guile/doc/ref/vm.texi

@c -*-texinfo-*-
@c This is part of the GNU Guile Reference Manual.
@c Copyright (C)  2008,2009,2010
@c   Free Software Foundation, Inc.
@c See the file guile.texi for copying conditions.

@node A Virtual Machine for Guile
@section A Virtual Machine for Guile

Guile has both an interpreter and a compiler. To a user, the difference
is transparent---interpreted and compiled procedures can call each other
as they please.

The difference is that the compiler creates and interprets bytecode
for a custom virtual machine, instead of interpreting the
S-expressions directly. Loading and running compiled code is faster
than loading and running source code.

The virtual machine that does the bytecode interpretation is a part of
Guile itself. This section describes the nature of Guile's virtual
machine.

@menu
* Why a VM?::                   
* VM Concepts::                 
* Stack Layout::                
* Variables and the VM::                   
* VM Programs::         
* Instruction Set::
@end menu

@node Why a VM?
@subsection Why a VM?

@cindex interpreter
For a long time, Guile only had an interpreter. Guile's interpreter
operated directly on the S-expression representation of Scheme source
code.

But while the interpreter was highly optimized and hand-tuned, it still
performs many needless computations during the course of evaluating an
expression. For example, application of a function to arguments
needlessly consed up the arguments in a list. Evaluation of an
expression always had to figure out what the car of the expression is --
a procedure, a memoized form, or something else. All values have to be
allocated on the heap. Et cetera.

The solution to this problem was to compile the higher-level language,
Scheme, into a lower-level language for which all of the checks and
dispatching have already been done---the code is instead stripped to
the bare minimum needed to ``do the job''.

The question becomes then, what low-level language to choose? There
are many options. We could compile to native code directly, but that
poses portability problems for Guile, as it is a highly cross-platform
project.

So we want the performance gains that compilation provides, but we
also want to maintain the portability benefits of a single code path.
The obvious solution is to compile to a virtual machine that is
present on all Guile installations.

The easiest (and most fun) way to depend on a virtual machine is to
implement the virtual machine within Guile itself. This way the
virtual machine provides what Scheme needs (tail calls, multiple
values, @code{call/cc}) and can provide optimized inline instructions
for Guile (@code{cons}, @code{struct-ref}, etc.).

So this is what Guile does. The rest of this section describes that VM
that Guile implements, and the compiled procedures that run on it.

Before moving on, though, we should note that though we spoke of the
interpreter in the past tense, Guile still has an interpreter. The
difference is that before, it was Guile's main evaluator, and so was
implemented in highly optimized C; now, it is actually implemented in
Scheme, and compiled down to VM bytecode, just like any other program.
(There is still a C interpreter around, used to bootstrap the compiler,
but it is not normally used at runtime.)

The upside of implementing the interpreter in Scheme is that we preserve
tail calls and multiple-value handling between interpreted and compiled
code. The downside is that the interpreter in Guile 2.0 is slower than
the interpreter in 1.8. We hope the that the compiler's speed makes up
for the loss!

Also note that this decision to implement a bytecode compiler does not
preclude native compilation. We can compile from bytecode to native
code at runtime, or even do ahead of time compilation. More
possibilities are discussed in @ref{Extending the Compiler}.

@node VM Concepts
@subsection VM Concepts

Compiled code is run by a virtual machine (VM). Each thread has its own
VM. When a compiled procedure is run, Guile looks up the virtual machine
for the current thread and executes the procedure using that VM.

Guile's virtual machine is a stack machine---that is, it has few
registers, and the instructions defined in the VM operate by pushing
and popping values from a stack.

Stack memory is exclusive to the virtual machine that owns it. In
addition to their stacks, virtual machines also have access to the
global memory (modules, global bindings, etc) that is shared among
other parts of Guile, including other VMs.

A VM has generic instructions, such as those to reference local
variables, and instructions designed to support Guile's languages --
mathematical instructions that support the entire numerical tower, an
inlined implementation of @code{cons}, etc.

The registers that a VM has are as follows:

@itemize
@item ip - Instruction pointer
@item sp - Stack pointer
@item fp - Frame pointer
@end itemize

In other architectures, the instruction pointer is sometimes called
the ``program counter'' (pc). This set of registers is pretty typical
for stack machines; their exact meanings in the context of Guile's VM
are described in the next section.

@c wingo: The following is true, but I don't know in what context to
@c describe it. A documentation FIXME.

@c A VM may have one of three engines: reckless, regular, or debugging.
@c Reckless engine is fastest but dangerous.  Regular engine is normally
@c fail-safe and reasonably fast.  Debugging engine is safest and
@c functional but very slow.

@c (Actually we have just a regular and a debugging engine; normally
@c we use the latter, it's almost as fast as the ``regular'' engine.)

@node Stack Layout
@subsection Stack Layout

While not strictly necessary to understand how to work with the VM, it
is instructive and sometimes entertaining to consider the structure of
the VM stack.

Logically speaking, a VM stack is composed of ``frames''. Each frame
corresponds to the application of one compiled procedure, and contains
storage space for arguments, local variables, intermediate values, and
some bookkeeping information (such as what to do after the frame
computes its value).

While the compiler is free to do whatever it wants to, as long as the
semantics of a computation are preserved, in practice every time you
call a function, a new frame is created. (The notable exception of
course is the tail call case, @pxref{Tail Calls}.)

Within a frame, you have the data associated with the function
application itself, which is of a fixed size, and the stack space for
intermediate values. Sometimes only the former is referred to as the
``frame'', and the latter is the ``stack'', although all pending
application frames can have some intermediate computations interleaved
on the stack.

The structure of the fixed part of an application frame is as follows:

@example
             Stack
   | ...              |
   | Intermed. val. 0 | <- fp + bp->nargs + bp->nlocs = SCM_FRAME_UPPER_ADDRESS (fp)
   +==================+
   | Local variable 1 |
   | Local variable 0 | <- fp + bp->nargs
   | Argument 1       |
   | Argument 0       | <- fp
   | Program          | <- fp - 1
   +------------------+    
   | Return address   |
   | MV return address|
   | Dynamic link     | <- fp - 4 = SCM_FRAME_DATA_ADDRESS (fp) = SCM_FRAME_LOWER_ADDRESS (fp)
   +==================+
   |                  |
@end example

In the above drawing, the stack grows upward. The intermediate values
stored in the application of this frame are stored above
@code{SCM_FRAME_UPPER_ADDRESS (fp)}. @code{bp} refers to the
@code{struct scm_objcode} data associated with the program at
@code{fp - 1}. @code{nargs} and @code{nlocs} are properties of the
compiled procedure, which will be discussed later.

The individual fields of the frame are as follows:

@table @asis
@item Return address
The @code{ip} that was in effect before this program was applied. When
we return from this activation frame, we will jump back to this
@code{ip}.

@item MV return address
The @code{ip} to return to if this application returns multiple
values. For continuations that only accept one value, this value will
be @code{NULL}; for others, it will be an @code{ip} that points to a
multiple-value return address in the calling code. That code will
expect the top value on the stack to be an integer---the number of
values being returned---and that below that integer there are the
values being returned.

@item Dynamic link
This is the @code{fp} in effect before this program was applied. In
effect, this and the return address are the registers that are always
``saved''. The dynamic link links the current frame to the previous
frame; computing a stack trace involves traversing these frames.

@item Local variable @var{n}
Lambda-local variables that are all allocated as part of the frame.
This makes access to variables very cheap.

@item Argument @var{n}
The calling convention of the VM requires arguments of a function
application to be pushed on the stack, and here they are. References
to arguments dispatch to these locations on the stack.

@item Program
This is the program being applied. For more information on how
programs are implemented, @xref{VM Programs}.
@end table

@node Variables and the VM
@subsection Variables and the VM

Consider the following Scheme code as an example:

@example
  (define (foo a)
    (lambda (b) (list foo a b)))
@end example

Within the lambda expression, @code{foo} is a top-level variable, @code{a} is a
lexically captured variable, and @code{b} is a local variable.

Another way to refer to @code{a} and @code{b} is to say that @code{a}
is a ``free'' variable, since it is not defined within the lambda, and
@code{b} is a ``bound'' variable. These are the terms used in the
@dfn{lambda calculus}, a mathematical notation for describing
functions. The lambda calculus is useful because it allows one to
prove statements about functions. It is especially good at describing
scope relations, and it is for that reason that we mention it here.

Guile allocates all variables on the stack. When a lexically enclosed
procedure with free variables---a @dfn{closure}---is created, it copies
those variables into its free variable vector. References to free
variables are then redirected through the free variable vector.

If a variable is ever @code{set!}, however, it will need to be
heap-allocated instead of stack-allocated, so that different closures
that capture the same variable can see the same value. Also, this
allows continuations to capture a reference to the variable, instead
of to its value at one point in time. For these reasons, @code{set!}
variables are allocated in ``boxes''---actually, in variable cells.
@xref{Variables}, for more information. References to @code{set!}
variables are indirected through the boxes.

Thus perhaps counterintuitively, what would seem ``closer to the
metal'', viz @code{set!}, actually forces an extra memory allocation
and indirection.

Going back to our example, @code{b} may be allocated on the stack, as
it is never mutated.

@code{a} may also be allocated on the stack, as it too is never
mutated. Within the enclosed lambda, its value will be copied into
(and referenced from) the free variables vector.

@code{foo} is a top-level variable, because @code{foo} is not
lexically bound in this example.

@node VM Programs
@subsection Compiled Procedures are VM Programs

By default, when you enter in expressions at Guile's REPL, they are
first compiled to VM object code, then that VM object code is executed
to produce a value. If the expression evaluates to a procedure, the
result of this process is a compiled procedure.

A compiled procedure is a compound object, consisting of its bytecode,
a reference to any captured lexical variables, an object array, and
some metadata such as the procedure's arity, name, and documentation.
You can pick apart these pieces with the accessors in @code{(system vm
program)}. @xref{Compiled Procedures}, for a full API reference.

@cindex object table
@cindex object array
The object array of a compiled procedure, also known as the
@dfn{object table}, holds all Scheme objects whose values are known
not to change across invocations of the procedure: constant strings,
symbols, etc. The object table of a program is initialized right
before a program is loaded with @code{load-program}.
@xref{Loading Instructions}, for more information.

Variable objects are one such type of constant object: when a global
binding is defined, a variable object is associated to it and that
object will remain constant over time, even if the value bound to it
changes. Therefore, toplevel bindings only need to be looked up once.
Thereafter, references to the corresponding toplevel variables from
within the program are then performed via the @code{toplevel-ref}
instruction, which uses the object vector, and are almost as fast as
local variable references.

We can see how these concepts tie together by disassembling the
@code{foo} function we defined earlier to see what is going on:

@smallexample
scheme@@(guile-user)> (define (foo a) (lambda (b) (list foo a b)))
scheme@@(guile-user)> ,x foo
Disassembly of #<procedure foo (a)>:

   0    (assert-nargs-ee 0 1)           
   3    (reserve-locals 0 1)            
   6    (object-ref 1)                  ;; #<procedure 85bfec0 at <current input>:0:16 (b)>
   8    (local-ref 0)                   ;; `a'
  10    (make-closure 0 1)              
  13    (return)                        

----------------------------------------
Disassembly of #<procedure 85bfec0 at <current input>:0:16 (b)>:

   0    (assert-nargs-ee 0 1)           
   3    (reserve-locals 0 1)            
   6    (toplevel-ref 1)                ;; `foo'
   8    (free-ref 0)                    ;; (closure variable)
  10    (local-ref 0)                   ;; `b'
  12    (list 0 3)                      ;; 3 elements         at (unknown file):0:28
  15    (return)                        
@end smallexample

First there's some prelude, where @code{foo} checks that it was called with only
1 argument. Then at @code{ip} 6, we load up the compiled lambda. @code{Ip} 8
loads up `a', so that it can be captured into a closure by at @code{ip}
10---binding code (from the compiled lambda) with data (the free-variable
vector). Finally we return the closure.

The second stanza disassembles the compiled lambda. After the prelude, we note
that toplevel variables are resolved relative to the module that was current
when the procedure was created. This lookup occurs lazily, at the first time the
variable is actually referenced, and the location of the lookup is cached so
that future references are very cheap. @xref{Environment Control Instructions},
for more details.

Then we see a reference to a free variable, corresponding to @code{a}. The
disassembler doesn't have enough information to give a name to that variable, so
it just marks it as being a ``closure variable''. Finally we see the reference
to @code{b}, then the @code{list} opcode, an inline implementation of the
@code{list} scheme routine.

@node Instruction Set
@subsection Instruction Set

There are about 150 instructions in Guile's virtual machine. These
instructions represent atomic units of a program's execution. Ideally,
they perform one task without conditional branches, then dispatch to
the next instruction in the stream.

Instructions themselves are one byte long. Some instructions take
parameters, which follow the instruction byte in the instruction
stream.

Sometimes the compiler can figure out that it is compiling a special
case that can be run more efficiently. So, for example, while Guile
offers a generic test-and-branch instruction, it also offers specific
instructions for special cases, so that the following cases all have
their own test-and-branch instructions:

@example
(if pred then else)
(if (not pred) then else)
(if (null? l) then else)
(if (not (null? l)) then else)
@end example

In addition, some Scheme primitives have their own inline
implementations, e.g. @code{cons}, and @code{list}, as we saw in the
previous section.

So Guile's instruction set is a @emph{complete} instruction set, in
that it provides the instructions that are suited to the problem, and
is not concerned with making a minimal, orthogonal set of
instructions. More instructions may be added over time.

@menu
* Environment Control Instructions::  
* Branch Instructions::         
* Loading Instructions::  
* Procedural Instructions::  
* Data Control Instructions::   
* Miscellaneous Instructions::  
* Inlined Scheme Instructions::  
* Inlined Mathematical Instructions::  
* Inlined Bytevector Instructions::  
@end menu

@node Environment Control Instructions
@subsubsection Environment Control Instructions

These instructions access and mutate the environment of a compiled
procedure---the local bindings, the free (captured) bindings, and the
toplevel bindings.

Some of these instructions have @code{long-} variants, the difference
being that they take 16-bit arguments, encoded in big-endianness,
instead of the normal 8-bit range.

@deffn Instruction local-ref index
@deffnx Instruction long-local-ref index
Push onto the stack the value of the local variable located at
@var{index} within the current stack frame.

Note that arguments and local variables are all in one block. Thus the
first argument, if any, is at index 0, and local bindings follow the
arguments.
@end deffn

@deffn Instruction local-set index
@deffnx Instruction long-local-ref index
Pop the Scheme object located on top of the stack and make it the new
value of the local variable located at @var{index} within the current
stack frame.
@end deffn

@deffn Instruction free-ref index
Push the value of the captured variable located at position
@var{index} within the program's vector of captured variables.
@end deffn

@deffn Instruction free-boxed-ref index
@deffnx Instruction free-boxed-set index
Get or set a boxed free variable. Note that there is no free-set
instruction, as variables that are @code{set!} must be boxed.

These instructions assume that the value at position @var{index} in
the free variables vector is a variable.
@end deffn

@deffn Instruction make-closure
Pop a vector and a program object off the stack, in that order, and
push a new program object with the given free variables vector. The
new program object shares state with the original program.

At the time of this writing, the space overhead of closures is 4 words
per closure.
@end deffn

@deffn Instruction fix-closure index
Pop a vector off the stack, and set it as the @var{index}th local
variable's free variable vector. The @var{index}th local variable is
assumed to be a procedure.

This instruction is part of a hack for allocating mutually recursive
procedures. The hack is to first perform a @code{local-set} for all of
the recursive procedures, then fix up the procedures' free variable
bindings in place. This allows most @code{letrec}-bound procedures to
be allocated unboxed on the stack.

One could of course do a @code{local-ref}, then @code{make-closure},
then @code{local-set}, but this macroinstruction helps to speed up the
common case.
@end deffn

@deffn Instruction box index
Pop a value off the stack, and set the @var{index}nth local variable
to a box containing that value. A shortcut for @code{make-variable}
then @code{local-set}, used when binding boxed variables.
@end deffn

@deffn Instruction empty-box index
Set the @var{indext}h local variable to a box containing a variable
whose value is unbound. Used when compiling some @code{letrec}
expressions.
@end deffn

@deffn Instruction toplevel-ref index
@deffnx Instruction long-toplevel-ref index
Push the value of the toplevel binding whose location is stored in at
position @var{index} in the object table.

Initially, a cell in the object table that is used by
@code{toplevel-ref} is initialized to one of two forms. The normal
case is that the cell holds a symbol, whose binding will be looked up
relative to the module that was current when the current program was
created.

Alternately, the lookup may be performed relative to a particular
module, determined at compile-time (e.g. via @code{@@} or
@code{@@@@}). In that case, the cell in the object table holds a list:
@code{(@var{modname} @var{sym} @var{public?})}. The symbol @var{sym}
will be looked up in the module named @var{modname} (a list of
symbols). The lookup will be performed against the module's public
interface, unless @var{public?} is @code{#f}, which it is for example
when compiling @code{@@@@}.

In any case, if the symbol is unbound, an error is signalled.
Otherwise the initial form is replaced with the looked-up variable, an
in-place mutation of the object table. This mechanism provides for
lazy variable resolution, and an important cached fast-path once the
variable has been successfully resolved.

This instruction pushes the value of the variable onto the stack.
@end deffn

@deffn Instruction toplevel-set index
@deffnx Instruction long-toplevel-set index
Pop a value off the stack, and set it as the value of the toplevel
variable stored at @var{index} in the object table. If the variable
has not yet been looked up, we do the lookup as in
@code{toplevel-ref}.
@end deffn

@deffn Instruction define
Pop a symbol and a value from the stack, in that order. Look up its
binding in the current toplevel environment, creating the binding if
necessary. Set the variable to the value.
@end deffn

@deffn Instruction link-now
Pop a value, @var{x}, from the stack. Look up the binding for @var{x},
according to the rules for @code{toplevel-ref}, and push that variable
on the stack. If the lookup fails, an error will be signalled.

This instruction is mostly used when loading programs, because it can
do toplevel variable lookups without an object vector.
@end deffn

@deffn Instruction variable-ref
Dereference the variable object which is on top of the stack and
replace it by the value of the variable it represents.
@end deffn

@deffn Instruction variable-set
Pop off two objects from the stack, a variable and a value, and set
the variable to the value.
@end deffn

@deffn Instruction make-variable
Replace the top object on the stack with a variable containing it.
Used in some circumstances when compiling @code{letrec} expressions.
@end deffn

@deffn Instruction object-ref n
@deffnx Instruction long-object-ref n
Push @var{n}th value from the current program's object vector. The
``long'' variant has a 16-bit index instead of an 8-bit index.
@end deffn

@node Branch Instructions
@subsubsection Branch Instructions

All the conditional branch instructions described below work in the
same way:

@itemize
@item They pop off Scheme object(s) located on the stack for use in the
branch condition
@item If the condition is true, then the instruction pointer is
increased by the offset passed as an argument to the branch
instruction;
@item Program execution proceeds with the next instruction (that is,
the one to which the instruction pointer points).
@end itemize

Note that the offset passed to the instruction is encoded as three 8-bit
integers, in big-endian order, effectively giving Guile a 24-bit
relative address space.

@deffn Instruction br offset
Jump to @var{offset}. No values are popped.
@end deffn

@deffn Instruction br-if offset
Jump to @var{offset} if the object on the stack is not false.
@end deffn

@deffn Instruction br-if-not offset
Jump to @var{offset} if the object on the stack is false.
@end deffn

@deffn Instruction br-if-eq offset
Jump to @var{offset} if the two objects located on the stack are
equal in the sense of @var{eq?}.  Note that, for this instruction, the
stack pointer is decremented by two Scheme objects instead of only
one.
@end deffn

@deffn Instruction br-if-not-eq offset
Same as @var{br-if-eq} for non-@code{eq?} objects.
@end deffn

@deffn Instruction br-if-null offset
Jump to @var{offset} if the object on the stack is @code{'()}.
@end deffn

@deffn Instruction br-if-not-null offset
Jump to @var{offset} if the object on the stack is not @code{'()}.
@end deffn


@node Loading Instructions
@subsubsection Loading Instructions

In addition to VM instructions, an instruction stream may contain
variable-length data embedded within it. This data is always preceded
by special loading instructions, which interpret the data and advance
the instruction pointer to the next VM instruction.

All of these loading instructions have a @code{length} parameter,
indicating the size of the embedded data, in bytes. The length itself
is encoded in 3 bytes.

@deffn Instruction load-number length
Load an arbitrary number from the instruction stream. The number is
embedded in the stream as a string.
@end deffn
@deffn Instruction load-string length
Load a string from the instruction stream. The string is assumed to be
encoded in the ``latin1'' locale.
@end deffn
@deffn Instruction load-wide-string length
Load a UTF-32 string from the instruction stream. @var{length} is the
length in bytes, not in codepoints
@end deffn
@deffn Instruction load-symbol length
Load a symbol from the instruction stream. The symbol is assumed to be
encoded in the ``latin1'' locale. Symbols backed by wide strings may
be loaded via @code{load-wide-string} then @code{make-symbol}.
@end deffn
@deffn Instruction load-array length
Load a uniform array from the instruction stream. The shape and type
of the array are popped off the stack, in that order.
@end deffn

@deffn Instruction load-program
Load bytecode from the instruction stream, and push a compiled
procedure.

This instruction pops one value from the stack: the program's object
table, as a vector, or @code{#f} in the case that the program has no
object table. A program that does not reference toplevel bindings and
does not use @code{object-ref} does not need an object table.

This instruction is unlike the rest of the loading instructions,
because instead of parsing its data, it directly maps the instruction
stream onto a C structure, @code{struct scm_objcode}. @xref{Bytecode
and Objcode}, for more information.

The resulting compiled procedure will not have any free variables
captured, so it may be loaded only once but used many times to create
closures.
@end deffn

@node Procedural Instructions
@subsubsection Procedural Instructions

@deffn Instructions new-frame
Push a new frame on the stack, reserving space for the dynamic link,
return address, and the multiple-values return address. The frame
pointer is not yet updated, because the frame is not yet active -- it
has to be patched by a @code{call} instruction to get the return
address.
@end deffn

@deffn Instruction call nargs
Call the procedure located at @code{sp[-nargs]} with the @var{nargs}
arguments located from @code{sp[-nargs + 1]} to @code{sp[0]}.

This instruction requires that a new frame be pushed on the stack
before the procedure, via @code{new-frame}. @xref{Stack Layout}, for
more information. It patches up that frame with the current @code{ip}
as the return address, then dispatches to the first instruction in the
called procedure, relying on the called procedure to return one value
to the newly-created continuation. Because the new frame pointer will
point to sp[-nargs + 1], the arguments don't have to be shuffled
around -- they are already in place.

For non-compiled procedures (continuations, primitives, and
interpreted procedures), @code{call} will pop the frame, procedure,
and arguments off the stack, and push the result of calling
@code{scm_apply}.
@end deffn

@deffn Instruction tail-call nargs
Like @code{call}, but reusing the current continuation. This
instruction implements tail calls as required by RnRS.

For compiled procedures, that means that @code{tail-call} simply
shuffles down the procedure and arguments to the current stack frame.

For non-VM procedures, the result is the same, but the current VM
invocation remains on the C stack. True tail calls are not currently
possible between compiled and non-compiled procedures.
@end deffn

@deffn Instruction apply nargs
@deffnx Instruction tail-apply nargs
Like @code{call} and @code{tail-call}, except that the top item on the
stack must be a list. The elements of that list are then pushed on the
stack and treated as additional arguments, replacing the list itself,
then the procedure is invoked as usual.
@end deffn

@deffn Instruction call/nargs
@deffnx Instruction tail-call/nargs
These are like @code{call} and @code{tail-call}, except they take the
number of arguments from the stack instead of the instruction stream.
These instructions are used in the implementation of multiple value
returns, where the actual number of values is pushed on the stack.
@end deffn

@deffn Instruction mv-call nargs offset
Like @code{call}, except that a multiple-value continuation is created
in addition to a single-value continuation.

The offset (a two-byte value) is an offset within the instruction
stream; the multiple-value return address in the new frame
(@pxref{Stack Layout}) will be set to the normal return address plus
this offset. Instructions at that offset will expect the top value of
the stack to be the number of values, and below that values
themselves, pushed separately.
@end deffn

@deffn Instruction return
Free the program's frame, returning the top value from the stack to
the current continuation. (The stack should have exactly one value on
it.)

Specifically, the @code{sp} is decremented to one below the current
@code{fp}, the @code{ip} is reset to the current return address, the
@code{fp} is reset to the value of the current dynamic link, and then
the top item on the stack (formerly the procedure being applied) is
set to the returned value.
@end deffn

@deffn Instruction return/values nvalues
Return the top @var{nvalues} to the current continuation.

If the current continuation is a multiple-value continuation,
@code{return/values} pushes the number of values on the stack, then
returns as in @code{return}, but to the multiple-value return address.

Otherwise if the current continuation accepts only one value, i.e. the
multiple-value return address is @code{NULL}, then we assume the user
only wants one value, and we give them the first one. If there are no
values, an error is signaled.
@end deffn

@deffn Instruction return/values* nvalues
Like a combination of @code{apply} and @code{return/values}, in which
the top value on the stack is interpreted as a list of additional
values. This is an optimization for the common @code{(apply values
...)} case.
@end deffn

@deffn Instruction truncate-values nbinds nrest
Used in multiple-value continuations, this instruction takes the
values that are on the stack (including the number-of-values marker)
and truncates them for a binding construct.

For example, a call to @code{(receive (x y . z) (foo) ...)} would,
logically speaking, pop off the values returned from @code{(foo)} and
push them as three values, corresponding to @code{x}, @code{y}, and
@code{z}. In that case, @var{nbinds} would be 3, and @var{nrest} would
be 1 (to indicate that one of the bindings was a rest argument).

Signals an error if there is an insufficient number of values.
@end deffn

@deffn Instruction call/cc
@deffnx Instruction tail-call/cc
Capture the current continuation, and then call (or tail-call) the
procedure on the top of the stack, with the continuation as the
argument.

@code{call/cc} does not require a @code{new-frame} to be pushed on the
stack, as @code{call} does, because it needs to capture the stack
before the frame is pushed.

Both the VM continuation and the C continuation are captured.
@end deffn

@node Data Control Instructions
@subsubsection Data Control Instructions

These instructions push simple immediate values onto the stack, or
manipulate lists and vectors on the stack.

@deffn Instruction make-int8 value
Push @var{value}, an 8-bit integer, onto the stack.
@end deffn

@deffn Instruction make-int8:0
Push the immediate value @code{0} onto the stack.
@end deffn

@deffn Instruction make-int8:1
Push the immediate value @code{1} onto the stack.
@end deffn

@deffn Instruction make-int16 value
Push @var{value}, a 16-bit integer, onto the stack.
@end deffn

@deffn Instruction make-uint64 value
Push @var{value}, an unsigned 64-bit integer, onto the stack. The
value is encoded in 8 bytes, most significant byte first (big-endian).
@end deffn

@deffn Instruction make-int64 value
Push @var{value}, a signed 64-bit integer, onto the stack. The value
is encoded in 8 bytes, most significant byte first (big-endian), in
twos-complement arithmetic.
@end deffn

@deffn Instruction make-false
Push @code{#f} onto the stack.
@end deffn

@deffn Instruction make-true
Push @code{#t} onto the stack.
@end deffn

@deffn Instruction make-nil
Push @code{%nil} onto the stack.
@end deffn

@deffn Instruction make-eol
Push @code{'()} onto the stack.
@end deffn

@deffn Instruction make-char8 value
Push @var{value}, an 8-bit character, onto the stack.
@end deffn

@deffn Instruction make-char32 value
Push @var{value}, an 32-bit character, onto the stack. The value is
encoded in big-endian order.
@end deffn

@deffn Instruction make-symbol
Pops a string off the stack, and pushes a symbol.
@end deffn

@deffn Instruction make-keyword value
Pops a symbol off the stack, and pushes a keyword.
@end deffn

@deffn Instruction list n
Pops off the top @var{n} values off of the stack, consing them up into
a list, then pushes that list on the stack. What was the topmost value
will be the last element in the list. @var{n} is a two-byte value,
most significant byte first.
@end deffn

@deffn Instruction vector n
Create and fill a vector with the top @var{n} values from the stack,
popping off those values and pushing on the resulting vector. @var{n}
is a two-byte value, like in @code{vector}.
@end deffn

@node Miscellaneous Instructions
@subsubsection Miscellaneous Instructions

@deffn Instruction nop
Does nothing! Used for padding other instructions to certain
alignments.
@end deffn

@deffn Instruction halt
Exits the VM, returning a SCM value. Normally, this instruction is
only part of the ``bootstrap program'', a program run when a virtual
machine is first entered; compiled Scheme procedures will not contain
this instruction.

If multiple values have been returned, the SCM value will be a
multiple-values object (@pxref{Multiple Values}).
@end deffn

@deffn Instruction break
Does nothing, but invokes the break hook.
@end deffn

@deffn Instruction drop
Pops off the top value from the stack, throwing it away.
@end deffn

@deffn Instruction dup
Re-pushes the top value onto the stack.
@end deffn

@deffn Instruction void
Pushes ``the unspecified value'' onto the stack.
@end deffn

@node Inlined Scheme Instructions
@subsubsection Inlined Scheme Instructions

The Scheme compiler can recognize the application of standard Scheme
procedures. It tries to inline these small operations to avoid the
overhead of creating new stack frames.

Since most of these operations are historically implemented as C
primitives, not inlining them would entail constantly calling out from
the VM to the interpreter, which has some costs---registers must be
saved, the interpreter has to dispatch, called procedures have to do
much typechecking, etc. It's much more efficient to inline these
operations in the virtual machine itself.

All of these instructions pop their arguments from the stack and push
their results, and take no parameters from the instruction stream.
Thus, unlike in the previous sections, these instruction definitions
show stack parameters instead of parameters from the instruction
stream.

@deffn Instruction not x
@deffnx Instruction not-not x
@deffnx Instruction eq? x y
@deffnx Instruction not-eq? x y
@deffnx Instruction null?
@deffnx Instruction not-null?
@deffnx Instruction eqv? x y
@deffnx Instruction equal? x y
@deffnx Instruction pair? x y
@deffnx Instruction list? x
@deffnx Instruction set-car! pair x
@deffnx Instruction set-cdr! pair x
@deffnx Instruction slot-ref struct n
@deffnx Instruction slot-set struct n x
@deffnx Instruction cons x y
@deffnx Instruction car x
@deffnx Instruction cdr x
@deffnx Instruction vector-ref x y
@deffnx Instruction vector-set x n y
Inlined implementations of their Scheme equivalents.
@end deffn

Note that @code{caddr} and friends compile to a series of @code{car}
and @code{cdr} instructions.

@node Inlined Mathematical Instructions
@subsubsection Inlined Mathematical Instructions

Inlining mathematical operations has the obvious advantage of handling
fixnums without function calls or allocations. The trick, of course,
is knowing when the result of an operation will be a fixnum, and there
might be a couple bugs here.

More instructions could be added here over time.

As in the previous section, the definitions below show stack
parameters instead of instruction stream parameters.

@deffn Instruction add x y
@deffnx Instruction add1 x
@deffnx Instruction sub x y
@deffnx Instruction sub1 x
@deffnx Instruction mul x y
@deffnx Instruction div x y
@deffnx Instruction quo x y
@deffnx Instruction rem x y
@deffnx Instruction mod x y
@deffnx Instruction ee? x y
@deffnx Instruction lt? x y
@deffnx Instruction gt? x y
@deffnx Instruction le? x y
@deffnx Instruction ge? x y
Inlined implementations of the corresponding mathematical operations.
@end deffn

@node Inlined Bytevector Instructions
@subsubsection Inlined Bytevector Instructions

Bytevector operations correspond closely to what the current hardware
can do, so it makes sense to inline them to VM instructions, providing
a clear path for eventual native compilation. Without this, Scheme
programs would need other primitives for accessing raw bytes -- but
these primitives are as good as any.

As in the previous section, the definitions below show stack
parameters instead of instruction stream parameters.

The multibyte formats (@code{u16}, @code{f64}, etc) take an extra
endianness argument. Only aligned native accesses are currently
fast-pathed in Guile's VM.

@deffn Instruction bv-u8-ref bv n
@deffnx Instruction bv-s8-ref bv n
@deffnx Instruction bv-u16-native-ref bv n
@deffnx Instruction bv-s16-native-ref bv n
@deffnx Instruction bv-u32-native-ref bv n
@deffnx Instruction bv-s32-native-ref bv n
@deffnx Instruction bv-u64-native-ref bv n
@deffnx Instruction bv-s64-native-ref bv n
@deffnx Instruction bv-f32-native-ref bv n
@deffnx Instruction bv-f64-native-ref bv n
@deffnx Instruction bv-u16-ref bv n endianness
@deffnx Instruction bv-s16-ref bv n endianness
@deffnx Instruction bv-u32-ref bv n endianness
@deffnx Instruction bv-s32-ref bv n endianness
@deffnx Instruction bv-u64-ref bv n endianness
@deffnx Instruction bv-s64-ref bv n endianness
@deffnx Instruction bv-f32-ref bv n endianness
@deffnx Instruction bv-f64-ref bv n endianness
@deffnx Instruction bv-u8-set bv n val
@deffnx Instruction bv-s8-set bv n val
@deffnx Instruction bv-u16-native-set bv n val
@deffnx Instruction bv-s16-native-set bv n val
@deffnx Instruction bv-u32-native-set bv n val
@deffnx Instruction bv-s32-native-set bv n val
@deffnx Instruction bv-u64-native-set bv n val
@deffnx Instruction bv-s64-native-set bv n val
@deffnx Instruction bv-f32-native-set bv n val
@deffnx Instruction bv-f64-native-set bv n val
@deffnx Instruction bv-u16-set bv n val endianness
@deffnx Instruction bv-s16-set bv n val endianness
@deffnx Instruction bv-u32-set bv n val endianness
@deffnx Instruction bv-s32-set bv n val endianness
@deffnx Instruction bv-u64-set bv n val endianness
@deffnx Instruction bv-s64-set bv n val endianness
@deffnx Instruction bv-f32-set bv n val endianness
@deffnx Instruction bv-f64-set bv n val endianness
Inlined implementations of the corresponding bytevector operations.
@end deffn