3 @c % /**-----------------------------------------------------------------**
5 @c % **-----------------------------------------------------------------**
7 @c % **-----------------------------------------------------------------**
8 @c % ** First version: july 6th 2002 **
9 @c % **-----------------------------------------------------------------**/
11 @c % release 1.0: September 17th 2002
12 @c % release 1.1: December 5th 2002
13 @c % release 1.2: April 22th 2003
14 @c % release 2.0: November 21th 2005 (and now in texinfo instead of LaTeX)
15 @c % release 2.1: October 15th 2007
17 @c %/**************************************************************************
18 @c % * CLooG : the Chunky Loop Generator (experimental) *
19 @c % **************************************************************************/
20 @c %/* CAUTION: the English used is probably the worst you ever read, please
21 @c % * feel free to correct and improve it !
24 @c %\textit{"I found the ultimate transformation functions, optimization for
25 @c %static control programs is now a closed problem, I have \textnormal{just}
26 @c %to generate the target code !"}
30 @c % /*************************************************************************
31 @c % * PART I: HEADER *
32 @c % *************************************************************************/
34 @setfilename cloog.info
35 @settitle CLooG - a loop generator for scanning polyhedra
38 @include gitversion.texi
39 @set UPDATED October 15th 2007
40 @setchapternewpage odd
44 @c % /*************************************************************************
45 @c % * PART II: SUMMARY DESCRIPTION AND COPYRIGHT *
46 @c % *************************************************************************/
49 This manual is for CLooG version @value{VERSION}, a software
50 which generates loops for scanning Z-polyhedra. That is, CLooG produces a
51 code visiting each integral point of a union of parametrized
52 polyhedra. CLooG is designed to avoid control overhead and to produce a very
55 It would be quite kind to refer the following paper in any publication that
56 results from the use of the CLooG software or its library:
59 @@InProceedings@{Bas04,
60 @ @ author =@ @ @ @ @{C. Bastoul@},
61 @ @ title =@ @ @ @ @ @{Code Generation in the Polyhedral Model
62 @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ Is Easier Than You Think@},
63 @ @ booktitle = @{PACT'13 IEEE International Conference on
64 @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ Parallel Architecture and Compilation Techniques@},
65 @ @ year =@ @ @ @ @ @ 2004,
66 @ @ pages =@ @ @ @ @ @{7--16@},
67 @ @ month =@ @ @ @ @ @{september@},
68 @ @ address =@ @ @ @{Juan-les-Pins@}
72 Copyright @copyright{} 2002-2005 C@'edric Bastoul.
75 Permission is granted to copy, distribute and/or modify this document under
76 the terms of the GNU Free Documentation License, Version 1.2
77 published by the Free Software Foundation. To receive a copy of the
78 GNU Free Documentation License, write to the Free Software Foundation, Inc.,
79 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
83 @c % /*************************************************************************
84 @c % * PART III: TITLEPAGE, CONTENTS, COPYRIGHT *
85 @c % *************************************************************************/
88 @subtitle A Loop Generator For Scanning Polyhedra
89 @subtitle Edition @value{EDITION}, for CLooG @value{VERSION}
90 @subtitle @value{UPDATED}
91 @author C@'edric Bastoul
93 @c The following two commands start the copyright page.
95 @noindent (September 2001)
97 @item C@'edric Bastoul
98 SCHEDULES GENERATE !!! I just need to apply them now, where can I find
99 a good code generator ?!
102 Hmmm. I fear that if you want something powerful enough, you'll have to
106 @vskip 0pt plus 1filll
110 @c Output the table of contents at the beginning.
113 @c % /*************************************************************************
114 @c % * PART IV: TOP NODE AND MASTER MENU *
115 @c % *************************************************************************/
135 @c % /*************************************************************************
136 @c % * PART V: BODY OF THE DOCUMENT *
137 @c % *************************************************************************/
139 @c % ****************************** INTRODUCTION ******************************
141 @chapter Introduction
142 CLooG is a free software and library generating loops for scanning Z-polyhedra.
143 That is, it finds a code (e.g. in C, FORTRAN...) that reaches each integral
144 point of one or more parameterized polyhedra. CLooG has been originally
145 written to solve the code generation problem for optimizing compilers based on
146 the polytope model. Nevertheless it is used now in various area, e.g., to build
147 control automata for high-level synthesis or to find the best polynomial
148 approximation of a function. CLooG may help in any situation where scanning
149 polyhedra matters. It uses the best state-of-the-art code generation
150 algorithm known as the Quiller@'e et al. algorithm (@pxref{Qui00})
151 with our own improvements and extensions (@pxref{Bas04}).
152 The user has full control on generated code quality.
153 On one hand, generated code size has to be tuned for sake of
154 readability or instruction cache use. On the other hand, we must ensure that
155 a bad control management does not hamper performance of the generated code,
156 for instance by producing redundant guards or complex loop bounds.
157 CLooG is specially designed to avoid control overhead and to produce a very
160 CLooG stands for @emph{Chunky Loop Generator}: it is a part of the Chunky
161 project, a research tool for data locality improvement (@pxref{Bas03a}).
163 also to be the back-end of automatic parallelizers like LooPo (@pxref{Gri04}).
165 compilable code oriented and provides powerful program transformation
166 facilities. Mainly, it allows the user to specify very general schedules where,
167 e.g., unimodularity or invertibility of the transformation doesn't matter.
169 The current version is still under
170 evaluation, and there is no guarantee that the upward compatibility
171 will be respected (but the previous API has been stable for two years,
172 we hope this one will be as successful -and we believe it-).
173 A lot of reports are necessary to freeze the library
174 API and the input file shape. Most API changes from 0.12.x to 0.14.x
175 have been requested by the users themselves.
176 Thus you are very welcome and encouraged
177 to post reports on bugs, wishes, critics, comments, suggestions or
178 successful experiences in the forum of @code{http://www.CLooG.org}
179 or to send them to cedric.bastoul@@inria.fr directly.
187 @section Basically, what's the point ?
188 If you want to use CLooG, this is because you want to scan or to find
189 something inside the integral points of a set of polyhedra. There are many
190 reasons for that. Maybe you need the generated code itself because it
191 actually implements a very smart program transformation you found.
192 Maybe you want to use the generated code
193 because you know that the solution of your problem belongs to the integral
194 points of those damned polyhedra and you don't know which one. Maybe you just
195 want to know if a polyhedron has integral points depending on some parameters,
196 which is the lexicographic minimum, maximum, the third on the basis of the
197 left etc. Probably you have your own reasons to use CLooG.
199 Let us illustrate a basic use of CLooG. Suppose we have a set of affine
200 constraints that describes a part of a whatever-dimensional space,
201 called a @strong{domain}, and we
202 want to scan it. Let us consider for instance the following set of constraints
204 and @samp{j} are the unknown (the two dimensions of the space) and
205 @samp{m} and @samp{n} are the parameters (some symbolic constants):
213 Let us also consider that we have a partial knowledge of the parameter values,
214 called the @strong{context}, expressed as affine constraints as well,
222 Note that using parameters is optional, if you are not comfortable with
223 parameter manipulation, just replace them with any scalar value that fits
224 @code{m>=2} and @code{n>=2}.
225 A graphical representation of this part of the 2-dimensional space, where
226 the integral points are represented using heavy dots would be for instance:
228 @image{images/basic,6cm}
230 The affine constraints of both the domain and the context are what we will
231 provide to CLooG as input (in a particular shape that will be described later).
232 The output of CLooG is a pseudo-code to scan the integral points of the
233 input domain according to the context:
236 for (i=2;i<=n;i++) @{
237 for (j=2;j<=min(m,-i+n+2);j++) @{
243 If you felt such a basic example is yet interesting, there is a good chance
244 that CLooG is appropriate for you. CLooG can do much more: scanning several
245 polyhedra or unions of polyhedra at the same time, applying general affine
246 transformations to the polyhedra, generate compilable code etc. Welcome
247 to the CLooG's user's guide !
250 @section Defining a Scanning Order: Scattering Functions
251 In CLooG, domains only define the set of integral points to scan and their
252 coordinates. In particular, CLooG is free to choose the scanning order for
253 generating the most efficient code. This means, for optimizing/parallelizing
254 compiler people, that CLooG doesn't make any speculation on dependences on and
255 between statements (by the way, it's not its job !).
256 For instance, if an user give to
257 CLooG only two domains @code{S1:1<=i<=n}, @code{S2:1<=i<=n} and the context
258 @code{n>=1}, the following pseudo-codes are considered to be equivalent:
262 /* A convenient target pseudo-code. */
263 for (i=1;i<=N;i++) @{
266 for (i=1;i<=N;i++) @{
274 /* Another convenient target pseudo-code. */
275 for (i=1;i<=N;i++) @{
282 The default behaviour
283 of CLooG is to generate the second one, since it is optimized in control.
284 It is right if there are no data dependences
285 between @code{S1} and @code{S2}, but wrong otherwise.
287 Thus it is often useful to force scanning to respect a given order. This can be
288 done in CLooG by using @strong{scattering functions}. Scattering is a
289 shortcut for scheduling, allocation, chunking functions and the like we can
290 find in the restructuring compilation literature. There are a lot of reasons
291 to scatter the integral points of the domains (i.e. the statement instances
292 of a program, for compilation people), parallelization or optimization are good
293 examples. For instance, if the user wants for any reason to set some
294 precedence constraints between the statements of our example above
295 in order to force the generation of the
296 first code, he can do it easily by setting (for example) the following
297 scheduling functions:
300 $$\theta _{S1}(i) = (1)$$
301 $$\theta _{S2}(j) = (2)$$
313 This scattering means that each integral point of the domain @code{S1}
314 is scanned at logical date @code{1} while each integral point of the domain
315 @code{S2} is scanned at logical date @code{2}. As a result, the whole
316 domain @code{S1} is scanned before domain @code{S2} and the first code in our
317 example is generated.
319 The user can set every kind of affine scanning order thanks to the
320 scattering functions. Each domain has its own scattering function and
321 each scattering function may be multi-dimensional. A multi-dimensional logical
322 date may be seen as classical date (year,month,day,hour,minute,etc.) where
323 the first dimensions are the most significant. Each scattering dimension
324 may depend linearly on the original dimensions (e.g., @code{i}), the
325 parameters (e.g., @code{n}) ans scalars (e.g., @code{2}).
327 A very useful example of multi-dimensional scattering functions is, for
328 compilation people, the scheduling of the original program.
329 The basic data to use for code generation are statement iteration domains.
330 As we saw, these data are not sufficient to rebuild the original
331 program (what is the ordering between instances of different statements ?).
332 The missing data can be put in the scattering functions as the original
333 scheduling. The method to compute it is quite simple (@pxref{Fea92}). The idea is to
334 build an abstract syntax tree of the program and to read the scheduling for
335 each statement. For instance, let us consider the following implementation of
336 a Cholesky factorization:
340 /* A Cholesky factorization kernel. */
341 for (i=1;i<=N;i++) @{
342 for (j=1;j<=i-1;j++) @{
343 a[i][i] -= a[i][j] ; /* S1 */
345 a[i][i] = sqrt(a[i][i]) ; /* S2 */
346 for (j=i+1;j<=N;j++) @{
347 for (k=1;k<=i-1;k++) @{
348 a[j][i] -= a[j][k]*a[i][k] ; /* S3 */
350 a[j][i] /= a[i][i] ; /* S4 */
357 The corresponding abstract syntax tree is given in the following figure.
358 It directly gives the scattering functions (schedules) for all the
359 statements of the program.
361 @image{images/tree,6cm}
365 \hbox{$ \cases{ \theta _{S1}(i,j)^T &$= (0,i,0,j,0)^T$\cr
366 \theta _{S2}(i) &$= (0,i,1)^T$\cr
367 \theta _{S3}(i,j,k)^T &$= (0,i,2,j,0,k,0)^T$\cr
368 \theta _{S4}(i,j)^T &$= (0,i,2,j,1)^T$}$}
375 T_S1(i,j)^T = (0,i,0,j,0)^T
377 T_S3(i,j,k)^T = (0,i,2,j,0,k,0)^T
378 T_S4(i,j)^T = (0,i,2,j,1)^T
383 These schedules depend on the iterators and give for each instance of each
384 statement a unique execution date. Using such scattering functions allow
385 CLooG to re-generate the input code.
391 @c % ***********************Using the CLooG Software **************************
393 @chapter Using the CLooG Software
398 * Writing The Input File::
404 @c %/*************************************************************************
405 @c % * A FIRST EXAMPLE *
406 @c % *************************************************************************/
407 @node A First Example
408 @section A First Example
409 CLooG takes as input a file that must be written accordingly to a grammar
410 described in depth in a further section (@pxref{Writing The Input File}).
411 Moreover it supports many options to tune the target code presentation or
412 quality as discussed in a dedicated section (@pxref{Calling CLooG}).
414 of CLooG is not very complex and we present in this section how to generate the
415 code corresponding to a basic example discussed earlier (@pxref{Basics}).
417 The problem is to find the code that scans a 2-dimensional polyhedron
418 where @samp{i} and @samp{j} are the unknown (the two dimensions of the space)
419 and @samp{m} and @samp{n} are the parameters (the symbolic constants),
420 defined by the following set of constraints:
428 @noindent We also consider a partial knowledge of the parameter values,
429 expressed thanks to the following affine constraints:
437 An input file that corresponds to this problem, and asks for a generated
438 code in C, may be the following. Note that we do not describe here precisely
439 the structure and the components of this file (@pxref{Writing The Input File}
440 for such information, if you feel it necessary):
443 # ---------------------- CONTEXT ----------------------
446 # Context (constraints on two parameters)
447 2 4 # 2 lines and 4 columns
448 # eq/in m n 1 eq/in: 1 for inequality >=0, 0 for equality =0
449 1 1 0 -2 # 1*m + 0*n -2*1 >= 0, i.e. m>=2
450 1 0 1 -2 # 0*m + 1*n -2*1 >= 0, i.e. n>=2
452 1 # We want to set manually the parameter names
453 m n # parameter names
455 # --------------------- STATEMENTS --------------------
456 1 # Number of statements
458 1 # First statement: one domain
460 5 6 # 5 lines and 6 columns
462 1 1 0 0 0 -2 # i >= 2
463 1 -1 0 0 1 0 # i <= n
464 1 0 1 0 0 -2 # j >= 2
465 1 0 -1 1 0 0 # j <= m
466 1 -1 -1 0 1 2 # n+2-i>=j
467 0 0 0 # for future options
469 1 # We want to set manually the iterator names
472 # --------------------- SCATTERING --------------------
473 0 # No scattering functions
476 This file may be called @samp{basic.cloog}
477 (this example is provided in the CLooG distribution as
478 @code{test/manual_basic.cloog}) and we can ask CLooG to process it
479 and to generate the code by a simple calling to CLooG with this file as input:
480 @samp{cloog basic.cloog}. By default, CLooG will print the generated code in
485 /* Generated by CLooG v@value{VERSION} in 0.00s. */
486 for (i=2;i<=n;i++) @{
487 for (j=2;j<=min(m,-i+n+2);j++) @{
494 @c %/*************************************************************************
496 @c % *************************************************************************/
497 @node Writing The Input File
498 @section Writing The Input File
499 The input text file contains a problem description, i.e. the context,
500 the domains and the scattering functions.
501 Because CLooG is very 'compilable code generation oriented', we can associate
502 some additional informations to each domain. We call this association a
503 @emph{statement}. The set of all informations is
504 called a @emph{program}. The input file respects the grammar below
505 (terminals are preceded by "_"):
509 Program ::= Context Statements Scattering
510 Context ::= Language Domain_union Naming
511 Statements ::= Nb_statements Statement_list Naming
512 Scatterings ::= Nb_functions Scattering_list Naming
513 Naming ::= Option Name_list
514 Name_list ::= _String Name_list | (void)
515 Statement_list ::= Statement Statement_list | (void)
516 Domain_list ::= _Domain Domain_list | (void)
517 Scattering_list ::= Domain_union Scattering_list | (void)
518 Statement ::= Iteration_domain 0 0 0
519 Iteration_domain ::= Domain_union
520 Domain_union ::= Nb_domains Domain_list
523 Nb_statements ::= _Integer
524 Nb_domains ::= _Integer
525 Nb_functions ::= _Integer
528 Note: if there is only one domain in a @samp{Domain_union},
529 i.e., if @samp{Nb_domains} is 1, then this 1 may be omitted.
532 @item @samp{Context} represents the informations that are
533 shared by all the statements. It consists on
534 the language used (which can be @samp{c} for C or @samp{f} for FORTRAN 90)
535 and the global constraints on parameters.
536 These constraints are essential
537 since they give to CLooG the number of parameters. If there is no
538 parameter or no constraints on parameters, just give a constraint
539 always satisfied like @math{1 \geq 0}. @samp{Naming} sets the parameter
541 If the naming option @samp{Option} is 1, parameter names will be read
542 on the next line. There must be exactly as many names as parameters.
543 If the naming option @samp{Option} is 0, parameter names are
544 automatically generated. The name of the first parameter will
545 be @samp{M}, and the name of the @math{(n+1)^{th}} parameter directly
546 follows the name of the @math{n^{th}} parameter in ASCII code.
547 It is the user responsibility to ensure that parameter names,
548 iterators and scattering dimension names are different.
549 @item @samp{Statements} represents the informations on the statements.
550 @samp{Nb_statements} is the number of statements in the program,
551 i.e. the number of @samp{Statement} items in the @samp{Statement_list}.
552 @samp{Statement} represents the informations on a given statement.
553 To each statement is associated a domain
554 (the statement iteration domain: @samp{Iteration_domain}) and three
555 zeroes that represents future options.
556 @samp{Naming} sets the iterator names. If the naming option
557 @samp{Option} is 1, the iterator names
558 will be read on the next line. There must be exactly as many names as
559 nesting level in the deepest iteration domain. If the naming option
560 @samp{Option} is 0, iterator names are automatically generated.
561 The iterator name of the outermost loop will be @samp{i}, and the
562 iterator name of the loop at level @math{n+1} directly follows the
563 iterator name of the loop at level @math{n} in ASCII code.
564 @item @samp{Scatterings} represents the informations on scattering functions.
565 @samp{Nb_functions} is the number of functions (it must be
566 equal to the number of statements or 0 if there is no scattering
567 function). The functions themselves are represented through
568 @samp{Scattering_list}.
569 @samp{Naming} sets the scattering dimension names. If the naming option
570 @samp{Option} is 1, the scattering dimension names will be read on the
572 There must be exactly as many names as scattering dimensions. If the
573 naming option @samp{Option} is 0, scattering dimension names are automatically
574 generated. The name of the @math{n^{th}} scattering dimension
579 * Domain Representation::
580 * Scattering Representation::
583 @node Domain Representation
584 @subsection Domain Representation
585 As shown by the grammar, the input file describes the various informations
586 thanks to characters, integers and domains. Each domain is defined by a set of
587 constraints in the PolyLib format (@pxref{Wil93}). They have the
590 @item some optional comment lines beginning with @samp{#},
591 @item the row and column numbers, possibly followed by comments,
592 @item the constraint rows, each row corresponds to a constraint the
593 domain have to satisfy. Each row must be on a single line and is possibly
594 followed by comments. The constraint is an equality @math{p(x) = 0} if the
595 first element is 0, an inequality @math{p(x) \geq 0} if the first element
596 is 1. The next elements are the unknown coefficients, followed by
597 the parameter coefficients. The last element is the constant factor.
599 For instance, assuming that @samp{i}, @samp{j} and @samp{k} are iterators and
600 @samp{m} and @samp{n} are parameters, the domain defined by the following
605 \hbox{$ \cases{ -i + m &$\geq 0$\cr
607 i + j - k &$\geq 0$}$}
621 @noindent can be written in the input file as follows :
626 3 7 # 3 lines and 7 columns
628 1 -1 0 0 1 0 0 # -i + m >= 0
629 1 0 -1 0 0 1 0 # -j + n >= 0
630 1 1 1 -1 0 0 0 # i + j - k >= 0
634 Each iteration domain @samp{Iteration_domain} of a given statement
635 is a union of polyhedra
636 @samp{Domain_union}. A union is defined by its number of elements
637 @samp{Nb_domains} and the elements themselves @samp{Domain_list}.
638 For instance, let us consider the following pseudo-code:
642 for (i=1;i<=n;i++) @{
643 if ((i >= m) || (i <= 2*m))
651 @noindent The iteration domain of @samp{S1} can be divided into two
652 polyhedra and written in the input file as follows:
656 2 # Number of polyhedra in the union
658 3 5 # 3 lines and 5 columns
664 3 5 # 3 lines and 5 columns
668 1 -1 2 0 0 # i <= 2*m
672 @node Scattering Representation
673 @subsection Scattering Function Representation
674 Scattering functions are depicted in the input file thanks a representation
675 very close to the domain one.
676 An integer gives the number of functions @samp{Nb_functions} and each function
677 is represented by a domain. Each line of the domain corresponds to an equality
678 defining a dimension of the function. Note that at present
679 (CLooG @value{VERSION})
680 @strong{all functions must have the same scattering dimension number}. If a
681 user wants to set scattering functions with different dimensionality, he has
682 to complete the smaller one with zeroes to reach the maximum dimensionality.
683 For instance, let us consider the following code and
684 scheduling functions:
688 for (i=1;i<=n;i++) @{
689 if ((i >= m) || (i <= 2*m))
699 \hbox{$ \cases{ \theta _{S1}(i) &$= (i,0)^T$\cr
700 \theta _{S2}(i,j)^T &$= (n,i+j)^T$}$}
708 T_S2(i,j)^T = (n,i+j)^T
714 @noindent This scheduling can be written in the input file as follows:
718 2 # Number of scattering functions
720 2 7 # 2 lines and 7 columns
721 # eq/in c1 c2 i m n 1
722 0 1 0 -1 0 0 0 # c1 = i
723 0 0 1 0 0 0 0 # c2 = 0
725 2 8 # 2 lines and 8 columns
726 # eq/in c1 c2 i j m n 1
727 0 1 0 0 0 0 -1 0 # c1 = n
728 0 0 1 -1 -1 0 0 0 # c2 = i+j
731 The complete input file for the user who wants to generate the code for this
732 example with the preceding scheduling would be
733 (this file is provided in the CLooG distribution
734 as @code{test/manual_scattering.cloog}:
737 # ---------------------- CONTEXT ----------------------
740 # Context (no constraints on two parameters)
741 1 4 # 1 lines and 4 columns
743 1 0 0 0 # 0 >= 0, always true
745 1 # We want to set manually the parameter names
746 m n # parameter names
748 # --------------------- STATEMENTS --------------------
749 2 # Number of statements
751 2 # First statement: two domains
753 3 5 # 3 lines and 5 columns
759 3 5 # 3 lines and 5 columns
763 1 -1 2 0 0 # i <= 2*m
764 0 0 0 # for future options
766 1 # Second statement: one domain
767 4 6 # 4 lines and 6 columns
769 1 1 0 0 0 -1 # i >= 1
770 1 -1 0 0 1 0 # i <= n
771 1 -1 1 0 0 -1 # j >= i+1
772 1 0 -1 1 0 0 # j <= m
773 0 0 0 # for future options
775 1 # We want to set manually the iterator names
778 # --------------------- SCATTERING --------------------
779 2 # Scattering functions
781 2 7 # 2 lines and 7 columns
782 # eq/in p1 p2 i m n 1
783 0 1 0 -1 0 0 0 # p1 = i
784 0 0 1 0 0 0 0 # p2 = 0
786 2 8 # 2 lines and 8 columns
787 # eq/in p1 p2 i j m n 1
788 0 1 0 0 0 0 -1 0 # p1 = n
789 0 0 1 -1 -1 0 0 0 # p2 = i+j
791 1 # We want to set manually the scattering dimension names
792 p1 p2 # scattering dimension names
796 @c %/*************************************************************************
797 @c % * Calling CLooG *
798 @c % *************************************************************************/
800 @section Calling CLooG
801 CLooG is called by the following command:
803 cloog [ options | file ]
805 The default behavior of CLooG is to read the input informations from a file and
806 to print the generated code or pseudo-code on the standard output.
807 CLooG's behavior and the output code shape is under the user control thanks
808 to many options which are detailed a further section (@pxref{CLooG Options}).
809 @code{file} is the input file. @code{stdin} is a special value: when used,
810 input is standard input. For instance, we can call CLooG to treat the
811 input file @code{basic.cloog} with default options by typing:
812 @code{cloog basic.cloog} or @code{more basic.cloog | cloog stdin}.
814 @c %/*************************************************************************
815 @c % * CLooG Options *
816 @c % *************************************************************************/
818 @section CLooG Options
821 * Last Depth to Optimize Control::
822 * First Depth to Optimize Control::
823 * Simplify Convex Hull::
824 * Once Time Loop Elimination::
825 * Equality Spreading::
826 * First Level for Spreading::
836 @node Last Depth to Optimize Control
837 @subsection Last Depth to Optimize Control @code{-l <depth>}
839 @code{-l <depth>}: this option sets the last loop depth to be optimized in
840 control. The higher this depth, the less control overhead.
841 For instance, with some input file, a user can generate
842 different pseudo-codes with different @code{depth} values as shown below.
845 /* Generated using a given input file and @strong{option -l 1} */
846 for (i=0;i<=M;i++) @{
848 for (j=0;j<=N;j++) @{
851 for (j=0;j<=N;j++) @{
860 /* Generated using the same input file but @strong{option -l 2} */
861 for (i=0;i<=M;i++) @{
863 for (j=0;j<=N;j++) @{
871 In this example we can see that this option can change the operation
872 execution order between statements. Let us remind that CLooG does not
873 make any speculation on dependences between statements
874 (@pxref{Scattering}). Thus if nothing (i.e. scattering functions)
875 forbids this, CLooG considers the above codes to be equivalent.
876 If there is no scattering functions, the minimum value for @code{depth}
877 is 1 (in the case of 0, the user doesn't really need a loop generator !),
878 and the number of scattering dimensions otherwise (CLooG will warn the
879 user if he doesn't respect such constraint).
880 The maximum value for depth is -1 (infinity).
881 Default value is infinity.
883 @node First Depth to Optimize Control
884 @subsection First Depth to Optimize Control @code{-f <depth>}
886 @code{-f <depth>}: this option sets the first loop depth to be optimized
887 in control. The lower this depth, the less control overhead (and the longer
888 the generated code). For instance, with some input file, a user
889 can generate different pseudo-codes with different @code{depth} values
891 The minimum value for @code{depth} is 1, and the
892 maximum value is -1 (infinity).
896 /* Generated using a given input file and @strong{option -f 3} */
897 for (i=1;i<=N;i++) @{
898 for (j=1;j<=M;j++) @{
909 /* Generated using the same input file but @strong{option -f 2} */
910 for (i=1;i<=N;i++) @{
911 for (j=1;j<=9;j++) @{
914 for (j=10;j<=M;j++) @{
922 @node Simple Convex Hull
923 @subsection Simple Convex Hull @code{-sh <boolean>}
925 @code{-sh <boolean>}: this option enables (@code{boolean=1})
926 or forbids (@code{boolean=0}) the use of an overapproximation
927 of the convex hull that may be easier to compute
928 (especially in the isl backend) and that may result in
930 This option works only for generated code without
931 code duplication (it means, you have to tune @code{-f} and
932 @code{-l} options first to generate only a loop nest with internal
933 guards). For instance, with the input file @code{test/union.cloog}, a user
934 can generate different pseudo-codes as shown below.
938 /* Generated using test/union.cloog and @strong{option -f -1 -l 2 -override} */
939 for (i=0;i<=11;i++) @{
940 for (j=max(0,5*i-50);j<=min(15,5*i+10);j++) @{
941 if ((i <= 10) && (j <= 10)) @{
944 if ((i >= 1) && (j >= 5)) @{
953 /* Generated using the same input file but @strong{option -sh 1 -f -1 -l 2 -override} */
954 for (i=0;i<=11;i++) @{
955 for (j=0;j<=15;j++) @{
956 if ((i <= 10) && (j <= 10)) @{
959 if ((i >= 1) && (j >= 5)) @{
967 @node Once Time Loop Elimination
968 @subsection Once Time Loop Elimination @code{-otl <boolean>}
970 @code{-otl <boolean>}: this option allows (@code{boolean=1}) or
971 forbids (@code{boolean=0}) the simplification of loops running
972 once. Default value is 1.
975 /* Generated using a given input file and @strong{option -otl 0} */
976 for (j=i+1;j<=i+1;j++) @{
983 /* Generated using the same input file but @strong{option -otl 1} */
990 @node Equality Spreading
991 @subsection Equality Spreading @code{-esp <boolean>}
993 @code{-esp <boolean>}: this option allows (@code{boolean=1}) or
994 forbids (@code{boolean=0}) values spreading when there
995 are equalities. Default value is 1.
998 /* Generated using a given input file and @strong{option -esp 0} */
1001 for (k=i;k<=j+M;k++) @{
1008 /* Generated using the same input file but @strong{option -esp 1} */
1009 for (k=M+2;k<=N+M;k++) @{
1010 S1(i = M+2, j = N) ;
1016 @node First Level for Spreading
1017 @subsection First Level for Spreading @code{-fsp <level>}
1019 @code{-fsp <level>}: it can be useful to set a
1020 first level to begin equality spreading. Particularly when using
1021 scattering functions, the user may want to see the scattering dimension
1022 values instead of spreading or hiding them. If user has set a
1023 spreading, @code{level} is
1024 the first level to start it. Default value is 1.
1027 /* Generated using a given input file and @strong{option -fsp 1} */
1028 for (j=0;j<=N+M;j++) @{
1031 for (j=0;j<=N+M;j++) @{
1038 /* Generated using the same input file but @strong{option -fsp 2} */
1040 for (j=0;j<=c1+M;j++) @{
1044 for (j=0;j<=N+c1;j++) @{
1051 @node Statement Block
1052 @subsection Statement Block @code{-block <boolean>}
1054 @code{-block <boolean>}: this option allows (@code{boolean=1}) to
1055 create a statement block for each new iterator, even if there is only
1056 an equality. This can be useful in order to parse the generated
1057 pseudo-code. When @code{boolean} is set to 0 or when the generation
1058 language is FORTRAN, this feature is disabled. Default value is 0.
1061 /* Generated using a given input file and @strong{option -block 0} */
1069 /* Generated using the same input file but @strong{option -block 1} */
1080 @subsection Loop Strides @code{-strides <boolean>}
1082 @code{-strides <boolean>}: this options allows (@code{boolean=1}) to
1083 handle non-unit strides for loop increments. This can remove a lot of
1084 guards and make the generated code more efficient. Default value is 0.
1087 /* Generated using a given input file and @strong{option -strides 0} */
1088 for (i=1;i<=n;i++) @{
1100 /* Generated using the same input file but @strong{option -strides 1} */
1101 for (i=2;i<=n;i+=2) @{
1110 @node Compilable Code
1111 @subsection Compilable Code @code{-compilable <value>}
1113 @code{-compilable <value>}: this options allows (@code{value} is not 0)
1114 to generate a compilable code where all parameters have the integral value
1115 @code{value}. This option creates a macro for each statement. Since
1116 CLooG do not know anything about the statement sources, it fills the
1117 macros with a basic increment that computes the total number of
1118 scanned integral points. The user may change easily the macros according
1119 to his own needs. This option is possible only if the generated code is
1120 in C. Default value is 0.
1123 /* Generated using a given input file and @strong{option -compilable 0} */
1124 for (i=0;i<=n;i++) @{
1125 for (j=0;j<=n;j++) @{
1134 /* Generated using the same input file but @strong{option -compilable 10} */
1135 /* DON'T FORGET TO USE -lm OPTION TO COMPILE. */
1137 /* Useful headers. */
1142 /* Parameter value. */
1145 /* Statement macros (please set). */
1146 #define S1(i,j) @{total++;@}
1147 #define S2(i,j) @{total++;@}
1148 #define S3(i) @{total++;@}
1151 /* Original iterators. */
1154 int n=PARVAL, total=0 ;
1156 for (i=0;i<=n;i++) @{
1157 for (j=0;j<=n;j++) @{
1164 printf("Number of integral points: %d.\n",total) ;
1170 @subsection Callable Code @code{-callable <boolean>}
1172 @code{-callable <boolean>}: if @code{boolean=1}, then a @code{test}
1173 function will be generated that has the parameters as arguments.
1174 Similarly to the @code{-compilable} option,
1175 a macro for each statement is generated. The generated definitions of
1176 these macros are as used during the correctness testing, but they
1177 can easily be changed by the user to suit her own needs.
1178 This option is only available if the target language is C.
1179 The default value is 0.
1182 /* Generated from double.cloog with @strong{option -callable 0} */
1183 for (i=0;i<=M;i++) @{
1185 for (j=0;j<=N;j++) @{
1193 /* Generated from double.cloog with @strong{option -callable 1} */
1194 extern void hash(int);
1196 /* Useful macros. */
1197 #define floord(n,d) (((n)<0) ? ((n)-(d)+1)/(d) : (n)/(d))
1198 #define ceild(n,d) (((n)<0) ? (n)/(d) : ((n)+(d)+1)/(d))
1199 #define max(x,y) ((x) > (y) ? (x) : (y))
1200 #define min(x,y) ((x) < (y) ? (x) : (y))
1202 #define S1(i) @{ hash(1); hash(i); @}
1203 #define S2(i,j) @{ hash(2); hash(i); hash(j); @}
1204 #define S3(i,j) @{ hash(3); hash(i); hash(j); @}
1205 #define S4(i) @{ hash(4); hash(i); @}
1207 void test(int M, int N)
1209 /* Original iterators. */
1211 for (i=0;i<=M;i++) @{
1213 for (j=0;j<=N;j++) @{
1223 @subsection Output @code{-o <output>}
1225 @code{-o <output>}: this option sets the output file. @code{stdout} is a
1226 special value: when used, output is standard output.
1227 Default value is @code{stdout}.
1230 @subsection Help @code{--help} or @code{-h}
1232 @code{--help} or @code{-h}: this option ask CLooG to print a short help.
1235 @subsection Version @code{--version} or @code{-v}
1237 @code{--version} or @code{-v}: this option ask CLooG to print some version
1241 @subsection Quiet @code{--quiet} or @code{-q}
1243 @code{--quiet} or @code{-q}: this option tells CLooG not to print
1244 any informational messages.
1247 @c %/*************************************************************************
1248 @c % * A Full Example *
1249 @c % *************************************************************************/
1251 @section A Full Example
1253 Let us consider the allocation problem of a Gaussian elimination, i.e. we want
1254 to distribute the various statement instances of the compute kernel onto
1255 different processors. The original code is the following:
1258 for (i=1;j<=N-1;i++) @{
1259 for (j=i+1;j<=N;j++) @{
1260 c[i][j] = a[j][i]/a[i][i] ; /* S1 */
1261 for (k=i+1;k<=N;k++) @{
1262 a[j][k] -= c[i][j]*a[i][k] ; /* S2 */
1269 @noindent The best affine allocation functions can be found by any good automatic
1270 parallelizer like LooPo (@pxref{Gri04}):
1274 \hbox{$ \cases{ \theta _{S1}(i,j)^T &$= (i)$\cr
1275 \theta _{S2}(i,j,k)^T &$= (k)$}$}
1288 @noindent To ensure that on each processor, the set of statement instances is
1289 executed according to the original ordering, we add as minor scattering
1290 dimensions the original scheduling (@pxref{Scattering}):
1294 \hbox{$ \cases{ \theta _{S1}(i,j)^T &$= (i,0,i,0,j,0)^T$\cr
1295 \theta _{S2}(i,j,k)^T &$= (k,0,i,0,j,1,k,0)^T$}$}
1302 T_S1(i,j)^T = (i,0,i,0,j,0)^T
1303 T_S2(i,j,k)^T = (k,0,i,0,j,1,k,0)^T
1308 @noindent To ensure that the scattering functions have the same dimensionality, we
1309 complete the first function with zeroes
1310 (this is a CLooG @value{VERSION} and previous versions requirement,
1311 it should be removed in a future version, don't worry it's absolutely legal !):
1315 \hbox{$ \cases{ \theta _{S1}(i,j)^T &$= (i,0,i,0,j,0,0,0)^T$\cr
1316 \theta _{S2}(i,j,k)^T &$= (k,0,i,0,j,1,k,0)^T$}$}
1323 T_S1(i,j)^T = (i,0,i,0,j,0,0,0)^T
1324 T_S2(i,j,k)^T = (k,0,i,0,j,1,k,0)^T
1329 @noindent The input file corresponding to this code generation problem
1330 could be (this file is provided in the CLooG distribution
1331 as @code{test/manual_gauss.cloog}:
1334 # ---------------------- CONTEXT ----------------------
1337 # Context (no constraints on one parameter)
1338 1 3 # 1 line and 3 columns
1340 1 0 0 # 0 >= 0, always true
1342 1 # We want to set manually the parameter name
1345 # --------------------- STATEMENTS --------------------
1346 2 # Number of statements
1348 1 # First statement: one domain
1349 4 5 # 4 lines and 3 columns
1352 1 -1 0 1 -1 # i <= n-1
1353 1 -1 1 0 -1 # j >= i+1
1355 0 0 0 # for future options
1358 # Second statement: one domain
1359 6 6 # 6 lines and 3 columns
1361 1 1 0 0 0 -1 # i >= 1
1362 1 -1 0 0 1 -1 # i <= n-1
1363 1 -1 1 0 0 -1 # j >= i+1
1364 1 0 -1 0 1 0 # j <= n
1365 1 -1 0 1 0 -1 # k >= i+1
1366 1 0 0 -1 1 0 # k <= n
1367 0 0 0 # for future options
1369 0 # We let CLooG set the iterator names
1371 # --------------------- SCATTERING --------------------
1372 2 # Scattering functions
1374 8 13 # 3 lines and 3 columns
1375 # eq/in p1 p2 p3 p4 p5 p6 p7 p8 i j n 1
1376 0 1 0 0 0 0 0 0 0 -1 0 0 0 # p1 = i
1377 0 0 1 0 0 0 0 0 0 0 0 0 0 # p2 = 0
1378 0 0 0 1 0 0 0 0 0 -1 0 0 0 # p3 = i
1379 0 0 0 0 1 0 0 0 0 0 0 0 0 # p4 = 0
1380 0 0 0 0 0 1 0 0 0 0 -1 0 0 # p5 = j
1381 0 0 0 0 0 0 1 0 0 0 0 0 0 # p6 = 0
1382 0 0 0 0 0 0 0 1 0 0 0 0 0 # p7 = 0
1383 0 0 0 0 0 0 0 0 1 0 0 0 0 # p8 = 0
1385 8 14 # 3 lines and 3 columns
1386 # eq/in p1 p2 p3 p4 p5 p6 p7 p8 i j k n 1
1387 0 1 0 0 0 0 0 0 0 0 0 -1 0 0 # p1 = k
1388 0 0 1 0 0 0 0 0 0 0 0 0 0 0 # p2 = 0
1389 0 0 0 1 0 0 0 0 0 -1 0 0 0 0 # p3 = i
1390 0 0 0 0 1 0 0 0 0 0 0 0 0 0 # p4 = 0
1391 0 0 0 0 0 1 0 0 0 0 -1 0 0 0 # p5 = j
1392 0 0 0 0 0 0 1 0 0 0 0 0 0 -1 # p6 = 1
1393 0 0 0 0 0 0 0 1 0 0 0 -1 0 0 # p7 = k
1394 0 0 0 0 0 0 0 0 1 0 0 0 0 0 # p8 = 0
1396 1 # We want to set manually the scattering dimension names
1397 p1 p2 p3 p4 p5 p6 p7 p8 # scattering dimension names
1400 Calling CLooG, with for instance the command line
1401 @code{cloog -fsp 2 gauss.cloog} for a better view
1402 of the allocation (the processor number is given by @code{p1}),
1403 will result on the following target code that actually implements
1404 the transformation. A minor processing on the dimension @code{p1}
1405 to implement, e.g., MPI calls, which is not shown here may
1406 result in dramatic speedups !
1411 for (p5=2;p5<=n;p5++) @{
1415 for (p1=2;p1<=n-1;p1++) @{
1416 for (p3=1;p3<=p1-1;p3++) @{
1417 for (p5=p3+1;p5<=n;p5++) @{
1418 S2(i = p3,j = p5,k = p1) ;
1421 for (p5=p1+1;p5<=n;p5++) @{
1427 for (p3=1;p3<=n-1;p3++) @{
1428 for (p5=p3+1;p5<=n;p5++) @{
1429 S2(i = p3,j = p5,k = n) ;
1436 @c %/*************************************************************************
1437 @c % * A Full Example *
1438 @c % *************************************************************************/
1440 @chapter Using the CLooG Library
1441 The CLooG Library was implemented to allow the user to call CLooG
1442 directly from his programs, without file accesses or system calls. The
1443 user only needs to link his programs with C libraries. The CLooG
1444 library mainly provides one function (@code{cloog_clast_create_from_input})
1445 which takes as input the problem
1446 description with some options, and returns the data structure corresponding
1447 to the generated code (a @code{struct clast_stmt} structure)
1448 which is more or less an abstract syntax tree.
1449 The user can work with this data structure and/or use
1450 our pretty printing function to write the final code in either C or FORTRAN.
1451 Some other functions are provided for convenience reasons.
1452 These functions as well as the data structures are described in this section.
1455 * CLooG Data Structures::
1457 * Retrieving version information::
1458 * Example of Library Utilization::
1462 @node CLooG Data Structures
1463 @section CLooG Data Structures Description
1464 In this section, we describe the data structures used by the loop
1465 generator to represent and to process a code generation problem.
1472 * CloogUnionDomain::
1480 @subsection CloogState
1483 CloogState *cloog_state_malloc(void);
1484 void cloog_state_free(CloogState *state);
1488 @noindent The @code{CloogState} structure is (implicitly) needed to perform
1489 any CLooG operation. It should be created using @code{cloog_state_malloc}
1490 before any other CLooG objects are created and destroyed using
1491 @code{cloog_state_free} after all objects have been freed.
1492 It is allowed to use more than one @code{CloogState} structure at
1493 the same time, but an object created within the state of a one
1494 @code{CloogState} structure is not allowed to interact with an object
1495 created within the state of an other @code{CloogState} structure.
1499 @subsection CloogMatrix
1501 @noindent The @code{CloogMatrix} structure is equivalent to the PolyLib
1502 @code{Matrix} data structure (@pxref{Wil93}). This structure is devoted to
1503 represent a set of constraints.
1508 @{ unsigned NbRows ; /* Number of rows. */
1509 unsigned NbColumns ; /* Number of columns. */
1510 cloog_int_t **p; /* Array of pointers to the matrix rows. */
1511 cloog_int_t *p_Init; /* Matrix rows contiguously in memory. */
1513 typedef struct cloogmatrix CloogMatrix;
1515 CloogMatrix *cloog_matrix_alloc(unsigned NbRows, unsigned NbColumns);
1516 void cloog_matrix_print(FILE *foo, CloogMatrix *m);
1517 void cloog_matrix_free(CloogMatrix *matrix);
1521 @noindent The whole matrix is stored in memory row after row at the
1522 @code{p_Init} address. @code{p} is an array of pointers where
1523 @code{p[i]} points to the first element of the @math{i^{th}} row.
1524 @code{NbRows} and @code{NbColumns} are respectively the number of
1525 rows and columns of the matrix.
1526 Each row corresponds to a constraint. The first element of each row is an
1527 equality/inequality tag. The
1528 constraint is an equality @math{p(x) = 0} if the first element is 0, but it is
1529 an inequality @math{p(x) \geq 0} if the first element is 1.
1530 The next elements are the coefficients of the unknowns,
1531 followed by the coefficients of the parameters, and finally the constant term.
1532 For instance, the following three constraints:
1536 \hbox{$ \cases{ -i + m &$= 0$\cr
1538 j + i - k &$\geq 0$}$}
1552 @noindent would be represented by the following rows:
1556 # eq/in i j k m n cst
1563 @noindent To be able to provide different precision version (CLooG
1564 supports 32 bits, 64 bits and arbitrary precision through the GMP library),
1565 the @code{cloog_int_t} type depends on the configuration options (it may be
1566 @code{long int} for 32 bits version, @code{long long int} for 64 bits version,
1567 and @code{mpz_t} for multiple precision version).
1570 @subsection CloogDomain
1573 CloogDomain *cloog_domain_union_read(CloogState *state,
1574 FILE *input, int nb_parameters);
1575 CloogDomain *cloog_domain_from_cloog_matrix(CloogState *state,
1576 CloogMatrix *matrix, int nb_par);
1577 void cloog_domain_free(CloogDomain *domain);
1581 @noindent @code{CloogDomain} is an opaque type representing a polyhedral
1582 domain (a union of polyhedra).
1583 A @code{CloogDomain} can be read
1584 from a file using @code{cloog_domain_union_read} or
1585 converted from a @code{CloogMatrix}.
1586 The input format for @code{cloog_domain_union_read}
1587 is that of @ref{Domain Representation}.
1588 The function @code{cloog_domain_from_cloog_matrix} takes a @code{CloogState}, a
1589 @code{CloogMatrix} and @code{int} as input and returns a pointer to a
1590 @code{CloogDomain}. @code{matrix} describes the domain and @code{nb_par} is the
1591 number of parameters in this domain. The input data structures are neither
1593 The @code{CloogDomain} can be freed using @code{cloog_domain_free}.
1594 There are also some backend dependent functions for creating
1595 @code{CloogDomain}s.
1598 * CloogDomain/PolyLib::
1602 @node CloogDomain/PolyLib
1603 @subsubsection PolyLib
1606 #include <cloog/polylib/cloog.h>
1607 CloogDomain *cloog_domain_from_polylib_polyhedron(CloogState *state,
1608 Polyhedron *, int nb_par);
1611 The function @code{cloog_domain_from_polylib_polyhedron} takes a PolyLib
1612 @code{Polyhedron} as input and returns a pointer to a @code{CloogDomain}.
1613 The @code{nb_par} parameter indicates the number of parameters
1614 in the domain. The input data structure if neither modified nor freed.
1616 @node CloogDomain/isl
1620 #include <cloog/isl/cloog.h>
1621 CloogDomain *cloog_domain_from_isl_set(struct isl_set *set);
1622 __isl_give isl_set *isl_set_from_cloog_domain(CloogDomain *domain);
1625 The function @code{cloog_domain_from_isl_set} takes a
1626 @code{struct isl_set} as input and returns a pointer to a @code{CloogDomain}.
1627 The function consumes a reference to the given @code{struct isl_set}.
1628 Similarly, @code{isl_set_from_cloog_domain} consumes a reference
1629 to a @code{CloogDomain} and returns an @code{isl_set}.
1632 @node CloogScattering
1633 @subsection CloogScattering
1636 CloogScattering *cloog_domain_read_scattering(CloogDomain *domain,
1638 CloogScattering *cloog_scattering_from_cloog_matrix(CloogState *state,
1639 CloogMatrix *matrix, int nb_scat, int nb_par);
1640 void cloog_scattering_free(CloogScattering *);
1645 The @code{CloogScattering} type represents a scattering function.
1646 A @code{CloogScattering} for a given @code{CloogDomain} can be read
1647 from a file using @code{cloog_scattering_read} or converted
1648 from a @code{CloogMatrix} using @code{cloog_scattering_from_cloog_matrix}.
1649 The function @code{cloog_scattering_from_cloog_matrix} takes a
1650 @code{CloogState}, a @code{CloogMatrix} and two @code{int}s as input and
1652 pointer to a @code{CloogScattering}.
1653 @code{matrix} describes the scattering, while @code{nb_scat} and
1654 @code{nb_par} are the number of scattering dimensions and
1655 the number of parameters, respectively. The input data structures are
1656 neither modified nor freed.
1657 A @code{CloogScattering} can be freed using @code{cloog_scattering_free}.
1658 There are also some backend dependent functions for creating
1659 @code{CloogScattering}s.
1662 * CloogScattering/PolyLib::
1663 * CloogScattering/isl::
1666 @node CloogScattering/PolyLib
1667 @subsubsection PolyLib
1670 #include <cloog/polylib/cloog.h>
1671 CloogScattering *cloog_scattering_from_polylib_polyhedron(
1672 CloogState *state, Polyhedron *polyhedron, int nb_par);
1675 The function @code{cloog_scattering_from_polylib_polyhedron} takes a PolyLib
1676 @code{Polyhedron} as input and returns a pointer to a @code{CloogScattering}.
1677 The @code{nb_par} parameter indicates the number of parameters
1678 in the domain. The input data structure if neither modified nor freed.
1680 @node CloogScattering/isl
1684 #include <cloog/isl/cloog.h>
1685 CloogScattering *cloog_scattering_from_isl_map(struct isl_map *map);
1688 The function @code{cloog_scattering_from_isl_map} takes a
1689 @code{struct isl_map} as input and returns a pointer to a @code{CloogScattering}.
1690 The output dimensions of the @code{struct isl_map} correspond to the
1691 scattering dimensions, while the input dimensions correspond to the
1693 The function consumes a reference to the given @code{struct isl_map}.
1696 @node CloogUnionDomain
1697 @subsection CloogUnionDomain
1700 enum cloog_dim_type @{ CLOOG_PARAM, CLOOG_ITER, CLOOG_SCAT @};
1702 CloogUnionDomain *cloog_union_domain_alloc(int nb_par);
1703 CloogUnionDomain *cloog_union_domain_add_domain(CloogUnionDomain *ud,
1704 const char *name, CloogDomain *domain,
1705 CloogScattering *scattering, void *usr);
1706 CloogUnionDomain *cloog_union_domain_set_name(CloogUnionDomain *ud,
1707 enum cloog_dim_type type, int index, const char *name);
1708 void cloog_union_domain_free(CloogUnionDomain *ud);
1712 @noindent A @code{CloogUnionDomain} structure represents a union
1713 of scattered named domains. A @code{CloogUnionDomain} is
1714 initialized by a call to @code{cloog_union_domain_alloc},
1715 after which domains can be added using @code{cloog_union_domain_add_domain}.
1717 @code{cloog_union_domain_alloc} takes the number of parameters as input.
1718 @code{cloog_union_domain_add_domain} takes a previously created
1719 @code{CloogUnionDomain} as input along with an optional name,
1720 a domain, an optional scattering function and a user pointer.
1721 The name may be @code{NULL} and is duplicated if it is not.
1722 If no name is specified, then the statements will be named according
1723 to the order in which they were added.
1724 @code{domain} and @code{scattering} are taken over
1725 by the @code{CloogUnionDomain}. @code{scattering} may be @code{NULL},
1726 but it must be consistently @code{NULL} or not over all calls
1727 to @code{cloog_union_domain_add_domain}.
1728 @code{cloog_union_domain_set_name} can be used to set the names
1729 of parameters, iterators and scattering dimensions.
1730 The names of iterators and scattering dimensions can only be set
1731 after all domains have been added.
1733 There is also a backend dependent function for creating
1734 @code{CloogUnionDomain}s.
1737 * CloogUnionDomain/isl::
1740 @node CloogUnionDomain/isl
1744 #include <cloog/isl/cloog.h>
1745 CloogUnionDomain *cloog_union_domain_from_isl_union_map(
1746 __isl_take isl_union_map *umap);
1747 CloogUnionDomain *cloog_union_domain_from_isl_union_set(
1748 __isl_take isl_union_set *uset);
1751 The function @code{cloog_union_domain_from_isl_union_map} takes a
1752 @code{isl_union_map} as input and returns a pointer
1753 to a @code{CloogUnionDomain}.
1754 The input is a mapping from different
1755 spaces (different tuple names and possibly different dimensions)
1756 to a common space. The iteration domains are set to the domains
1757 in each space. The statement names are set to the names of the
1758 spaces. The parameter names of the result are set to those of
1759 the input, but the iterator and scattering dimension names are
1761 The function consumes a reference to the given @code{isl_union_map}.
1762 The function @code{cloog_union_domain_from_isl_union_set} is similar,
1763 but takes unscattered domains as input.
1766 @node CloogStatement
1767 @subsection CloogStatement
1770 struct cloogstatement
1771 @{ int number ; /* The statement unique number. */
1772 char *name; /* Name of the statement. */
1773 void * usr ; /* Pointer for user's convenience. */
1774 struct cloogstatement * next ;/* Next element of the linked list. */
1776 typedef struct cloogstatement CloogStatement ;
1778 CloogStatement *cloog_statement_malloc(CloogState *state);
1779 void cloog_statement_print(FILE *, CloogStatement *);
1780 void cloog_statement_free(CloogStatement *);
1784 @noindent The @code{CloogStatement} structure represents a @code{NULL}
1786 list of statements. In CLooG, a statement is only defined by its unique
1787 number (@code{number}). The user can use the pointer @code{usr} for his
1788 own convenience to link his own statement representation to the
1789 corresponding @code{CloogStatement} structure. The whole management of the
1790 @code{usr} pointer is under the responsibility of the user, in particular,
1791 CLooG never tries to print, to allocate or to free a memory block pointed
1797 @subsection CloogOptions
1801 @{ int l ; /* -l option. */
1802 int f ; /* -f option. */
1803 int strides ; /* -strides option. */
1804 int sh ; /* -sh option. */
1805 int esp ; /* -esp option. */
1806 int fsp ; /* -fsp option. */
1807 int otl ; /* -otl option. */
1808 int block ; /* -block option. */
1809 int compilable ; /* -compilable option. */
1810 int language; /* LANGUAGE_C or LANGUAGE_FORTRAN */
1811 int save_domains; /* Save unsimplified copy of domain. */
1813 typedef struct cloogoptions CloogOptions ;
1815 CloogOptions *cloog_options_malloc(CloogState *state);
1816 void cloog_options_print(FILE *foo, CloogOptions *options);
1817 void cloog_options_free(CloogOptions *options);
1821 @noindent The @code{CloogOptions} structure contains all the possible options to
1822 rule CLooG's behaviour (@pxref{Calling CLooG}).
1823 As a reminder, the default values are:
1825 @item @math{l = -1} (optimize control until the innermost loops),
1826 @item @math{f = 1} (optimize control from the outermost loops),
1827 @item @math{strides = 0} (use only unit strides),
1828 @item @math{sh = 0} (do not compute simple convex hulls),
1829 @item @math{esp = 1} (spread complex equalities),
1830 @item @math{fsp = 1} (start to spread from the first iterators),
1831 @item @math{otl = 1} (simplify loops running only once).
1832 @item @math{block = 0} (do not make statement blocks when not necessary).
1833 @item @math{compilable = 0} (do not generate a compilable code).
1836 The @code{save_domains} option is only useful for users of the CLooG
1837 library. This option defaults to 0, but when it is set, the @code{domain}
1838 field of each @code{clast_user_stmt} will be set to the set of values for the
1839 scattering dimensions for which this instance of the user statement is executed.
1840 The @code{domain} field of each @code{clast_for} contains the set of values for
1841 the scattering dimensions for which an instance of a user statement is executed
1842 inside the @code{clast_for}. It is only available if the @code{clast_for}
1843 enumerates a scattering dimension.
1846 @subsection CloogInput
1849 CloogInput *cloog_input_read(FILE *file, CloogOptions *options);
1850 CloogInput *cloog_input_alloc(CloogDomain *context,
1851 CloogUnionDomain *ud);
1852 void cloog_input_free(CloogInput *input);
1854 void cloog_input_dump_cloog(FILE *, CloogInput *, CloogOptions *);
1858 @noindent A @code{CloogInput} structure represents the input to CLooG.
1859 It is essentially a @code{CloogUnionDomain} along with a context
1860 @code{CloogDomain}. A @code{CloogInput} can be created from
1861 a @code{CloogDomain} and a @code{CloogUnionDomains} using
1862 @code{cloog_input_alloc}, or it can be read from a CLooG input
1863 file using @code{cloog_input_read}. The latter also modifies
1864 the @code{language} field of the @code{CloogOptions} structure.
1865 The constructed @code{CloogInput} can be used as input
1866 to a @code{cloog_clast_create_from_input} call.
1868 A @code{CloogInput} data structure and a @code{CloogOptions} contain
1869 the same information as a .cloog file. This function dumps the .cloog
1870 description of the given data structures into a file.
1872 @node Dump CLooG Input File Function
1873 @subsection Dump CLooG Input File Function
1878 @section CLooG Output
1881 Given a description of the input,
1882 an AST corresponding to the @code{CloogInput} can be constructed
1883 using @code{cloog_clast_create_from_input} and destroyed using
1884 @code{free_clast_stmt}.
1886 struct clast_stmt *cloog_clast_create_from_input(CloogInput *input,
1887 CloogOptions *options);
1888 void free_clast_stmt(struct clast_stmt *s);
1891 @code{clast_stmt} represents a linked list of ``statements''.
1893 struct clast_stmt @{
1894 const struct clast_stmt_op *op;
1895 struct clast_stmt *next;
1899 The entries in the list are not of type @code{clast_stmt} itself,
1900 but of some larger type. The following statement types are defined
1904 struct clast_root @{
1905 struct clast_stmt stmt;
1908 struct clast_root *new_clast_root(CloogNames *names);
1910 struct clast_assignment @{
1911 struct clast_stmt stmt;
1913 struct clast_expr * RHS;
1915 struct clast_assignment *new_clast_assignment(const char *lhs,
1916 struct clast_expr *rhs);
1918 struct clast_block @{
1919 struct clast_stmt stmt;
1920 struct clast_stmt * body;
1922 struct clast_block *new_clast_block(void);
1924 struct clast_user_stmt @{
1925 struct clast_stmt stmt;
1926 CloogDomain * domain;
1927 CloogStatement * statement;
1928 struct clast_stmt * substitutions;
1930 struct clast_user_stmt *new_clast_user_stmt(CloogDomain *domain,
1931 CloogStatement *stmt, struct clast_stmt *subs);
1934 struct clast_stmt stmt;
1935 CloogDomain * domain;
1936 const char * iterator;
1937 struct clast_expr * LB;
1938 struct clast_expr * UB;
1940 struct clast_stmt * body;
1942 struct clast_for *new_clast_for(CloogDomain *domain, const char *it,
1943 struct clast_expr *LB, struct clast_expr *UB,
1944 cloog_int_t stride);
1946 struct clast_guard @{
1947 struct clast_stmt stmt;
1948 struct clast_stmt * then;
1950 struct clast_equation eq[1];
1952 struct clast_guard *new_clast_guard(int n);
1955 The @code{clast_stmt} returned by @code{cloog_clast_create}
1956 is a @code{clast_root}.
1957 It contains a placeholder for all the variable names that appear
1958 in the AST and a (list of) nested statement(s).
1961 A @code{clast_assignment} assigns the value given by
1962 the @code{clast_expr} @code{RHS} to a variable named @code{LHS}.
1965 A @code{clast_block} groups a list of statements into one statement.
1966 These statements are only generated if the @code{block} option is set,
1967 @pxref{Statement Block} and @ref{CloogOptions}.
1970 A @code{clast_user_stmt} represents a call to a statement specified
1971 by the user, @pxref{CloogStatement}.
1972 @code{substitutions} is a list of @code{clast_assignment} statements
1973 assigning an expression in terms of the scattering dimensions to
1974 each of the original iterators in the original order.
1975 The @code{LHS}s of these assignments are left blank (@code{NULL}).
1976 The @code{domain} is set to @code{NULL} if the @code{save_domains} option
1977 is not set. Otherwise, it is set to the set
1978 of values for the scattering dimensions
1979 for which this instance of the user statement is executed.
1980 Note that unless the @code{noscalars} option has been set, the
1981 constant scattering dimensions may have been removed from this set.
1984 A @code{clast_for} represents a for loop, iterating @code{body} for each
1985 value of @code{iterator} between @code{LB} and @code{UB} in steps
1986 of size @code{stride}.
1987 The @code{domain} is set to @code{NULL} if the @code{save_domains} option is not
1988 set. Otherwise, it is set to the set of values for the scattering dimensions
1989 for which a user statement is executed inside this @code{clast_for}. Note that
1990 unless the @code{noscalars} option has been set, the constant scattering
1991 dimensions may have been removed from this set.
1994 A @code{clast_guard} represents the guarded execution of the @code{then}
1995 (list of) statement(s) by a conjunction of @code{n} (in)equalities.
1996 Each (in)equality is represented by a @code{clast_equation}.
1998 struct clast_equation @{
1999 struct clast_expr * LHS;
2000 struct clast_expr * RHS;
2005 The condition expressed by a @code{clast_equation} is
2006 @code{LHS <= RHS}, @code{LHS == RHS} or @code{LHS >= RHS}
2007 depending on whether @code{sign} is less than zero, equal
2008 to zero, or greater than zero.
2010 The dynamic type of a @code{clast_stmt} can be determined
2011 using the macro @code{CLAST_STMT_IS_A(stmt,type)},
2012 where @code{stmt} is a pointer to a @code{clast_stmt}
2013 and @code{type} is one of @code{stmt_root}, @code{stmt_ass},
2014 @code{stmt_user}, @code{stmt_block}, @code{stmt_for} or
2016 Users are allowed to define their own statement types by
2017 assigning the @code{op} field of the statements a pointer
2018 to a @code{clast_stmt_op} structure.
2020 struct clast_stmt_op @{
2021 void (*free)(struct clast_stmt *);
2025 The @code{free} field of this structure should point
2026 to a function that frees the user defined statement.
2029 A @code{clast_expr} can be an identifier, a term,
2030 a binary expression or a reduction.
2032 enum clast_expr_type @{
2038 struct clast_expr @{
2039 enum clast_expr_type type;
2041 void free_clast_expr(struct clast_expr *e);
2045 Identifiers are of subtype @code{clast_name}.
2047 struct clast_name @{
2048 struct clast_expr expr;
2051 struct clast_name *new_clast_name(const char *name);
2052 void free_clast_name(struct clast_name *t);
2055 The character string pointed to by @code{name} is
2056 assumed to be part of the @code{CloogNames} structure
2057 in the root of the clast as is therefore not copied.
2060 Terms are of type @code{clast_term}.
2062 struct clast_term @{
2063 struct clast_expr expr;
2065 struct clast_expr *var;
2067 struct clast_term *new_clast_term(cloog_int_t c, struct clast_expr *v);
2068 void free_clast_term(struct clast_term *t);
2071 If @code{var} is set to @code{NULL}, then the term represents
2072 the integer value @code{val}. Otherwise, it represents
2073 the term @code{val * var}.
2074 @code{new_clast_term} simply copies the @code{v} pointer
2075 without copying the underlying @code{clast_expr}.
2076 @code{free_clast_term}, on the other hand, recursively frees
2080 Binary expressions are of type @code{clast_bin_type} and
2081 represent either the floor of a division (fdiv),
2082 the ceil of a division (cdiv), an exact division or
2083 the remainder of an fdiv.
2085 enum clast_bin_type @{ clast_bin_fdiv, clast_bin_cdiv,
2086 clast_bin_div, clast_bin_mod @};
2087 struct clast_binary @{
2088 struct clast_expr expr;
2089 enum clast_bin_type type;
2090 struct clast_expr* LHS;
2093 struct clast_binary *new_clast_binary(enum clast_bin_type t,
2094 struct clast_expr *lhs, cloog_int_t rhs);
2095 void free_clast_binary(struct clast_binary *b);
2099 Reductions are of type @code{clast_reduction} and
2100 can represent either the sum, the minimum or the maximum
2103 enum clast_red_type @{ clast_red_sum, clast_red_min, clast_red_max @};
2104 struct clast_reduction @{
2105 struct clast_expr expr;
2106 enum clast_red_type type;
2108 struct clast_expr* elts[1];
2110 struct clast_reduction *new_clast_reduction(enum clast_red_type t,
2112 void free_clast_reduction(struct clast_reduction *r);
2115 @node Retrieving version information
2116 @section Retrieving version information
2117 CLooG provides static and dynamic version checks to assist on
2118 including a compatible version of the library.
2119 A static version check at compile time can be achieved by
2120 querying the version constants defined in @code{version.h}:
2123 @item @code{CLOOG_VERSION_MAJOR}
2124 @item @code{CLOOG_VERSION_MINOR}
2125 @item @code{CLOOG_VERSION_REVISION}
2128 This way it is possible to ensure the included headers are of the
2129 correct version. It is still possible that the installed CLooG
2130 library version differs from the installed headers.
2131 In order to avoid this, a dynamic version check is provided with
2136 int cloog_version_major(void);
2137 int cloog_version_minor(void);
2138 int cloog_version_revision(void);
2142 By using both the static and the dynamic version check, it is possible
2143 to match CLooG's header version with the library's version.
2145 @node Example of Library Utilization
2146 @section Example of Library Utilization
2147 Here is a basic example showing how it is possible to use the CLooG library,
2148 assuming that a standard installation has been done.
2149 The following C program reads a CLooG input file on the standard input,
2150 then prints the solution on the standard output.
2151 Options are preselected to the default values of the CLooG software.
2152 This example is provided in the @code{example} directory of the
2157 # include <cloog/cloog.h>
2163 CloogOptions * options ;
2164 struct clast_stmt *root;
2166 /* Setting options and reading program informations. */
2167 state = cloog_state_malloc();
2168 options = cloog_options_malloc(state);
2169 input = cloog_input_read(stdin, options);
2171 /* Generating and printing the code. */
2172 root = cloog_clast_create_from_input(input, options);
2173 clast_pprint(stdout, root, 0, options);
2175 cloog_clast_free(root);
2176 cloog_options_free(options) ;
2177 cloog_state_free(state);
2182 @noindent The compilation command could be:
2184 gcc example.c -lcloog -o example
2186 @noindent A calling command with the input file test.cloog could be:
2188 more test.cloog | ./example
2192 @c % ******************************** HACKING *********************************
2194 @c @chapter Hacking CLooG
2197 @c * Program organization::
2198 @c * Special Options::
2199 @c * CLooG Coding Standards::
2202 @c @node Program organization
2203 @c @section Program organization
2205 @c @node Special Options
2206 @c @section Special Options
2208 @c @node CLooG Coding Standards
2209 @c @section CLooG Coding Standards
2212 @c % ****************************** INSTALLING ********************************
2214 @chapter Installing CLooG
2219 * Basic Installation::
2220 * Optional Features::
2226 First of all, it would be very kind to refer the following paper in any
2227 publication that result from the use of the CLooG software or its library,
2228 @pxref{Bas04} (a bibtex entry is provided behind the title page of this
2229 manual, along with copyright notice, and in the CLooG home
2230 @code{http://www.CLooG.org}.
2232 This library is free software; you can redistribute it and/or
2233 modify it under the terms of the GNU Lesser General Public
2234 License as published by the Free Software Foundation; either
2235 version 2.1 of the License, or (at your option) any later version.
2236 This library is distributed in the hope that it will be useful,
2237 but WITHOUT ANY WARRANTY; without even the implied warranty of
2238 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
2239 Lesser General Public License for more details.
2240 @code{http://www.gnu.org/licenses/lgpl-2.1.html}
2242 Note, though, that if you link CLooG against a GPL library such
2243 as the PolyLib backend, then the combination becomes GPL too.
2244 In particular, a CLooG library based on the PolyLib backend
2245 is GPL version 2 only.
2246 Since the isl backend is LGPL, linking against it does not affect
2247 the license of CLooG.
2251 @section Requirements
2253 CLooG can be used with one of two possible backends,
2254 one using isl and one using PolyLib.
2255 The isl library is included in the CLooG distribution,
2256 while the PolyLib library needs to be obtained separately.
2257 On the other hand, isl requires GMP, while PolyLib can be
2258 compiled with or without the use of GMP.
2259 The user therefore needs to install at least one of
2269 @subsection PolyLib (optional)
2270 To successfully install CLooG with the PolyLib backend,
2271 the user first needs to install PolyLib
2272 version 5.22.1 or above (default 64 bits version is satisfying
2273 as well as 32 bits or GMP multiple precision version).
2274 Polylib can be downloaded freely
2275 at @code{http://icps.u-strasbg.fr/PolyLib/} or
2276 @code{http://www.irisa.fr/polylib/}. Once downloaded and unpacked
2277 (e.g. using the @samp{tar -zxvf polylib-5.22.3.tar.gz} command),
2278 the user can compile
2279 it by typing the following commands on the PolyLib's root directory:
2282 @item @code{./configure}
2284 @item And as root: @code{make install}
2287 Alternatively, the latest development version can be obtained from the
2290 @item @code{git clone git://repo.or.cz/polylib.git}
2291 @item @code{cd polylib}
2292 @item @code{./autogen.sh}
2293 @item @code{./configure}
2295 @item And as root: @code{make install}
2298 The PolyLib default installation is @code{/usr/local}. This directory may
2299 not be inside your library path. To fix the problem, the user should set
2301 export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib
2303 @noindent if your shell is, e.g., bash or
2305 setenv LD_LIBRARY_PATH $LD_LIBRARY_PATH:/usr/local/lib
2307 @noindent if your shell is, e.g., tcsh. Add the line to your .bashrc or .tcshrc (or
2308 whatever convenient file) to make this change permanent. Another solution
2309 is to ask PolyLib to install in the standard path by using the prefix
2310 option of the configure script:
2311 @samp{./configure --prefix=/usr}.
2313 CLooG makes intensive calls to polyhedral operations, and PolyLib
2314 functions do the job. Polylib is a free library written in C for the
2315 manipulation of polyhedra. The library is operating on objects like
2316 vectors, matrices, lattices, polyhedra, Z-polyhedra, unions of
2317 polyhedra and a lot of other intermediary structures. It provides
2318 functions for all the important operations on these structures.
2321 @subsection GMP Library (optional)
2323 To be able to deal with insanely large coefficient, the user will need to
2324 install the GNU Multiple Precision Library (GMP for short) version 4.1.4
2325 or above. It can be freely downloaded from @code{http://www.swox.com/gmp}.
2326 Note that the isl backend currently requires GMP.
2327 The user can compile GMP by typing the following commands on the GMP root
2331 @item @code{./configure}
2333 @item And as root: @code{make install}
2336 The GMP default installation is @code{/usr/local}, the same method to
2337 fix a library path problem applies as with PolyLib (@pxref{PolyLib}).
2339 If you want to use the PolyLib backend, then
2340 PolyLib has to be built using the GMP library by specifying the option
2341 @samp{--with-libgmp=PATH_TO_GMP} to the PolyLib configure script
2342 (where @code{PATH_TO_GMP} is @code{/usr/local} if you did not change the GMP
2343 installation directory). Then you have to set the convenient CLooG configure
2344 script options to build the GMP version (@pxref{Optional Features}).
2347 @node Basic Installation
2348 @section CLooG Basic Installation
2350 Once downloaded and unpacked
2351 (e.g. using the @samp{tar -zxvf cloog-@value{VERSION}.tar.gz} command),
2352 you can compile CLooG by typing the following commands on the CLooG's root
2356 @item @code{./configure}
2358 @item And as root: @code{make install}
2361 Alternatively, the latest development version can be obtained from the
2364 @item @code{git clone git://repo.or.cz/cloog.git}
2365 @item @code{cd cloog}
2366 @item @code{./get_submodules.sh}
2367 @item @code{./autogen.sh}
2368 @item @code{./configure}
2370 @item And as root: @code{make install}
2373 Depending on which backend you want to use and where they
2374 are located, you may need to pass some
2375 options to the configure script, @pxref{Optional Features}.
2377 The program binaries and object files can be removed from the
2378 source code directory by typing @code{make clean}. To also remove the
2379 files that the @code{configure} script created (so you can compile the
2380 package for a different kind of computer) type @code{make distclean}.
2382 Both the CLooG software and library have been successfully compiled
2383 on the following systems:
2385 @item PC's under Linux, with the @code{gcc} compiler,
2386 @item PC's under Windows (Cygwin), with the @code{gcc} compiler,
2387 @item Sparc and UltraSparc Stations, with the @code{gcc} compiler.
2390 @node Optional Features
2391 @section Optional Features
2392 The @code{configure} shell script attempts to guess correct values for
2393 various system-dependent variables and user options used during compilation.
2394 It uses those values to create the @code{Makefile}. Various user options
2395 are provided by the CLooG's configure script. They are summarized in the
2396 following list and may be printed by typing @code{./configure --help} in the
2397 CLooG top-level directory.
2400 @item By default, the installation directory is @code{/usr/local}:
2401 @code{make install} will install the package's files in
2402 @code{/usr/local/bin}, @code{/usr/local/lib} and @code{/usr/local/include}.
2403 The user can specify an installation prefix other than @code{/usr/local} by
2404 giving @code{configure} the option @code{--prefix=PATH}.
2406 @item By default, the isl backend will use the version of isl
2407 that is @code{bundled} together with CLooG.
2408 Using the @code{--with-isl} option of @code{configure}
2409 the user can specify that @code{no} isl,
2410 a previously installed (@code{system}) isl or a @code{build} isl
2412 In the latter case, the user should also specify the build location
2413 using @code{--with-isl-builddir=PATH}.
2414 In case of an installed isl,
2415 the installation location can be specified using the
2416 @code{--with-isl-prefix=PATH} and
2417 @code{--with-isl-exec-prefix=PATH} options of @code{configure}.
2419 @item By default, the PolyLib backend will use an installed
2420 (@code{system}) PolyLib, if any.
2421 The installation location can be specified using the
2422 @code{--with-polylib-prefix=PATH} and
2423 @code{--with-polylib-exec-prefix=PATH} options of @code{configure}.
2424 Using the @code{--with-polylib} option of @code{configure}
2425 the user can specify that @code{no} PolyLib or a @code{build} PolyLib
2427 In the latter case, the user should also specify the build location
2428 using @code{--with-polylib-builddir=PATH}.
2430 @item By default, the PolyLib backend of CLooG is built
2431 in 64bits version if such version of the
2432 PolyLib is found by @code{configure}. If the only existing version of the
2433 PolyLib is the 32bits or if the user give to @code{configure} the option
2434 @code{--with-bits=32}, the 32bits version of CLooG will be compiled. In the
2435 same way, the option @code{--with-bits=gmp} have to be used to build
2436 the multiple precision version.
2438 @item By default, @code{configure} will look for the GMP library
2439 (necessary to build the multiple precision version) in standard
2440 locations. If necessary, the user can specify the GMP path by giving
2441 @code{configure} the option @code{--with-gmp-prefix=PATH} and/or
2442 @code{--with-gmp-exec-prefix=PATH}.
2445 @node Uninstallation
2446 @section Uninstallation
2447 The user can easily remove the CLooG software and library from his system
2448 by typing (as root if necessary) from the CLooG top-level directory
2449 @code{make uninstall}.
2451 @c % **************************** DOCUMENTATION ******************************
2453 @chapter Documentation
2454 The CLooG distribution provides several documentation sources. First, the
2455 source code itself is as documented as possible. The code comments use a
2456 Doxygen-compatible presentation (something similar to what JavaDoc does for
2457 JAVA). The user may install Doxygen
2458 (see @code{http://www.stack.nl/~dimitri/doxygen}) to automatically
2459 generate a technical documentation by typing @code{make doc} or
2460 @code{doxygen ./autoconf/Doxyfile} at the CLooG top-level directory after
2461 running the configure script (@pxref{Installing}). Doxygen will generate
2462 documentation sources (in HTML, LaTeX and man) in the @code{doc/source}
2463 directory of the CLooG distribution.
2465 The Texinfo sources of the present document are also provided in the @code{doc}
2466 directory. You can build it in either DVI format (by typing
2467 @code{texi2dvi cloog.texi}) or PDF format
2468 (by typing @code{texi2pdf cloog.texi}) or HTML format
2469 (by typing @code{makeinfo --html cloog.texi}, using @code{--no-split}
2470 option to generate a single HTML file) or info format
2471 (by typing @code{makeinfo cloog.texi}).
2473 @c % ****************************** REFERENCES ********************************
2479 @anchor{Bas03a}[Bas03a] C. Bastoul, P. Feautrier. Improving data locality
2480 by chunking. CC'12 International Conference on Compiler Construction,
2481 LNCS 2622, pages 320-335, Warsaw, april 2003.
2484 @anchor{Bas03b}[Bas03b] C. Bastoul. Efficient code generation for automatic
2485 parallelization and optimization. ISPDC'03 IEEE International Symposium on
2486 Parallel and Distributed Computing, pages 23-30, Ljubljana, october 2003.
2489 @anchor{Bas04}[Bas04] C. Bastoul. Code Generation in the Polyhedral Model
2490 Is Easier Than You Think. PACT'13 IEEE International Conference on Parallel
2491 Architecture and Compilation Techniques, pages 7-16, Juan-les-Pins,
2495 @anchor{Fea92}[Fea92] P. Feautrier Some efficient solutions to the affine
2496 scheduling problem, part II: multidimensional time.
2497 International Journal of Parallel Programming, 21(6):389--420, December 1992.
2500 @anchor{Gri04}[Gri04] M. Griebl. Automatic parallelization of loop programs
2501 for distributed memory architectures. Habilitation Thesis. Facult@"at f@"ur
2502 Mathematik und Informatik, Universit@"at Passau, 2004.
2503 @emph{http://www.infosun.fmi.uni-passau.de/cl/loopo/}
2506 @anchor{Qui00}[Qui00] F. Quiller@'e, S. Rajopadhye, and D. Wilde.
2507 Generation of efficient nested loops from polyhedra.
2508 International Journal of Parallel Programming, 28(5):469-498,
2512 @anchor{Wil93}[Wil93] Doran K. Wilde.
2513 A library for doing polyhedral operations.
2514 Technical Report 785, IRISA, Rennes, France, 1993.
2521 @c % /*************************************************************************
2522 @c % * PART VI: END OF THE DOCUMENT *
2523 @c % *************************************************************************/
2524 @c @unnumbered Index