You can get a printable version of this document on
http://www.cri.ensmp.fr/pips/tpips-user-manual.htdoc/tpips-user-manual.pdf and a HTML version on http://www.cri.ensmp.fr/pips/tpips-user-manual.htdoc.

Contents

1 Introduction

tpips is the line interface and scripting language of the PIPS project. All PIPS functionalities are available, with on-line help and automatic completion.

tpips is more efficient than the pips based scripts (Init, Display,...) because the database is not open, read, written and closed for each command. tpips is less user-friendly than wpips and epips¹ , although it provides the unique opportunity to apply a transformation or display some information on several modules with one command, using pipsmake metavariables, %CALLERS and %CALLEES, or environment user defined variables such as $MODULES (see Example 4.2). The jpips future user interface will provide these pips/tpips functionalities too.

tpips is recommended for large benchmarks and experiments, when analysis and optimization results do not require interactive working. tpips is used for the validation of each new PIPS version.

tpips can be used to automatically replay a manual wpips or tpips session by using logfile_to_tpips. The logfile_to_tpips command derives a tpips script from a logfile in a workspace (see Example 4.2). A PIPS beginner should start with a window interface, wpips or epips² , and not with tpips. But the simple tools of section 10 can be useful for simple tasks.

Access to Unix Shell and to PIPS properties is provided by tpips. PIPS properties can be modified from tpips , and only from tpips during a PIPS session (see Example 4.3). On the other hand, wpips reads at the beginning of the session the default properties and the new ones (re)defined in the properties.rc file, but no change can be done afterwards. Resetting properties during a session can be used to rebuild an application when some modules have been updated by PIPS, e.g. after partial evaluation.

The tpips commands are listed on-line by the help command. Possible arguments are completed or suggested by automatic completion (TAB and TAB/TAB). A command may be spread on several lines by guarding the line feed characters by backslashes.

2 Syntax

Different functions can be activated with the tpips commands. We decompose these functions in those related to the workspace, the analysis and transformation requests, the PIPS properties and options and the the environment. Functions and their arguments are presented in the next subsections.

2.1 Workspace

The commands for workspaces are: create a new workspace (see pips command Init), open an existing workspace, close the current workspace, and delete the current workspace or a specific one (see pips command Delete).

These commands can be used several times during a session or within a tpips script file.

create <workspace-name> <file-name>

creates and opens a new workspace from a source file — or set of source files

open <workspace-name>

opens an existing workspace

close <workspace-name>

closes an opened workspace

delete <workspace-name>

deletes a closed workspace

quit

quits tpips

2.2 Requests

To apply a transformation or request a particular resource, use the following tpips commands.

For a list of the available resources and their descriptions, have a look to http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc or http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc/pipsmake-rc.pdf.

apply <rule-descriptor>

applies a specified rule and produces the associated resources.

capply <rule-descriptor>

applies concurrently a specified rule on the different OWNERs of the <rule-descriptor> and produces the associated resources (see Example 4.4).

display <resource-descriptor>

makes a printable resource (if it does not exist) and prints it out. A printed resource is a ACSII file with the _FILE extension.

make <resource-descriptor>

makes a resource according to default rules

Generic example:

A resource, i.e. any piece of information computed by PIPS, can be required by command:

or

which use interproceduraly and recursively default rules (see pips command Build). Or it is produced by a specific rule:

which will require all its resource arguments by recursively calling make (see pips command Perform).

The owner of a resource is the module it is attached to. By default, the owner is assumed to be the current module.

File resources can be displayed on screen:

as with the pips Display script.

2.3 PIPS properties and options

For a list of the available properties and their descriptions, have a look to http://www.cri.ensmp.fr/pips/properties-rc.htdoc or http://www.cri.ensmp.fr/pips/properties-rc.htdoc/properties-rc.pdf.

activate <rule-name>

activates a rule. Several rules can be used to produce a resource. At any time one of these rules is the default rule.

module <module-name>

specifies the current module to work on.

getproperty <property-name>

prints the value of a PIPS property.

setproperty <property-name> <value>

sets the value of a property — after type check. Possible types are: boolean, int and string.

checkpoint <number>

checkpoints the current workspace every <number> pipsmake operations.

Generic example:

A specific current module is selected as current module with:

When several rules, i.e. functions, can be used to compute the same resource, for instance a parallel version of a module, the rule selection command is:

It is equivalent to the pips Select command.

2.4 Environment

The tpips Unix-related commands are: cd, setenv, getenv, echo, and shell.

shell <sh-command> <parameters>

Executes the line as a shell command. With no arguments, run an inferior shell;

source <filenames...>

reads tpips commands from files;

cd <directory-name>

changes the default directory;

echo <string>

prints <string>;

getenv <variable-name>

prints the value of <variable-name> (equivalent to echo ${<variable-name>});

setenv <variable-name> = <value>

sets <value> to <variable-name>;

help <command>

prints out general help or detailed help about an item;

info <name>

print information about <name>. Here <name> can be the module, modules, workspace or directory words.

2.5 Parameters

The parameters of the different commands are :

<file-name>

Unix path and file name.

Shell syntax, such as *.f, can be used when s option is specified or TPIPS_IS_A_SHELL variable is set;

<workspace-name>

Unix file name — no path;

<module-name>

Fortran function name in capital letters or a C function name;

<rule-name>

rule name (see Section 5);

<resource-descriptor>

It is a resource[OWNER] expression: a pipsmake resource name to be computed for each module associated to OWNER (see Sections 6 and 7);

<rule-descriptor>

It is a rulename[OWNER] expression: a pipsmake instantiated rule to be applied on each module associated to OWNER (see Section 5);

the OWNER

— used in the previous <rule-descriptor> and <resource-descriptor> — can take these formats:

module-name

the target module name;

%PROGRAM

the whole program (for global resource);

%MAIN

the main module of the program;

%MODULE

the current module;

${MODULES}

the modules associated to the Shell variable MODULES. The different names are SPACE or/and TAB separated;

%ALL

All workspace modules, including C compilation units³ ;

%ALLFUNC

All workspace functions, excluding C compilation units;

%CALLEES

All the callees of the current selected module;

%CALLERS

All the callers of the current selected module.

If no OWNER is specified, then the current module is assumed. For generic owners, a sign (%) has been added to differentiate them from existing module names and environment variables.

3 Simple session

A simple tpips session is made of a few steps. First, a workspace must be created:

and a subdirectory of the current directory, here work.database, is created to store and retrieve code and information computed by PIPS about subroutines and functions found in source1.f or source2.f.

Then, you have to select a module, let say FOO, by typing:

unless there is only one function or subroutine in the Fortran file(s) which were passed to the create command, in which case it is automatically selected.

To keep it simple the first time, you should then try to display the sequential code of routine FOO by typing:

Other resource names let you visualize the parallelized code (PARALLELPRINTED_FILE), or the call graph (CALLGRAPH_FILE)... Names for printable resources usually contain the string FILE.

The standard command to leave tpips is quit but a control-D or end-of-file condition are valid too.

Interactive analysis of module MATMUL from source code pgmsrc.f, in workspace pgm:

4 Examples

4.1 Call Graph

To visualize the call graph of the application, the following tpips session can be executed. The call graph of each application routine is displayed.

This example presents a use of:

• environment variable: setenv PIPS_SRCPATH SRC
• shell path file name expression that will be expanded: *.f,
• a <resource-descriptor> : CALLGRAPH_FILE[%ALL]

This tpips example is useful to test PIPS parsing on each new benchmark.

4.2 Dependences

A programmer that would like to see the routines dependence graph to parallelize interprocedurally the application can executed the following tpips commands.

This example presents some PIPS analyses and use of

• module selection: module TEST
• <rule-name>: MUST_REGIONS, REGION_CHAINS
• logfile_to_tpips command

To store this simple session in a tpips file, use the command:

that generates the following tpips commands in the file foo.tpips.

ote that expression %ALL has been expanded by all program routines, in that example there is only one TEST routine in essai.f.

4.3 Transformations

The following transformations are designed for application optimization. Specialization and code reduction eliminating useless instructions and declarations are performed.

This example presents some PIPS transformations and use of

• program restructuration, cloning, dead code elimination, useless declaration elimination, and code regeneration;
• pips property: PREPROCESSOR_MISSING_FILE_HANDLING "generate";
• transformation application : apply CLONE_ON_ARGUMENT[FUNCT].
• <rule-name> : PRINT_CODE...
• <rule-descriptor>: SUPPRESS_DEAD_CODE[${modules}], UNSPLIT[%PROGRAM]...
• <resource-descriptor> : PRINTED_FILE[FUNCT], CALLGRAPH_FILE[%ALL]...

4.4 Analyses

This example presents some classical PIPS analyzes:

• computation of preconditions;
• transformers;
• effects;
• regions

and a concurrent transformation application.

capply PARTIAL_EVAL[%ALL] applies a partial evaluation on all routines concurrently without any verification process between two transformations. The partial evaluation of a routine does not modify other routine predicates, even if the code (one routine) has been modified by the transformation. It is not necessary to ask for a sequential application that will impose the analysis and re-computation of all routine predicates between two partial evaluations.

4.5 Graphical Graph View

This example presents a call to a graphical call graph view.

4.6 Parallelization

This example presents some classical analyses and transformations designed to parallelize benchmarks, for instance the adm program of the Perfect club.

5 Rules — Pips options

Rules are used to compute resources. Several rules can be activated to compute a single resource. For instance a parallel version of a module may be computed from the interprocedural precondition (PRECONDITIONS_INTER_FULL), interprocedural transformer (TRANSFORMERS_INTER_FULL) and accurate dependencies (RICE_SEMANTICS_DEPENDENCE_GRAPH). Rules are selected by activate or selected and applied by apply and capply. More rules and more information on each particular rule are presented in the pipsmake documentation, http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc or http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc/pipsmake-rc.pdf.

6 Resources

The names of the current useful resources are given. These resources are computed by the make and display commands. These non-printable resources are encoded in internal data structures. The corresponding printable resources are listed in the following section. These resources and others are detailed the pipsmake documentation, http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc or http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc/pipsmake-rc.pdf.

7 Printable resources

The printable resources usually contain the string FILE. They can be computed by invoking the display tpips command. They are stored as human readable ASCII file in the workspace. The current resources which are printable follow, for more information see the pipsmake documentation, http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc or http://www.cri.ensmp.fr/pips/pipsmake-rc.htdoc/pipsmake-rc.pdf. .

8 Options

Usage:
tpips [-nscvh?jw] [-l logfile ] [-r rcfile ] [-e tpips-cmds ] tpips-scripts

-n

no execution mode. Just check the script syntax;

-s

behaves like a shell. tpips commands simply extend a shell;

-c

behaves like a command, not a shell (it is the default option);

-v

displays version and architecture information;

-h or -?

provides some help;

-j

jpips special mode;

-w

starts with a wrapper (jpips special again)...

-l

<logfile>: log to <logfile>;

-r

<rcfile>: source the <rcfile> file (default $HOME/.tpipsrc);

-e

tpips-cmds: executes the <tpips-cmds> commands.

9 Environment

Before using tpips, you need to add a PIPS root directory to your path and to set some PIPS environment variables. In order to do that, you can source from the PIPS distribution the shell script pipsrc.sh for any sh compatible shell, e.g. ksh or bash, or pipsrc.csh for any csh compatible shell, e.g. tcsh.

10 Small tools

tpips is the way to go with PIPS but it may be too complex to do some simple tasks used for some demonstrations or some well defined tasks.

10.1 pips_c2openmp

This tools parallelizes the C source files given in parameters and leaves the OpenMP generated files in the locally created database.

It is close to an automation of the example at section 4.6, but for C.

10.2 pips_f2openmp

This tools parallelizes the Fortran source files given in parameters and leaves the OpenMP generated files in the locally created database.

It is close to an automation of the example at section 4.6.