The CPPC Project

• Zero-blocks (blocks of memory of 4112 bytes each containing only zeros) are now automatically compressed during runtime.
• Fixed bug that caused the generation of incorrect incremental checkpoints during a restart.
• Fixed several bugs in the Fortran compiler.
• Modified the control flow analysis to follow the dependence graph. This should improve registration places for global variables.
• Removed debug output from the compiler and the runtime system.

General

• Added incremental checkpointing support.
• Greatly reduced memory footprint during checkpoint file creation.
• Multithreading is always used and is no longer a configuration option.
• Fixed bug with open files (FILE*) during restart.
• Removed support for HDF5 versions older than 1.8.
• Added support for OpenMPI v1.4.
• The library now requires OpenSSL, used for memory hashing.

General

• Added support for HDF-5 v1.8.
• General bugfixes to the build system to improve portability.

Compiler

• Added support for empty strings in Fortran 77.
• Added support for the C "putenv()" function.
• Fixed the calculation of cost heuristics for if statements.
• Fixed a bug related to heuristic analysis of loop statements.

Runtime

• Fixed bug related to the insertion of non-portable statements inside loops.

General

• Integration with GNU Autotools. This should reduce compilation and portability issues to a minimum.

Compiler

• Improved support for automatic checkpoint insertion and communication analysis. The analyses should run significantly faster, too.
• The number of processes that will execute the parallel code can now be supplied to the compiler via the "-nodes" command-line parameter. This improves the communication analysis.
• MPI_Waitall() now supported by the compiler.
• Added support for COMPLEX Fortran 77 datatypes.
• Added semantic support for the following Fortran 77 built-in functions:
  • ALOG
  • ALOG10
  • AMAX1
  • AMIN1
  • DACOS
  • DASIN
  • DCMPLX
  • DCONJG
  • DCOS
  • DEXP
  • DIMAG
  • DSIN
  • FSEEK
  • MAX0
  • RAND
  • SRAND
• Added semantic info for the following Linux functions:
  • ctime
  • fgets
  • localtime
  • memcpy
  • memset
  • realloc
  • scanf
  • sscanf
  • strftime
  • strstr
  • toupper
• Added semantic info for the following MPI functions:
  • MPI_Abort
  • MPI_Error_string

Library

• The code should now compile without errors on a wider range of C++ compilers, including GCC v4.3 (which introduces FSEEK support in gfortran).

General

• The parser now allows function calls with no parameters to be made using the "CALL NAME()" format, in addition to "CALL NAME".
• Several miscellaneous bugfixes.
• Slightly improved build system. The integration with the GNU Autotools is not yet complete.

Compiler

• cppc_cc now accepts multiple C files as input.
• Added semantic support for the following Linux functions:
  • cosh
  • exp
  • feof
  • fflush
  • fgetc
  • fread
  • isunordered
  • log
  • pow
  • sinh
  • sleep
  • sqrt
  • srand48
  • strcat
  • strcmp
  • strcpy
  • strncat
  • system
  • time
• Fixed bug that could cause a ClassCastException when analyzing C switch statements.
• Added compiler support for escape literals (e.g. '\0').
• Fixed bug that could cause live arrays not to be registered if they were passed to a function that modified some, but not all, of their elements.
• Added support for Fortran computed gotos.
• Addressed bug present in some JDKs that could cause a NumberFormatException when parsing integer literals beginning with "+".
• Fixed bug that caused the compiler to perform incorrect transformations when dealing with file-opening functions using expressions to represent the file path.
• Fixed bug that would raise a NullPointerException when trying to register a C global variable, defined outside of any procedure scope.
• Fixed bug that could raise a ClassCastException when compiling Fortran 77 applications with substantially different definitions of the same COMMON block.

Runtime

• Fixed code generation routines and HDF-5 plugin bugs which could make the re-execution flow enter an infinite loop.
• Fixed bug which could cause files not to be correctly reopened upon application restart.
• Fixed bug which could cause incorrect recovery on big endian machines.

General

• Added preliminary support for user-defined datatypes. Currently, typedefs renaming basic types are recognized.

Compiler

• Changed Cetus infrastructure to version 0.5. This improves the C compiler.
• Improved appearance of Fortran generated code.

Parser

• Fixed a bug in fortran-cppc-modules.xml which could cause an exception in initializer error.
• Added support for computed goto's to the Fortran Compiler.
• Added support for generating DO WHILE...END DO structures in Fortran (which were accepted by the frontend but not correctly written by the backend).

Analysis

• Improved data flow analyses to avoid registering arrays which are fully generated before being used (via pointer assignment in C, or a loop matching their size in both C and Fortran).
• Improved data flow analyses to avoid registering full a whole common block when only some of its variables are live.
• Improved global variables management to avoid double registrations, which could cause a segmentation fault at runtime.
• Corrected bug in data flow analyses which could cause a variable generated inside a conditional to be considered generated outside the conditional block (which could cause necessary registers not to be inserted).

Compile time

• Fixed bug which could cause the library to try to delete /usr/local when running "make install".
• Fixed bug which could cause errors when linking with the HDF5 shared library.

Runtime

• Fixed minor memory leaks.
• Added user-friendly messages regarding errors when trying to restart an application.

Parser

• Added support for numeric literals starting with "+" (e.g., +1.00) to the Fortran Compiler.
• Corrected bug which could insert labels in empty lines when generating Fortran 77 code.
• Corrected placement of the INCLUDE directives, which could cause compilation errors due to code being preparsed before the IMPLICIT statement.

Analysis

• Added semantic information for the Fortran 77 version of the MPI_RECV function.
• Added a check for duplicate entry points (multiple "main" methods), a situation that could cause the compiler to malfunction.
• Improved variable registration routines when checkpoints are placed in a subroutine different from the "main" procedure. The number of registers inserted might grow up a bit, but the set of registered variables is much more reliable.
• Added support for data flow analysis of pre-increment (++var) and pre-decrement (--var) C operators.
• Added support for data flow analysis of "break;" and "switch( ... )" C statements.
• Fixed a bug that could cause a null pointer exception upon analysis of C "for" loops with complex conditions/steps.
• Further removed debugging output.

Compile time

• Fixed a Makefile bug which could cause interface compilation to complain about missing definitions.

Runtime

• Corrected a bug that could cause an Internal Error on certain Java VMs.
• Tuned the memory reserved by the Java VM to avoid heap overflow exceptions and/or long runtimes when parsing large codes (over around 5,000 lines of code). This dramatically improves compile time when dealing with such applications.
• Fixed a bug which could cause the library to enter an infinite loop when restarting an application if the MPI_Init function call was placed in a non-main subroutine.

Parser

• Automatic insertion of execute ... end execute directives introduced. The compiler uses semantic information to detect functions with non portable outcome, and logs their parameters to achieve an exact copy of the function call upon application restart. Note that this "parameter logging" is a "by value" version of the variable registration, which works "by reference" (storing the address for value dumping upon checkpoint instead of directly storing its value).
• The compiler code has been tuned to achieve a much better efficiency. Execution times reduced a 97%.
• Added support for Fortran DO WHILE .... END DO loops.
• Added support for "double" literals to the Fortran compiler (i.e., literals like "1.0d0" were translated as "1.0e0" in previous versions which caused wrong results in some applications).
• Added support for ENDIF and ENDDO terminations.
• Code generation improved, avoiding the insertion of conditional jumps and jump labels between adjacent blocks of code.

Analysis

• Improved register detection routine for correctly detecting variables needed when a "parent" procedure makes a call to another one which contains the checkpoint directive. Variables needed upon return of the "children" function are now registered.
• Improved register detection routine for correctly recognizing variable values generated at variable initialization.
• Improved register creation routine to avoid placing registers inside loops, improving execution overhead.
• Improved the analysis for insertion of the CPPC initialization routines. Now the library should correctly deal with procedures not having their comm init (e.g., MPI_Init) at the beginning of the entry procedure.
• Improved the file tracking analyses to check the status of a file opening if the checkpointed application checks it.
• The Fortran compiler will now automatically place an "include" declaration for the CPPC headers on relevant subroutines.
• Removed debugging output to allow for a more "user-friendly" execution.
• Corrected bug that could cause an IllegalArgumentException to be thrown when inserting the shutdown directive.
• Corrected bug that could cause wrong conditional jump/jump label placement.
• Corrected bug that could cause wrong declaration statement placement.

Library

• The library now uses execution context tracing to associate registers and checkpoints. This system supports complex function call sequences (including any level of recursive calls) and is designed to add a very low overhead and only when a call contains CPPC code. Calls to functions not CPPC-related are not traced.
• The library interface code automatically detects the mangling scheme used by the Fortran compiler, avoiding linking errors due to symbols not being found.
• Corrected bug that could cause problems when removing aliasing registers.
• Removed binary writer from the distribution, since it didn't have any reasonable advantage over the provided HDF-5 version.

Runtime

• The Fortran compiler (cppc_fc) now supports -I <dir> (adds <dir> to the include search path) and -o <dir> (places the output files in directory <dir>).

• Adapted runtime library code to work with GCC 4

• A portable system and format for storing pointers is now used in the runtime library. It detects and supports aliasing between registered variables. This format allows for a recovery that preserves the original data overlap.
• Improved library routines and structures for dealing with open files in a portable way.
• Removed the synchronization barrier when calling CPPC_Shutdown().
• Added code for supporting semantic analysis based on information provided through XML descriptions of functions.
• Added semantic analysis for automatically detecting and dealing with open files in C and Fortran 77.
• Added semantic analysis for automatically inserting the init directive.
• Added syntactic analysis for automatically inserting the shutdown directive.
• General improvements to Fortran 77 code structure.
• The precompiler has now an option to begin its execution using the 'jdb' debugger.
• The Fortran 77 precompiler does now recognize the 'REAL' intrinsic.
• Removed meaningless and syntactically incorrect unregister directives.
• The precompiler does no longer insert empty RECs before execute or checkpoint directives when no variables need to be registered.
• The precompiler does no longer insert an empty label in the end of annotated procedures in Fortran 77.
• Corrected a bug that made the runtime library overwrite the checkpoint file used for a restart when another state file was created.
• Corrected a bug that could cause data corruption on state files depending on the size of the data being stored.
• Corrected live-variable analysis when a checkpoint is placed inside a loop.
• Corrected an issue with jump indexing when automatically inserting code in Fortran 77.
• Corrected misplacement of CPPC_IERROR_VAR when the init directive was the first Statement in a Fortran 77 code.
• Corrected a bug when dealing with arrays of strings in Fortran 77.

• Added a plain configuration file parser. Xerces is not a constraint now.
• Enabled automatic analysis in the precompiler.
• Automatic registration of global variables now works.
• Corrected minor bug in the Fortran 77 parser when recognising Fortran 77 built-in functions.
• Corrected bug with character arrays in the C Variable-size-analyzer in the precompiler.
• Corrected bug in the precompiler which could cause a SIGSEGV under certain circumstances.

• First public release