jumps to <destination-label>.  Otherwise,
execution proceeds normally at the next opcode.  See the :xxx-wrapper
operands.

 (:jmp <function>)

fcn contains a function to call.  This must be a compiled function,
which had the PCL-FAST-CALL declaration in it.  The call should be "tail
recursive" in that whatever values it returns should be returned
immediately from the LAP closure itself.

Method lookup is implemented by finding a function in the cache and then
using :JMP to call it.  The various kinds of traps are implemented by
using :JMP to call a trap function that was stored in one of the closure
variables.

 (:return <value>)

Return immediately from the LAP closure.  <value> is the single value to
return.

In certain cases of method lookup when all the methods are accessor methods,
there is an important optimization which implements the slot access in the
discriminating function itself.  This opcode is used in that case.

 (:label <label>)

Identifies a point in the lap code.  <label> is a symbol.  This can be
the target of any of the control transfer opcodes (:GO, :EQ, :NEQ,
:FIX=, :STD-INSTANCE-P, :FSC-INSTANCE-P, :STRUCTURE-INSTANCE-P,
:BUILT-IN-INSTANCE-P)

 (:go <label>)

Jump to the label <label>.  <label> is a symbol.

*** Examples ***

Here is an example of the use of the abstract LAP mechanism.  It doesn't
use all operands or opcodes, but it is representative of the LAP
sequences PCL will use.

Several things are worth noting:
  * This is, I believe, just about the simplest such sequence.  There
    are some others of comparable simplicity, but none much simpler.

  * A total of only 5 registers are used.  I haven't checked, but I
    am pretty sure most all such code sequences will get by with 16
    or less and many will stay under 8.

(defun make-1-arg-std-dfun ()
  (let ((cg
      (making-lap-closure-generator
        (initialize-lap-cvars '(cache mask reflect trap))
        (initialize-lap-args '(a0))
        (initialize-lap-registers 4 4 3)
        (let ((cache (allocate-lap-register 'simple-vector))
          (count (allocate-lap-register))
          (wrapper (allocate-lap-register))
          (index (allocate-lap-register 'index))
          (t1 (allocate-lap-register))
          (t2 (allocate-lap-register))
          (i1 (allocate-lap-register 'index))
          (i2 (allocate-lap-register 'index)))

          (opcode :move (operand :cvar 'cache) cache)
          (opcode :move (operand :arg 'a0) t1)
          (opcode :std-instance-p t1 'std-instance)
          (opcode :go 'trap)
          (opcode :label 'std-instance)
          ;;
          ;; The stable register pattern for the rest of the code is:
          ;;   cache    Cache
          ;;   count    Cache count
          ;;   wrapper  Wrapper
          ;;   index    index under consideration
          ;;
          ;; Whenever we jump to HIT, this pattern must be in force.
          ;;
          (opcode :move (operand :std-wrapper t1) wrapper);
          (opcode :move (operand :cref cache 0) count)    ;
          (opcode :move (operand :cref wrapper 0) i2)     ;wrapper-no -> i2
          (opcode :move (operand :cvar 'mask) i1)          ;mask       -> i1
          (opcode :move (operand :ilogand i1 i2) index)   ;
                                      ;
          (opcode :move (operand :iref cache index) t1)   ;key        -> t1
          (opcode :eq t1 wrapper 'hit)                    ;
          (opcode :move (operand :cvar 'reflect) i1)      ;reflect    -> i1
          (opcode :move index i2)                      ;index      -> i2
          (opcode :move (operand :i- i1 i2) index)          ;
          (opcode :move (operand :iref cache index) t1)   ;key        -> t1
          (opcode :eq t1 wrapper 'hit)                    ;
          (opcode :go 'trap)

          ;;
          ;; To do the hit, we use registers as follows:
          ;;   0   cache comes in here
          ;;   1   cache count comes in here
          ;;   2   use this for the function
          ;;   3   index comes in here
          ;;   4   1+ index, then new count
          ;;
          (opcode :label 'hit)
          (opcode :move (operand :i1+ index) i1)
          (opcode :move (operand :iref cache i1) t1)

          (opcode :move (operand :cref cache 0) t2)
          (opcode :fix= count t2 'call)
          (opcode :go 'trap)

          (opcode :label 'call)
          (opcode :jmp t1)

          (opcode :label 'trap)
          (opcode :move (operand :cvar 'trap) t1)
          (opcode :jmp t1)))))

    (funcall cg
         (make-array 16)
         (make-index-mask 16 2)
         (- 16 2)
         #'(lambda (a)
         (declare (pcl-fast-call) (ignore a))
         (break "Trap")))))