Fix make-array transforms.

[sbcl.git] / doc / manual / efficiency.texinfo

@node Efficiency
@comment  node-name,  next,  previous,  up
@chapter Efficiency
@cindex Efficiency

@menu
* Slot access::
* Dynamic-extent allocation::
* Modular arithmetic::
* Global and Always-Bound variables::
* Miscellaneous Efficiency Issues::
@end menu

@node  Slot access
@comment  node-name,  next,  previous,  up
@section Slot access
@cindex Slot access

@subsection Structure object slot access

Structure slot accessors are efficient only if the compiler is able to
open code them: compiling a call to a structure slot accessor before
the structure is defined, declaring one @code{notinline}, or passing
it as a functional argument to another function causes severe
performance degradation.

@subsection Standard object slot access

The most efficient way to access a slot of a @code{standard-object} is
by using @code{slot-value} with a constant slot name argument inside a
@code{defmethod} body, where the variable holding the instance is a
specializer parameter of the method and is never assigned to. The cost
is roughly 1.6 times that of an open coded structure slot accessor.

Second most efficient way is to use a CLOS slot accessor, or
@code{slot-value} with a constant slot name argument, but in
circumstances other than specified above. This may be up to 3 times as
slow as the method described above.

Example:

@lisp
(defclass foo () ((bar)))

;; Fast: specializer and never assigned to
(defmethod quux ((foo foo) new)
  (let ((old (slot-value foo 'bar)))
    (setf (slot-value foo 'bar) new)
    old))

;; Slow: not a specializer
(defmethod quux ((foo foo) new)
  (let* ((temp foo)
         (old (slot-value temp 'bar)))
    (setf (slot-value temp 'bar) new)
    old))

;; Slow: assignment to FOO
(defmethod quux ((foo foo) new)
  (let ((old (slot-value foo 'bar)))
    (setf (slot-value foo 'bar) new)
    (setf foo new)
    old))
@end lisp

Note that when profiling code such as this, the first few calls to the
generic function are not representative, as the dispatch mechanism is
lazily set up during those calls.

@node  Dynamic-extent allocation
@comment  node-name,  next,  previous,  up
@section Dynamic-extent allocation
@cindex @code{dynamic-extent} declaration
@cindex declaration, @code{dynamic-extent}

SBCL has fairly extensive support for performing allocation on the
stack when a variable is declared @code{dynamic-extent}. The
@code{dynamic-extent} declarations are not verified, but are simply
trusted as long as @code{sb-ext:*stack-allocate-dynamic-extent*} is
true.

@include var-sb-ext-star-stack-allocate-dynamic-extent-star.texinfo

If dynamic extent constraints specified in the Common Lisp standard
are violated, the best that can happen is for the program to have
garbage in variables and return values; more commonly, the system will
crash.

In particular, it is important to realize that dynamic extend is
contagious:

@lisp
(let* ((a (list 1 2 3))
       (b (cons a a)))
   (declare (dynamic-extent b))
   ;; Unless A is accessed elsewhere as well, SBCL will consider
   ;; it to be otherwise inaccessible -- it can only be accessed
   ;; through B, after all -- and stack allocate it as well.
   ;;
   ;; Hence returning (CAR B) here is unsafe.
   ...)
@end lisp

This allows stack allocation of complex structures. As a notable
exception to this, SBCL does not as of 1.0.48.21 propagate
dynamic-extentness through @code{&rest} arguments -- but another
conforming implementation might, so portable code should not rely on
this.

@lisp
(declaim (inline foo))
(defun foo (fun &rest arguments)
  (declare (dynamic-extent arguments))
  (apply fun arguments))

(defun bar (a)
  ;; SBCL will heap allocate the result of (LIST A), and stack allocate
  ;; only the spine of the &rest list -- so this is safe, but unportable.
  ;;
  ;; Another implementation, including earlier versions of SBCL might consider
  ;; (LIST A) to be otherwise inaccessible and stack-allocate it as well!
  (foo #'car (list a)))
@end lisp

There are many cases when @code{dynamic-extent} declarations could be
useful. At present, SBCL implements stack allocation for

@itemize

@item
@code{&rest} lists, when these are declared @code{dynamic-extent}.

@item
@findex @cl{cons}
@findex @cl{list}
@findex @cl{list*}
@findex @cl{vector}
@code{cons}, @code{list}, @code{list*}, and @code{vector} when the
result is bound to a variable declared @code{dynamic-extent}.

@item
@findex @cl{make-array}
simple forms of @code{make-array}, whose result is bound to a variable
declared @code{dynamic-extent}: stack allocation is possible only if
the resulting array is known to be both simple and one-dimensional,
and has a constant @code{:element-type}.

@cindex Safety optimization quality
@strong{Note}: stack space is limited, so allocation of a large vector
may cause stack overflow. For this reason potentially large vectors,
which might circumvent stack overflow detection, are stack allocated
only in zero @code{safety} policies.

@item
@findex @cl{flet}
@findex @cl{labels}
@cindex @code{safety} optimization quality
@cindex optimization quality, @code{safety}
closures defined with @code{flet} or @code{labels}, with a bound
@code{dynamic-extent} declaration. Blocks and tags are also allocated
on the heap, unless all non-local control transfers to them are
compiled with zero @code{safety}.

@item
user-defined structures when the structure constructor defined using
@code{defstruct} has been declared @code{inline} and the result of the
call to the constructor is bound to a variable declared
@code{dynamic-extent}.

@strong{Note}: structures with ``raw'' slots can currently be
stack-allocated only on x86 and x86-64.

@item
all of the above when they appear as initial parts of another
stack-allocated object.

@end itemize

Examples:

@lisp
;;; Declaiming a structure constructor inline before definition makes
;;; stack allocation possible.
(declaim (inline make-thing))
(defstruct thing obj next)

;;; Stack allocation of various objects bound to DYNAMIC-EXTENT
;;; variables.
(let* ((list (list 1 2 3))
       (nested (cons (list 1 2) (list* 3 4 (list 5))))
       (vector (make-array 3 :element-type 'single-float))
       (thing (make-thing :obj list
                          :next (make-thing :obj (make-array 3)))))
  (declare (dynamic-extent list nested vector thing))
  ...)

;;; Stack allocation of arguments to a local function is equivalent
;;; to stack allocation of local variable values.
(flet ((f (x)
         (declare (dynamic-extent x))
         ...))
  ...
  (f (list 1 2 3))
  (f (cons (cons 1 2) (cons 3 4)))
  ...)

;;; Stack allocation of &REST lists
(defun foo (&rest args)
  (declare (dynamic-extent args))
  ...)
@end lisp

Future plans include

@itemize

@item
Automatic detection of the common idiom of calling quantifiers with a
closure, even when the closure is not declared @code{dynamic-extent}.

@end itemize

@node  Modular arithmetic
@comment  node-name,  next,  previous,  up
@section Modular arithmetic
@cindex Modular arithmetic
@cindex Arithmetic, modular
@cindex Arithmetic, hardware
@findex @cl{logand}
Some numeric functions have a property: @var{N} lower bits of the
result depend only on @var{N} lower bits of (all or some)
arguments. If the compiler sees an expression of form @code{(logand
@var{exp} @var{mask})}, where @var{exp} is a tree of such ``good''
functions and @var{mask} is known to be of type @code{(unsigned-byte
@var{w})}, where @var{w} is a ``good'' width, all intermediate results
will be cut to @var{w} bits (but it is not done for variables and
constants!). This often results in an ability to use simple machine
instructions for the functions.

Consider an example.

@lisp
(defun i (x y)
  (declare (type (unsigned-byte 32) x y))
  (ldb (byte 32 0) (logxor x (lognot y))))
@end lisp

The result of @code{(lognot y)} will be negative and of type
@code{(signed-byte 33)}, so a naive implementation on a 32-bit
platform is unable to use 32-bit arithmetic here. But modular
arithmetic optimizer is able to do it: because the result is cut down
to 32 bits, the compiler will replace @code{logxor} and @code{lognot}
with versions cutting results to 32 bits, and because terminals
(here---expressions @code{x} and @code{y}) are also of type
@code{(unsigned-byte 32)}, 32-bit machine arithmetic can be used.

As of SBCL 0.8.5 ``good'' functions are @code{+}, @code{-};
@code{logand}, @code{logior}, @code{logxor}, @code{lognot} and their
combinations; and @code{ash} with the positive second
argument. ``Good'' widths are 32 on HPPA, MIPS, PPC, Sparc and x86 and
64 on Alpha.  While it is possible to support smaller widths as well,
currently this is not implemented.

@node  Global and Always-Bound variables
@comment  node-name,  next,  previous,  up
@section Global and Always-Bound variables

@include macro-sb-ext-defglobal.texinfo

@deffn {Declaration} @sbext{global}

Syntax: @code{(sb-ext:global symbol*)}

Only valid as a global proclamation.

Specifies that the named symbols cannot be proclaimed or locally
declared @code{special}. Proclaiming an already special or constant
variable name as @code{global} signal an error. Allows more efficient
value lookup in threaded environments in addition to expressing
programmer intention.
@end deffn

@deffn {Declaration} @sbext{always-bound}

Syntax: @code{(sb-ext:always-bound symbol*)}

Only valid as a global proclamation.

Specifies that the named symbols are always bound. Inhibits
@code{makunbound} of the named symbols. Proclaiming an unbound symbol
as @code{always-bound} signals an error. Allows the compiler to elide
boundness checks from value lookups.
@end deffn

@node  Miscellaneous Efficiency Issues
@comment  node-name,  next,  previous,  up
@section Miscellaneous Efficiency Issues

FIXME: The material in the CMUCL manual about getting good
performance from the compiler should be reviewed, reformatted in
Texinfo, lightly edited for SBCL, and substituted into this
manual. In the meantime, the original CMUCL manual is still 95+%
correct for the SBCL version of the Python compiler. See the
sections

@itemize
@item Advanced Compiler Use and Efficiency Hints
@item Advanced Compiler Introduction
@item More About Types in Python
@item Type Inference
@item Source Optimization
@item Tail Recursion
@item Local Call
@item Block Compilation
@item Inline Expansion
@item Object Representation
@item Numbers
@item General Efficiency Hints
@item Efficiency Notes
@end itemize

Besides this information from the CMUCL manual, there are a few other
points to keep in mind.

@itemize

@item
@findex @cl{let}
@findex @cl{let*}
@findex @cl{setq}
@findex @cl{setf}
The CMUCL manual doesn't seem to state it explicitly, but Python has a
mental block about type inference when assignment is involved. Python
is very aggressive and clever about inferring the types of values
bound with @code{let}, @code{let*}, inline function call, and so
forth. However, it's much more passive and dumb about inferring the
types of values assigned with @code{setq}, @code{setf}, and
friends. It would be nice to fix this, but in the meantime don't
expect that just because it's very smart about types in most respects
it will be smart about types involved in assignments.  (This doesn't
affect its ability to benefit from explicit type declarations
involving the assigned variables, only its ability to get by without
explicit type declarations.)

@c <!-- FIXME: Python dislikes assignments, but not in type
@c     inference. The real problems are loop induction, closed over
@c     variables and aliases. -->

@item
Since the time the CMUCL manual was written, CMUCL (and thus SBCL) has
gotten a generational garbage collector. This means that there are
some efficiency implications of various patterns of memory usage which
aren't discussed in the CMUCL manual. (Some new material should be
written about this.)

@item
SBCL has some important known efficiency problems.  Perhaps the most
important are

@itemize @minus

@item
The garbage collector is not particularly efficient, at least on
platforms without the generational collector (as of SBCL 0.8.9, all
except x86).

@item
Various aspects of the PCL implementation of CLOS are more inefficient
than necessary.

@end itemize

@end itemize

Finally, note that Common Lisp defines many constructs which, in
the infamous phrase, ``could be compiled efficiently by a
sufficiently smart compiler''. The phrase is infamous because
making a compiler which actually is sufficiently smart to find all
these optimizations systematically is well beyond the state of the art
of current compiler technology. Instead, they're optimized on a
case-by-case basis by hand-written code, or not optimized at all if
the appropriate case hasn't been hand-coded. Some cases where no such
hand-coding has been done as of SBCL version 0.6.3 include

@itemize

@item
@code{(reduce #'f x)} where the type of @code{x} is known at compile
time

@item
various bit vector operations, e.g.  @code{(position 0
some-bit-vector)}

@item
specialized sequence idioms, e.g.  @code{(remove item list :count 1)}

@item
cases where local compilation policy does not require excessive type
checking, e.g.  @code{(locally (declare (safety 1)) (assoc item
list))} (which currently performs safe @code{endp} checking internal
to assoc).

@end itemize

If your system's performance is suffering because of some construct
which could in principle be compiled efficiently, but which the SBCL
compiler can't in practice compile efficiently, consider writing a
patch to the compiler and submitting it for inclusion in the main
sources. Such code is often reasonably straightforward to write;
search the sources for the string ``@code{deftransform}'' to find many
examples (some straightforward, some less so).