LIBPQTYPES CHANGE LOG
---------------------

Initially, libpqtypes was a libpq patch named (PQparam/PGparam) and can
be found on the -hackers and -patches mailing lists.  This change log
begins at the point libpqtypes was put on pgfoundry as its own project.

To gain access to the "(NOT RELEASED)" code, do a cvs checkout.

Jun 27, 2013 v1.5.1
- fixed issue in spec.c where the type name was not being lower cased
  when it was required.
- Moved AUTHORS, ChangeLog, INSTALL, LICENSE and README to EXTRA_DIST
  - Jozef Mlich
- Modified configure.ac and Makefile.am to properly do SO versioning.
  - Jozef Mlich
- error.c uses the GCC contructor attribute so it can create the the
  TLS key for each thread's error structure.  For whatever reason, 
  calling the constructor function 'my_init' was causing it to never
  get called?!?  Renamed it and all is well.  This becomes apparent
  when libpqtypes is linked with apps using pthread_key_create.
- Removed -no-undefined from LDFLAGS in Makefile.am.  This proved to be
  annoying.
- Updated groff2html so it no longer includes an inline style for
  pre tag width.
- Updated pqt-handlers.3, the PGtypeHandler structure was out of date.
- pqt-composites.3 man page had a bug when registering the simple and complex
  composite.  It was only indicating 1 type instead of 2.

Apr 11, 2011 v1.5.0
- Bumped version to 1.5.0 due to many API removals and additions.
- Added -avoid-version and -no-undefined to Makefile.am.
- Added PQinitTypes to replace PQtypesRegister.  The new function
  adds no new behavior, just a name change.
- Deprecated PQtypesRegister, use PQinitTypes in its place.
- PQclearTypes was added to allow one to clear all registered type handlers.
- Added an internal API pqt_cleartypes since clearing the type data
  is now called from PQclearTypes and pqt_eventproc.
- Added a PQclearSpecs function that allows clearing all prepared specifiers
  in a single call.  Currently, specs can be cleared one at a time by
  setting PQspecPrepare's 3rd 'format' argument to NULL.  This is useful
  but makes clearing all specs difficult because it is not always known
  what specs have been prepared by all modules of an application.
- Removed "automatic" re-registering of types during PQresetXXX.  This
  fails miserably during asynchronous resets.  Replaced with the
  folowing call sequence: PQresetXXX, PQclearTypes, PQregisterTypes.
  NOTE: It is not required to re-register type handlers after a PQresetXXX.
    It is only useful if there is concern that registered types may have
    gone stale.
  WARNING: If an application is reliant upon automatic re-registration of
    type handlers during a PQresetXXX, you will have to update your
    code to manually do a PQclearTypes followed by PQregisterTypes just
    after the reset call.
- Removed the orig_xxx members and regtype from PGtypeHandler structure.
  These members were used by the reset event which was yanked.
- Finally 86'd PQregisterTypeHandler, which has been deprecated since
  v1.4.0, around 18 months as of Apr 2011.
- Updated copyright date from 2009 to 2011, affected source code and
  man pages.
- Updated AUTHORS file to include a list of maintainers and some verbage.
- Deprecated PQregisterSubClasses, PQregisterComposites and
  PQregisterUserDefinedTypes in favor of PQregisterTypes.  This new
  function takes a 'which' argument that indicates if PQT_SUBCLASS,
  PQT_COMPOSITE or PQT_USERDEFINED types are being registered.  This
  function provides full support for asynchronous type registration via
  a simple 'async' boolean argument.  The typical PQconsumeInput, PQisBusy
  and PQgetResult should be used to obtain a result.  After a result is
  obtained, the new PQregisterResult function needs to be called
  to complete the registration.
- Added PQregisterResult which registers the types found within
  the result set.  Added to support asynchronous type registration and
  to avoid type lookups for every new connection by caching the PGresult
  type data for use with this function.
- Add PQregisterTypes.3 man page and added it to the make system install.
- Add PQregisterResult.3 man page and added it to the make system install.
- Removed referneces in documentation and code examples to PQregisterSubClasses,
  PQregisterComposites and PQregisterUserDefinedTypes.  Replaced with
  new PQregisterTypes call.
- Regression test now uses PQregisterTypes.  NOTE: no regression test
  for PQregisterResult or asynchronous type registration yet.
- Marked PQregisterSubClasses.3, PQregisterComposites.3 and
  PQregisterUserDefinedTypes.3 man pages as deprecated.  They are still
  installed with a `make install` and will be until removed.
- Added PQparamDup, which given a PGparam will make an exact duplicate.
  This is useful in cases where you want to queue qeuries to execute
  at a later time, like a connection pooler.  The problem is PGparamCreate
  requires a connection object, which may not be available when attempting
  to enqueue a query with its PGparam object.  Instead, a PGparam object
  can used for the sole purpose of creating duplicates while there are
  no available PGconn objects.
- Updated PGparamCreate man page to include documentation about PGparamDup.

Feb 26, 2010 v1.4.3
- Fixed bug in handler.c. The type lookup query was matching typename
  its array based on the array name being '_' plus the typename. This
  was slower and unsafe.

Oct 03, 2010 v1.4.2
- Fixed bug in events.c.  When clearing a type handler list the 'max' count
  (amount allocated) was not set to zero.  This caused a crash during
  a PQreset call.
- Updated PQgetf documentation, clarified how to handle errors when
  getting data types that require memory management: PGresult and PGarray.
- Moved libpq-events.h out of libpqtypes.h and into libpqtypes-int.h.
  Added libpq-fe to libpqtypes.h.

Apr 06, 2010 v1.4.1
- Fix memory leak in events.c during PGEVT_CONNRESET.  Not a big deal as the
  leak occurs within an out-of-memory error case, so the process is most
  likely toast.  Either way, leak has been plugged.
- pqt_allowsptr was returning FALSE if the type's schema was not provided.
  It now returns TRUE as it should.
- Fixed how missing types are handled during register functions.  They were
  not properly indicating errors in all cases.

Nov 10, 2009 v1.4.0
- Added PQgetvf to compliment PQgetf.  PQgetf is now a light wrapper
  PQgetvf.  This did not change the behavior of PQgetf in any way.
- Fixed bug in IPv6 regression_test.  When AF_INET6 is available, the
  'hints' parameter to getaddrinfo should be supplied with the ai_family
  set to AF_INET6.
- 63% Performance improvement to the type handling duplication which occurs
  whenever a result is copied or a PGparam is created.  Previously, the
  type handlers contained allocated strings for the identifier names:
  like type and schema name.  When duplicating the handlers, the strdup
  cost added up to a noticeable amount of time (mostly seen during
  PQgetf(a_composite_or_array) which duplicates the handlers when generating
  a new PGresult).  These strings were changed to fixed length buffers
  65 bytes wide; 64 max char identifier plus terminating NUL.  The names
  no longer need to be deep copied, they are instead copied as part of a
  larger memcpy.  Additionally, the attDescs array now uses a fixed length
  buffer when the number of compsoite attrs is <= 16 (common case).  When
  larger, allocation is used.  'freeAttDescs' flag was added to the
  PGtypeHandler structure to indicate if the attDescs requires freeing.
  In total, these changes made type handling duplication 63% faster.
  NOTE:
    This increased the size of PGtypeHandler by around 434 bytes.  The
    size of PGrecordAttDesc increased by 65 bytes.  This change modified
    2 public structures so please ensure your libpqtypes.h is current.
- Split PQregisterTypeHandler into three different functions:
  PQregisterComposites, PQregisterUserDefinedTypes and PQregisterSubClasses.
  This was done for two reasons: needed the ability to register multiple
  types at the same time, avoid round trip issues, and simplify the interface.
- PQregisterTypeHandler is now deprecated and should no longer be used.  It
  exists only for compatibility reasons.  It is a wrapper to
  PQregisterComposites, PQregisterUserDefinedTypes and PQregisterSubClasses.
- The PGEVT_REGISTER event was not be handled properly by libpqtypes.  It
  neglected to initialize the internal type formatting information.
- Error case in pqt_parse() was accessing a NULL pointer.
- Bug in network.c when getting text inet/cidr types.  Apparently, windows
  requires a few hints about the address family when using getaddrinfo:
  both IPv4 and IPv6.  This must be a recent Microsoft winsock update as
  this code used to work.

Sep 15, 2009 v1.3.4
- In v1.2b, a caching system for type spec format strings was introduced.
  PQputf and PQgetf would cache the last spec format seen, like "%int4".
  This avoided constant type handling lookups when using arrays, since
  one must put/get every element.  This version expands on this concept
  by adding an additional API call: PQspecPrepare.  The application can
  now prepare an unlimited number of spec format strings.
  NOTE:
    This change does not effect the behavior of putf or getf in older versions.
    Although, some libpqtypes code may run slower since type spec format
    strings are no longer automatically cached.  If you have loops calling
    putf or getf with the same spec format string, typically arrays, you
    would probably benefit from PQspecPrepare.
- Added a PQsendf() and PQsendvf(), same as PQexecf and PQexecvf except they
  are asynchronous versions.  Created manual pages for these functions as well.
- Fixed a bug in PQexecf and PQexevf documentation, the prototypes were
  documented as returning an int rather than a PGresult*.
- Fixed spelling error in PQspecPrepare.3 man page and added PQsendf and
  PQsendvf to the list of supported functions.
- Fixed memory leak in the error handling code.  Multi-threaded windows
  builds were not cleaning up the internal error structure in error.c.
  This is because windows has no automatic cleanup for TLS storage.  The
  only solution was to use fixed length buffers and truncate unusually
  long error messages; including error fields.

May 29, 2009 v1.3.3
- Added PQexecf and PQexecvf which allows one to pack parameters and
  execute a command all in one statement.  Example:
    PQexecf(conn, "SELECT %int4 + %int4", 1, 1);
- Updated the PQputf.3 man page execf documentation which stated that
  execf should be implemented by the application ... not libpqtypes.
  Apparently, we no longer feel that way :)
- Fixed bug in datetime.c that accessed index 7 of a 7 byte array, ooppps.
  Thanks to GCC version 4.3.2 for catching this one:
    "src/datetime.c:930: warning: array subscript is above array bounds"

Apr 27, 2009 v1.3.2
- Slight change in the behavior of PQputf and PQputvf.  These functions
  will now silently auto-fill unset composite fields with NULLs.  This only
  pertains to the last N fields: if there are 10 fields in a compsoite
  and only 7 are put, the last 3 fields will be auto-filled with SQL NULLs.
  Previously, an error was raised indicating the client and server have
  different ideas about the number of fields in the composite.  That error
  is still raised if the number of params exceeds the number of composite
  fields.  Auto-filling only occurs when the number of params falls short.
- User-defined CFLAGS were being lost due to a dumb assigment at the
  top of the script.  It now properly appends flags to the ones supplied
  by the user.
- Updated PQgetf man page to include a warning about getting arrays and
  composites.  If getf fails after an array or composite was retrieved,
  the PGresult representing the array/composite must still be cleared.
  The solution is to get arrays/composites by themselves or to always
  make them the last field in the getf call (the latter only works
  if there is only one array or composite in the field list).
- Fixed SEGFAULT when getting a NULL compsoite array item.

Feb 11, 2009 v1.3.1
- PQgetf of an empty array was always returning a result with a single
  tuple (array item), eventhough the PGarray members were correctly set
  for an empty array.  This has been fixed.
- Added empty array test to regression-test.c.
- getaddrinfo is missing on Windows 2000 and earlier versions.  Wspiapi.h
  and ws2tcpip.h must be included to provide a replacement.  libpqtypes-int.h
  neglected to include Wspiapi.h, which it now does.
  Microsoft docs on issue: http://support.microsoft.com/kb/955045

Feb 03, 2009 v1.3.0
- The internal put_int2 and put_int4 type handlers were passing a va_arg
  call as an argument to the pqt_buf_putint2 and pqt_buf_putint4 macros.
  This corrupted the value being put on UnixWare 7.1.4.  The return
  of va_arg is now stored in a local variable that is passed to
  one of the putintX macros.
- Renamed PGinet sa_len member since it conflicts with some systems.
  For instance: SCO UnixWare defines a sa_len macro in sys/socket.h
  #define sa_len sa_un.sa_s.sa_saus_len
  Renamed it to sa_buf_len and changed all references.
- network.c had an issue with struct sockaddr not being zero'd on AIX.
  This caused all inet/cidr tests to fail.  Update regresion-test.c
  with similar fixes.
- AI_NUMERICHOST defined if not available in network.c.  If getaddrinfo
  fails with EAI_BADFLAGS, its tried again w/o numeric host.
- AIX defaults to an unsigned char, which broke "char" type handling.
  Changed PGchar typedef to signed char and added the GCC option
  -fsigned-char to configure.ac.
- Changed the return type of PQgeterror from a const char* to a char*.
  Having it be const served no purpose.
- Added PQparamCount which returns the number of parameters in a PGparam.
  Added PQparamCount.3 man page and updated Makefile.
- Made errFields implementation more memory efficient by allocating all
  fields as a single reusable buffer.
- Reformatted code to match PostgreSQL Project.  This changed every
  source file so take note when looking at CVS diffs.
- Defined HAVE_VSNPRINTF in port.c when compiling for mingw or cygwin.
- Fixed man page errors in PQputf and pqt-handlers.
- Updated INSTALL file to include the 'install' and uninstall' targets.
- Makefile.am (unix make) now installs PQgetErrorField man page.
- Update copyright notices to reflect 2009: source files, LICENSE
  and man pages.
- Removed typeargs state flag PUT_STATE_FREEOUT from param.c since it
  is no longer used.

Dec 03, 2008 v1.2e
- Added support for PQresultErrorField as PQgetErrorField.  PQparamExec and
  PQparamExecPrepared clear resutls in a failure status, thus removing
  the abillity to get error fields.  PQgetErrorField can now be used and
  it doesn't require a result.

Dec 02, 2008 v1.2d
- Fixed bug when putting a NULL array item; %null or %pqt.null.  array.c
  was not adding in the 4 byte length when the length's value was NULL_LEN.
  This resulted in "ERROR:  insufficient data left in message" errors from
  the server during a PQparamExec() call.
- Added a %null test in regression-test.c for composites and arrays.

Nov 18, 2008 v1.2c
- removed GCC __thread modifier from error.c.  Only pthread TLS keys
  are now used.  The TLS key for the error system is now intialized
  with the GCC function attribute "constructor"; which replaced _init().
  This GCC attribute has been around since GCC version 2.7.0, old enough.
  The error system used to intialize the TLS key via pthread_once(),
  but this function is broken on all version of Solaris prior to 10
  (its actually a stub function the returns 0).  This means libpqtypes
  versions prior to this one cannot use thread-safe mode on Solaris 9 or
  below.  Windows continues to use __declspec(thread).
- The PATH geo type was completely broken (SIGBUS) on many RISC processors,
  we tested a couple of UltraSparcs, Itanium, PA-RISC and MIPS.  This is
  due to memory alignment issues (boy, spoiled by x86!).  A configure
  check was added to detect when strict memory alignment is required
  (STRICT_MEMORY_ALIGNMENT) and the source is toggled appropriately.
- Add pqt_buf_xxx functions to abstract reading and writing data
  to/from buffers.  Another memory alignment change.
- fixed regression-test.c on hpux which was referencing unknown macros:
  LLONG_MAX and LLONG_MIN.  HPUX uses LONG_LONG_MAX.
- cleaned up a few things in Makefile.am (unix make file)
- win32.mak for MSVC had an issue with include variable.  created INC2
  to solve this.

Nov 11, 2008 v1.2b
- Added a type specifier caching system.  Whenever putf or getf are
  called, it compares the specifier string with the last one libpqtypes
  saw.  If it is a match, the parsing and type handler lookup
  stage can be bipassed, giving a 25% performance increase for arrays
  and large result sets.  This also makes libpqtypes noticably faster
  than using libpq in text mode ... PQexec.
- preprocessing bug in port.c, #elif HAVE_VSNPRINTF instead of
  #elif defined(HAVE_VSNPRINTF).
- windows was completely broken from a change in v1.2a.  Apparantley
  windows has different memory addresses for DLL exported functions;
  the address seen outside the DLL and the address seen from within the DLL.
  This broke the PQtypesRegister macro that was referencing PQtypesEventProc,
  since the external address was being used to register the event proc
  but the internal address was being used by libpqtypes.  This problem was
  solved by converting PQtypesRegister to a function, which forces the
  event proc to always be referenced from within libpqtypes.
  PQtypesEventProc was completely removed form the public interface.
- Updated make system to be more configurable, allows setting things
  like CC, CFLAGS, etc.. at the prompt (see INSTALL).

Sept 26, 2008 v1.2a
- libpqtypes uses the libpq event system, an approaved patch for 8.4.
  The object hooks patch was redesigned and than renamed to libpq events.
- bug fix in timestamp code
- memory leak fix in error system, didn't clean up thread memory.
- pqytpe's PGEventProc is now a public function named PQtypesEventProc

May 19, 2008 v1.0c
- Bug fix in datetime.c.  When performing a time value conversion, 1-based
  and 0-based month values were being used/confused.
- libpqtypes using objecthooks, a proposed patch for libpq allowing
  outside apps, like libpqtypes, to hook into libpq.

May 09, 2008 v1.0b
- Added copyright notice to all source files.
- Updated several man pages.
- Added an include for stdarg.h to libpqtypes.h
- Updated makefile and groff2html

April 28, 2008 v1.0
- libpqtypes project was created on pgfoundry
- make system was put into place (mingw, cygwin and msvc support)