File r34/doc/scope.rof artifact 3a048cc7f0 part of check-in 5f584e9b52


.\" NOTE: This file can be processed with eqn <filename> | tbl | troff -ms
.hw unknowns
.hw FORTRAN
.hw FORMAC
.hw REDUCE
.hw GENTRAN
.nr PS 11
.ps 11
.nr VS 15
.EQ
tdefine !eq '~ = back 80 / ~'
ndefine !eq '~ = "./" ~'
tdefine empty % size +1 \(es %
ndefine empty % O./ %
tdefine !member ' ^ \(mo back 70 / ^ '
ndefine !member % C.-./ %
tdefine => % ^ = back 40 = back 40 > ^ %
ndefine => % "==>" %
tdefine !=> % ^ = back 40 / back 30 = back 60 > ^ %
tdefine <=> % ^ < back 25 = back 40 = back 60 > ^ %
ndefine <=> % "<=>" %
tdefine exists %"\s-3\h'.2m'\v'.2m'\z\(em\v'-.5m'\z\(em\v'-.5m'\z\(em\
\v'.85m'\h'.9m'\v'-.3m'\z\(br\h'.01m'\(br\v'.3m'\h'.02m'\v'-.05m'\s0" ~%
tdefine !exist %"\s-3\h'.2m'\v'.2m'\z\(em\v'-.5m'\z\(em\v'-.5m'\z\(em\
\v'.85m'\h'.9m'\v'-.3m'\z\(br\h'.01m'\(br\v'.3m'\h'.02m'\v'-.05m'\
\s0" { back 65 down 10 size +3 / } ^%
tdefine forall % "\f1\s+1\zV\zV\v'-.2m'\h'.22m'\z--\v'.2m'\h'.05m'\s0\fP" %
ndefine forall % V.- %
tdefine orsign % ^ "\s-2\h'.05m'\v'.15m'\z\e\e\h'-.08m'\z\(sl\
\(sl\h'-.1m'\v'-.15m'\s0" ^ %
ndefine orsign % \e / %
tdefine andsign % ^ "\s-2\v'.15m'\z\(sl\(sl\h'-.18m'\z\e\e\v'-.15m'\s0" ^ %
ndefine andsign % / \e %
delim @@
.EN
.de Sq
.pl +\\n(.v
.nr tL \\n(.l-\\w'\(sq'
.lt 0
.lt +\\n(.lu
.br
.if !\\n(.n-\\n(tL .sp -1
.tl '''\fR\(sq\fP'
.nr tL 0
.pl -\\n(.v
..
..
\f3SCOPE, a Source-Code Optimization PackagE for REDUCE - User's Manual\f1
.sp 2
\f2Preliminary Version\f1
.sp 2
\f2J.A. van Hulzen \f1
.sp 4
\f3Table of Contents\f1
.LP
.nr vs 13
.TS
expand,tab(/);
l n.
_
.sp
\f31. Introduction\f1/2
.sp .3
\f32. Source-Code Optimization : The Basic Facilities\f1/4
.sp .3
\ \ \  \f32.1. The Strategy\f1/4
.sp .3
\ \ \  \f32.2. The Facilities\f1/6
.sp .3
\f33. Preprocessing Possibilities\f1/17
.sp .3
\f34. Generation of Declarations\f1/22
.sp .3
\f35. File Management and Optimization Strategies\f1/23
.sp .3
\f36. Some Possible Shortcomings and Future Versions\f1/26
.sp .3
\f3Acknowledgements\f1/26
.sp .3
\f3References\f1/27
.sp .3
\f3Appendix 1: How to install the code\f1/29
.sp 
_
.TE
.nr VS 15
.vs 15
.sp 10
\f3Abstract\f1
.sp .4
A survey of the strategy behind and the facilities of SCOPE, a Source-Code 
Optimization PackagE for REDUCE is given. We avoid a detailed discussion of the
different algorithms and concentrate on the user aspects of the package.
Examples of straightforward and more advanced usage are shown.
A combined use of GENTRAN and SCOPE is not yet discussed in this preliminary
version of the SCOPE manual.
.bp
.NH
Introduction
.LP
An important application of computer algebra systems is the generation
of code for numerical purposes via automatic or semi-automatic program
generation. GENTRAN [3,4], a flexible general-purpose package, was 
especially developed to assist in such a task, when using MACSYMA or
REDUCE.
.br
Attendant to automatic program generation is the problem of automatic
source-code optimization. This is a crucial aspect because code generated
from symbolic computations often tends to be made up of lengthy arithmetic
expressions. One of our test examples contained, for instance, 20534
additions and subtractions, 41746 multiplications,
12473 integer exponentiations and 7990 other operations, such as function
calls.
These lengthy expressions will be grouped together in blocks of
straightline code in a program for numerical purposes. The main objective
of source-code optimization is to minimize the number of (elementary)
arithmetic operations in such blocks.
This form of optimization is often helpful in reducing redundancy in
expressions. Simplification algorithms applied on expressions
viewed as entities, neither guarantee complete structure preservation nor
allow improvements inside expressions by renaming sub-expressions.
.br
Optimizing compilers ought to deal effectively and efficiently with the
average, hand coded program. The enormous, arithmetic intensive
expressions, easily producable by a computer algebra system, fall outside the
range of the FORTRAN-programs, once analyzed and discussed by Knuth [15].
He suggested that optimization of the arithmetic in such
a program is slightly overdone. This may explain why even in reasonably
recent literature, such as [1], optimization of arithmetic code is hardly
discussed. The dag-models, usually employed for optimization of arithmetic
code, hardly allow the application of any algebraic identity (see for instance
[6]). These models often force constants to act as if they were
indeterminates and powers
as objects requiring function calls, i.e. they force to think of
@2a~+~3b@ and @4a~+~6b@ or of @a sup 2@, @a sup 4@ and @a sup 6@ as being 
different entities.
Our optimization strategy however, requires the validity of some
elementary algebraic laws. We employ heuristic techniques to reduce the
arithmetic complexity of the given representation of a set of input expressions
@ roman E sub in@, thus producing a set of output expressions @roman E sub
out@. The optimized version of the earlier mentioned test example contains
"only" 4316 additions and subtractions, 4919 multiplications,
13 integer exponentiations and 60 other operations.
@roman E sub in@ and @roman E sub out@ define blocks of code, which would
compute the same exact values for the same exact inputs, thus implicitly
proving the correctness of the underlying software.  Obviously the use
of @roman E sub out@ saves a considerable amount of execution time in
comparison with the application of @roman E sub in@.
Johnson e.a. [14] suggest that such transformations do not destabilize the
computations. However this is only apparent after a careful error analysis
of both @roman E sub in@ and @roman E sub out@. In view of the 
size of both @roman E sub in@ and @roman E sub out@ such an analysis has
to be automatized as well. Work in this direction is in progress.
.LP 
The current version of SCOPE, our Source-Code Optimization PackagE, 
is written in RLISP. It can be used as an extension of REDUCE [9]. 
It allows to subject
almost any set of proper REDUCE assignment statements to a heuristic 
search for common (sub)expressions (cse's).
The output is obtained as a sequence of assignment statements,
by default given in REDUCE syntax.
.br
The first version of the package was designed to optimize the description of
REDUCE-statements, generated by NETFORM [17,18]. This version was
tailored to a restrictive class of problems, occurring mainly in 
electrical network theory, thus implying that the 
righthand sides (rhs's) in the input were limited to elements of
@{roman bold Z} sub 2@[V], where V is a set of identifiers.
The second version [12] allowed rhs's from \f3Z\f1[V]. For both versions the
validity of the commutative and the associative law was assumed.
A third version evolved from the latter package by allowing to apply the
distributive law, i.e. by replacing (sub)expressions like @a.b~+~a.c@
by @a.(b~+~c)@ whenever possible. 
But the range of possible applications of this version was really
enlarged by redefining V as a set of kernels, implying
that, at least by that time, almost any proper REDUCE expression could
function as a rhs. 
The mathematical capabilities of this version are shortly summarized in [19],
in the context of code generation and optimization for finite-element analysis.
It is used in combination with GENTRAN, for the construction
of Jacobians and Hessians [10] and also in experiments with strategies for
code vectorization [5].
It still assumes constant coefficients to be elements of \f3Z\f1.
The user-interface of the present version relies on some GENTRAN 
facilities.
.sp
.LP
.ps 11
In [11,12] we described the overall optimization-strategy used for SCOPE 
as a composite function @{roman R sup -1}~\(de~{roman T}~\(de~{roman R}@. 
The function R
defines how to store the input @{roman E} sub 0@ in an expression data base
@{roman D} sub 0@. This @{roman D} sub 0@ is formed by two matrix-structures
and a function table.
The incidence matrices represent @{roman E} sub 0@, a set of arithmetic 
expressions, in a two-dimensional structure where the rows represent expression
or sub-expression references and the columns represent identifier references
such as variable and function names. 
The function names are taken from the function table, 
consisting of a list of pairs of function applications occurring in 
@{roman E} sub 0@, and system selected names functioning as their placeholders
during the optimization process. Arguments of functions are similarly
entered in the matrix-structures when ever relevant.
A given sub-expression will be entered in one
of two types of incidence matrices, one for sums and one for products, depending
on the nature of the arithmetic operation at the top level of the expression.
The two matrices are correlated by auxiliary predecessor-successor information
at the row level for every sub-expression reference. The actual entries in the
matrices are either multiplicative numerical coefficients for the sums matrix
or powers for the products matrix. 
The inverse function @{roman R} sup {-1}@ defines the output-production.
The function T defines the optimization process itself. It essentially
consists of a heuristic remodeling of the (extendable) matrices in
combination with storing information required for a fast retrieval and
correct insertion of the detected cse's in the output.
This is accomplished by an iteratively applied search, resulting
in a stepwise reduction of the arithmetic complexity of the input set, using
an extended version of Breuer's grow factor algorithm [2,11,12]. It is applied until
no further profit is gained, i.e. until the reduction in arithmetic 
complexity stops. Before producing output, a finishing touch can be performed
to further reduce the arithmetic complexity with some locally applied
techniques. The overall process can be summarized as follows:
.ps 10
.EQ 
{roman R}~mark :~{{roman E} sub 0}~->~({{roman D} sub 0},{{roman profit} sub 0})
.EN
.EQ 
{{roman T} sub beta}~mark :~({{roman D} sub i},{{roman profit} sub i})~->~
({{roman D} sub {i+1}},{{roman profit} sub {i+1}})~,~{roman i}~=~0,..., lambda -1.
.EN
.EQ 
{roman F}~mark :~({{roman D} sub lambda},{{roman profit} sub lambda})~->~
{D sub lambda}
.EN
.EQ 
{{roman R} sup {-1}}~mark :~{D sub lambda}~->~{{roman E} sub lambda}
.EN
.ps 11
@{roman D} sub 0@ is created as a result of an R-application
performed on input @{roman E} sub 0@.
The termination condition depends on some profit criterion related
to the arithmetic complexity of the latest version of the input,
@{{roman D} sub i}@. Hence we assume @{{roman profit} sub i}~=~true@ for
@i~=~0,..., lambda -1@ and @{{roman profit} sub lambda}~=~false@.
The function T is composite as well, and defined by @{roman T}~=~{roman F}~\(de~
{{{roman T} sub beta} sup lambda}@. @{roman T} sub beta@
defines one iteration step, i.e one
application of the extended version of Breuer's algorithm. The function F 
defines a finishing touch, resulting in the final
version @D sub lambda@ of @{roman D} sub 0@, used to produce the 
output @{roman E} sub lambda@. We omit a further discussion of the different
algorithms used for optimization; this can be found in [11,12], 
for instance.
.LP
.ps 11
.LP
The present version makes use of some GENTRAN facilities to translate
its input into  LISP prefix-forms. This approach can be seen as a form
of preprocessing, i.e. @{roman E} sub 0@, the input for R can be considered
as a list of \f3setq\f1-applications
.LP
.ps 11
The GENTRAN-SCOPE Interface, allows other preprocessing activities.
We introduced the optional use of GENTRAN's \f3declare\f1-statement, 
thus allowing
to specify the type of some or all of the lhs's and of the identifiers
used to construct the rhs's. In addition to the prefixlist,
a list of declarations in the Target Language can be produced,
based on default assumptions concerning un-typed lhs's and identifiers
in the input.
This facility is based on the use of GENTRAN's symbol-table.
.br
Before optimizing rhs's it might be attractive to rewrite them using a
generalized form of Horner's rule. We designed such a command,
which does not necessarily have to be used in the context
of SCOPE. It can operate on a set
of assignment statements and it can deliver the result in the form of a sequence
of prefix-forms, defining the rewritten statements. Subjecting such a
sequence of prefix-forms to a SCOPE-application implies that the
GENTRAN-approach is not directly applicable. The GENTRAN := and :=:
assignment operators define literal translation or rhs-simplification,
respectively. Therefore we extended our Interface with special facilities,
allowing SCOPE to accept the result of the application of such a
command literally. Besides the \f3g\f1(eneralized)\f3horner\f1(rule) we have a
command, generalizing the impact of the \f3structr\f1-command to a set of
assignment statements. 
.LP
.ps 11
We discuss and illustrate a straightforward use of SCOPE in section 2.
In section 3 we introduce the special commands \f3ghorner\f1 and \f3gstructr\f1
and show
how to use them, also in combination with SCOPE. We use section 4 to
discuss the declaration-facilities and section 5 to show the different
file-handling possibilities and modes of operation.
Section 6 is finally used to indicate future work.
Guidelines for installing the package are given in Appendix 1.
.NH
Source-Code Optimization : The Basic Facilities
.NH 2
The Strategy
.LP
.ps 11
Before illustrating the effect of applying SCOPE, we shortly 
describe the operations, covered by the functions @{roman T} sub beta@ and F,
mentioned in the previous section.
.br
The function R accepts assignment statements given in prefix-form.
We can  divide these forms in three categories using
their leading operator. We distinguish between PLUS, TIMES and OTHER-operators.
Leaving aside the OTHER-operators for awhile, we reduce the
structure of possible rhs's to those of not necessarily expanded multivariate
polynomials with integer coefficients. Assuming the leading operator is PLUS,
the operands, being terms of a polynomial (for instance
@3a~+~2b~+~3 {b sup 2} c (3a~+~2b){(c~+~d) sup 2}@),
can either be primitive or composite.
A primitive term is an integer, an identifier or the product of an integer
and an identifier.
Hence the primitive terms of a sum form an (eventually empty) linear
expression (@3a~+~2b@).
Composite terms are products, which cannot be qualified as a primitive
term (@3 {b sup 2} c (3a~+~2b) {(c~+~d)} sup 2@).
Like sums, prefix-forms with a TIMES-operator, can have
a primitive and/or composite part. The primitive part of a product is an
(eventually empty) power product(@{b sup 2} c@).
The composite part is a product of sums and/or powers of sums
(@(3a~+~2b) {(c~+~d) sup 2}@).
Observe that our expression-structure discussed so far is still too simple. 
Powers of sums have EXPT as their leading operator (@{(c~+~d)} sup 2@).
Similarly, a product can have a integer coefficient (@3 {b sup 2} c@).
.br
This description suggests, as already indicated in section 1,
that we can consider any set of rhs's as
being built with linear expressions and power products only. This allows
to map such a set onto two incidence matrices: One defining the linear
expressions, using the coefficients, and another defining the power products,
using the exponents. The rows of these matrices can be associated with the
(sub)expressions under consideration and the columns with the identifiers,
used to construct these expressions. This is why we need to assume the
validity of the commutative and associative law. To be able to retrieve
the structure of the assignment statements forming the input set,
we need to combine additional information with
the rows and columns of these matrices. Essential is, for instance, storage
of the exponents of sums and of the coefficients of products.
Equally important is storage of the lhs's, which are the rhs-recognizers.
Details are given in [12]. The examples 2.2.1 and 2.2.2  provide
illustrations of these data structures.
.br
When introducing kernels, i.e. when assuming the set of OTHER-operators 
to be not empty, we have to store lists of non-commutable arguments.
Therefore a function table of pairs is made, formed by the
kernels and their internally created names. These names are entered
in the matrices as new identifiers. The arguments of such a kernel
can be arbitrary REDUCE-expressions, which also have to be incorporated in
the matrices.  Hence the function table is created recursively.
.LP
What is a cse and how do we locate its occurrences?
A (sub)expression is common when it occurs repeatedly in the 
input. The occurrences are, as part of the input, distributed over the matrices,
and shown as equivalent integer (sub)patterns.
In fact, we repeatedly search for completely dense (sub)matrices of rank 1.
The expression @2a~+~3c@ is a cse of
@{e sub 1}~=~2a~+~4b~+~3c@ and @{e sub 2}~=~4a~+~6c~+~5d@, representable
by (2,4,3,0) and (4,0,6,5), respectively.
We indeed have to assume commutativity, so as
to be able to produce new patterns (2,0,3,0,0), (0,4,0,0,1) and (0,0,0,5,2),
representing @s~=~2a~+~3c@, @{e sub 1}~=~4b~+~s@ and @{e sub 2}~=~5d~+~2s@,
respectively, and thus saving one addition and one multiplication.
Such an additive cse can be a factor in a composite (sub)product,
which in turn can be reduced to a primitive product, when the cse is replaced by
a new symbol.
Therefore an essential part of an optimization step is regrouping
of information. This migration of information between the matrices is performed
if the Breuer-searches are temporarily completed. 
After this regrouping the distributive law is applied, eventually also leading
to a further regrouping. If at least one of these actions leads to a
rearrangement of information the function @{roman T} sub beta@ is again applied.
Observe that this @{roman T} sub beta@ is also a composite function.
In view of the iterative character of the optimization process
we always accept minimal profits.
.br
A similar search is performed to detect multiplicative cse's, for instance
occuring in @{e sub 1}~=~{a sup 2} {b sup 4} {c sup 3}@ and
@{e sub 2}~=~{a sup 4} {c sup 6} {d sup 5}@.
However, given a power product @prod from i=1 to m {x sub i} sup {a sub i}@,
any product @prod from i=1 to m {x sub i} sup {b sub i}@, such that some
@{b sub i}~<~{a sub i}@, for i = 1(1)m, can function as a cse.
We therefore extend the search for multiplicative cse's by employing this
property, and as indicated in [12].
.br
The function F -defining the finishing touch- performs one-row and/or
one-column searches. Once the extended Breuer-searches do not lead to 
further reduction in the arithmetic complexity we try -applying it-
to improve what is left.
The integer coefficients in (sub)sums can have, possibly locally,
a gcd, which can be factored out.
One-column operations serve to discover and properly replace integer
multiples of identifiers.
As part of the output-process we subject all exponentiations left -
at most one for each identifier - to an addition chain algorithm. 
Another action, covered by F is therefore replacement by a new symbol
of those (sub)sums, which are raised to some integer power.
.NH 2
The Facilities
.LP
.ps 11
REDUCE allows, roughly speaking, two modes of operation: ON EXP or OFF EXP.
The first alternative is the default setting leading to expanded forms.
The latter gives unexpanded forms, as discussed by Hearn  in some
detail [7,8]. It is obvious that the OFF EXP setting is in general preferable
over the ON EXP setting when attempting to optimize the description of
a set of assignment statements.
.br
Starting a REDUCE session gives the initial state. All switches have their
initial value: ON EXP, PERIOD and OFF FORT, for instance.
When loading SCOPE we create a new operating environment, without
disturbing the current state of the system. 
.br
The result of an application of SCOPE can be influenced by the use
of certain REDUCE- or SCOPE-switches. The influence of EXP is obvious.
By default the switch ACINFO is turned on. This guarantees an echo of the form
in which the assignment statements are consumed by SCOPE. It also
guarantees tables with the numbers of arithmetic operations,
occuring in @{{roman E} sub 0}@ and @{roman E} sub lambda@, respectively,
to be printed.
Some switches are available to obtain information about the process itself.
They were introduced to assist in debugging and testing.
PRIMAT can be used to visualize both @{roman D} sub 0@ and @D sub lambda@.
PRIALL is a switch which combines not only the effect of ACINFO and PRIMAT,
but also allows to obtain timings of the different sub-algorithms of SCOPE.
.br
Output is by default given in REDUCE syntax, but FORTRAN syntax is possible
in the usual way. The switch PREFIX can be used to obtain the prefixlist itself
as output. 
.br
A SCOPE action is easily performed.
Either the command "\f3optimize\f1 @<@object@>@;" or the command
"\f3optimize\f1 @<@object@>@ \f3iname\f1 @<@cse-prefix@>@;" suffices.
The @<@object@>@ to be elaborated is either one assignment statement or
a list of such statements, all obeying the GENTRAN rules.
The @<@cse-prefix@>@ is an identifier, used to generate the cse-names, by 
extending it with an integer part. The \f3gensym\f1-function is applied
when the \f3iname\f1-extension is omitted.
.LP
We now illustrate the use of SCOPE through some small examples,
by showing parts of REDUCE sessions.
.sp
\f3Example 2.2.1\f1
.LP
The multivariate polynomial Z is a sum of 6 composite terms. These terms,
monomials, are constant multiples of primitive products.
A picture of @{roman D} sub 0@
is shown after the input echo. The sums-matrix consists of only one row,
identifiable by its Fa(the)r Z, the lhs. Its exponent is given in the
E(xponent or )C(oefficient) field. The 6 monomials are stored in the 
products-matrix. The coefficients are stored in the EC-fields and the 
predecessor row index, 0, is given in the Far-field. Before the @D sub lambda@
picture is given the effect of the optimization process, the output
and the operator counts are shown. The optimized form of Z is obtained by
applying the distributive law. The output also shows applications of an addition
chain algorithm ([16], 441-466) as part of @{roman R} sup {-1}@, although
its use in example 2.2.3 is more apparent.
.br
Observe that the output illustrates the heuristic character of the optimization
process: In this particular case the rhs can be written as a polynomial in
S3, thus saving one extra multiplication.
.LP
.in +3
.vs 10
.nf
\fC\s8

on primat$

optimize z:=a^2*b^2+10*a^2*m^6+a^2*m^2+2*a*b*m^4+2*b^2*m^6+b^2*m^2 iname s;

      2  2       2  6    2  2          4      2  6    2  2
Z := A *B  + 10*A *M  + A *M  + 2*A*B*M  + 2*B *M  + B *M



Sumscheme :

   || EC|Far
------------
  0||  1| Z
------------


Productscheme :

   |  0  1  2| EC|Far
---------------------
  1|     2  2|  1| 0
  2|  6     2| 10| 0
  3|  2     2|  1| 0
  4|  4  1  1|  2| 0
  5|  6  2   |  2| 0
  6|  2  2   |  1| 0
---------------------
0  : M
1  : B
2  : A


Number of operations in the input is:

Number of (+,-)-operations : 5
Number of (*)-operations : 10
Number of integer exponentiations : 11
Number of other operations : 0


S0 := B*A
S4 := M*M
S8 := B*B
S1 := S4*S8
S9 := A*A
S2 := S4*S9
S3 := S4*S4
Z := S1 + S2 + S0*(2*S3 + S0) + S3*(2*S1 + 10*S2)


Number of operations after optimization is:

Number of (+,-)-operations : 5
Number of (*)-operations : 12
Number of integer exponentiations : 0
Number of other operations : 0



Sumscheme :

   |  0  3  4  5| EC|Far
------------------------
  0|        1  1|  1| Z
 15|        2 10|  1| 14
 17|  2  1      |  1| 16
------------------------
0  : S3
3  : S0
4  : S1
5  : S2

Productscheme :

   |  8  9 10 11 17 18 19 20| EC|Far
------------------------------------
  7|                    1  1|  1| S0
  8|  1                 2   |  1| S1
  9|  1                    2|  1| S2
 10|  2                     |  1| S3
 11|                 2      |  1| S4
 14|     1                  |  1| 0
 16|              1         |  1| 0
------------------------------------
8  : S4
9  : S3
10 : S2
11 : S1
17 : S0
18 : M
19 : B
20 : A


\f1\s0
.fi
.in -3
.vs 15
.Sq
.LP
\f3Example 2.2.2\f1
.LP
The input echo below shows the literal copy of the first assignment,
in accordance with the GENTRAN := operator.
The second assignment, again in accordance with the GENTRAN operator ::=:,
has a rhs in expanded form.
.br
The @{roman D} sub 0@ picture shows that during parsing string matching
of kernels in prefix-form already contributes to optimization :
S2 = C*X + D and S3 =SIN(S2) are stored once.
.br
Application of the distributive law gives the original structure of A(1,1)
back.
.in +3
.vs 10
.nf
\fC\s8

on primat$
operator a$
k:=j:=1$
u:=c*x+d$
v:=sin(u)$

optimize {a(k,j) := v*(v^2*cos(u)^2+u),
          a(k,j)  ::=:v*(v^2*cos(u)^2+u)} iname s;


              2       2
A(K,J) := V*(V *COS(U)  + U)

                      2             3
A(1,1) := COS(C*X + D) *SIN(C*X + D)  + SIN(C*X + D)*C*X + SIN(C*X + D)*D


Sumscheme :

   |  7  8| EC|Far
------------------
  1|  1   |  1| 0
  3|      |  1| A(1,1)
  5|     1|  1| S2
------------------
7  : U
8  : D

Productscheme :

   |  0  1  2  3  4  5  6| EC|Far
---------------------------------
  0|                    1|  1| A(K,J)
  2|                 2  2|  1| 1
  4|     3  2            |  1| 3
  6|           1  1      |  1| 5
  7|     1     1  1      |  1| 3
  8|  1  1               |  1| 3
---------------------------------
0  : D
1  : S3=SIN(S2)
2  : S1=COS(S2)
3  : X
4  : C
5  : S0=COS(U)
6  : V

Number of operations in the input is:

Number of (+,-)-operations : 7
Number of (*)-operations : 10
Number of integer exponentiations : 4
Number of other operations : 5

S6 := COS(U)*V
S9 := S6*S6
A(K,J) := V*(U + S9)
S2 := D + X*C
S3 := SIN(S2)
S7 := S3*COS(S2)
S8 := S7*S7
A(1,1) := S3*(S2 + S8)

Number of operations after optimization is:

Number of (+,-)-operations : 3
Number of (*)-operations : 7
Number of integer exponentiations : 0
Number of other operations : 3

Sumscheme :

   |  2 12 13| EC|Far
---------------------
  1|     1   |  1| 0
  3|         |  1| A(1,1)
  5|        1|  1| S2
 11|  1      |  1| 10
---------------------
2  : S2
12 : U
13 : D

Productscheme :

   |  0  1  5  6  7  8  9 10 11| EC|Far
---------------------------------------
  0|                          1|  1| A(K,J)
  2|     2                     |  1| 1
  4|  2                        |  1| 11
  9|                 1  1      |  1| 5
 10|           1               |  1| 3
 13|                       1  1|  1| S6
 14|           1  1            |  1| S7
---------------------------------------
0  : S7
1  : S6
5  : D
6  : S3=SIN(S2)
7  : COS(S2)
8  : X
9  : C
10 : COS(U)
11 : V
\f1\s0
.fi
.in -3
.vs 15
.Sq
.LP
\f3Example 2.2.3\f1
.LP
The effect is shown of a finishing touch application, in combination
with FORTRAN output.
During output preparation @{roman S0} sup 13@ is rewritten, using the
earlier mentioned addition chain algorithm.
.in +3
.vs 10
.nf
\fC\s8
on fort$
off acinfo,period$

optimize z:=96*a+18*b+9*c+3*d+6*e+18*f+6*g+5*h+5*k+3)^13 iname s;

      S0=5*(H+K)+3*(3*C+D+1+6*(B+F)+2*(A+E+G))
      S4=S0*S0
      S3=S0*S4
      S2=S3*S3
      S1=S2*S2
      Z=S0*S1
\f1\s0
.fi
.in -3
.vs 15
.Sq
\f3Example 2.2.4\f1
.LP
Recovery of repeatedly occurring integer multiples of identifiers,
as part of the finishing touch, is illustrated. The switch ACINFO
is turned off.
.LP
.in +3
.vs 10
.nf
\fC\s8
optimize {x:=3*a*p,
          y:=3*a*q,
          z:=6*a*r+2*b*p,
          u:=6*a*d+2*b*q,
          v:=9*a*c+4*b*d,
          w:=4*b} iname s;

S1 := 3*A
X := S1*P
Y := S1*Q
S2 := 6*A
S3 := 2*B
Z := S3*P + S2*R
U := S3*Q + S2*D
S0 := 4*B
V := S0*D + 9*A*C
W := S0
\f1\s0
.fi
.in -3
.vs 15
.Sq
.bp
\f3Example 2.2.5\f1
.LP
.ps 11
The effect of ON EXP or OFF EXP on the result of a 
SCOPE-application is now shown by optimizing the representation of the
determinant of a symmetric (3,3) matrix M. Besides differences in computing time
we also observe that the arithmetic complexity of the optimized version of the
expanded representation of the determinant is about the same as the not
optimized form of the unexpanded representation.
.LP
.in +3
.vs 10
.nf
\fC\s8

matrix M(3,3)$

m(1,1):=18*cos(q3)*cos(q2)*m30*p^2-sin(q3)^2*j30y+sin(q3)^2*j30z-
        9*sin(q3)^2*m30*p^2+j1oy+j30y+m10*p^2+18*m30*p^2$
m(2,1):=
m(1,2):=9*cos(q3)*cos(q2)*m30*p^2-sin(q3)^2*j30y+sin(q3)^2*j30z-
        9*sin(q3)^2*m30*p^2+j30y+9*m30*p^2$
m(3,1):=
m(1,3):=-9*sin(q3)*sin(q2)*m30*p^2$
m(2,2):=-sin(q3)^2*j30y+sin(q3)^2*j30z-9*sin(q3)^2*m30*p^2+j30y+9*m30*p^2$
m(3,2):=
m(2,3):=0$
m(3,3):=9*m30*p^2+j30x$

optimize detm:=:det(M) iname s;

                   4        2  6    3             4        2  4    2
DETM := 729*SIN(Q3) *SIN(Q2) *P *M30  + 81*SIN(Q3) *SIN(Q2) *P *M30 *

                 4        2  4    2                   2        2  6
J30Y - 81*SIN(Q3) *SIN(Q2) *P *M30 *J30Z - 729*SIN(Q3) *SIN(Q2) *P *

   3             2        2  4    2                   2  6    3
M30  - 81*SIN(Q3) *SIN(Q2) *P *M30 *J30Y - 729*SIN(Q3) *P *M30  - 81*

       2  6    2                 2  4    2                  2  4    2
SIN(Q3) *P *M30 *M10 - 81*SIN(Q3) *P *M30 *J30Y + 81*SIN(Q3) *P *M30

                  2  4    2                  2  4    2
*J30Z - 81*SIN(Q3) *P *M30 *J1OY - 81*SIN(Q3) *P *M30 *J30X - 9*

       2  4                         2  4                         2  4
SIN(Q3) *P *M30*J30Y*M10 + 9*SIN(Q3) *P *M30*J30Z*M10 - 9*SIN(Q3) *P

                         2  2                          2  2
*M30*M10*J30X - 9*SIN(Q3) *P *M30*J30Y*J1OY - 9*SIN(Q3) *P *M30*J30Y*

                2  2                          2  2
J30X + 9*SIN(Q3) *P *M30*J30Z*J1OY + 9*SIN(Q3) *P *M30*J30Z*J30X - 9*

       2  2                        2  2                        2  2
SIN(Q3) *P *M30*J1OY*J30X - SIN(Q3) *P *J30Y*M10*J30X + SIN(Q3) *P *

                       2                         2
J30Z*M10*J30X - SIN(Q3) *J30Y*J1OY*J30X + SIN(Q3) *J30Z*J1OY*J30X - 

           2        2  6    3             2        2  4    2
729*COS(Q3) *COS(Q2) *P *M30  - 81*COS(Q3) *COS(Q2) *P *M30 *J30X + 

     6    3       6    2           4    2            4    2
729*P *M30  + 81*P *M30 *M10 + 81*P *M30 *J30Y + 81*P *M30 *J1OY + 81

  4    2           4                   4                   2
*P *M30 *J30X + 9*P *M30*J30Y*M10 + 9*P *M30*M10*J30X + 9*P *M30*J30Y

           2                    2                  2
*J1OY + 9*P *M30*J30Y*J30X + 9*P *M30*J1OY*J30X + P *J30Y*M10*J30X + 

J30Y*J1OY*J30X

Number of operations in the input is: 

Number of (+,-)-operations : 36
Number of (*)-operations : 148
Number of integer exponentiations : 84
Number of other operations : 32


S0 := SIN(Q3)
S30 := S0*S0
S1 := SIN(Q2)
S34 := S1*S1
S35 := P*P
S7 := S35*M30
S33 := S7*S7
S5 := S33*J30Y
S6 := S30*S7
S8 := S30*M10
S49 := COS(Q2)*COS(Q3)
S9 := S49*S49
S11 := S34*S30*S30
S22 := S35*S7
S14 := S30*J30Z
S19 := S35*J30X
S23 := J30X*J1OY
S31 := S33*S7
S47 := 81*S33*J30X
S39 :=  - S47 - S23*J30Y - 81*S33*J1OY
S40 :=  - 81*S30*S5 - 729*S33*S6
S45 := 9*S6*J30Z
S46 := 9*S6
S48 := 81*S5
DETM := S48 + S40 - S39 + 729*S31 + ( - J1OY - J30X)*(9*(S6*J30Y - S7

      *J30Y) - S45) + (J30Z - J30Y)*(9*S22*S8 + S19*S8) + 9*(M10 - S8

   )*(S22*J30X + 9*S22*S7) + M10*J30Y*(9*S22 + S19) + S23*(S14 + 9*S7

    - S46) + S39*S30 + S31*(729*(S11 - S9)) + S34*(S40 - S46*S45) - 

S47*S9 + 81*S33*S14 + S48*S11


Number of operations after optimization is:

Number of (+,-)-operations : 29
Number of (*)-operations : 58
Number of integer exponentiations : 0
Number of other operations : 4


off exp$

optimize detm:=:det(M) iname s;

             2                           2                   2
DETM := ((9*P *M30 + J30Y - J30Z)*SIN(Q3)  - (18*M30 + M10)*P  - 18*

                          2                         2
         COS(Q3)*COS(Q2)*P *M30 - J30Y - J1OY)*((9*P *M30 + J30Y - 

                   2      2                 2
      J30Z)*SIN(Q3)  - 9*P *M30 - J30Y)*(9*P *M30 + J30X) - 

     2                           2                      2          2
((9*P *M30 + J30Y - J30Z)*SIN(Q3)  - 9*COS(Q3)*COS(Q2)*P *M30 - 9*P *

            2     2                       2
 M30 - J30Y) *(9*P *M30 + J30X) + 81*((9*P *M30 + J30Y - J30Z)*

          2      2                    2        2  4    2
   SIN(Q3)  - 9*P *M30 - J30Y)*SIN(Q3) *SIN(Q2) *P *M30


Number of operations in the input is: 

Number of (+,-)-operations : 24
Number of (*)-operations : 42
Number of integer exponentiations : 21
Number of other operations : 10

S0 := SIN(Q3)
S9 := S0*S0
S8 := P*P
S5 := S8*M30
S6 := S5*COS(Q2)*COS(Q3)
S15 := 9*S5
S13 := (S15 + J30Y - J30Z)*S9
S14 := S13 - S15 - J30Y
S3 := S14 - 9*S6
S4 := SIN(Q2)
DETM := (S15 + J30X)*(S14*(S13 - 18*S6 - J30Y - J1OY - S8*(18*M30 + 

         M10)) - S3*S3) + 9*S15*S14*S9*S5*S4*S4


Number of operations after optimization is:

Number of (+,-)-operations : 13
Number of (*)-operations : 20
Number of integer exponentiations : 0
Number of other operations : 4

\f1\s0
.fi
.in -3
.vs 15
.LP
.ps 11
We can also use this example to show that correctness of the results can easily
be verified. When turning off the switch NAT and storing the result of
a SCOPE application in a file, it is of course possible to read the result 
in again. But we then operate in a normal REDUCE-like way. This
implies that all cse-names are automatically replaced by their values.
We show the "correctness" of SCOPE by storing the optimized version of 
the expanded form of the determinant of M, called detm1 in file out1 and
the result of a SCOPE-application on the unexpanded form, detm2, in file out2,
followed by reading both files and by subtracting detm2 from detm1, resulting
in the value 0. This is of course an ad hoc correctness-proof for one specific
example. It is in fact another way of testing the code of the package.
So, assuming SCOPE is loaded and the matrix M is known to the system, all we
have to do is:
.LP
.in +3
.vs 10
.nf
\fC\s8

2: off acinfo,nat$

3: out out1$
4: optimize detm1:=:det(M) iname s;
5: write "end$"$
6: shut "out1"$

7: off exp$

8: out out2$
9: optimize detm2:=:det(M) iname t;
10: write "end$"$
11: shut out2$

12: on nat$

13: in out1;

S0 := SIN(Q3)$
S30 := S0*S0$
S1 := SIN(Q2)$
S34 := S1*S1$
S35 := P*P$
S7 := S35*M30$
S33 := S7*S7$
S5 := S33*J30Y$
S6 := S30*S7$
S8 := S30*M10$
S49 := COS(Q2)*COS(Q3)$
S9 := S49*S49$
S11 := S34*S30*S30$
S22 := S35*S7$
S14 := S30*J30Z$
S19 := S35*J30X$
S23 := J30X*J1OY$
S31 := S33*S7$
S47 := 81*S33*J30X$
S39 :=  - S47 - S23*J30Y - 81*S33*J1OY$
S40 :=  - 81*S30*S5 - 729*S33*S6$
S45 := 9*S6*J30Z$
S46 := 9*S6$
S48 := 81*S5$
DETM1 := S48 + S40 - S39 + 729*S31 + ( - J1OY - J30X)*(9*(S6*J30Y - S7*J30Y) -

S45) + (J30Z - J30Y)*(9*S22*S8 + S19*S8) + 9*(M10 - S8)*(S22*J30X + 9

*S22*S7) + M10*J30Y*(9*S22 + S19) + S23*(S14 + 9*S7 - S46) + S39*S30

+ S31*(729*(S11 - S9)) + S34*(S40 - S46*S45) - S47*S9 + 81*S33*S14 +

S48*S11$

end$

14: in out2;

T0 := SIN(Q3)$
T9 := T0*T0$
T8 := P*P$
T5 := T8*M30$
T6 := T5*COS(Q2)*COS(Q3)$
T15 := 9*T5$
T13 := (T15 + J30Y - J30Z)*T9$
T14 := T13 - T15 - J30Y$
T3 := T14 - 9*T6$
T4 := SIN(Q2)$
DETM2 := (T15 + J30X)*(T14*(T13 - 18*T6 - J30Y - J1OY - T8*(18*M30 +

 M10)) - T3*T3) + 9*T15*T14*T9*T5*T4*T4$

end$

15: detm1-detm2;

0

\f1\s0
.fi
.in -3
.vs 15
.Sq
\f3Example 2.2.6\f1
.LP
.ps 11
This example serves to show how SCOPE deals with rational exponents.
All rational exponents of a variable are collected. The least common
multiple lcm of the denominators of these rational exponents is computed 
and the variable is replaced
by a possibly newly selected variable name, denoting the variable raised
to the power 1/lcm. This facility is only efficient for what we believe
to be problems occurring in computational practice. This is easily verified by
extending the sum we are elaborating here with some extra terms.
.br
.ps 11
Producing FORTRAN-output shows an implied danger, due to a shortcoming in 
GENTRAN. This rational exponent will in practice act as if it were 0.
.LP
.ps 11
This example is also used to show the effect of turning on the switch PRIALL.
.LP
.in +3
.vs 10
.nf
\fC\s8

on fort,priall$
 
optimize z:=:for j:=2:6 sum q^(1/j) iname s;


      1/6    1/5    1/4    1/3
Z := Q    + Q    + Q    + Q    + SQRT(Q)


Sumscheme :

   || EC|Far
------------
  0||  1| Z
------------

Productscheme :

   |  0| EC|Far
---------------
  1| 10|  1| 0
  2| 12|  1| 0
  3| 15|  1| 0
  4| 20|  1| 0
  5| 30|  1| 0
---------------
0  : Q


Number of operations in the input is: 

Number of (+,-)-operations : 4
Number of (*)-operations : 0
Number of integer exponentiations : 0
Number of other operations : 5


Time: 2992 ms

Breuer search : 
Time: 867 ms

Removal of different names for identical cse's : 
Time: 17 ms

Change Scheme : 
Time: 0 ms

Local Factorization : 
Time: 34 ms

Breuer search : 
Time: 204 ms

Removal of different names for identical cse's : 
Time: 0 ms

Change Scheme : 
Time: 17 ms

Local Factorization : 
Time: 0 ms

Breuer search : 
Time: 187 ms

Removal of different names for identical cse's : 
Time: 0 ms

Change Scheme : 
Time: 17 ms

Local Factorization : 
Time: 0 ms

Breuer search : 
Time: 119 ms

Removal of different names for identical cse's : 
Time: 0 ms

Change Scheme : 
Time: 17 ms

Local Factorization : 
Time: 0 ms

Additional optimization during finishing touch :
Time: 34 ms


      Q=Q**(1/60)
      S7=Q*Q
      S6=S7*Q
      S4=S7*S6
      S2=S4*S4
      S1=S7*S2
      S0=S6*S1
      S3=S4*S0
      Z=S3+S0+S1+S2+S3*S2



Number of operations after optimization is:

Number of (+,-)-operations : 4
Number of (*)-operations : 8
Number of integer exponentiations : 0
Number of other operations : 1


Sumscheme :

   |  3  4  5  6| EC|Far
------------------------
  0|  1  1  1  1|  1| Z
------------------------
3  : S3
4  : S0
5  : S1
6  : S2


Productscheme :

   |  9 10 12 13 14 15 16 22| EC|Far
------------------------------------
  5|           1  1         |  1| 0
  6|     1           1      |  1| S0
  7|  1           1         |  1| S1
  8|        2               |  1| S2
  9|        1           1   |  1| S3
 10|  1  1                  |  1| S4
 12|  1                    1|  1| S6
 13|                       2|  1| S7
------------------------------------
9  : S7
10 : S6
12 : S4
13 : S3
14 : S2
15 : S1
16 : S0
22 : Q

Time: 459 ms
\f1\s0
.fi
.in -3
.vs 15
.Sq
.NH
Preprocessing Possibilities
.LP
.ps 11
It may happen that structure is obviously visible in the 
rhs's of a set of assignment statements, which we want to optimize.
One can think of a set of partial derivatives of products. 
Or one may consider the application of Horner-rules.
Such facilities may be attractive, independent of the question if a 
SCOPE-application will be performed on its result.
Therefore we first discuss these facilities and show their effect, again
by using simple examples, before we continue with a combined use of SCOPE
and these possibilities.
.br
.ps 11
The first alternative demands a generalized \f3structr\f1-command.
We implemented such a facility. Its syntax is straightforward:
"\f3gstructr\f1 @<@object@>@ \f3name\f1 @<@id@>@;" The @<@object@>@ 
to be elaborated is one assignment statement or a set of such
statements, separated by semicolons and grouped together between the
special symbols @<<@ and @>>@. In stead of a statement a matrix-name
is also allowed. Then all non-zero matrix elements are incorporated in the
search for obvious cse's. The @<@id@>@ of the optional \f3name\f1-part,
being an identifier, is used to identify 
the subexpressions, produced via the application of a \f3gstructr\f1
command. When the switch ALGPRI is on -the default setting- 
the output is given in REDUCE syntax, while turning it off leads to output
in prefix-form.
The latter is employed by the function R, used to store SCOPE-input 
in @{roman D} sub 0@.
It is also possible to get FORTRAN-output by turning off the switch PERIOD and 
turning on the switch  FORTRAN.
The input remains unchanged when the switch EXP is on.
.LP
\f3Example 3.1\f1
.LP
We show part of a REDUCE session.
.vs 10
.in +3
.nf
\fC\s8

off exp$

matrix a(2,2)$

a(1,1) := x+y+z$
a(1,2) := x*y$
a(2,1) := (x+y)*x*y$
a(2,2) := (x+2*y+3)^3-x$

on fort$
off period$

load struct$

gstructr << a;
            b:=(x+y)^2;
            c:=(x+y)*(y+z);
            d:=(x+2*y)*(y+z)*(z+x)^2
         >> name v;

      V1=X+Y+Z
      A(1,1)=V1
      A(1,2)=X*Y
      V2=X+Y
      A(2,1)=V2*X*Y
      V3=X+2*Y+3
      V4=V3**3-X
      A(2,2)=V4
      B=V2**2
      V5=Y+Z
      C=V2*V5
      V6=X+2*Y
      V7=X+Z
      D=V6*V7**2*V5

\f1\s0
.fi
.in -3
.vs 15
.LP
.ps 11
Observe that V1, V3, V4, V6 and V7 only occur once in this result
of a \f3gstructr\f1-application. When applied as part of a SCOPE-operation
these redundancies will be removed before the actual optimization process
is performed, as shown in example 3.3.
.Sq
.sp
.LP
.ps 11
The syntax for the \f3ghorner\f1-command is very similar.
The application of a Horner-rule assumes an ordering of the identifiers.
We allow instead of the \f3name\f1-part of the \f3gstructr\f1 command an
optional \f3vorder\f1 @<@list of id.s@>@ extension.
The @<@list of id.s@>@ consists
of at least one identifier. This list overrules, in the order given, the
current identifier ordering of the system. The rhs's are considered as
polynomials in the leftmost element of the \f3vorder\f1-list. The thus created
coefficients are in turn considered as polynomials in the second element 
of this list. And so on. When the \f3vorder\f1-extension is omitted the
current system identifier ordering is applied.
The internal switch ALGPRI is again applicable  and has the same meaning as for
\f3gstructr\f1.
.br
Some optimizing compilers apply Horner-rules when possible. Our optimization
strategy is based on an all levels, all expressions search. This contradicts
the Horner-mechanism. To avoid destabilizing side-effects of
Horner-rule applications we decided to bring such a facility under user-control.
.LP
\f3Example 3.2
.LP
Some Taylor-expansions are shown. 
.LP
.vs 10
.in +3
.nf
\fC\s8

algebraic procedure taylor(fx,x,x0,n);
sub(x=x0,fx)+for k:=1:n sum (sub(x=x0,df(fx,x,k))*(x-x0)^k/
                                     (if k<3 then k else for j:=2:k product j))$

let x^4=0,y^7=0$

f1:=(taylor(e^x,x,0,4)*taylor(cos y,y,0,6))^2;


             3  6       3  4        3  2        3       2  6       2
F1 :=  - (8*X *Y  - 60*X *Y  + 180*X *Y  - 180*X  + 12*X *Y  - 90*X *

           4        2  2        2         6         4          2
          Y  + 270*X *Y  - 270*X  + 12*X*Y  - 90*X*Y  + 270*X*Y  - 

                     6       4        2
          270*X + 6*Y  - 45*Y  + 135*Y  - 135)/135

load horner$

ghorner << f1:=f1;
           g1:=taylor(e^x,x,0,4);
           h1:=taylor(cos y,y,0,6);
           f1:=(g1*h1)^2 >> vorder y,x;

                                            2
F1 := ((135 + X*(270 + X*(270 + X*180))) + Y *(( - 135 + X*( - 270 + 

                                           2
                X*( - 270 + X*(-180)))) + Y *((45 + X*(90 + X*(90 + X

                                2
                      *60))) + Y *( - 6 + X*( - 12 + X*( - 12 + X*(-8

                      )))))))/135

       6 + X*(6 + X*(3 + X))
G1 := -----------------------
                 6

              2            2        2
       720 + Y *( - 360 + Y *(30 + Y *(-1)))
H1 := ---------------------------------------
                        720


                              2              2         2           2   2
       (6 + X*(6 + X*(3 + X)))  * ( - 720 + Y *(360 + Y *( - 30 + Y )))
F1 := -------------------------------------------------------------------
                               18662400
\f1\s0
.fi
.in -3
.vs 15
.Sq
.LP
Both commands can be used inside an \f3optimize\f1-command. We advise to compile
both facilities and SCOPE separately (see also Appendix 1).
.LP
.ps 11
To be able to order the application of either a \f3gstructr\f1-command or a
\f3ghorner\f1-rewrite instruction inside the definition of a SCOPE-operation
we have to extend the rules given in section 2.2. The permissible structures
for the <object>'s to be elaborated by SCOPE are simply extended with 
syntactically correct \f3ghorner\f1- and \f3gstructr\f1-commands. Hence the
structure of an \f3optimize\f1-command is not altered, as is shown by the
following two examples.
.LP
.ps 11
\f3Example 3.3\f1
.LP
We show the effect of an application of the \f3optimize\f1-command
on the \f3gstructr\f1-command of example 3.1.
Observe that the cse-names produced during optimization  begin with an S,
while \f3gstructr\f1 created names start with a V.
.in +3
.vs 10
.nf
\fC\s8

on fort,acinfo$
off exp,period$

optimize gstructr << a;
                     b:=(x+y)^2;
                     c:=(x+y)*(y+z);
                     d:=(x+2*y)*(y+z)*(z+x)^2
                  >> name v
         iname s;


A(1,1) := X + Y + Z

A(1,2) := X*Y

V2 := X + Y

A(2,1) := V2*X*Y

                       3
A(2,2) := (X + 2*Y + 3)  - X

       2
B := V2

V5 := Y + Z

C := V2*V5

                      2
D := (X + 2*Y)*(X + Z) *V5

Number of operations in the input is: 

Number of (+,-)-operations : 9
Number of (*)-operations : 8
Number of integer exponentiations : 3
Number of other operations : 0


      S5=X+Z
      A(1,1)=S5+Y
      S8=Y*X
      A(1,2)=S8
      V2=X+Y
      A(2,1)=S8*V2
      S6=X+2*Y
      S4=S6+3
      A(2,2)=S4*S4*S4-X
      B=V2*V2
      V5=Y+Z
      C=V5*V2
      D=S6*S5*S5*V5


Number of operations after optimization is:

Number of (+,-)-operations : 7
Number of (*)-operations : 10
Number of integer exponentiations : 0
Number of other operations : 0

\f1\s0
.fi
.in -3
.vs 15
.Sq
\f3Example 3.4\f1
.LP
.ps 11
For completeness we also show how to use the Horner-facilities inside an 
\f3optimize\f1-command. Due to the structure of the method -we operate
internally on expanded forms- both representations of h1, and thus also
of the corresponding prefix representations used to built @{roman D} sub 0@
slightly differ. The consequences are visualized in the results of the
SCOPE-application.
.LP
.in +3
.vs 10
.nf
\fC\s8

load scope$

optimize ghorner <<h1:=taylor(cos y,y,0,6);
                   f1:=(taylor(e^x,x,0,4)*h1)^2>> vorder y,x
         iname s;


              2            2        2
       720 + Y *( - 360 + Y *(30 + Y *(-1)))
H1 := ---------------------------------------
                        720

                              2              2         2           2   2
       (6 + X*(6 + X*(3 + X)))  * ( - 720 + Y *(360 + Y *( - 30 + Y )))
F1 := -------------------------------------------------------------------
                               18662400

Number of operations in the input is: 

Number of (+,-)-operations : 9
Number of (*)-operations : 8
Number of integer exponentiations : 8
Number of other operations : 2


S6 := Y*Y
       720 + S6*(S6*(30 - 1*S6) - 360)
H1 := ---------------------------------
                     720
S7 := (S6*(360 + S6*(S6 - 30)) - 720)*(6 + X*(6 + X*(3 + X)))
        S7*S7
F1 := ----------
       18662400


Number of operations after optimization is:

Number of (+,-)-operations : 9
Number of (*)-operations : 10
Number of integer exponentiations : 0
Number of other operations : 2

\f1\s0
.fi
.in -3
.vs 15
.Sq
.NH
Generation of Declarations
.LP
The GENTRAN \f3declare\f1-statement can also be used as an optional extension of
the \f3optimize\f1-command. An illustration of this facility is given in 
example 4.1. The syntax of such a statement is in accordance with the
GENTRAN-rules [3]. We also use the symbol-table of GENTRAN.
During parsing, the declared identifiers and/or  array- and matrix-names
are entered in the symbol-table. Once optimization is finished
all relevant information for completing the declarations is known,
and collected in the prefixlist, which is used for output-production.
This prefixlist is employed to decide which not yet 
typed identifiers and system selected cse-names have to be 
entered in the symbol-table. We make use of already known information, 
expression-structure and the normal hierarchy in data types. 
The strategy to achieve this is essentially based on chapter 6 of [1].
Once this table is completed a list of declarations is produced
if the switch OPTDECS is turned on.
SCOPE-output is by default given in REDUCE syntax. Alternative output is
obtained by assigning a relevant value to the global identifier Optlang!*.
This causes GENTRAN to take over the output preparation, as shown in:
.LP
\f3Example 4.1\f1
.LP
.in +3
.vs 10
.nf
\fC\s8
on optdecs$
off acinfo$
optlang!*:='fortran$

optimize {x(i+1):=a(i+1,i-1)+b(i),y(i-1):=a(i-1,i+1)-b(i)} 
         iname s
         declare << a(4,4),x(4),y(5):real; b(5):integer >>;


      INTEGER B(5),I,S1,S3
      DOUBLE PRECISION A(4,4),S4,X(4),Y(5)
      S1=I+1
      S3=I-1
      S4=B(I)
      X(S1)=A(S1,S3)+S4
      Y(S3)=A(S3,S1)-S4

optlang!*:='c$

optimize {x(i+1):=a(i+1,i-1)+b(i),y(i-1):=a(i-1,i+1)-b(i)} 
         iname s
         declare << a(4,4),x(4),y(5):real; b(5):integer >>;


LONG B[6],I,S1,S3;
DOUBLE A[5][5],S4,X[5],Y[6];
{
    S1=I+1;
    S3=I-1;
    S4=B[I];
    X[S1]=A[S1][S3]+S4;
    Y[S3]=A[S3][S1]-S4;
}

optlang!*:='pascal$

optimize {x(i+1):=a(i+1,i-1)+b(i),y(i-1):=a(i-1,i+1)-b(i)} 
         iname s
         declare << a(4,4),x(4),y(5):real; b(5):integer >>;


VAR
    B[0..5],I,S1,S3: INTEGER;
    A[0..4,0..4],S4,X[0..4],Y[0..5]: REAL;
BEGIN
    S1:=I+1;
    S3:=I-1;
    S4:=B[I];
    X[S1]:=A[S1,S3]+S4;
    Y[S3]:=A[S3,S1]-S4
END;

optlang!*:='ratfor$

optimize {x(i+1):=a(i+1,i-1)+b(i),y(i-1):=a(i-1,i+1)-b(i)} 
         iname s
         declare << a(4,4),x(4),y(5):real; b(5):integer >>;


INTEGER B(5),I,S1,S3
DOUBLE PRECISION A(4,4),S4,X(4),Y(5)
{
    S1=I+1
    S3=I-1
    S4=B(I)
    X(S1)=A(S1,S3)+S4
    Y(S3)=A(S3,S1)-S4
}

%%% The following command restores the initial situation. %%%

optlang!*:='nil$
\f1\s0
.fi
.in -3
.vs 15
.Sq
.sp
.LP
.NH
File Management and Optimization Strategies
.LP
Another alternative for the <object>'s to be optimized is formed by the
sequence \f3in\f1 @{roman file} sub 1@, @{roman file} sub 2@, ..., 
@{roman file} sub n@, @n~>=~1@.  Each of these files is assumed to contain 
one or a list of more assignment statements, obeying the GENTRAN-assignment 
rules.  A SCOPE application results in a unified sequence of assignment
statements in the target language. This is illustrated by the following
example, where each file fi contains one assignment statement of the form
ei := some expression.
.LP
.ps 11
\f3Example 5.1\f1
.LP
.ps 11
.LP
.in +3
.vs 10
.nf
\fC\s8

3: optimize in f1,f2,f3 iname s;


                                2
                2        (X + Y)              8        2      2
       2*(SIN(X)  - COS(E        ) + 3*COS(X)) *(X + Y)  + 4*Y  + 4*Y
E1 := ----------------------------------------------------------------
                                 3*X + 2*Y

E2 := (4*

                              2
              2        (X + Y)              2        3       2
       (SIN(X)  - COS(E        ) + 2*COS(X)) *(X + Y)  + (4*X  - 4*Y)

                   2
        - 6*X)/(8*X  + 3*Y - 2*X)

                         2
                  (X + Y)
E3 := (4*SIN(COS(E        )) + SIN(X + Y) + 

           2           2
       (4*X  - X + 2*Y) )/(3*Y + F(X,G( - COS(X))))


Number of operations in the input is: 

Number of (+,-)-operations : 21
Number of (*)-operations : 20
Number of integer exponentiations : 12
Number of other operations : 16


S3 := SIN(X)
S7 := X + Y
S6 := S7*S7
           S6
S4 := COS(E  )
S8 := COS(X)
S28 := S3*S3 - S4
S2 := S28 + 3*S8
S36 := S2*S2
S35 := S36*S36
S30 := 2*Y
S9 := S30 + 3*X
       2*(2*Y + S30*Y + S6*S35*S35)
E1 := ------------------------------
                    S9
S12 := S28 + 2*S8
S29 := 4*X*X
S27 := S29 - X
S31 := 3*Y
       S29 - 2*S9 + 4*S6*S12*S12*S7
E2 := ------------------------------
               S31 + 2*S27
S18 := S30 + S27
       4*SIN(S4) + SIN(S7) + S18*S18
E3 := -------------------------------
            S31 + F(X,G( - S8))



Number of operations after optimization is:

Number of (+,-)-operations : 15
Number of (*)-operations : 24
Number of integer exponentiations : 0
Number of other operations : 11

\f1\s0
.fi
.in -3
.vs 15
.Sq
.Lp
.ps 11
.br
However a switch is available for stepwise performing the optimization of
a set of assignment statements, distributed over different files.
When turning on this AGAIN switch the finishing touch is not done.
Moreover,  the system is instructed to save relevant internal information
in combination with the result of the present optimization run. The
thus extended output is assumed to be stored in a file. When the 
optimization task is continued during another session this file is
assumed to be read before all other remaining files. 
This mode of operation is illustrated in 
.LP
.ps 11
\f3Example 5.2\f1
.LP
.ps 11
.LP
.in +3
.vs 10
.nf
\fC\s8
2: off acinfo$
3: on again$

4: out f5$
5: optimize in f1,f2 iname s;
6: write "end$"$
7: shut f5$

8: off again$
9: on acinfo$
10: optimize in f5,f3 iname t;

S7 := X + Y

        2
S6 := S7

S8 := COS(X)

             2        S6
S18 := SIN(X)  - COS(E  )

S9 := 3*X + 2*Y

                2                    8
       4*Y + 4*Y  + 2*S6*(S18 + 3*S8)
E1 := ---------------------------------
                     S9

        2
S15 := X

                                       2
       4*S15 - 2*S9 + 4*S6*(S18 + 2*S8) *S7
E2 := --------------------------------------
                8*S15 - 2*X + 3*Y

                         2
                  (X + Y)
E3 := (4*SIN(COS(E        )) + SIN(X + Y) + 

           2           2
       (4*X  - X + 2*Y) )/(3*Y + F(X,G( - COS(X))))

Number of operations in the total input, i.e. in the 2 input sets is: 

Number of (+,-)-operations : 22
Number of (*)-operations : 20
Number of integer exponentiations : 13
Number of other operations : 17


T17 := X + Y
T16 := T17*T17
S8 := COS(X)
T1 := SIN(X)
           T16
T2 := COS(E   )
S18 := T1*T1 - T2
T28 := 2*Y
S9 := T28 + 3*X
T6 := S18 + 3*S8
T36 := T6*T6
T35 := T36*T36
       2*(2*Y + T28*Y + T35*T35*T16)
E1 := -------------------------------
                    S9
S15 := X*X
T9 := S18 + 2*S8
T30 := 4*S15
T26 := T30 - X
T29 := 3*Y
       T30 - 2*S9 + 4*T17*T9*T9*T16
E2 := ------------------------------
               T29 + 2*T26
T19 := T28 + T26
       4*SIN(T2) + SIN(T17) + T19*T19
E3 := --------------------------------
            T29 + F(X,G( - S8))



Number of operations after optimization is:

Number of (+,-)-operations : 15
Number of (*)-operations : 24
Number of integer exponentiations : 0
Number of other operations : 11

\f1\s0
.fi
.in -3
.vs 15
.Sq
.LP
.ps 11
Since the construction of declarations in combination with some optimization
activity is based on a quite specific use of GENTRAN's symbol table, one has
to operate carefully when optimizing input in different sessions. A correct
list of declarations is only guaranteed, when the last optimization-command
is extended with the required declaration-information.
.NH
Some Possible Shortcomings and Future Versions
.LP
.ps 11
The present version of SCOPE may have some shortcomings and possibly also
some inefficiencies. However, since we are working on a second version,
as stated in [13], we do not have the intention to largely modify the
present version. 
However, we intend to improve one special aspect of the present SCOPE-version:
The combined use of SCOPE and GENTRAN. This preliminary version of the manual
will shortly be extended with the description of these combined features.
.br
Bugs and obvious deficiencies will of course be removed. 
.sp
\f3Acknowledgements\f1
.LP
.ps 11
The many discussions I had over the past years with  Barbara L. Gates,
Victor V. Goldman, Anthony C. Hearn, Jaap Smit and Paul S. Wang about the 
symbolic-numeric aspects of computer algebra have been very stimulating and 
valuable. They also contributed to the present status of SCOPE.
.br
Completion of the code would have been impossible without the dedicated
assistance of my students and the frequent discussions we had.
I certainly want to mention Ben Hulshof, Pim van den Heuvel, Marcel van 
Heerwaarden, Anco Smit, Johan de Boer and Jan Verheul.
.bp
\f3References\f1
.LP
.KS
.IP [1]
Aho, A.V., Sethi, R., Ullman, J.D.: \f2Compilers Principles, Techniques
and Tools\f1. Reading, Mass.: Addison-Wesley (1986).
.KE
.KS
.IP [2]
Breuer, M.A.: "Generation of optimal code for expressions via factorization",
\f2Comm. ACM\f1 \f312, 6\f1, 330-340 (1969).
.KE
.KS
.IP [3]
Gates, B.L.: "GENTRAN: An automatic code generation facility for REDUCE",
\f2A.C.M. SIGSAM Bulletin\f1 \f319, 3\f1, 24-42. New York: A.C.M. (1985).
.KE
.KS
.IP [4]
Gates, B.L., Wang,P.S.: "A LISP-based RATFOR code generator", \f2Proceedings
1984 MACSYMA User's Conference\f1 (V.E. Golden, ed.), 319-329.
Schenectady, N.Y.: Gen. El. (1984).
.KE
.KS
.IP [5]
Goldman, V.V., van Hulzen, J.A.: "Automatic code vectorization of arithmetic
expressions by bottom-up structure recognition", \f2Computer Algebra and 
Parallelism\f1 (J. Della Dora and J. Fitch, ed.s), 119-132. London: Academic
Press (1989).
.KE
.KS
.IP [6]
Gonzales, T., Ja' Ja', J.: "Evaluation of arithmetic expressions with 
algebraic identities", \f2SIAM J. Comp.\f1, \f311\f1, \f34\f1, 633-662
(1982).
.KE
.KS
.IP [7]
Hearn, A.C.: "Structure: The key to improved algebraic computation",
\f2Proceedings RSYMSAC2\f1 (N. Inada en T. Soma, ed.'s), 215-230.
Singapore: World Scientific Publ. (1985).
.KE
.KS
.IP [8]
Hearn, A.C.: "Optimal evaluation of algebraic expressions", \f2Proceedings
AAECC-3\f1 (J. Calmet, ed.), Springer LNCS series nr \f3229\f1, 392-403.
Heidelberg: Springer Verlag (1986).
.KE
.KS
.IP [9]
Hearn, A.C.: \f2REDUCE user's manual, version 3.3\f1. Santa Monica, CA:
The Rand Corporation (1988).
.KE
.KS
.IP [10]
van den Heuvel, P., van Hulzen, J.A., Goldman, V.V.: "Automatic generation
of FORTRAN-coded Jacobians and Hessians", \f2Proceedings EUROCAL '87\f1 (J.H.
Davenport, ed.), Springer LNCS-series nr \f3378\f1, 120-131. Heidelberg:
Springer Verlag (1989).
.KE
.KS
.IP [11]
van Hulzen, J.A.: "Breuer's grow factor algorithm in computer 
algebra", \f2Proceedings SYMSAC '81\f1 (P.S. Wang, ed.), 100-104. New York:
ACM-Press (1981).
.KE
.KS
.IP [12]
van Hulzen, J.A.: "Code optimization of multivariate polynomial schemes:
A pragmatic approach", \f2Proceedings EUROCAL '83\f1, (J.A. van Hulzen, ed.),
Springer LNCS series \f3162\f1, 286-300. Heidelberg: Springer Verlag (1983).
.KE
.KS
.IP [13]
van Hulzen, J.A.: "Current trends in source-code optimization", \f2Proceedings
JINR IV Conference on Computer Algebra and its Applications in Theoretical
Physics\f1, Dubna, May 20-25, 1990 (to appear). Also available as Memorandum 
\f3INF-90-41\f1, Department of Computer Science, Uniersity Twente (1990).
.KE
.KS
.IP [14]
Johnson, D.B., Miller, W., Minnihan, B., Wrathall, C.: "Reducibility among
floating-point graphs", \f2J. ACM\f1 \f326\f1, \f34\f1, 739-760 (1979).
.KE
.KS
.IP [15]
Knuth, D.E.: "An empirical study of Fortran programs", \f2Software Practice
and Experience\f1 \f31\f1, 105-133 (1971).
.KE
.KS
.IP [16]
Knuth, D.E.: \f2The art of computer programming, Vol.\f1 \f32\f1
(Second Edition). Reading, Mass.: Addison-Wesley (1980).
.KE
.KS
.IP [17]
Smit, J., van Hulzen, J.A., Hulshof, B.J.A.: "NETFORM and code optimizer
manual", \f2A.C.M. SIGSAM Bulletin\f1 \f315,4\f1, 23-32.
New York: A.C.M. (1981).
.KE
.KS
.IP [18]
Smit, J., van Hulzen, J.A.: "Symbolic-numeric methods in microwave
technology", \f2Proceedings EUROCAM '82\f1 (J. Calmet, ed.), Springer
LNCS series \f3144\f1, 281-288. Heidelberg: Springer Verlag (1982).
.KE
.KS
.IP [19]
Wang, P.S., Chang, T.Y.P., van Hulzen, J.A.: "Code generation and
optimization for finite element analysis", \f2Proceedings EUROSAM '84
\f1, (J.P. Fitch ed.), Springer LNCS series \f3174\f1,
237-247. Heidelberg: Springer Verlag (1984).
.KE
.bp
\f3Appendix 1: How to install the Code\f1
.sp
.LP
.ps 11
The code consists of a number of modules, collected in five files. Two of
these modules play a special role and can best be compiled separately:
gstructr, defining the \f3gstructr\f1 facilities, and ghorner, containing
the code for the Horner-rules.
.LP
.ps 11
The other modules form SCOPE. Since @{roman D} sub 0@ and all operations on it
and on its later versions @{roman D} sub i@ are defined using \f3smacros's\f1
it is essential to read in the module cosmac, containing these \f3smacro's\f1,
first. Since we also use part of
the GENTRAN code care have to be taken that GENTRAN is loaded when 
compiling the code.


REDUCE Historical
REDUCE Sourceforge Project | Historical SVN Repository | GitHub Mirror | SourceHut Mirror | NotABug Mirror | Chisel Mirror | Chisel RSS ]