%%
%% $Id$
%%
%% Copyright 1989-2016 MINES ParisTech
%%
%% This file is part of PIPS.
%%
%% PIPS is free software: you can redistribute it and/or modify it
%% under the terms of the GNU General Public License as published by
%% the Free Software Foundation, either version 3 of the License, or
%% any later version.
%%
%% PIPS is distributed in the hope that it will be useful, but WITHOUT ANY
%% WARRANTY; without even the implied warranty of MERCHANTABILITY or
%% FITNESS FOR A PARTICULAR PURPOSE.
%%
%% See the GNU General Public License for more details.
%%
%% You should have received a copy of the GNU General Public License
%% along with PIPS. If not, see .
%%
\documentstyle[12pt]{article}
%\input{/usr/share/local/lib/tex/macroslocales/Dimensions.tex}
\newcommand{\titre}{PIPS BUGS and WISHES \\
Version 1.1 \\
Draft (November 21, 1991)}
\newcommand{\auteur}{Franc,ois Irigoin\\
et al.}
\newcommand{\docdate}{10 January 1992}
\newcommand{\numero}{Exxx}
\begin{document}
%%\input{/usr/share/local/lib/tex/macroslocales/PageTitre.tex}
\tableofcontents
\newpage
\section*{Introduction}
Should be augmented with meeting results: solved problems and new
problems and new priorities.
\paragraph{Priorities as defined in May 1991:}
\begin{itemize}
\item programming environment \marginpar{(PJ/FI/BB)}
\item parallelization correctness \marginpar{(PJ)}
\item parser robustness \marginpar{(FI)}
\item pips process size: pipsdbm and pipsmake
\item evaluation and quantification and benchmarking \marginpar{(YY)}
\end{itemize}
\paragraph{Goals for 1992:}
\begin{itemize}
\item PIPS-CRAY
\item Yi-Qing's PhD;
\item Lei's PhD;
\item CEA's goals;
\item PIPS evaluation: interprocedural analysis, interprocedural
parallelization;
\end{itemize}
Non-regression: please fill in Validate and Bugs in a systematic way!
\section{Atomizer (Alexis Platonoff)}
\section{Bootstrap}
\begin{enumerate}
\item intrinsic \verb+LOG+ is missing; 27 February 1992; to be added
in effects library too; fixed by FI, 10 March 1992;
\item same thing for \verb+LOG10+; fixed by FI, 10 March 1992;
\item intrinsics with a varying number of arguments are encoded
as having 0 arguments; no good solution has yet been found;
FI, 18 March 1991 (this would also solve the next problem);
\item \verb+STOP+ was declared as always having one parameter;
this was changed to {\em zero} parameters; modified by
FI, 19 March 1992
\end{enumerate}
\subsection{Wishes}
\begin{itemize}
\item information about intrinsics, scattered in syntax, ri-util,
bootstrap and effects should be centralized, probably in
ri-util; FI, 27 March 1992;
\end{itemize}
\section{Call Graph (Lei Zhou)}
\begin{enumerate}
\item make a difference between the Call Graph based on symbol tables
and the elided interprocedural Control Flow Graph, which is
currently called {\em call graph}
\item present\marginpar{(LZ/FI)} {\em call graph} is broken by
unstructured; no relation with the source program; use {\em
parsed\_code} instead? Faster... but not as fast as a real call
graph based on declarations;
\item property \verb+ALL_DO+ to see nesting levels regardless of procedure
calls; i.e. display another kind of squeletton for the ICFG:
all loop control structure, DO-loops and while, every control
structures, including IFs;
\item unstructured are not displayed and so WHILE loops are not
visible (i.e. cycles in the control graph);
\item function calls are not taken into account; FI, 29 July 1992;
\end{enumerate}
\section{Chains}
Pierre Jouvelot.
\subsection{Bugs}
\begin{enumerate}
\item local variables should be handled as global ones; else,
statement grouped by a private declaration may be distributed;
this doesn't explain the erronneous behavior observed; problems with
B should also occur;
\item \verb+KEEP_READ_READ_DEPENDENCE+: \verb+pips_error in dept_type+;
\item test added so as never not taking into account proper effects or
regions of blocks; this is due to semantics difference between
proper effects and regions which look a bit more like cumulated
effects; FI, 5 August 1992;
\end{enumerate}
\subsection{Modifications}
\begin{enumerate}
\item two new routines added to compute base chain computation on
proper effects (as before) or on regions; FI, 5 August 1992;
\end{enumerate}
\section{Code-Generate}
This library was merged into library Transformations by Bruno Baron.
\subsection{Wishes}
\begin{enumerate}
\item Perform priv is required to detect that the inner loop is
parallel in the following case:
\begin{verbatim}
program dist
real a(100,100)
do 10 i=1,100
m = j
do 20 j=1,m
a(i,j) = m
20 continue
10 continue
end
\end{verbatim}
The fast privatize routine is unable to detect that J is local. But
this may not be correct. Depends of the semantics of DOALL.
\end{enumerate}
\section{Complexity}
The complexity library is managed by Lei.
\subsection{Bugs}
\begin{enumerate}
\item Do not substract negative execution time for loops with
non-satisfied conditions
\item The callee's complexity result which contains the COMMON
variables as its component CANNOT be added to the caller's
complexity which contains the same COMMON variables.
LZ thinks it is beacuse of the restriction of module name.
like: FOO:IL and TOTO:IL , they can't be added. 24/02/92
\item pips core dumps when writing ZPROAX.COMPLEXITY after an
interprocedural analysis;
see \verb+zproax.f+ in \verb+Bugs/Complexity+;
FI, 9 August 1992;
\item Fail to pass FFT algorithm:
The sample version is kept in Bugs/Complexity/fstr2.f .
There are two problems: 1) N is unknown in Loop 12, that is to
say, the precondition is not available.
2) MMAX is unknown, it is no surprise beacuse it is updated
on each iteration. Then MMAX has the value ``UR'' as final
complexity results. Loop 13 and Loop 12 both have MMAX as
loop bound and ``UR'' is kept out of C3/polynomial library,
so two polynomials can not be added, let alone multiplication.
What we can get is ``Bad internal polynomial representation''.
LZ, Oct. 8, 92
\item For tmines.f program in ~pips/Pips/Expe*/Complexity,
when {\bf setenv COMPLEXITY\_DEBUG\_LEVEL 9} is set,
{\bf Display -m calcg comp} core dumped.
If the debug level is not set, it works.
LZ, 08/06/04
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item The loop bound cannot be the value of an array.
When the inner loop's bound are independent, we have no way to
evaluate the outer loop's body complexity.
That is, we cannot use a polynomial form to represent the inner
loop's complexity.
Though it's not a normal case. FI: Example?
\item How should target machines be modeled? Define a NewGen structure
and an external format? See \verb+model.rc+ in WP65.
\item Where should model files be stored? How
should they be accessed? Everything should be in Documentation
in a LATEX form...
\item Evaluate parallel complexity for different kind of processors
(unbounded PRAM machine, multiprocessor, vector processor,...)
\item Evaluate space and communication complexities
\item Evaluate dynamic complexity: how much does each statement contribute
to the total complexity? Proper complexity times number of executions.
\item Propagate complexities downwards so as to know how many times a
statement is executed
\item set up cost tables to compute the number of (visible) floating point
operations and/or memory requests (loads and stores); intrinsics
like \verb+SIN+are not taken into account;
\item define a mechanism to select a particular cost table;
\item build and test a cost table for Sparcstation 1;
\end{enumerate}
\section{Control}
Build the control graph and generate the \verb+CONTROLIZED_CODE+.
\begin{enumerate}
\item FORMATs which are unreachable are dead-code eliminated;
see \verb+Bug/Control/tformat.f+
\item unstructured code seems to unstructure nearby structured code;
see \verb+dyna/vt.f+; the bug hasn't been reproduced
on smaller cases; FI, 5 August 1992;
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item \marginpar{PJ} make unstructureds as small as possible;
see TMINES and its semantic analysis for instance or \verb+split.f+
and the complexity evaluation; BAS should be preserved; prologue
and postlogue should be moved out of the unstructured;
\item recognize structured constructs like
IF/THEN/ELSE? Should be performed by an external tool:
\item STOPs in subroutines and functions are not handled at
all (core dumped); control effects?
\end{enumerate}
\section{Conversion}
Who's responsible?
\section{Dg-util (Utilities for the dependence graph)}
\section{Effects}
\subsection{Bugs}
\begin{enumerate}
\item improve \marginpar{(FI)} reshaping handling for numerical array
bounds: T(961) and U(31,31) in spec77 for instance
\item recompute
effects a second time, using information about constant
propagation \item document the origin of {\em bad array reshaping}
\item
add a property to cope silently with * dimension declaration; CEA
suggestion; is it possible to trust the user better?
\item bug with spec77/del4 (dim transformed in foodim):
\verb+pips_assertion_failed in reference_ith_index() +
\item use order on cumulated effects to
destroy useless read effects; should be also useful for proper effects
of composite instruction like DO; may have been done since
\verb=Bugs/Effects/effect.f= is correctly parallelized (FI, 17 January
1992)
\item add an array of logical
units (this entity should be created at parse time) and generate
write effects on this entity for IO statements
\item core dumps in PerfectClub/W.WS with function {\em gfidi}:\\
\verb+pips assertion failed in reference_ith_index+
\item demo/ifp-pips.f, module onde24 with {\tt PIPSDBM\_DEBUG\_LEVEL=5}:
\\
\verb+[db_put_resource] Checking Newgen Consistency for resource
PROPER_EFFECTS(ONDE24):+ \\
\verb+Newgen User Error -- gen_check: Type clash (expecting entity,
getting unit)+ \\
Could probably be illustrated with another code.
\item generate a \verb+MAY+ effect instead of a \verb+MUST+ effect when
an array element is referenced; this produce wrong \verb+array kill+
information for subroutine; see \verb+Bugs/Effects/effect2.f+
(FI, 6 May 1991);
\item WRITE has a \verb+read+ effects on its IO-list; see for instance
\verb+ker22.f+
\item \verb+TOP-LEVEL:PAUSE+ is not recognized by
\verb+proper_effects_of_call()+;
see \verb+Bugs/Syntax/pset.f+; FI, 7 February 1992;
the bug was in \verb+bootstrap+ where \verb+PAUSE+ was not
recognized as intrinsics; \verb+PAUSE+ was not recognized
in effects either;
\item control effects are ignored in \verb+PAUSE+ and \verb+STOP+;
FI, 7 February 1992
\item effects of IO lists are redundant; the same item is counted
n times, where n is its position in the io list;
this has a dramatic effect on use-def chain counts;
see \verb+Bugs/Effects/udread.f+; YY et FI, 23 June 1992;
fixed by FI, 23 June 1992;
\item typo for \verb+OPEN+ with \verb+ACCESS+ option; observed
with \verb+dyna.f+; fixed by FI in \verb+io.c+, 30 July 1992;
\item output buffers in \verb+store_text_line+ were filled in without
checking for overflow; they were allocated in the stack so
as to create havoc on overflow; observed and fixed - size
increase and overflow check - by FI,
31 July 1992;
\item print stars '*' for unknown array indices when prettyprinting
effects; FI, 2 August 1992;
\item This last modification seems to put the same array many times in
the same list; maybe, this was an old bug; when are the stars
added? When the effect is built? When the effect is printed?
This shows in cumulated effects: each proper effect is
transformed into a cumulated effect without reduction for
redundant cumulated effects; FI, 27 April 1993;
\item wrong reshaping on \verb+Y+ in \verb+zproax+; a macro was
called with a functional argument expansed twice: too bad
for the side effects; FI, 11 August 1992;
\item proper effects of DO's include a MAY-READ of the loop index
due to its incrementation; this does not let the MUST-WRITE
of the loop index kill use-def chains that should be killed
and legal privatizations are not possible and parallel loops
are kept serial; this is due to the compound nature of the PIPS DO
instruction;
It is difficult to define a semantics for the effects ordering
as the masking does not always occur (see test instructions for
instance).
It is hard to add a masking system to the effect computation.
It was not chosen to give a copy-in copy-out semantics to
private variables.
It was chosen to handle the DO as a special case and to
suppress the MAY-READ effect on the loop index.
See \verb+Validation/Effects/loop.f+. FI, RK, March 12, 1993.
\item the effect print buffer may be too small; size increased to
2048; overflow check added; FI, 26 May 1993;
\item \verb+effects_conflict_with_entity()+ modified to call
\verb+entity_conflict_p()+ instead of performing a pointer
comparison; \verb+effects_read_or_write_entity()+ added;
FI, 29 JUne 1993
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item bad reshaping wrongly found for \verb+A+ in \verb+zproax.f+; function
\verb+translate_array_effect+ does not handle arrays of dimension
greater than or equal to three because the direct test is not
linear; a new test should be developped to test that the first
k-1 dimensions have the same sized in the formal and actual arguments
and that formal dimension k is less then actual dimension k;
the formal dimension is k; the actual dimension must be greater
than or equal to k; equality between two expressions can be proved
by checking that their difference evaluates to 0 (\verb+sc_eval+);
FI, 11 August 1992;
\item effects should be alphabetically sorted by variable names to
improve readability; FI, 1 Dec. 1992;
\item \verb+MUST+ effects should mask \verb+MAY+ effects;
for instance, the same
statement should not have a \verb+MUST WRITE T+ and a
\verb+MAY WRITE T+ simultaneously; the \verb+MAY+ effect
should be deleted; let say that different kind of effects could
be computed;
\item control effects (stop/exit) should also be computed;
\item effects should take into account the declarations; see
\$PIPSDIR/Test/Bugs/Semantics/contrex3.f for an example. FI, BA,
9 february 1994.
\end{enumerate}
\section{Evaluation of PIPS}
\begin{enumerate}
\item see what happens in OASSEM (cr2cnf) after privatization;
\item parser core dumps on SPHOT (PerfectClub)
\item PerfectClub
\item Riceps
\item SIDOLO (YY)
\end{enumerate}
\section{Generation}
Who's responsible?
\section{Hyperplane}
Hyperplane method, meaning linear scheduling, i.e. unimodular transformations.
Include loop interchange.
\begin{enumerate}
\item \marginpar{YY} should include a test for dependences
\item \marginpar{BB} should include a machine dependent cost
function for interchange decision
\item should handle not perfectly nested loops
\end{enumerate}
\section{ICFG}
Interprocedural Control Flow Graph.
\section{Linear}
\verb+linear.h+ was suppressed. It was redundant with ri.h because
of NewGen requirements for external types. It was not properly
installed as not derived from any particular library. FI, December 91.
\subsection{Vecteur}
\begin{enumerate}
\item integer overflow are usually not detected; one test was
added in \verb+vect_cl2()+
\end{enumerate}
\subsection{Contrainte}
\begin{enumerate}
\item \verb+contrainte_normalize+ did not handle correctly constant terms
in inequalities; the new bound was off by 1, in the safe direction; fixed
by FI, October 91;
\item coefficients in pairwise elimination were not reduced by their
GCD; fixed by FI, December 91;
\end{enumerate}
\subsection{Syste`me de Contraintes}
\begin{enumerate}
\item \verb+sc_rm_empty_constraints()+ did not update the numbers of
equalities and inequalities (FI, Oct. 30, 1991);
\end{enumerate}
\begin{enumerate}
\item modify the grammar so as to allow comments linear system
description files (FI, 8 August 1992)
\end{enumerate}
\subsection{Polye`dres}
\subsubsection{Bugs}
\begin{enumerate}
\item the convex hull operator core dumps with (Alexis Platonoff,
6 August 1992, using CEA version):
\begin{verbatim}
{PHI3<=60, 1<=PHI3, PHI2<=21, 1<=PHI2, PHI1<=52, 1<=PHI1,
PHI3==1, PHI2==4, PHI1==4}
{1<=PHI1, PHI1<=52, 1<=PHI2, PHI2<=21, 1<=PHI3, PHI3<=60,
PHI1==5, PHI2==5, PHI3==1}
Assertion failed: file "binaires.c", line 181
/java/home3/pips/Pips/Utilities/Pips: 14764 Abort - core dumped
Exit code for pips: 134; exiting /java/home3/pips/Pips/Utilities/Pips
\end{verbatim}
\item the convex hull operator is too slow for large systems as produced
by semantic analysis on \verb+aile.f+; a new version
\verb+sc_fast_convex_hull+ was introduced, but it's slower
for small systems; FI, 9 August 1992;
\item the convex hull operator does not eliminate redundancy properly;
see for instance the precondition for the call to \verb+NORM+
in \verb+geom.f+ (part of \verb+aile.f+)
\end{enumerate}
\subsubsection{Wishes}
\begin{enumerate}
\item reproduce optimization for small systems that are in
\verb+sc_enveloppe+ into \verb+sc_fast_convex_hull+;
FI, 9 August 1992;
\end{enumerate}
\section{Link(er)}
\begin{enumerate}
\item could \marginpar{(BB/LZ)} it be executed before the controlizer
to get a nicer interprocedural control flow graph? Remi may have decided
to activate the controlizer just after the parser to save a load/unload
sequence?
\item LZ: bug linked to variable name change from T to Q
\item calls in unreachable code seems are not linked; see
\verb+Bugs/Newgen/simple4.f+; this bug may disappear because the
link is going to be executed in the parser;
Franc,ois Irigoin, 3~February~1992;
\item core dumps with an illegal branching inside a loop; see
\verb+Bugs/Link/branch.f+; Bruno Baron, 31 mars 1991.
Pips is not supposed to work on illegal fortran programs.
AL, 12 august 1992.
\end{enumerate}
\section{Loop Distribution}
No library. Included in Allen \& Kennedy parallelization algorithm and
as such coded in \verb+rice+.
\begin{enumerate}
\item induction variables definitions are distributed when they appear
in inner loop bounds; difficult to correct because of Allen \&
Kennedy algorithm; PMACS: lots of dependence arcs are added between
the index definition and the loop body; it was decided not to do
that in PIPS for performance reasons;
\item definition of privates variables are sometimes distributed wrt
their uses; see example mailed to Pierre; corrected
\item useless privatization after loop distribution; Re'mi is likely to
reuse the same DO statement...
\end{enumerate}
\section{Loop\_normalize}
\section{Loops\_Interchange}
Mispelled...
\section{Misc(ellaneous)}
\subsection{Bugs}
\begin{enumerate}
\item \verb+concatenate+ does not check for overflows; observed by FI,
31 July 1992; fixed by FI, 1 August 1992; wish: \verb+concatenate+
should automatically malloc'ed a larger buffer when necessary...
\end{enumerate}
Is it a library or a special section?
\begin{enumerate}
\item define an installation procedure for Paris; how is the PIPS account
doing? Look at make-tape?
\item PIPS documentation has to be translated into English...
\item cycles between libraries: misc (get\_bool\_property) and properties
(concatenate and pips\_error); what the best solution with these
low-level libraries? Fuse them? Duplicate some functions with
diferent names in different libraries? Temporary fix: these libraries
are called twice;
Bruno suggests to fuse the two libraries. Properties is a low level
library. It could be used regardless of PIPS. It contains only
one file and a few functions. Any disagreement?
Another solution would be to use a \verb+getenv()+ in \verb+misc+
to decided whether or not to print messages.
\end{enumerate}
\section{Movements}
\section{NewGen}
Pierre is responsible for NewGen maintenance.
\subsection{Bugs}
\begin{enumerate}
\item support for mappings over untabulated types?
\item proper handling of double quotes (onde24.f); fixed, PJ;
\item consistent code is written using \verb+gen_write()+ and is re-read
and found not consistent anymore; see onde24.f; summarized
in \verb+Bugs/Newgen/tformat.f+; further summarized in
\verb+Bugs/Newgen/simpl4.f+; FI, January 1992; the problem
is due to Newgen handling of forward referencex; a fixed order
is now imposed for file input: tabulated domain have to be
read first; fixed by PJ, February 1992;
\item SPACE characters are lost in strings delimited by double quotes;
see the parsed code for tformat (with double quotes used in
format 1100; FI, January 1992; the bug may be due to PIPS parser;
\item \verb+gen_free()+ does not go across persistant arcs but tests for
sharing beyond them; this is highly inefficient when effects
are freed; fixed by PJ, February 1992;
\item \verb+gen_read+ does not accept forward references in a tabulated
domain, entity, with a recursive data structure; FI, 10 February 1992;
fixed by PJ;
\item \verb+gen_read+ does accept forward references in a tabulated
domain, entity, with a recursive data structure; but it does
not check that they are all filled in at the end; FI, February 1992;
fixed by PJ by adding \verb+gen_read_and_check_tabulated+; FI:
I do not understand why this facility is not available at the
\verb+gen_read+ level? Because the check is already available for
non-tabulated domain?
\item \verb+gen_free()+ does dereference a free area and core dumps;
fixed by PJ, July 1992;
\item \verb+read.l+ does not check \verb+yytext+ overflows; fixed, FI,
July 1992;
\item untested overflow for tabulated domain in \verb+enter_tabulated_def()+;
FI, 28 July 1992;
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item support for mappings over untabulated types?
\item efficient handling of persistant arcs in \verb+gen_free()+
(and in its call to shared pointers).
\item better density for NewGen files; for instance, \verb+hash_tables+ are
stored as dense arrays and, as a consequence, half of entity files
are likely to be full of \verb+#]null+ to represent empty buckets;
\item check for disk full condition;
\item mapping.h : MAKE\_STATEMENT\_MAPPING should take no argument in
order to be coherent with the make functions usually generated
by NewGen; fixed, BA, August 1993.
\end{enumerate}
\section{Normalize}
Should be merged into Ri-util (20 October 1991). FI in charge? Done,
Dec. 12, 1991.
\newpage
\section{Pips}
Batch interface. The program is named \verb+pips+ and the script shell
used to make it more user-friendly is named \verb+Pips+.
Provides support for \verb+Init+\verb+Select+n, \verb+Build+ and
\verb+Perform+. See also \verb+Delete+. See man(l) for details about options.
\subsection{Bugs:}
\begin{itemize}
\item varying exit codes; no explicit \verb+exit()+ was executed;
fixed by Bruno Baron, 27 March 1992.
\item illegal instruction in exit() avec ifp.f; Bruno Baron, 16 June 96;
the bug seems to have been an overflow in \verb+Newgen/read.l+;
the code generated by Lex does not check for overflows or
\verb=yytext=;
fixed by FI, 19 July 1992;
\item no \verb+-v+ option for Perform; non-uniform interface;
fixed by FI, 23 July 1992;
\item \verb+Display+ advertizes a non-existing \verb+para+ option
and do not display the existing \verb+cray+ option;
found and fixed by FI, 23 July 1992;
\item modified to accept many \verb+-b+ options so as to compute
more than one result per call (\verb+tpips+ should be revived
for large applications); FI, 4 August 1992;
\end{itemize}
\section{Pipsdbm}
\begin{enumerate}
\item there is no need for absolute file names; they should be relative
to the location of the database schema file;
\item xxx.schema and xxx.pipsmake should be hidden in xxx.database.
( Done by LZ Sep.91 )
\item pipsdbm should be rewritten; who is interested in it? BB.
\item builder\_map: should be derived automatically from pipsmake-rc.f.tex
\item shall we keep the PURE/IMPURE request parameter or move the problem
upwards in pipsmake? Is it possible? When an IMPURE request is made,
pipsdbm copy the existing resource on disk, read it back and
return a pointer towards the copy. This means that all data structures
based on it are potentially flawed. As an example, suppose you
get a module code with the impure flag set up. The dependence
graph that was built before on the module code is not consistent
with the copy: its nodes point towards the initial copy of the
code and its arcs point towards the initial references.
In the same way, statement mappings are likely to be unusable
after an IMPURE request.
The solution to this problem is linked to a NewGen extension to
support absolute naming and some kind of type (or object?) hierarchy.
All object should be ...?!?
Fixes (kludges!):
\begin{enumerate}
\item to avoid problems with references in \verb+rice+, a special
test
is performed in \verb+db_unload()+ so as not to free the
statements
of a module;
\item \verb+Wp65+ modifies the code but performs a \verb+pure+
get which is OK to generate code in batch mode but kills
\verb+wpips+
\item \verb+Privatize+
\item \verb+Loop Interchange+
\item \verb+Partial Eval+
\end{enumerate}
\item \verb+put+ with \verb+PIPSDBM_DEBUG_LEVEL+ higher than 5 calls
\verb+gen_consistent_p+ on \verb+DBR_PRECONDITIONS+ which is
not a Newgen object; fixed, FI, 11 February 1992;
\verb+DBM_SUMMARY_PRECONDITION+ is also not tested (although
it should theoretically be tested);
\item empty lists seem not to be well handled; pipsdbm behaves as
if the resource was non-existent; see \verb+ifp.f+ and
\verb+openfile+ effects: the latter are always re-computed
whenever they are needed; this seems to be due to the fact
that an empty list is represented by NIL; FI, 22 July 1992;
problem with \verb+string_undefined+; fixed, FI, 3 August 1992;
\item calls to unput() from within a MAP over the resources;
the current cons in the MAP is destroyed by unput()...
fixed by FI, 20 July 1992;
\item \verb+REGIONS+ and \verb+SUMMARY_REGIONS+ added; FI, AL,
4 August 1992;
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item add a dirty bit to resource description to avoid useless copy back;
this will introduce a higher risk of non-consistency since
resources can be modified although pipsdbm is not notified;
FI/BB, 21 February 1992;
\end{enumerate}
\newpage
\section{Pipsmake}
Modifications of \verb+pipsmake-rc.f.tex+ are recorded here along with
information regarding library \verb+pipsmake+.
\subsection{Bugs}
\begin{enumerate}
\item unique type {\tt text} for all displayable objects; why should
we have special cases like parallel code and call graph?!?
automatic display as soon as a function producing a text object
is selected (no more {\tt view}?)? Why should we have to distinguish
between them? Because they might be parsed again later? One
reason is to generate wpips menus;
\item direct display of text objects: no double selection like
{\em statement + transformer} and {\em view sequential}; it's
not practical
\item hash tables on statements are wrong as soon as the code is saved
on disk (i.e. after a program transformation or a
parallelization); the code is saved on disk when pipsdbm is called
with an IMPURE spec. and reloaded in memory at different addresses;
one could think of unloading every data based on code when code
is modified (could it be infered automatically by an extended
NewGen with mappings?); one could also use statement\_ordering as
a storage invariant key in the statement\_mapping package;
it is more a pipsdbm problem than a pipsmake problem;
\item cope with formal functions by assuming that they have no side
effects?
\item use network parallelism?
\item blockdata exploitation?
\item vfork() for close\_module()? It's a pain to demo pips each time
you've used the main! Bug generating feature?
\item a way to make macros, i.e. to chain privatization and
parallelization would be interesting for some demonstration and
for naive users; how could we get:
\begin{quote}
parallelize(privatise(MODULE.code))
\end{quote}
without having to write C code; we need a o operator! on
pourrait ajouter a pips une option -c pour donner une liste de
Commandes (select, build et perform);
pipsmake is orthogonal wrt Parafrase pass list;
\item we need a compile-make to write (temporarily) useless objects
onto disk and to minimize memory need (see aile.f); similar to
life span and register allocation?
\item correctness bug spotted by Bruno; rules should be checked when expansed
to make sure they still expand in the same way (CALLEES might have
changed for instance)
\item correctness problem with \verb+SUMMARY_PRECONDITION+; \verb+pips_make+
should explictly destroy any resource detected as inconsistent; this
would prevent \verb+pips_dbm+ from providing stale resources; this
should never happen with a {\em clean} make mechanism, but we have
to be not so clean when resources like summary preconditions are
incrementally computed; see \verb+Bugs/Semantics/eval.f+, unless
it's in \verb+demo/Vienna+: preconditions are not accurately recomputed
after a \verb+Perform eval+ which linearized part of the code;
FI, 6 August 1992;
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item incremental source file addition: very frustrating when three
lines are missing out of 4000; at least, a list of missing sources
should be provided; there is a need for some user interaction:
\verb+query_user+; FI, each time a new program is analyzed!
\item user file restoration (same file names, same content, same order);
keep some additional information: user\_file/source\_file/rank;
MODULE.user\_file
\item \verb+builder_map+ should be automatically derived from
\verb+pipsmake-rc.f.tex+
\end{enumerate}
\subsection{Modifications}
\begin{enumerate}
\item old rule \verb+chains+ replaced by \verb+atomic_chains+ (which
is equivalent and \verb+region_chains+ which let pips build the
use-def chains and the dependence graph on top of regions;
FI, 5 August 1992;
\item rules for icfg are modified to enforce recursive calls
\end{enumerate}
\section{Prettyprint(er)}
Managed by Lei Zhou, \verb++.
\subsection{Bugs}
\begin{enumerate}
\item SUBROUTINE and FUNCTION calls with no arguments should not be
displayed with a parenthesized empty argument list.
\item redundant parentheses should not be displayed; unbearable with
complex expressions (i.e. intricate);
AC in charge, April 90; fixed by LZ, November 91;
\item some COMMENTs disappear in onde24.f; maybe a parser problem,
or a controlizer problem, or...
\item a function to print all declarations is missing; necessary when some
code is generated (i.e. Atomizer, wp65); done: new version of the
prettyprinter written by Lei Zhou (October 1991);
\item there seems to be no way to know whether a module is a PROGRAM
or a SUBROUTINE; the problem is usually avoided by keeping all
declarations in a string called \verb+decls_text+; in doubt,
SUBROUTINEs with no arguments are deemed to be PROGRAMs (Lei Zhou,
October 1991);
\item wrong user line numbering (detected by Bruno Baron, 23 October 1991);
observed again by BB, November 12; use {\em validate} to get
the bug; moved to \verb+Syntax+ (FI, 3 January 1992);
\item \marginpar{LZ}
no \verb+ENDDO+ after \verb+DOALL+; fixed by LZ 17/01/92.
Initial loop label should only be preserved in comments;
( Note: when the loop has to be distributed , loop label and
\verb+CONTINUE+ have been suppressed in
\verb+PARALLELIZED_CODE+, How can
we pick these things up to make them into comments.)
\item \marginpar{LZ} missing \verb+ENDDO+ after \verb+DOALL+;
( Same thing, See previous item LZ , 17/01/92 )
\item \marginpar{LZ} option needed to avoid printing statement number in
columns~73-80 (November 8, 1991);
\item loop increments equal to 1 are not printed anymore (LZ)
\item statement ordering missed before the IF statement.
observed by LZ. 20/02/92 in \verb+Bugs/Prettyprint/intpl.f+;
likely to be a bug in PJ's stuff to prettyprint controls (FI);
\item remove the confusion between para, para77 and para90; para should
disappear from pipsmake.rc; Lei is in charge (17 January 1992);
\item RETURN statements are kept implicit whenever possible;
FI, AL, 16 July 1992;
\item the priority between the integer division and the integer
multiplication is not observed; useful parentheses as in
\verb+2*(I/2)+ disappear; FI, 10 July 1993
\end{enumerate}
\section{Privatize}
\begin{enumerate}
\item with nested loops internal private variables are also privatized
in the outer loop (to be checked);
\item privatization at the block level, or at least at the procedure
level would be useful for wp65; at the procedure level, interprocedural
effects might be enough;
\end{enumerate}
\section{Programming Environment}
See report dret146.
\begin{enumerate}
\item get PIPS\_LIBS from the shell environment and not from the
config.makefile; done (PJ);
\item add a restore: entry in standard PIPS makefile; i.e. modify install
to preserve the current files in a Backup sub-tree; install:
mv Production/xxx Backup/xxx; mkdir Production/xxx
\item archivage des versions successives?
\item develop new makefiles/shell script in Production to try option
-O2 or -O3 and -pg; problem with external libraries NewGen and Linear;
problem with /usr/debug/malloc (add a new environment variable,
PIPS\_MALLOC?);
see make-tape for preliminary cleanup; conflicts with
development activities (available libraries won't be dbx
compatible); FI, BB: new utility added, {\tt make-pips}.
\item set core size limit to 0 for demonstration to avoid wasting time;
how can we do that automatically? On the {\em dret} account? see
\verb+$PIPSDIR/Tests/demo/demo.rc+;
\item avoid copying binary codes and libraries twice on {\em install};
install vs quick-install?!? BB: pips-makemake is modified
\item add a {\em local pips} entry to library makefile to test library
changes; the present situation is unbearable if two people work
at the same time; PJ: a new entry, {\tt test}, is added.
\item non-regression test: \verb+validate+ added by PJ (November 91);
bug: validate executes in the Validation directory and
uses only \verb+Pips/Bin/pips+ (i.e. the pips version visible
from the Validation directory); fixed by FI, 11 February 1992:
validate now runs in the current directory;
\item validate: specific test script shell must be written so as
to be executable from anywhere, i.e. the user file names must
be prefixed by \verb+$PIPSDIR/Tests/Validation/xxx+. As a benefit,
the Validation directories can be write-protected since there
is no good reason to run validate there. 12 February 1992.
\item validate generates unpredictable results if two validate
processes are executed simultaneously in the same directory;
FI, 12 February 1992.
\item validate may generate a positive results although some user
defined tests failed; we do not know if the problem is with
shell script tests not returning a proper exit code or
with validate failing to test that a result file exists;
FI, BB: 12 February 1992.
\item validate does not always destroy temporary files like diff.nnnnn,
validate.nnnnn
and errors.nnnnn in /tmp; maybe interruptions are not trapped;
FI, 28 February 1992;
\end{enumerate}
\subsection{Wishes:}
\begin{enumerate}
\item add an option ``-l'' to validate to force \verb+PIPSDBM_DEBUG_LEVEL+ to
0 (default), 5 or 9.
\item move PIPS into a SUNOS 4.1 environment
\end{enumerate}
\section{Properties}
\begin{enumerate}
\item Should they be included in the pipsmake mechanism as variables?
Wpips could generate automatically their potential values.
Consistency problem since properties are indirectly used to implement
pipsmake functions.
\item \verb+properties-rc.f.tex+ has to be filled in (17 January 1992);
Lei has to put information about the complexity options; Yi-Qing
has to explain how her statistics are organized; the present
content is too sketchy;
\end{enumerate}
\newpage
\section{Regions}
Managed by Beatrice Apvrille, \verb++.
\subsection{Bugs}
\begin{enumerate}
\item debug Alexis's code; done, June 1992;
\item elementary statements are decorated with proper regions
instead of cumulated regions; FI, 2 August 1992;
useless information is kept but it's hard to decide whether
a predicate is useful or not (connected components?); {\em
Franc,ois has found a counter-example that shows that all the
information from the predicate must be kept.\/}
redundant information should be eliminated at least;
\item three intrinsics were missing: \verb+LOG+, \verb+LOG10+ and
\verb+PAUSE+; too bad that so many different tables define
them; FI, 2 August 1992;
\item \verb+make_phi_entity()+ generated entities with
\verb+undefined_storage+; changed to \verb+rom+ to avoid
later problems; FI, 3 August 1992;
\item the TCST entity was added in a \verb+sc+ basis in \verb+add_phi_xxx+;
test added to avoid it; FI, 3 August 1992;
\item regions could not be computed for io statement with more than
one element in the io-list; same problem in library effects;
dealt with in a different way; the keyword \verb+IOLIST==+
seems ot prefix {\em every} io-list element; FI, 3 August 1992;
\item debug statements directly and indirectly called by \verb+regions+
\verb+summary_regions+ do require a value mapping regardless of
the debug level because arguments of \verb+debug()+ are evaluated
even if debug does do anything with them; \verb+debug()+ is not
a macro because varying arguments are not supported by the
preprocessor; FI, 4 August 1992;
\item prettyprint modified (as well as \verb+pipsmake-rc.f.tex+)
to display the summary region at the beginning of the listing;
FI, 4 August 1992;
\item no indices are displayed for regions imported from a called
procedure; this makes their prettyprinting unreadable and
the dependence test wrong; fixed by FI and AL, 5 August 1992;
\item user subscript expressions are substituted by phi variables
and a polyhedron just before the regions are stored in the
statement mapping; this is too late for proper combining
between two (or more) references in the same statement and/or
expression; see Alexis' mail; FI, 6 August 1992;
\item dynamic buffer overflow in \verb+words_predicate+; partially
fixed (buffer size increased, buffer made static and a
posteriori check); FI, 9 August 1992;
\item improper profile for function \verb+regions_of_implied_do()+;
this function expects an expression and an action tag as
arguments; this is what's passed when it is called from
\verb+regions_of_iolist()+; but an entity and a list of
expressions are passed when it is called from
\verb+regions_of_intrinsics()+; I cannot think of any easy
solution to retrieve the action tag, but to set it in a static
variable when the first call is issued from \verb+regions_of_iolist+;
see \verb+norm+ in \verb+aile.f+; FI, 9 August 1992;
I must have fixed it when I debugged {\tt regions\_of\_implied\_do} in
september or october 1993. But I don't remember how.
BA, 10 February 1994.
\item inadequate optimization in region translation was losing one
region out of two: each array could have either one READ region
or one WRITE region; FI, 9 August 1992;
\item preconditions were not added to regions computed for a user
call; this bug is likely to have been introduced when FI
decided to add context to reference predicates as early
as possible; FI, 7 December 1992; see \verb+Bugs/Regions/vt.f+
\item the region given for the expression
\verb+A(1,1)+ + \verb+A(3,3)+
is a MUST region; it must be a MAY region because the convex
hull of the two elementary regions is not equal to their strict
union. BA, September 17, 1993.
fixed, BA, September 20, 1993 : each time a convex hull is
calculated, the resulting region is set to MAY.
\item the regions given for an implied do are wrong :
\begin{verbatim}
C
READ (UNIT=5, FMT=*) (A(I), I=1,5)
\end{verbatim}
is displayed instead of :
\begin{verbatim}
C
READ (UNIT=5, FMT=*) (A(I), I=1,5)
\end{verbatim}
BA, September 23, 1993
fixed, even for nested IMPLIED\_DO; BA, September 27, 1993
\item {\tt sc\_projection} returns a {\tt SC\_EMPTY} predicate when an
unfeasible system is given as argument. This induces a bug in {\tt
enlarge\_statement\_regions} because a {\tt sc\_rn} is expected (bug
detected in {\tt regions\_exact\_convex\_hull}). Fixed, BA,
September 28, 1993.
\item there is a pb during the translation of regions. see mm.f in
Validation/Regions. BA, February 10, 1994.
\end{enumerate}
\subsection{Whishes}
\begin{enumerate}
\item use the dependence graph to decide whether or not two elementary
regions shoud or should not be fused;
\item the handling of reshaping could benefit from preconditions;
see \verb+zproax+ in \verb+cr2cnf.f+; Franc,ois Irigoin,
13 July 1992;
\item use {Corinne}'s conditions to derive MUST regions? How about MAY/MUST
preconditions? When preconditions are added within a loop they
can only *restrain* the region; as long as no (unsafe) projection
and no convex hull is performed, the region is a {\em must} region
and it can be used for an array kill if the inclusion of the
read-may into the write-must can be proved (we could try to use
the rational generating system, check that it is integer, and
verify that all vertices are meet the other region constraints);
this could not be done with the present implementation of regions
because proper regions are only available for elementary statements
and not for compound statements; instead of using the \verb+arguments+
field of transformers, the write effects of the compound statement
should be used (as for transformers and preconditions);
FI, 4 August 1992;
\item in the line with previous paragraph, change the projection
policy for each control structure: block, test, loop and
unstructured; the projection should no longer be performed
uniformely at the statement level in \verb+regions_of_statement+;
this may create havoc as a major design assumption is changed;
FI, 5 August 1992; Done, see report E/181/CRI. BA, February 10, 1994.
\item replace \verb+SC_UNDEFINED+ by \verb+transformer_undefined+ as
context for scalar variables. Replced by {\tt sc\_rn(..}). BA,
February 10, 1994.
\end{enumerate}
Note: array kill information is necessary to get rid of dependences on
temporary arrays. The restriction to true dependences does not eliminate
false write-read dependences because there is no way to prove that an
intermediate write rewrites the whole array.
\subsection{Modifications}
\begin{enumerate}
\item associate to statement cumulated regions instead of summary regions;
i.e. delay the projection of modified variables as much as
possible; FI, 5 August 1992; a first step was carried out for
elementary instruction (i.e. \verb+call+ instruction); well, I
prettyprint proper regions for elementary statements, but I
compute cumulated regions to have the regions of a block. See
report E/181/CRI. BA, February 10, 1994.
\item better computation of MUST vs MAY regions; a new projection procedure
should be used to know whether the result is exact or not; only
loop indices can be safely projected (because of the loop semantics);
loop bounds have to be checked to make sure they are both used
in the region context before the region is declared MUST; projection
of variables that are not enclosing loop indices always result
in a MAY region, regardless of the projection exactness;
FI, 6 August 1992;
YY wrote \verb+combiner_with_test(sc, v)+ to check exactness;
it's in \verb+ricedg/testdep_util.c+ and is cluttered with
measurement statements; YY, 6 August 1992;
\end{enumerate}
\section{Ri-util}
\begin{enumerate}
\item break the ri-util.c file into smaller files, one per data type
(entity, expression, etc...) or at least by data type SCC; we
need an entity.c for sure; done;
\item \verb+sc_gen_write+ and \verb+sc_gen_read+ could not handle
\verb+SC_UNDEFINED+; fixed, FI, 4 August 1992;
botched: the representation chosen for {\tt SC\_UNDEFINED} is a
linear system with no equalities and no inequalities over a
0-dimensional space; this is just the system representing an
identity transformer;
suggestion: do not use {\tt SC\_UNDEFINED} in valid data structure;
modify library region to use \verb+transformer_undefined+ as context
for scalar variables;
other suggestion: choose a more complicated scheme to store
undefined \verb+Psysteme+; for instance \verb+(SC_UNDEFINED)+;
this implies a test after the opening parenthesis but seems
pretty easy to implement; FI, 8 August 1992;
\item dynamic and static areas are declared when a module is created
\item intrinsic \verb+PAUSE+ was not recognized as an intrinsics:
there is still another intrinsic table in
\verb+ri-util/prettyprint.c+;
FI, 12 March 1993
\item \verb+entity_conflict_p()+ was deadly wrong; it returned almost
always false; it contains a static hash table that is going to
cause trouble in wpips sone day! FI, 29 June 1993;
\item NORMALIZE\_EXPRESSION is now used instead of
NormalizeExpression to recurse in normalize.c. RK, FI, 3/12/93.
\end{enumerate}
\section{Rice (parallelization phases)}
\begin{enumerate}
\item take control dependences into account? Use properly the hierarchical
control graph? To be studied by BB;
\item \marginpar{PJ} option to only use data flow dependences
(\verb+RICE_DG_DATAFLOW_DEPENDENCE_ONLY+); seems to be bugged: only
output dependencies are ignored (FI);
\item \marginpar{YY}
quantify/summarize parallelization results; comparison with FPP
should be made easier;
\item Distribution doesn't work if the loop has a label. pips\_error()
in local\_name() instead. First loop should keep the label,
shouldn't it? (BB, 3.1.92)
\item some parallel loops seem not to be detected in onde24.f and
its simplification, Wp65/seismic.f
\item bug in Tests/demo/gewpp.f: combination of tests and loops;
GEWPP: Gaussian Elimination With Partial Pivoting;
a \verb/+1/ had
been commented out for unknown reasons; fixed by YY, Nov. 8,
1991;
\item bug in distribution because of control dependences: if inner
loop range depends on a statement in the outer loop, it doesn't
inhibit distribution! See Bugs/Rice/ctrl.f. Bruno Baron,
6~April~92; this bug was initially detected by Nahid Emad in 1990;
Basically, the proper effects of statements inside loops now include
some kind of "control" effects that force sequentialization when loop
bounds depend on previously defined variables. The statements' effects
within a loop have an added read effect on the (possible) variables
appearing in the loop bounds. The loop body, even if it is a block has
a proper read effect on these variables. Fix by Pierre Jouvelot, 21 May
1992;
\item bug in INTERF from mdg.f; rice does not handle labelled DO loops
like:
\begin{quote}
100 DO 200 I = 1, N
\end{quote}
Label 100 should be reported on an additional previous CONTINUE
statement, because the Allen\&Kennedy algorithm may duplicate loops.
FI, PJ: 16 June 1993; see above, comments from BB on 3 Jan. 92;
\end{enumerate}
\newpage
\section{Ricedg (Dependence Graph Computation)}
This library is managed by Yi-Qing, FI and BA.
\subsection{Bugs}
\begin{enumerate}
\item \marginpar{YY} normalization problem for:
\begin{quote}
\begin{verbatim}
A(2*I+4*J) = A(2*I+4*J+1)
\end{verbatim}
\end{quote}
Fixed by YY, Nov. 7, 1991; a normalization step was added before
a call to \verb+sc_minmax_of_variable()+;
however, normalization steps
were later added in \verb+sc_minmax_of_variable()+ by CA for PUMA;
this should be cleaned up;
\item \verb+quick_privatize_graph()+ did not work very well. I have
modified it slightly. It works well now for Wolfe's Cholesky. I can't assure
it always works for any other programs. If it fails, use
\verb+Perform priv+ before; (YY)
\item the types {\tt arc\_label} and {\tt vertex\_label} were not
defined. They were declared as externals in graph.hand used to
define vertex; {\tt dg\_vertex\_label} and {\tt dg\_arc\_label} are
defined in dg.h using vertex. It didn't appear until PJ added
castings in dg.h. fixed by FI and BA, August 1993, by defining the
types {\tt arc\_label} and {\tt vertex\_label} in {\tt
ricedg-local.h}.
\end{enumerate}
\subsection{Wishes}
\begin{enumerate}
\item ddv \marginpar{(YY/FI/PJ)}implementation; necessary for
PIPS-CRAY and for PUMA; how can we preserve distances (necessary for
Benoit)? NewGen ddv representation; can we replace $<$ with
+1 and say we define a generator of an over-approximation of the
dependence set? Decision: dependence cones are implemented;
\item \marginpar{YY} dependences should be prettyprintable,
filterable, and so on; Dominique Cle'ment and visitors in
Brussels;
\item the dependence graph should be saved on disk; this is likely
to become necessary to cooperate with CEA (it was planned
in FIPEX); it is also very inconvenient when one switches
from parallel Fortran to Fortran 90 as output style: everything
has to be recomputed; also, it would make testing easier: now,
bugs in the dependence graph are likely to be seen as bug in the
parallelizer; two problems have to be solved, unless one is
ready to save the statements with it (which may be the easiest
way): replace pointers from vertices to statements by statement
ordering (as in mappings), replace pointers from arcs to
references by some kind of numbering (number of the reference
in the statement + statement ordering?);
\item \marginpar{YY}
an ASCII version of the dependence graph is necessary to validate
its computation regardless of its use (paralellization, PUMA, Cray);
its also useful as a debugging tool and as a way of comparison with
other systems like PAF;
\item should \marginpar{(PJ)}we introduce control dependences?
Or rely on our hierarchical control flow graph? should it be
stored on disk? No NewGen support?
\item \marginpar{YY} loops with control structures, structured or not,
are not analyzed; Remi probably skipped them to gain time because he
knew that his parallelizer, based on A\&K, would not handle them;
however, we now need dg information about these loops for PUMA,
PIPS-CRAY, Yi-Qing's PhD and CEA; the easiest change would be to add a
property
\verb+RICEDG_ANALYZE_SIMPLE_LOOPS+, which should be set to
\verb+true+ to obtain Remi's behavior; when set to \verb+false+,
its default value, the whole dg would be analyzed; Yi-Qing is in
charge of writing the missing pieces of code;
The new version was bugged because statements could not be ordered
any more using the low weight part of \verb+statement_ordering+ (YY);
PJ suggests to consider non comparable statements as belonging to
the same compound statement (wich is true). He also suggests to
handle compound statements like user calls, but the use-def chains
are not build that way (so does FI believe).
\item code related to statistics gathering should not be intimately
mixed with functional code; FI, 5 August 1992;
\item the partial evaluation of expression should be used for the
symbolic loop increment; (YY, 08/92)
\end{enumerate}
\subsection{Modifications and improvements}
\begin{enumerate}
\item the dependence cones are implemented which are represented by
generating system from \verb+C3/Linear+; (YY, finished in 08/91)
\item ricedg did not contain the tests of faisabilities for the
initial system of dependence before the projection to the Di system.The
modification has been done. (YY, 18/10/91)
\item The module computing
the levels of dependences gave a false level when di$\leq$0.
Thus the case
of di = NotPositif has been considered in ricedg now. (YY, 10/91)
\item auxiliary difference variables have been introduced for all
variables modified in the current environment (i.e. current
loop nest); so in this case the dependence test is correct now;
(YY, 08/92)
\item the symbolic increment can be accepted by ricedg; the
independence can be found by GCD test and Di ``all zeros'' test
in this case; (YY, 08/92)
\item statistics \marginpar{(YY)}
about dependence tests :number of tests, number of dependences, etc...(see
properties-rc.f.tex also) have been done; its results are stocked
in a file by setting the property (\verb+RICEDG_PROVIDE_STATISTICS+) TRUE;
(YY, 08/92)
\item the dependence graph is prettyprintable with a option ``dg''
(ex. Display module dg); (YY, 08/92)
\item ricedg was (slightly) modified to use the \verb+context+ field
of effects decorating the dependence graph; this is possible
with the \verb+rice+ option but not with \verb+rfulldg+ or
\verb+rsdg+ as the additional information they would brought
is already contained in the context; to use this new option,
you have to select the \verb+region_chains+ option to compute
the use-def chains; FI, 6 August 1992;
\end{enumerate}
\section{SDFI}
Summary Data Flow Information.
\begin{enumerate}
\item \verb+same_sdfi_p()+ was bugged; this was shown on
\verb+ONERA/aile.f+; fixed, FI, 7 August 1992;
\end{enumerate}
\section{Semantics}
\begin{enumerate}
\item Karr like fix-point operator for induction variables; not implemented;
\item some integer scalar variables are ignored; JH in EXTR;
IX, IY, INCX, INCY in SDOT; see transformers and effects
\item use declarations to derive constraints like N.ge.1; use these
extra-constraints to get MUST effects information for loop bodies?
see section {\em Regions};
\item exploit intrinsics like MOD, MAX, MIN and /; keep it simple;
do not try to use temporary variables for sub-expressions;
\item check that common variables are properly handled; they should
be translated like formal parameters; done, FI, July 1992;
\item EQUIVALENCE equations are lost
\item change transformer\_convex\_hull() to take into account distinct
basis; a 0-dimensional space should be a neutral element (always?); how
partial subspaces could be dealt with? Is transformer\_identity\_p()
satisfying?
\item display summary\_transformer before module declarations when
prettyprinting transformers;
\item value\_mappings should be reset in module\_name\_to\_transformers,
module\_name\_to\_preconditions, prettyprint\_transformers,
prettyprint\_preconditions
\item does the name {\em t} of the Fortran program have the magic
effects? we can't get the effects of the Fortran program that
has the name {\em t}. when we change the name, everything works
fine. Strange indeed ?! ( added by lz 05/06/91 )
\item improve handling of test like \verb+i.eq.1+ to cope with Wegman/Zadeck
example (see \verb+zadeck.f+);
useful for switches too: if-elseif-elseif....
\item propagate call site preconditions for functions (only call to
subroutines are handled);
\item integer variables stored in COMMON are not handled interprocedurally;
the translation routines wrongly assumed that they were named as
\verb+common_name:variable_name+; in fact, they are named like
any other variable as \verb+module_name:variable_name+; fixed,
FI, July 1992;
\item a new bug cropped up for equivalenced variables; it appeared
the first time for ICS'91; it seems that \verb+entity_conflict_p+
was modified: equivalenced variables do not conflict anymore...
see \verb+equi.f+;
\item integer user functions like {\tt m = ifoo()} are not exploited because
the expression is not linear; FI: I think it is very unusual to
know the value returned by a function; LZ: I have a case in
Hermann Mierendorf code; December 1991;
\item preconditions are not properly computed after a privatization
in demo/choles.f:
\item in func.f : information about function values is lost when
transformers are computed: see interprocedural preconditions of func.
\item Pips doesn't work for the empty THEN or ELSE block, needs a
CONTINUE to be passed by pips. LZ 24/02/93
\begin{verbatim}
baron@isatis Tests/demo : Init -d -f choles.f ch
baron@isatis Tests/demo : Perform -m choles priv
baron@isatis Tests/demo : Select rsdg
baron@isatis Tests/demo : Display -v
Workspace ch (default)
Module CHOLES (default)
pips -m CHOLES -s PRINT_PARALLELIZED_CODE -b PARALLELPRINTED_FILE ch
Workspace ch opened
Selecting rule: PRINT_PARALLELIZED_CODE
Request: build resource PARALLELPRINTED_FILE for module CHOLES.
PROPER_EFFECTS building PROPER_EFFECTS(CHOLES)
CUMULATED_EFFECTS building CUMULATED_EFFECTS(CHOLES)
CHAINS building CHAINS(CHOLES)
SUMMARY_EFFECTS building SUMMARY_EFFECTS(CHOLES)
TRANSFORMERS_INTRA_F building TRANSFORMERS(CHOLES)
PRECONDITIONS_INTRA building PRECONDITIONS(CHOLES)user warning in call_site
_count: not implemented yet, always returns 1
pips error in entity_to_new_value: unbounded entity K
\end{verbatim}
Fixed by FI, 1 August 1992; two changes were performed; (1) conditions
on loop indices are added only if the loop index has a value;
(2) privatized scalar integer variables are declared as having
a value although they do not have any effect; see \verb+mappings.c+
\item \verb+Bugs/Semantics/lmult.f+: assertion violated in sommet.c;
this routine computes values close to MAXINT using parentheses
to avoid overflows; YY/FI: 18 February 1992;
\item \verb+add_or_kill_equivalenced_variable()+ warns about return
storage of \verb+isamax+ in \verb+isamax+; this warning does
seem to be useful, but is it really useful? FI, 23 July 1992
\item summary preconditions are not always correct; there is no way
to reset them and they are always increasing when convex hulls
are performed; so if a summary precondition has been set equal
to the whole space at some time, it can never be made more
precise afterwards; see \verb+eval.f+; FI, BB, 30 July 1992;
\item prettyprint modified (as well as \verb+pipsmake-rc.f.tex+)
to display the summary precondition or transformer
at the beginning of the listing;
FI, 4 August 1992;
\item exponential time and space for interprocedural preconditions
of \verb+onera/aile.f+; heuristics used to trade accuracy
vs speed; FI, 13 Jan. 1993;
\item global value translation: an actual parameter was a COMMON
variable in the caller, and just a formal in the callee;
translation from the caller into the callee was attempted;
see \verb+translate_global_value+, FI, 13 Jan. 93;
\item seems to loop forever on LMULT in LGS.f (QCD Perfect Benchmark);
likely origin: subtle additions of 2**31... FI, 5 May 1993
\item value entity \verb+PROFYL:IS1ST#init+ is found with fields type,
storage and value undefined (\verb+undefined_xxx+) when
computing preconditions for \verb+TRACK+; the core dump occured
when the summary precondition of \verb+THINIT+ was computed;
the bug was not reproducable; it seems due to a long use of
a workspace; I applied \verb+stf-workspace+ after a first run;
no idea; FI, 18 June 1993;
\item call to \verb+sc_minmax_of_variable+ added in
\verb+add_good_loop_conditions+ to handle symbolic steps
when their signs are known; suggested by Kathryn McKinley;
FI, 28 June 1993;
\item \verb+translate_global_value+ and
\verb+precondition_intra_to_inter+ are modified to let invisible
information pass thru intermediate procedures; e.g. A calls B
and B calls C; A and C share common variables, which are not
seen from B; however, B summary transformer and precondition
must take them into account; FI, 29 June 1993;
\item code added for MIN0 and MAX0 intrinsics; FI, 18 August 1993;
\item code added for STOP intrinsic; FI, 3 February 1993; this may be
devastating in a main; STOP is abstracted as the non-feasible
transformer;
\item prettyprint of transformers and preconditions with the parsed
code instead of the code to avoid display troubles introduced by
the controlizer; FI, 3 February 1994
\item in {\tt statement\_to\_transformer}, the transformer is not
copied before being stored in the statement mapping and is then
returned to the caller. {\tt loop\_to\_transformer} modifies it by
side effect. The transformer and the preconditions computed with it
where wrong in some cases (see {\tt
Validation/Semantics/li92\_1.f}. BC, oct. 94.
\end{enumerate}
\subsection{Wishes:}
\begin{enumerate}
\item the precondition list should not be reversed from each statement
to the next;
\item equalities with negative constants are not nicely printed:
\verb+I==-1+ is shown as \verb/I+1==0/;
\item it would be useful to compute the buffer size necessary to
read a precondition p (i.e. a Pvecteur) each time p is
written;
\item it would be nice to accept a user file as an initializer for
summary preconditions, at least for the main; if the C3/Linear
format is used, a parser is already available; \verb+user_request+
could be used to enter a file name; FI, 30 July 1992
\item LOGICAL variables should be handled; see bdna.f (perfect Club)
as an example; this is also requested by Stein and Fox in CPE,
Dec. 93; FI, 10 February 1994
\end{enumerate}
\section{Syntax}
Lexical and syntactic analyzer.
% Should be called \verb+parser+. BTW, to
% see debug messages, set \verb+PARSER_DEBUG_LEVEL+ instead of
% \verb+SYNTAX_DEBUG_LEVEL+!
%
% No more: I changed PARSER_DEBUG_LEVEL into SYNTAX_DEBUG_LEVEL; FI, 15
% September 1993
\subsection{Bugs}
\begin{enumerate}
\item \label{name-spaces}
unique space for COMMON names and variable names; modify lexical
analyzer and parser to keep slashes around common names so as to
distinguish them textually from variables? See MakeCommon and the
prettyprinter. Or use a general scheme to distinguish external
variables from internal variables.
bug in Bugs/Syntax/decl2.f: the common is declared before the variable;
a nice {\em user} error message is printed; fixed by FI;
bug in CEA benchmark \verb+matrie.f+
(also stored in \verb+Bugs/Syntax/matrie.f+):
double use of \verb+ZZZZ+ as
a logical variable name and as a common name; PIPS used to core dump
because no doucle declaration check was performed when the
variable was declared before the common; fixed (if I may say so!)
by FI, 16 January 1992: a proper {\em user} error message
is printed before PIPS stops...
The PIPS grammar \verb+gram.y+ and \verb+MakeIntrinsic+ were
modified to directly prefix global names
with \verb+TOP-LEVEL+ instead of \verb+CurrentModule+.
\verb+MakeAtom+ should also be modified to update entity names
when it finds out that it is an external entity; \verb+MakeAtom+
modified mainly through \verb+MakeExternalFunction+; FI, 9 March 1992
\item unique space for INTRINSIC names and variable names;
intrinsics
detection could (should) be delayed to link time, but their names
should not appear as callees (done); else the request for source file
will abort; check that the linker does not use callees to build
its link hash table; how is the INTRINSIC statement handled? The
EXTERNAL one?
more or less fixed by FI, 21 September 1991; there is still a unique
name space for intrinsics and variables but a variable can have
as name the name of an intrinsic;
further fixed by FI, 16 January 1992; see previous
bug \ref{name-spaces};
\item Pips' parser bug : type problem.
It is observed when running ``pips'' on BLAS3 subroutines for the
control-flow graph.
The problem comes from the following declaration:
\begin{verbatim}
CHARACTER*( * ) NAME, OPTS
\end{verbatim}
pips' parser fails to give a good type to \verb+NAME+ and
\verb1OPTS1.
dbx detects that
MakeFortranType(t = 9, v = 0x1ee1b8),line 755 in syntax/declaration.c
where 9 is too large for a Fortran type. ( Supposed to be 1 to 7 ).
The sample program is kept in ~pips/Pips/Tests/Bugs/Syntax/ilaenc.f .
Reported by LZ, Oct. 9, 92.
Fixed; FI, 15 September 1993
\item missing declarations observed by Lei Zhou when building the call
graph; might be due to name captures?
\item see if the call to FindOrCreateEntity() should/could not be delayed
in gram.y; when an entity is recognized, only its name should
be passed in \$\$ instead of a pointer to an entity whose full name
is still unknown; entities should be effectively created when the
context is better known, COMMON statement or not, declaration
or instruction part of the module.
In fact, this is dealt with in bug \ref{name-spaces}.
\item accept BLOCKDATA, even if they are not exploited, to avoid core
dumps on real programs; try to do that when writing a pips fsplit
in awk (FI); but awk cannot open more than 7 files and is stuck
when more than 6 subroutines have to be split;
BLOCKDATA are parsed like other modules; FI, 15 September 1993;
\item run PIPS on SPEC77 (PerfectClub); look at \verb+del4+, \verb+gloop+ and \verb+gwater+;
these files are now properly parsed; parallelization bug from
Yi-qing prevents interprocedural parallelization in \verb+del4+; effects
bug in \verb+gfidi+ prevents interprocedural parallelization in \verb+gloop+;
effects bug in \verb+lrgscl+ prevents interprocedural parallelization in
gwater;
\item ParserError() should accept varargs; at least errors should be better
documented using user\_error() before;
\item {\em conflict for symbol ...}: no line number specified;
\item move as many modules as possible into ri-util; they are going to
be used to generate code automatically in PIPS-CRAY and PUMA;
Corinne has already taken a look at this; Bruno too;
\item EXTERNAL declaration in onde24.f causes a core dump (not a syntax
error); fixed by FI, 21 September 1991;
\item functional formal parameter should be detected by the parser and
processing should be stopped with a proper warning;
Theoretically fixed by FI, 18 January 1992, but not checked.
Checked with formel.f and installed in Validation/Syntax,
FI, 12 February 1992.
\item some comments seem to be lost in onde24.f, although they are
not linked to control statement (i.e. \verb+GO_TO+);
\item constant \verb+MAXBLOCK+ controls the nesting level; since
\verb+ELSEIF+ are desugared in nested \verb+IF+, its value of
10 was too small; it was increased to 20 (FI, 2 December 1991)
\item wrong user line numbering (detected by Bruno Baron, 23 October 1991);
observed again by BB, November 12; use {\em validate} to get
the bug; \verb+stat_num+ was statically initialized; statement
number depended on the module ordering for parsing (FI, 3
January 1991); this does not explain the problem with validate
since everything should be redone in the same order unless
databases are not destroyed;
\item unique space for {\em intrinsics} and user names; keywords like
\verb+RETURN+ are handled as intrinsics and are misinterpreted
as calls to a user function, it is defined:
\begin{verbatim}
subroutine return
return
end
\end{verbatim}
PIPS core dumps because this is encoded as recursive calls to
user function \verb+return+. There is no difference in the RI
for \verb+RETURN+ and \verb+CALL RETURN+.
This may have now changed because external names are directly prefixed
by \verb+TOP-LEVEL:+. Such user function names should
be refused as redefinition of an intrinsic.
See the correction of the first Syntax bug.
To be checked (FI, 22 January 1992).
\item \label{complex-constant}
complex constants are not parsed: \verb+syntax error near ,+;
discovered because Dongarra parallelizer benchmark was not
accepted; see Bugs/Syntax/complex.f; FI, 08/01/92;
partially fixed at the lexical level: pairs of floating point
numbers are recongized but not pairs of integer numbers or
integer/floating point combinations (see paragraph 4.6.1 in
the standard); FI, 9 January 1992; this does not work: pairs
of floating point numbers can appear in a function call like
\verb+FOO(1.,2.)+;back
to the previous version;
Fixed by modifying \verb+gram.y+. See comments there.
FI, 18 January 1992;
This introduced a bug related to IO list (see below
\ref{implicit-do}). Complex
constants are not recognized in PIPS Fortran. They have to
be replaced by a call to \verb+CMPLX+.
\item COMMON length are likely to be varying from a module to another.
They are statically allocated as \verb+TOP-LEVEL:FOO+ instead
of \verb+CURRENT:FOO+ and common variables are systematically
added at the current end of the common.
This bugs was fixed a first time by generating {\em local}
common names, to be later linked (more or less?). This caused
problems when a variable and a common have the same name. This
seems to be more the rule than the exception: programmers
name their common like the most important variable of the
common.
It should be fixed by adding a \verb+common_size_map+ in the
parser to remember the current size of the common wrt. the
current module. The consistence between the different
declarations of a given common could be checked at the
end of each module syntactic analysis.
A new bug file should be added to \verb+Bug/Syntax+.
Fixed by FI, February 1992.
\item character constants are not handled in a uniform and consistent
way in statements and in FORMATs. See \verb+Bug/Syntax/string.f+
\item \label{implicit-do}
implicit DO in IO lists seem to be refused by the parser; see
\verb+Bugs/Syntax/edition.f+ from \verb+aile.f+ and
\verb+Bugs/Syntax/grandg.f+ from \verb+cfc2.f+ (SIDOLO);
Yi-Qing bumped into the problem many other times in Sidolo
and ONERA benchmarks; this used to work; 29 January 1992;
this might be linked to the new rule introduced to parse
complex constants (see above \ref{complex-constant})
No simple solution seems to exist; complex constants have
to be not recognized for the time being; FI, 4 February 1992.
\item blank commons are not handled properly; see \verb+Lib/syntax/common3.f+
FI, 6 February 1992;
\item long lists of values in DATA are not accepted: yacc stack overflow;
see \verb+Bugs/Syntax/hit.f+; Yi-Qing, 5 February 1992;
fixed, FI, 6 February 1992;
\item long list of variables and values found inconsistent by PIPS in
\verb+AnalyzeData+:
\begin{quote}
Too few initializers
\end{quote}
but not by f77; see \verb+Bugs/Syntax/calavs.f+, Yi-Qing,
6 February 1992
Same problem seems to occur in \verb+Bugs/Syntax/rffti1.f+ which
is extracted from \verb+APS.f+; the context is easier;
FI, 14 February 1992;
MakeDataVar() used to assume that a whole array was initialized
each time an array element was initialized; AnalyzeData() seemed
to have a wrong test for the error message \verb+too few initializer+;
fixed (?), FI, 18 February 1992;
\item long lists of arguments in CALL are not accepted: yacc stack overflow;
see \verb+Bugs/Syntax/pset.f+; Lei Zhou, 6 February 1992;
fixed, FI, 6 February 1992;
\item \verb+fsplit+ does not indentify subroutines whose names appear
only in a continuation line; FI, 7 February 1992
\item empty programs having just one END statement are declared illegal;
FI, 7 February 1992 (almost a feature; this is due to the grammar;
PJ tried to update it but failed);
\item the substring operator is recognized but not accepted; it is not
possible to generate a regular call to the substring function
because it may appear as a lhs; however, in SIDOLO, the program
where we bumped into te problem, all substring operations seem
to appear in expressions (i.e. rhs)
FI, June 1991 and 11 February 92.
\item long lists of formal parameters are not accepted: yacc stack overflow;
see \verb+run.f+ in Perfect Club benchmark \verb+APS.f+;
Francois Irigoin, 12 February 1992;
\item no overflow check for the comment buffer; array \verb+keywidx+ was thus
overwritten when parsing \verb+RUN+ in
Perfect Club benchmark \verb+APS.f+; Francois Irigoin, 12
February 1992; overflow check added (and tested) and then
buffer size doubled; FI, 12 February 1992.
\item \verb+implicit_type_p()+ seems to be to strict;
see \verb+Bugs/Syntax/imprif.f+; because the implicit statement
occurs (necessarily) after the formal parameter declaration, they
do not have their current implicit type when they are redeclared;
see variable \verb+CN+ in \verb+imprif.f+; LZ, FI: 12 February 1992;
the only short term fix I can think of is to move formal parameter
declarations {\em before} the \verb+IMPLICIT+ statement;
\item one more (unidentified) bug with \verb+RUN+ in
Perfect Club benchmark \verb+APS.f+;
\begin{quote}
user error in MakeExternalFunction: parser error between lines 303
and 304 \\
Formal functions are not supported
\end{quote}
Franc,ois Irigoin, 12 February 1992
In fact, it was a bug in \verb+APS.f+;
YY had commented out declarations in \verb+run.f+ thus making regular
array (formal parameter) look like formal functions; the error
message should be more informative; fixed;
\item TABs are not properly expanded; see \verb+Bugs/Syntax/tabs.f+;
they seem to be systematically replaced by 7 spaces
in \verb+GetChar()+; FI, 21 February 1992; fixed, FI, 26
February 1992 (Column was replaced by col in the test)
\item pips core dumps with too many ENDIFs instead of issuing a user
error; see \verb+Bugs/Syntax/endif.f+; missing check in
\verb+MakeEndif()+; FI, 21 February 1992
\item SPACEs in empty lines are taken into account to compute the
truncation at column 73 in the next instruction; FI, 27 February
1992; fixed by reinitializing col one more time, FI, 27 February 1992;
\item on \verb+user_error+ the database is not saved; \verb+sserror+
(i.e. \verb+yyerror+ in gram.y) ends with a regular exit;
this exit was replaced by a call to \verb+pips_error+; FI, BB,
27 February 1992;
\item \verb+LOG+ and \verb+LOG10+ are standard generic intrinsics;
they were forgotten in bootstrap and effects;
see \verb+Validation/Syntax/testlog.f+
fixed, FI, 27 February 1992;
\item all symbols declared in an EXTERNAL statement are considered
FUNCTIONS and their results are given implicit types;
this clashes when it is discovered later that is is a SUBROUTINE;
the type consistency test was soften to allow implicit types
to be redefined; why is the \verb+unknown_type+ not used? See
above the problems with types of formal parameters. FI, 9 March 1992.
\item comment lines starting with an asterisk are not recognized
although this is part of the standard (section 3.2.1);
see \verb+Bugs/Syntax/asterisk.f+, FI, 10 March 1992;
fixed, FI, 12 March 1992;
\item initial lines with a zero in continuation column were mistaken
as continuation lines; see \verb+Validation/Syntax/cont0.f+;
FI, 12 March 1992;
\item PIPS core dumps on the following string declaration:
\verb+CHARACTER*(*) MESAGE+; see \verb+/Bugs/Syntax/errors.f+; YQ,
6 Apr. 1992; Fixed, FI, 15 September 1993;
\item type reconstruction error in \verb+MakeExternalFunction+; see
\verb+Validation/Syntax/linear.f+ and \verb+formel.f+; the bug
was introduced by FI when trying to perform link at parse time;
fixed; FI, 12 July 1992;
\item TABs were not properly expanded; see \verb+Validation/Syntax/sdot.f+;
fixed; FI, 14 July 1992;
\item global \verb+SAVE+ statement without argument list is not recognized;
FI, 28 July 1992; fixed in gram.y but no action implemented, FI,
28 July 1992;
\item calls to function without arguments are not taken into account
in \verb+CALLEES+; see \verb+Bugs/syntax/funcall.f+;FI, 30 July 1992;
\item comments carried by \verb+CONTINUE+ statements seem to disappear;
this may not be due to the parser but to the controlizer;
FI, 5 August 1992;
\item comments carried by instruction formated on multiple lines seem
to disappear; see \verb+Bugs/syntax/contcom.f+; FI, 30 December 1992;
\item parameterless function calls are analyzed as reference to
variables; see \verb+Bugs/Syntax/time.f+; FI, 11 March 1993;
fix on March 11 showed a bug in ri-util on PAUSE;
\item buffer size increased for very long declarations in ocean.f,
PerfectClub; FI, 21 June 1993;
\item {\tt free\_entity()} added in procedure.c when a assumed variable turns
out to be a function and when that variable has been explictly
declared before as in:
\begin{quote}
INTEGER*4 CLOCK
I = CLOCK()
\end{quote}
This does not solve all problems if LOG = LOG(3) is allowed
by Fortran standard. See comments in procedure.c.
FI, 21 June 1993
\item bug in oa118.f because of the previous modification; the freed
entity is not removed from the module declaration list; YY, FI,
24 June 1993; new function \verb+remove_variable_entity()+
added in \verb+ricedg+ to clean decalrations better; FI, 25 June
1993
\item bug in validate: formal functional parameter are no longer
detected; previous modification is refined according to the
storage in MakeExternalFunction()+; the bug was not apparent
because the validation shell scripts were bugged; fixed; FI, 29
June 1993
\item bug in ??? (renault.f) in a statement containing an expression looking like
\begin{quote}
ALOG(0.5*ALOG(X))
\end{quote}
\verb+remove_variable_entity()+ is called on the fly and one of
the two pointers to \verb+MODULE:ALOG+ is left dangling; a list
of variables to remove should be built and non-existent
variables should be removed at the end of the parsing phase;
CA, FI: 8 July 1993; fixed the same day;
\item intrinsics \verb+BACKSPACE+ was only partially recognized in the
parser, and forgotten in bootstrap/bootstrap.c and
effects/intrinsics.c; CA, FI: 8 July 1993
\item character constants are unduly capitalized (see standard section
4.8: {\em any character capable of representation in the processor});
FC: July 1993
\item computed GOTO are desugared; FI, 15 September1993;
\item field offset is fixed at the end of the declaration section; it
was wrong when the type was declared after the first occurence
of the variable; this bug introduced pseudo-conflict between
variables in the same common (see TRACK for instance); FI, 1 December 1993
\item commons already in the symbol table are now properly declared as
visible in the current module declaration field; FI, 1 December 1993
\item bug on offset computation for variables of different modules
fixed; FI, 2 December 1993
\item CleanLocalVariables() improved to clean up the layout field of
commons; possible explanantion for a problem with free added;
FI, 2 December 1993
\item bug on area layout when types of variables are redeclared: the
type redeclaration modifies the length used by the variable in
the area but not its offset; observed on track-fi.f; 1 February
1994; problem avoided by ignoring type redefinition for
variables whose store is already defined; FI, 1 February 1994;
\item bug on area layout (again); previous modification was a mistake;
bug in {\tt update\_common\_layout()} fixed; validation with retype.f;
FI, 7 February 1994;
\end{enumerate}
\subsection{Wishes:}
\begin{enumerate}
\item recognize and handle Fortran macros
\item recognize and handle ENTRY
\end{enumerate}
\section{Text and Text-Util}
Low-level data structure used by prettyprint oriented functions.
\subsection{Bugs}
\begin{enumerate}
\item likely to be responsible for many memory leaks (FI); intermediate
storage is not freed, just in case it might be shared, which happens
sometimes! Please no more code of that kind in PIPS\verb+:-)+.
\item the layout is Fortran dependent; information from ri-util,
namely \verb+STATEMENT_NUMBER_UNDEFINED+, should be used but is
not available according to the PIPS library hierarchy; this constant
is assumed {\em negative}; the bug was: multiple line statements
unconditionally had -001 as statement number for all lines but
the last one, that had the proper number; FI, 16 November 1991;
\item strings to long to fit on one line were not broken into parts
and continuation lines but just printed; code produced by PIPS
was not compilable; this showed mainly in FORMAT statements;
FI, 15 March 1993.
\end{enumerate}
\section{Tiling}
No library yet. See {\em wp65}.
\subsection{Wishes:}
\begin{enumerate}
\item handle not perfectly nested loops by adding guards and optimize
guards in generated code by moving as many of them as possible
in loop bounds
\item handle adjacent loops, like the ones in image processing, by
using a \verb+switch+-like statement
\end{enumerate}
\section{Top-Level}
\subsection{Bugs:}
\begin{enumerate}
\item can we assume there exist a split shell script? Is it easier
to modify than C code? Not compatible with file regeneration
when pips is stopped? Adapt \verb+top-level/source_file.c+
and test \verb+Utilities/split.awk+
\end{enumerate}
\section{Transformer}
Low level data structure for semantic analysis (see {\em Semantics}).
\subsection{Bugs:}
\begin{enumerate}
\item \verb+transformer_filter+ had to be extended to cope with
empty relations (non-feasable systems)
\item \verb+transformer_combine+ was extended to cope with dead code
by generating a non-satisfiable transformer; FI, 27 May 1993;
\item \verb+print_transformer+ was restricted to well-defined
transformers; i.e. a pips error is generated; FI, 27 May 1993;
\item \verb+transformer_convex_hulls+ was extended to cope with dead code
by generating a non-satisfiable transformer; FI, 27 May 1993;
\item \verb+arguments_union+ was not unioning its second argument;
FI, 27 May 1993
\item \verb+value_entity_p()+ added; FI, 29 June 1993
\end{enumerate}
\section{Utilities}
\subsection{Bugs}
\begin{enumerate}
\item \verb+validate+ does not complain directly
when it is run on a non-existing directory (it sends an e-mail message)
\item \verb+Perform+ does not check the existence of a rule before
starting Pips and pips; this leads to a core dump for a typo!
FI, 10 August 1992
\item \verb+pips-makemake+ modified to use sccs; RK, FI, June 1993
\end{enumerate}
\section{WP65}
Code written for the PUMA project: emulated shared memory scheme.
\begin{enumerate}
\item allocate scalar variables cyclically over the banks; how are we
going to remember on which bank is what? One more hash-table?
\item also allocate array starting cyclically over the banks; the initial
bank has to be remembered; this is going to make the address equation
more complicated;
\item se'parer clairement le calcul du nombre de copies
a` effectuer du calcul de leur dimension (?!?);
\item estimate the space complexity;
\item improve (!) the tiling decision making procedure;
\item improve the target machine description (see also Complexity and
Bruno's work);
\item \marginpar{CA}
the message lengths (\verb+LS+)
for \verb+WP65_SEND+ and \verb+WP65_RECEIVE+ should be
expressed in bytes instead of number of elements;
the underlying library has no way to guess the data type used;
(fixed by CA in \verb+build_sc_machine.c+ \&
\verb+tiling.c+, January 1992)
\item in the target machine description, the bank width (i.e. the
line size \verb+LS+) should be expressed in bytes;
(have to be fixed in each model.rc)
\item \verb+model.rc+ should be stored as \verb+model-rc.tex+ in
\verb+Documentation+ and derived automatically as for other
\verb+xxx.rc+ files
\item initial loop indices should be replaced by expressions and not
by local loop indices when they do not appear in subscript
expressions; example: \verb+MAT(I,J)= I + J+; general
expression:
\begin{quote}
\verb/I = P I_0 + O + L_I/
\end{quote}
\item parallelism over the channel requests cannot be expressed in
Fortran~3L; it's OK with sockets because IO's are buffered;
\item make WP65 code validation easier; add a call to generic (?) init
and validate procedures?
\item \verb+load code+ should be based on local declarations:
\begin{enumerate}
\item multiple references are almost always reads;
\item local declarations cannot contains elements that
do not exist in the emulated shared memory; to be proved,
using the fact that arrays are rectangular;
\item see with Corinne if the global problem cannot be broken
into pieces;
\item use \verb+sc_proj_test+ and other test routines to see
if thinks would be easier when the bank number or the processor
number are numerically known;
\end{enumerate}
\item \marginpar{LZ}
\verb+REAL+ variables are generated with type \verb+REAL*8+
instead of \verb+REAL*4+
\item \marginpar{CA} temporary local variables like loop indices are declared
more than once; (fixed by CA, january 92)
\end{enumerate}
\section{Wpips}
\begin{enumerate}
\item no problem with X-events sent to a stopped wpips at Rice nor at
ICASE; we should compile with XView3 available on font (waits
until /usr/lib/libc.so.1.6 is available: required for XView3);
\item do not create any database before the user have selected the
fortran files. This way, should no more core dump after: create
workspace in a directory without .f file, change directory,
create workspace again.
\item {\tt wpips\_user\_error} debugged (RK, January 1993)
\item numerous interface bugs corrected (RK, FI, January 1993)
\end{enumerate}
\subsection{Wishes:}
\begin{enumerate}
\item simpler module selection: direct input from the main window;
selection in any window (e.g. from a displayed call graph
or from some displayed code); first solution: the module
selection window can stay as long as you want;
\item try dynamic linking to reduce startup time
\item give a new exit choice to avoid storing lots of stuff in the
database; done;
\item new entry point like \verb+user_error+ to display information or query
the user; suppose we write a \verb+check_call_sites+ (as we should have
according to PIPS contract): how are we going to display binding
errors (parameter number, type, in/out mismatches)? How are we
going to display the problem location?
\item Parascope has a special {\tt snapshot} button to include screen
views in papers; also available here (xwd -frame | xpr -device
ps | lpr -Pnec); sa. xselection
\item delay calls to pipsmake() when a user selects a rule; he may change
his mind and resources may be unnecessarily destroyed; done by BB;
\item do not print {\em closing module} when nothing is done; thanks!
\item add a copyright notification (for instance an extra-window displayed
while wpips is dynamically linked...)
\item no documentation available yet; needed for ICS'91 in June 91;
\item \verb+char * top_level_query_user(char * prompt)+ needed for
interactive phases (prettyprint dependences, unrolling, etc...);
\end{enumerate}
\subsection{Modifications}
\begin{enumerate}
\item new buttons have been added in the menu for icfg, dependence
graph and various transformations (FI, January 1993)
\end{enumerate}
\end{document}
\end