Removed arg passing from frontend to backend functions.

[ragel.git] / ChangeLog

Ragel 6.3 - Aug 29, 2008
========================
 -Fixed an assertion that is too strong. In the condition code we need to copy
  transitions that have non-empty lmActionTable arrays so we don't assert
  emptiness in the constructor. Lift out the assertion and copy the array in
  the constructor.
 -Fixed and improved multiple include prevention. We now track the entire
  include history of a parser state to prevent duplicates.
 -Fixed crash on failed lookup of goto/call/etc target.

Ragel 6.2 - May 9, 2008
=======================
 -Bug fix: The lm_switch actions needs to set p from tokend when there is no
  user action.
 -Bug fix: when not using indicies we can't use a transitions's id to identify
  the eof transition to take. Instead add the transition to the end of the
  transition list and store its position in a new var called pos. The pos var
  is then used as the index.
 -Bug fix: an fnext followed by an fbreak in -G2 was not working. The fbreak
  was not aware that the fnext causes the cs variable to be forced active. In
  this case fbreak does not need to save cs because it is already current.
 -Bug fix: need to compute the low and high character-space keys from the
  condition-trans overlap when computing an expansion. Can't use the range
  supplied from the condition overlap since they may not match. An incorrect
  machine that accepted 1(!cond1, !cond2) was generated for the following
  grammar. This bug was reported by Tim Chklovski.
      c = 2 @matched_c;
      sc1 = 1..2 when cond1;
      sc2 = 1..2 when cond2;
      main := sc1 | c | sc2;
 -Bug fix: error messages in start label analysis of join operations were
  causing assertion failures because location info was not set. Fixed by
  adding locations.
 -Include and import file searching now searches for the file name given based
  on the location of the current file, not ragel's current path.
  Additional search locations can be given using the -I option.
 -Rubinius code generation was updated to the latest Rubinius. Patch from Evan
  Phoenix.
 -Switched from strcasecmp to strcmp for testing long arguments.
 -Applied a patch from Andrei Polushin for setting the error message format.
     --error-format=gnu (default)
     --error-fromat=msvc
 -Now using the _WIN32 define instead of _WIN32. Other MSVC compilation
  improvments from Andrei Polushin.
 -Added the hyperref package to the manual. 

Ragel 6.1 - Mar 26, 2008
========================
 -Scanners now ensure that any leaving actions at the end of a pattern are
  executed. They are always executed before the pattern action.
 -Added an option -d for turning off the removal of duplicate actions from
  actions lists.
 -Need to unset the final state status of the start state in kleene star if it
  is set. It is possible to crash ragel when the warning is ignored.
 -In the dot file generation we need to print any actions that are in
  State::eofTrans. These come from scanners only.
 -Use @docdir@ for the docdir Makefile variable.
 -Check for ar and ranlib in the configure script.

Ragel 6.0 - Jan 12, 2008
========================
 -Removed the 'noend' write option from examples/atoi.rl. This example is
  referenced a lot as a first example and as such it shouldn't contain a
  special purpose write option like 'noend'.
 -Introcuded the "eof" variable for indicating the end of file. The p variable
  is checked against eof when the processing loop reaches the end of a block.
  If p == eof at this time then the EOF actions are executed. The variable is
  required only when EOF actions have been emebedded.
 -The "write eof" command is no longer needed and was removed.
 -Scanners now use EOF actions to generate tokens. This eliminates the need to
  flush the last token.
 -Restructured the Java driver; a switch statement with fallthrough cases are
  now used to emulate gotos.
 -Ruby code generation was also restructured. Gotos are elmulated using a
  series of if tests. 
 -Went back to 3.X semantics for >, % and error actions. The > operator also
  embeds a leaving action/priority into the start state if it is final. If EOF
  happens in a state with a leaving operator then the leaving action is
  executed. If EOF happens in a non-final state that has an error action, the
  error action is executed.
 -The "ragel" program now executes frontend and backend processes separately,
  connecting them with a temporary file in the current directory. Without the
  -x option the "ragel" program marshals arguments and calls the frontend and
  backend. With the -x option the "ragel" program acts as the frontend only.
 -Added name finding for executables. If any forward slash is found in argv0
  then it is assumed that the path is explicit and the path to the backend
  executable should be derived from that. Whe check that location and also go
  up one then inside a directory of the same name in case we are executing
  from the source tree. If no forward slash is found it is assumed the file is
  being run from the installed location. The PREFIX supplied during
  configuration is used.
 -On windows GetModuleFileNameEx is used to find out where the the current
  process's binary is. That location is searched first. If that fails then we
  go up one directory and look for the executable inside a directory of the
  same name in case we are executing from the source tree.
 -Changed the -l option in rlgen-cd to -L because it is covered in the
  frontend. Added a passthrough in the frontend for the backend options.
 -Dot file generation can now be invoked using the -V option to ragel. We
  now require an input file. If standard in is used then we don't have a file
  name on which to base the output.
 -Able to build native windows executables using Cygwin+MinGW. 
 -Patch from David Waite: Large arrays are now created by copying in the data
  from smaller arrays using System.arraycopy(). This eliminates the debug data
  associated with explicit initialization statements. It is also much easier
  on the java compiler which can run out of memory compiling very large
  machines. The downside is that it takes slightly longer to initialize static
  data at run time.
 -The fbreak statement now advances p.
 -In the :> :>> and <: operators it was possible for the priority assignment
  to be bypassed via the zero length string. In :> this was fixed
  automatically with the semantics change to the entering priority operator.
  If the start state is final it now embeds a leaving action into it,
  preventing persistance through the zero length string. In :>> and <: this
  was fixed explicitly. With <: the entering priority operator was used and
  with :> a special exception was added. Since it uses the finishing
  transition operator it also adds a leaving priority to the start state if it
  is final.
 -Ranlib is now run on the archives. Patch from Kenny MacDermid.
 -The case statement syntax in ruby code generation used a form depreciated in
  Ruby 1.9. Updated it.
 -Made a number of fixes that eliminate warnings in GCC 4.3. Mostly concern
  the now depreciate automatic conversion of string contsants to "char*" type.
  Other fixes include adding parenthesis around && within ||.
 -The "tokstart" and "tokend" variables were changed to "ts" and "te".

Ragel 5.25 - Dec 24, 2007
=========================
 -Fixed segfault reported by Ryan Phelps. Affected Java and Ruby code
  generation. The dataExpr variable was not initialized.
 -Fixed incorrect case label in test/runtests. Caused Objective-C tests to be
  ignored.
 -Added missing include to common.cpp.

Ragel 5.24 - Sep 16, 2007
=========================
 -Applied patch from Victor Hugo Borja <vic@rubyforge.org>. This patch
  implements -T1 -F0 -F1 and -G0 in the ruby code generator. Goto-driven code
  generation is experimental and requires rubinius asm directives (specify
  with --rbx option). These code generators pass all the ruby tests.
 -If the condition embedding code runs out of available characters in the
  keyspace an error message is emitted.
 -The first example that appeared in the manual used the special-purpose
  'noend' write option. This caused confusion. Now a basic example appears
  first.
 -Added two new statements: prepush and postpop. These are code blocks that
  are written out during call and return statements. The prepush code is
  written immediately before pushing the current state to the state stack
  during a call. The postpop code is written immediately after popping the
  current state during return. These can be used to implement a dynamically
  resizable stack. 

Ragel 5.23 - Jul 24, 2007
=========================
 -Eliminated the use of callcc as an alternative to goto. Instead, the named
  breaks implementation used in the Java code generator is imitated using
  control flow variables.
 -Improved the error message given when there is a write statement but no
  machine instantiations and hence no state machine.
 -Documentation improvements: updates to "Machine Instantiation", "Write Init"
  and "Write Exports" sectons. Added the "Variables Used by Ragel" section.
 -Renamed "Entering Actions" to "Starting Actions."
 -Other documentation updates.

Ragel 5.22 - June 14, 2007
==========================
 -Bug fix: need to isolate the start state of a scanner before setting the
  to-state and from-state actions which clear and set tokstart. This affected
  very simple scanners only. Most scanners have an isolated start state due to
  the pattern structure.
 -Bug fix: when -S or -M was given the ragel version number was not emitted,
  causing the backend to reject the intermediate format. From Tim Potter.
 -The p varialbe is now set up at the beginning of a scanner action, rather
  than at the end. This leaves scanner actions free to manipulate p and
  removes the need for the special holdTE and execTE (TE for tokend) versions
  of hold and exec. It also removes the need to set p = tokend-1 immediately
  before any control flow. We loose the ability to determine where in the
  input stream a scanner action is executed, however this information is of
  little use because it is primarily an artifact of the scanner implementation
  (sometimes the last char, other times later on). The gains of this change
  are consistency and simplicity.
 -The "data" variable (used in Java and Ruby code generation only) can now be
  overridden using the variable statement.

Ragel 5.21 - May 9, 2007
========================
 -Fixed an inconsistency in the value of p following an error. In the C
  directly executable code (rlgen-cd -G2) p is left at the character where
  the error occurred, which is correct. In all other code generators it was
  left at the following character. This was fixed. Now in all code generators
  p is left at the character where the error occurred.
 -Bug fix: when fhold was used in scanner pattern actions which get executed
  on the last character of the pattern (pattern matches which do not require
  any lookahead), fhold was modifying p instead of tokend. This was fixed and
  the patact.rl test was modified to cover the case.
 -Fixed typos in the guide, improved the state action embedding operator
  section and added subsections on the variable, import, and export
  statements.
 -Implemented a better solution than the pri hack for resolving the '-'
  ambiguity: force a shortest match of term.
 -Fixed bugs in the binary searching for condition keys in both the Ruby and
  Java code generation. 
 -Can now embed the negative sense of a condition. Added a language-
  independent test case for this feature and the necessary transformation
  support.
 -Added new condition embedding syntax:
    expr inwhen cond   - The transitions into the machine (starting transitions).
    expr outwhen cond  - The pending transitions out of the machine.
 -The argument to the variable statement which affects the name of the current
  state variable was changed from "curstate" to "cs" (the default name used
  for the current state)
 -Implemented the other variables names in the variable statement. Now all
  variables (p, pe, cs, top, stack, act, tokstart, tokend) can be renamed.
 -Parse errors in the intermediate XML file now cause the backend to exit
  immediately rather then forge on. The recovery infrastructure isn't there
  and segfaults are likely.
 -When no input is given to the backend program, it should not print an error
  message, it should just return a non-zero exit status. The assumption is
  that the frontend printed an error.
 -The version number is now included in the intermediate file. An error is
  emitted if there is a mismatch.
 -The alphabet type is now communicated from the frontend to the backend using
  a one-word internal name instead of an array offset.
 -The Ruby host language types had been just copied from Java. Reduced them to
  two basic types: char and int, both signed with the usual C sizes.

Ragel 5.20 - Apr 7, 2007
========================
 -The cs variable is now always initialized, unless the "nocs" option is given
  to the write init command. If there is no main machine, cs is initialized to
  the entry point defined by the last machine instantiation.
 -A number of fixes were made to the Ruby code generator.
 -The frontend now scans ruby comments and regular expressions.
 -A transformation for Ruby was added to the language-independent test suite.
  The Ruby code generator passes on all the language-independent tests.
 -A new Ruby test and two language-independent tests were added.
 -Some portability fixes were made (Patches from Josef Goettgens and Aaron
  Campbell).
 -Fixed a make dependency bug which caused a problem for parallel building
  (Patch from Jeremy Hinegardner).

Ragel 5.19 - Mar 14, 2007
=========================
 -Added an import statement to ragel. This statement takes a literal string as
  an argument, interprets it as a file name, then scrapes the file for
  sequences of tokens that match the following forms. Tokens inside ragel
  sections are ignored. An example is in test/import1.rl
    name = number
    name = lit_string
    "define" name number
    "define" name lit_string
 -Added an export mechanism which writes defines for single character machines
  that have been tagged with the export keyword in their definition. Defines
  are used for C, ints for D, Java and Ruby. Examples of the export feature
  are in test/export*.rl.
 -All machine instantiations are now always generated, even if they are not
  referenced. In the backend, entry points for all instantiations are written
  out alongside start, error and first final states.
 -If the main machine is not present then do not emit an error. Generate the
  machine without a start state and do not initialize cs in the write init
  code.
 -Added an option -l to rlgen-cd which inhibits the writing of #line
  directives.
 -Added a new syntax for verbose embeddings. This adds parentheses: 
      $from(action_name);
  Verbose embeddings without parentheses can make code difficult to read
  because they force a space in the middle of an action embedding. There is a
  tendency to associtate spaces with concatenation. Without syntax
  highlighting to make it clear that the embedding type is a keyword, the
  problem is especially bad. The danger is that a verbose embedding could be
  read as an embedding of the keyword representing the empbedding type. With
  parentheses, verbose embeddings read much more clearly.
 -Conditions now have a forced order when more than one is executed on a
  single character. Previously ordering relied on pointers, which caused
  results to vary by compiler. Ordering is now done using conditon action
  declaration order. This fixes the failure of cond4.rl which occured with
  g++ 4.1 and other compiler versions.
 -In the port from flex to ragel, the name separator :: in Ragel code was
  lost. Added it back.
 -In the examples directory switched from rlcodegen to rlgen-cd. Silenced a
  warning in statechart.rl.
 -In the root makefile the distclean target was fixed. It was calling clean in
  the subdirs. In docs, the clean target was not deleting the new manpages for
  the rlgen-* programs. Fixed.
 -Portability and other fixes from Josef Goettgens were applied.
 -The @datadir@ and @mandir@ variables are made use of in doc/Makefile.in for
  specifying where documentation should be installed. Patch from Marcus
  Rueckert. 

Ragel 5.18 - Feb 13, 2007
=========================
 -There is now a 100% correspondence between state id numbers in the
  intermediate XML file, Graphviz dot files and generated code. This was
  achieved by moving code which determines if the error state is necessary
  into the frontend, and then assigning state numbers before writing out the
  intermediate file.
 -Backened class structure was reorganized to make it easier to add new code
  generators without having to also modify the existing code generators.
 -The C and D code generation executable was changed to rlgen-cd.
 -The Java code generation was split out into it's own exectuable (rlgen-java)
  to allow it to freely diverge from the C/D-based code generation.
 -The graphviz dot file generation was also split out to it's own executable
  (rlgen-dot).
 -The Ruby code generation patch from Victor Hugo Borja was added. This is
  highly experimental code and is not yet completely functional. It is in the
  executable rlgen-ruby.
 -The problem with large state machine machines in Java was fixed. This
  problem was discovered by Colin Fleming, who also contributed a patch.
  Rather than specify arrays as comma-separated lists of literals, array
  initialization is now done in a static function. This is the approach used
  by the Java compiler. Unlike the compiler Ragel is careful split large
  initilization functions.
 -The manual was expanded and reorganized somewhat.
 -Eliminated per-example directories in examples/.
 -Made some fixes to the pullscan.rl example.
 -In the frontend CR characters are now treated as whitespace.
 -Updated to the latest aapl. This completely eliminates the shallowCopy
  function. With that, a definitive memory leak is fixed.
 -Control codes with escape sequences are now printable characters (-p
  option). Also, the space character is now printed as SP.
 -Fixed the null dereference and consequential segfault which occurred when
  trying to create empty machines with [] and // and /a[]b/. 
 -Fixed the segfault which occured when a machine reference failed.
 -Discontinuing ragel.spec. It is more appropriate for this to be written by
  package maintenance developers.

Ragel 5.17 - Jan 28, 2007
=========================
 -The scanners and parsers in both the frontend and backend programs were
  completely rewritten using Ragel and Kelbt.
 -The '%when condition' syntax was functioning like '$when condition'. This
  was fixed.
 -In the Vim syntax file fixes to the matching of embedding operators were
  made. Also, improvements to the sync patterns were made.
 -Added pullscan.rl to the examples directory. It is an example of doing
  pull-based scanning. Also, xmlscan.rl in rlcodegen is a pull scanner.
 -The introduction chapter of the manual was improved. The manually-drawn
  figures for the examples were replaced with graphviz-drawn figures.
 
Ragel 5.16 - Nov 20, 2006
=========================
 -Policy change: the fhold and fexec directives did not function correctly in
  scanner pattern actions. In this context manipulations of p may be lost or
  made invalid. In the previous version of Ragel they were banned because of
  this. Instead of banning these directives they have been fixed. The fexec
  and fhold directives now manipulate tokend, which is now always used to
  update p when the action terminates.

Ragel 5.15 - Oct 31, 2006
=========================
 -A language independent test harness was introduced. Test cases can be
  written using a custom mini-language in the embedded actions. This
  mini-language is then translated to C, D and Java when generating the
  language-specific test cases.
 -Several existing tests have been ported to the language-independent format
  and a number of new language-independent test cases have been added.
 -The state-based embedding operators which access states that are not the
  start state and are not final (the 'middle' states) have changed. They
  were:
    <@/    eof action into middle states
    <@!    error action into middle states
    <@^    local error action into middle states
    <@~    to-state action into middle states
    <@*    from-state action into middle states
  They are now:
    <>/    eof action into middle states
    <>!    error action into middle states
    <>^    local error action into middle states
    <>~    to-state action into middle states
    <>*    from-state action into middle states
 -The verbose form of embeddings using the <- operator have been removed.
  This syntax was difficult to remember.
 -A new verbose form of state-based embedding operators have been added.
  These are like the symbol versions, except they replace the symbols:
    /  !  ^  ~  *
  with literal keywords:
    eof err lerr to from
 -The following words have been promoted to keywords:
    when eof err lerr to from
 -The write statment now gets its own lexical scope in the scanner to ensure
  that commands are passed through as is (not affected by keywords).
 -Bug fix: in the code generation of fret in scanner actions the adjustment to
  p that is needed in some cases (dependent on content of patterns) was not
  happening.
 -The fhold directive, which decrements p, cannot be permitted in the pattern
  action of a scanner item because it will not behave consistently. At the end
  of a pattern action p could be decremented, set to a new value or left
  alone. This depends on the contents of the scanner's patterns. The user
  cannot be expected to predict what will happen to p.
 -Conditions in D code require a cast to the widec type when computing widec.
 -Like Java, D code also needs if (true) branches for control flow in actions
  in order to fool the unreachable code detector. This is now abstracted in
  all code generators using the CTRL_FLOW() function.
 -The NULL_ITEM value in java code should be -1. This is needed for
  maintaining tokstart.

Ragel 5.14 - Oct 1, 2006
========================
 -Fixed the check for use of fcall in actions embedded within longest match
  items. It was emitting an error if an item's longest-match action had an
  fcall, which is allowed. This bug was introduced while fixing a segfault in
  version 5.8.
 -A new minimization option was added: MinimizeMostOps (-l). This option
  minimizes at every operation except on chains of expressions and chains of
  terms (eg, union and concat). On these chains it minimizes only at the last
  operation. This makes test cases with many states compile faster, without
  killing the performance on grammars like strings2.rl.
 -The -l minimiziation option was made the default.
 -Fixes to Java code: Use of the fc value did not work, now fixed. Static data
  is now declared with the final keyword. Patch from Colin Fleming. Conditions
  now work when generating Java code.
 -The option -p was added to rlcodegen which causes printable characters to be
  printed in GraphViz output. Patch from Colin Fleming.
 -The "element" keyword no longer exists, removed from vim syntax file.
  Updated keyword highlighting.
 -The host language selection is now made in the frontend.
 -Native host language types are now used when specifying the alphtype.
  Previously all languages used the set defined by C, and these were mapped to
  the appropriate type in the backend.

Ragel 5.13 - Sep 7, 2006
========================
 -Fixed a careless error which broke Java code generation.

Ragel 5.12 - Sep 7, 2006
========================
 -The -o flag did not work in combination with -V. This was fixed.
 -The split code generation format uses only the required number of digits
  when writing out the number in the file name of each part.
 -The -T0, -F0 and -G0 codegens should write out the action list iteration
  variables only when there are regular, to state or from state actions. The
  code gens should not use anyActions().
 -If two states have the same EOF actions, they are written out in the finish
  routine as one case.
 -The split and in-place goto formats would sometimes generate _out when it is
  not needed. This was fixed.
 -Improved the basic partitioning in the split code gen. The last partition
  would sometimes be empty. This was fixed.
 -Use of 'fcall *' was not causing top to be initialized. Fixed.
 -Implemented a Java backend, specified with -J. Only the table-based format
  is supported.
 -Implemented range compression in the frontend. This has no effect on the
  generated code, however it reduces the work of the backend and any programs
  that read the intermediate format.

Ragel 5.11 - Aug 10, 2006
=========================
 -Added a variable to the configure.in script which allows the building of
  the parsers to be turned off (BUILD_PARSERS). Parser building is off by
  default for released versions. 
 -Removed configure tests for bison defines header file. Use --defines=file
  instead. 
 -Configure script doesn't test for bison, flex and gperf when building of the
  parsers is turned off.
 -Removed check for YYLTYPE structure from configure script. Since shipped
  code will not build parsers by default, we don't need to be as accomodating
  of other versions of bison. 
 -Added a missing include that showed up with g++ 2.95.3.
 -Failed configure test for Objective-C compiler is now silent.

Ragel 5.10 - Jul 31, 2006
=========================
 -Moved the check for error state higher in the table-based processing loop.
 -Replaced naive implementations of condition searching with proper ones. In
  the table-based formats the searching is also table-based. In the directly
  executed formats the searching is also directly executable.
 -The minimization process was made aware of conditions.
 -A problem with the condition implementation was fixed. Previously we were
  taking pointers to transitions and then using them after a call to
  outTransCopy, which was a bad idea because they may be changed by the call.
 -Added test mailbox3.rl which is based on mailbox2.rl but includes conditions
  for restricting header and message body lengths.  
 -Eliminated the initial one-character backup of p just before resuming
  execution.
 -Added the -s option to the frontend for printing statistics. This currently
  includes just the number of states. 
 -Sped up the generation of the in-place goto-driven (-G2) code style.
 -Implemented a split version of in-place goto-driven code style. This code
  generation style is suitable for producing fast implementations of very
  large machines. Partitioning is currently naive. In the future a
  high-quality partitioning program will be employed. The flag for accessing
  this feature is -Pn, where n is the number of partitions.
 -Converted mailbox1.rl, strings2.rl and cppscan1.rl tests to support the
  split code generation.
 -Fixes and updates were made to the runtests script: added -c for compiling
  only, changed the -me option to -e, and added support for testing the split
  code style.
 
Ragel 5.9 - Jul 19, 2006
========================
 -Fixed a bug in the include system which caused malformed output from the
  frontend when the include was made from a multi-line machine spec and the
  included file ended in a single line spec (or vice versa).
 -Static data is now const.
 -Actions which referenced states but were not embedded caused the frontend to
  segfault, now fixed.
 -Manual now built with pdflatex.
 -The manual was reorganized and expanded. Chapter sequence is now:
  Introduction, Constructing Machines, Embedding Actions, Controlling
  Nondeterminism and Interfacing to the Host program.

Ragel 5.8 - Jun 17, 2006
========================
 -The internal representation of the alphabet type has been encapsulated
  into a class and all operations on it have been defined as C++ operators.
 -The condition implementation now supports range transitions. This allows
  conditions to be embedded into arbitrary machines. Conditions are still
  exprimental.
 -More condition embedding operators were added
   1. Isolate the start state and embed a condition into all transitions
      leaving it:   
      >when cond   OR   >?cond
   2. Embed a condition into all transitions:
      when cond    OR   $when cond   OR   $?cond
   3. Embed a condition into pending out transitions:
      %when cond   OR   %?cond
 -Improvements were made to the determinization process to support pending out
  conditions.
 -The Vim sytax file was fixed so that :> doesn't cause the match of a label.
 -The test suite was converted to a single-file format which uses less disk
  space than the old directory-per-test format.
 
Ragel 5.7 - May 14, 2006
========================
 -Conditions will not be embedded like actions because they involve a
  manipulation of the state machine they are specified in. They have therefore
  been taken out of the verbose action embedding form (using the <- compound
  symbol). A new syntax for specifying conditions has been created: 
  m = '\n' when {i==4};
 -Fixed a bug which prevented state machine commands like fcurs, fcall, fret,
  etc, from being accounted for in from-state actions and to-state actions.
  This prevented some necessary support code from being generated.
 -Implemented condition testing in remaining code generators.
 -Configure script now checks for gperf, which is required for building.
 -Added support for case-insensitive literal strings (in addition to regexes).
  A case-insensitive string is made by appending an 'i' to the literal, as in
  'cmd'i or "cmd"i.
 -Fixed a bug which caused all or expressions inside of all regular
  expressions to be case-insensitive. For example /[fo]o bar/ would make the
  [fo] part case-insensitive even though no 'i' was given following the
  regular expression.

Ragel 5.6 - Apr 1, 2006
=======================
 -Added a left-guarded concatenation operator. This operator <: is equivalent
  to ( expr1 $1 . expr2 >0 ). It is useful if you want to prefix a sequence
  with a sequence of a subset of the characters it matches. For example, one
  can consume leading whitespace before tokenizing a sequence of whitespace
  separated words: ( ' '* <: ( ' '+ | [a-z]+ )** )
 -Removed context embedding code, which has been dead since 5.0.
  
Ragel 5.5 - Mar 28, 2006
========================
 -Implemented a case-insensitive option for regular expressions: /get/i.
 -If no input file is given to the ragel program it reads from standard input.
 -The label of the start state has been changed from START to IN to save on
  required screen space.
 -Bug fix: \0 was not working in literal strings, due to a change that reduced
  memory usage by concatenating components of literal strings.  Token data
  length is now passed from the scanner to the paser so that we do not need to
  rely on null termination. 

Ragel 5.4 - Mar 12, 2006
========================
 -Eliminated the default transition from the frontend implementation. This
  default transition was a space-saving optimization that at best could reduce
  the number of allocated transitions by one half. Unfortunately it
  complicated the implementation and this stood in the way of introducing
  conditionals. The default transition may be reintroduced in the future.
 -Added entry-guarded concatenation. This operator :>, is syntactic sugar
  for expr1 $0 . expr >1. This operator terminates the matching of the first
  machine when a first character of the second machine is matched. For
  example in any* . ';' we never leave the any* machine. If we use any* :> ';'
  then the any* machine is terminiated upon matching the semi-colon.
 -Added finish-guarded concatenation. This operator :>>, is syntactic sugar
  for expr1 $0 . expr @1. This operator is like entry guarded concatenation
  except the first machine is terminated when the second machine enters a
  final state. This is useful for delaying the guard until a full pattern is
  matched. For example as in '/*' any* :>> '*/'.
 -Added strong subtraction. Where regular subtraction removes from the first
  machine any strings that are matched by the second machine, strong
  subtraction removes any strings from the first that contain any strings of
  the second as a substring. Strong subtraction is syntactic sugar for 
  expr1 - ( any* expr2 any* ).
 -Eliminated the use of priorities from the examples. Replaced with
  subtraction, guarded concatenation and longest-match kleene star.
 -Did some initial work on supporting conditional transitions. Far from
  complete and very buggy. This code will only be active when conditionals are
  used.

Ragel 5.3 - Jan 27, 2006
========================
 -Added missing semi-colons that cause the build to fail when using older
  versions of Bison.
 -Fix for D code: if the contents of an fexec is a single word, the generated
  code will get interpreted as a C-style cast. Adding two brackets prevents
  this. Can now turn eliminate the "access this.;" in cppscan5 that was used to
  get around this problem.
 -Improved some of the tag names in the intermediate format.
 -Added unsigned long to the list of supported alphabet types.
 -Added ids of actions and action lists to XML intermediate format. Makes it
  more human readable.
 -Updated to latest Aapl package.

Ragel 5.2 - Jan 6, 2006
========================
 -Ragel emits an error if the target of fentry, fcall, fgoto or fnext is inside
  a longest match operator, or if an action embedding in a longest match
  machine uses fcall. The fcall command can still be used in pattern actions.
 -Made improvements to the clang, rlscan, awkemu and cppscan examples.
 -Some fixes to generated label names: they should all be prefixed with _.
 -A fix to the Vim syntax highlighting script was made
 -Many fixes and updates to the documentation. All important features and
  concepts are now documented. A second chapter describing Ragel's use
  was added.
 
Ragel 5.1 - Dec 22, 2005
========================
 -Fixes to the matching of section delimiters in Vim syntax file.
 -If there is a longest match machine, the tokend var is now initialized by
  write init. This is not necessary for correct functionality, however
  prevents compiler warnings.
 -The rlscan example was ported to the longest match operator and changed to
  emit XML data.
 -Fix to the error handling in the frontend: if there are errors in the lookup
  of names at machine generation time then do not emit anything.
 -If not compiling the full machine in the frontend (by using -M), avoid
  errors and segfaults caused by names that are not part of the compiled
  machine.
 -Longest match bug fix: need to init tokstart when returing from fsm calls
  that are inside longest match actions.
 -In Graphviz drawing, the arrow into the start state is not a real
  transition, do not draw to-state actions on the label.
 -A bug fix to the handling of non-tag data within an XML tag was made.
 -Backend exit value fixed: since the parser now accepts nothing so as to
  avoid a redundant parse error when the frontend dies, we must force an
  error. The backend should now be properly reporting errors.
 -The longest match machine now has it's start state set final. An LM machine
  is in a final state when it has not matched anything, when it has matched
  and accepted a token and is ready for another, and when it has matched a
  token but is waiting for some lookahead before determining what to do about
  it (similar to kleene star).
 -Element statement removed from some tests.
 -Entry point names are propagated to the backend and used to label the entry
  point arrows in Graphviz output.

Ragel 5.0 - Dec 17, 2005
========================
 (additional details in V5 release notes)
 -Ragel has been split into two executables: A frontend which compiles
  machines and emits them in an XML format, and a backend which generates code
  or a Graphviz dot file from the XML input. The purpose of this split is to
  allow Ragel to interface with other tools by means of the XML intermediate
  format and to reduce complexity by strictly separating the previously
  entangled phases. The intermediate format will provide a better platform
  inspecting compiled machines and for extending Ragel to support other host
  languages.
 -The host language interface has been reduced significantly. Ragel no longer
  expects the machine to be implemented as a structure or class and does not
  generate functions corresponding to initialization, execution and EOF.
  Instead, Ragel just generates the code of these components, allowing all of
  them to be placed in a single function if desired. The user specifies a
  machine in the usual manner, then indicates at which place in the program
  text the state machine code is to be generated. This is done using the write
  statement. It is possible to specify to Ragel how it should access the
  variables it needs (such as the current state) using the access statement.
 -The host language embedding delimiters have been changed. Single line
  machines start with '%%' and end at newline. Multiline machines start with
  '%%{' and end with '}%%'. The machine name is given with the machine
  statement at the very beginning of the specification. This purpose of this
  change is to make it easier separate Ragel code from the host language. This
  will ease the addition of supported host languages.
 -The structure and class parsing which was previously able to extract a
  machine's name has been removed since this feature is dependent on the host
  language and inhibits the move towards a more language-independent frontend.
 -The init, element and interface statements have been made obsolete by the
  new host language interface and have been removed.
 -The fexec action statement has been changed to take only the new position to
  move to. This statement is more useful for moving backwards and reparsing
  input than for specifying a whole new buffer entirely and has been shifted
  to this new use. Giving it only one argument also simplifies the parsing of
  host code embedded in a Ragel specification. This will ease the addition of
  supported host languages.
 -Introduced the fbreak statement, which allows one to stop processing data
  immediately. The machine ends up in the state that the current transition
  was to go to. The current character is not changed.
 -Introduced the noend option for writing the execute code. This inhibits
  checking if we have reached pe. The machine will run until it goes into the
  error state or fbreak is hit. This allows one to parse null-terminate
  strings without first computing the length.
 -The execute code now breaks out of the processing loop when it moves into
  the error state. Previously it would run until pe was hit. Breaking out
  makes the noend option useful when an error is encountered and allows
  user code to determine where in the input the error occured. It also
  eliminates needlessly iterating the input buffer.
 -Introduced the noerror, nofinal and noprefix options for writing the machine
  data. The first two inhibit the writing of the error state and the
  first-final state should they not be needed. The noprefix eliminates the
  prefixing of the data items with the machine name.
 -Support for the D language has been added. This is specified in the backend
  with the -D switch.
 -Since the new host language interface has been reduced considerably, Ragel
  no longer needs to distinguish between C-based languages. Support for C, C++
  and Objective-C has been folded into one option in the backend: -C
 -The code generator has been made independent of the languages that it
  supports by pushing the language dependent apsects down into the lower
  levels of the code generator.
 -Many improvements to the longest match construction were made. It is no
  longer considered experimental. A longest match machine must appear at the
  top level of a machine instantiation. Since it does not generate a pure
  state machine (it may need to backtrack), it cannot be used as an operand to
  other operators.
 -References to the current character and current state are now completely
  banned in EOF actions.

Ragel 4.2 - Sep 16, 2005
========================
 (additional details in V4 release notes)
 -Fixed a bug in the longest match operator. In some states it's possible that
  we either match a token or match nothing at all. In these states we need to
  consult the LmSwitch on error so it must be prepared to execute an error
  handler. We therefore need to init act to this error value (which is zero).
  We can compute if we need to do this and the code generator emits the
  initialization only if necessary. 
 -Changed the definition of the token end of longest match actions. It now
  points to one past the last token. This makes computing the token length
  easier because you don't have to add one. The longest match variables token
  start, action identifier and token end are now properly initialized in
  generated code. They don't need to be initialized in the user's code.
 -Implemented to-state and from-state actions. These actions are executed on
  transitions into the state (after the in transition's actions) and on
  transitions out of the state (before the out transition's actions). See V4
  release notes for more information.
 -Since there are no longer any action embedding operators that embed both on
  transitions and on EOF, any actions that exist in both places will be there
  because the user has explicitly done so. Presuming this case is rare, and
  with code duplication in the hands of the user, we therefore give the EOF
  actions their own action switch in the finish() function. This is further
  motivated by the fact that the best solution is to do the same for to-state
  and from-state actions in the main loop.
 -Longest match actions can now be specified using a named action. Since a
  word following a longest match item conflicts with the concatenation of a
  named machine, the => symbol must come immediately before a named action.
 -The longest match operator permits action and machine definitions in the
  middle of a longest match construction. These are parsed as if they came
  before the machine definition they are contained in. Permitting action and
  machine definitions in a longest match construction allows objects to be
  defined closer to their use.
 -The longest match operator can now handle longest match items with no
  action, where previously Ragel segfaulted.
 -Updated to Aapl post 2.12.
 -Fixed a bug in epsilon transition name lookups. After doing a name lookup
  the result was stored in the parse tree. This is wrong because if a machine
  is used more than once, each time it may resolve to different targets,
  however it will be stored in the same place. We now store name resolutions
  in a separated data structure so that each walk of a parse tree uses the
  name resolved during the corresponding walk in the name lookup pass.
 -The operators used to embed context and actions into states have been
  modified. The V4 release notes contain the full details.
 -Added zlen builtin machine to represent the zero length machine. Eventually
  the name "null" will be phased out in favour of zlen because it is unclear
  whether null matches the zero length string or if it does not match any
  string at all (as does the empty builtin).
 -Added verbose versions of action, context and priority embedding. See the V4
  release notes for the full details. A small example: 
      machine <- all exec { foo(); } <- final eof act1
 -Bugfix for machines with epsilon ops, but no join operations. I had 
  wrongfully assumed that because epsilon ops can only increase connectivity,
  that no states are ever merged and therefore a call to fillInStates() is not
  necessary. In reality, epsilon transitions within one machine can induce the
  merging of states. In the following, state 2 follows two paths on 'i':
      main := 'h' -> i 'i h' i: 'i';
 -Changed the license of the guide from a custom "do not propagate modified
  versions of this document" license to the GPL.

Ragel 4.1 - Jun 26, 2005
========================
 (additional details in V4 release notes)
 -A bug in include processing was fixed. Surrounding code in an include file
  was being passed through to the output when it should be ignored. Includes
  are only for including portions of another machine into he current. This
  went unnoticed because all tested includes were wrapped in #ifndef ...
  #endif directives and so did not affect the compilation of the file making
  the include.
 -Fixes were made to Vim syntax highlighting file.
 -Duplicate actions are now removed from action lists.
 -The character-level negation operator ^ was added. This operator produces a
  machine that matches single characters that are not matched by the machine
  it is applied to. This unary prefix operator has the same precedence level
  as !.
 -The use of + to specify the a positive literal number was discontinued.
 -The parser now assigns the subtraction operator a higher precedence than
  the negation of literal number.
  
Ragel 4.0 - May 26, 2005
========================
 (additional details in V4 release notes)
 -Operators now strictly embed into a machine either on a specific class of
  characters or on EOF, but never both. This gives a cleaner association
  between the operators and the physical state machine entitites they operate
  on. This change is made up of several parts:
  1. '%' operator embeds only into leaving characters.
  2. All global and local error operators only embed on error character
     transitions, their action will not be triggerend on EOF in non-final
     states.
  3. EOF action embedding operators have been added for all classes of states
     to make up for functionality removed from other operators. These are 
     >/ $/ @/ %/.
  4. Start transition operator '>' no longer implicitly embeds into leaving
     transtions when start state is final.
 -Ragel now emits warnings about the improper use of statements and values in
  action code that is embedded as an EOF action. Warnings are emitted for fpc,
  fc, fexec, fbuf and fblen.
 -Added a longest match construction operator |* machine opt-action; ... *|.
  This is for repetition where an ability to revert to a shorter, previously
  matched item is required. This is the same behaviour as flex and re2c. The
  longest match operator is not a pure FSM construction, it introduces
  transitions that implicitly hold the current character or reset execution to
  a previous location in the input. Use of this operator requires the caller
  of the machine to occasionally hold onto data after a call to the exectute
  routine. Use of machines generated with this operator as the input to other
  operators may have undefined results. See examples/cppscan for an example.
  This is very experimental code.
 -Action ids are only assigned to actions that are referenced in the final
  constructed machine, preventing gaps in the action id sequence. Previously
  an action id was assigned if the action was referenced during parsing.
 -Machine specifications now begin with %% and are followed with an optional
  name and either a single Ragel statement or a sequence of statements
  enclosed in {}.
 -Ragel no longer generates the FSM's structure or class. It is up to the user
  to declare the structure and to give it a variable named curs of type
  integer. If the machine uses the call stack the user must also declare a
  array of integers named stack and an integer variable named top.
 -In the case of Objective-C, Ragel no longer generates the interface or
  implementation directives, allowing the user to declare additional methods.
 -If a machine specification does not have a name then Ragel tries to find a
  name for it by first checking if the specification is inside a struct, class
  or interface. If it is not then it uses the name of the previous machine
  specification. If still no name is found then this is an error.
 -Fsm specifications now persist in memory and statements accumulate.
 -Ragel now has an include statement for including the statements of a machine
  spec in another file (perhaps because it is the corresponding header file).
  The include statement can also be used to draw in the statements of another
  fsm spec in the current file.
 -The fstack statement is now obsolete and has been removed.
 -A new statement, simply 'interface;', indicates that ragel should generate
  the machine's interface. If Ragel sees the main machine it generates the
  code sections of the machine. Previously, the header portion was generated
  if the (now removed) struct statement was found and code was generated if
  any machine definition was found.
 -Fixed a bug in the resolution of fsm name references in actions. The name
  resolution code did not recurse into inline code items with children
  (fgoto*, fcall*, fnext*, and fexec), causing a segfault at code generation
  time.
 -Cleaned up the code generators. FsmCodeGen was made into a virtual base
  class allowing for the language/output-style specific classes to inherit
  both a language specific and style-specific base class while retaining only
  one copy of FsmCodeGen. Language specific output can now be moved into the
  language specific code generators, requiring less duplication of code in the
  language/output-style specific leaf classes.
 -Fixed bugs in fcall* implementation of IpgGoto code generation.
 -If the element type has not been defined Ragel now uses a constant version
  of the alphtype, not the exact alphtype. In most cases the data pointer of
  the execute routine should be const. A non-const element type can still be
  defined with the element statement.
 -The fc special value now uses getkey for retrieving the current char rather
  than *_p, which is wrong if the element type is a structure.
 -User guide converted to TeX and updated for new 4.0 syntax and semantics.

Ragel 3.7 - Oct 31, 2004
========================
 -Bug fix: unreferenced machine instantiations causing segfault due to name
  tree and parse tree walk becomming out of syncronization.
 -Rewrote representation of inline code blocks using a tree data structure.
  This allows special keywords such as fbuf to be used as the operatands of
  other fsm commands.
 -Documentation updates.
 -When deciding whether or not to generate machine instantiations, search the
  entire name tree beneath the instantiation for references, not just the
  root.
 -Removed stray ';' in keller2.rl
 -Added fexec for restarting the machine with new buffer data (state stays the
  same), fbuf for retrieving the the start of the buf, and fblen for
  retrieving the orig buffer length.
 -Implemented test/cppscan2 using fexec. This allows token emitting and restart
  to stay inside the execute routine, instead of leaving and re-entering on
  every token.
 -Changed examples/cppscan to use fexec and thereby go much faster.
 -Implemented flex and re2c versions of examples/cppscan. Ragel version
  goes faster than flex version but not as fast as re2c version.
 -Merged in Objective-C patch from Erich Ocean.
 -Turned off syncing with stdio in C++ tests to make them go faster.
 -Renamed C++ code generaion classes with the Cpp Prefix instead of CC to make
  them easier to read.
 -In the finish function emit fbuf as 0 cast to a pointer to the element type
  so it's type is not interpreted as an integer.
 -The number -128 underflows char alphabets on some architectures. Removed
  uses of it in tests.
 -Disabled the keller2 test because it causes problems on many architectures
  due to its large size and compilation requirements.

Ragel 3.6 - Jul 10, 2004
========================
 -Many documentation updates.
 -When resolving names, return a set of values so that a reference in an
  action block that is embedded more than once won't report distinct entry
  points that are actually the same.
 -Implemented flat tables. Stores a linear array of indicies into the
  transition array and only a low and high key value. Faster than binary
  searching for keys but not usable for large alphabets.
 -Fixed bug in deleting of transitions leftover from converstion from bst to
  list implementation of transitions. Other code cleanup.
 -In table based output calculate the cost of using an index. Don't use if
  cheaper. 
 -Changed fstate() value available in init and action code to to fentry() to
  reflect the fact that the values returned are intended to be used as targets
  in fgoto, fnext and fcall statements. The returned state is not a unique
  state representing the label. There can be any number of states representing
  a label.
 -Added keller2 test, C++ scanning tests and C++ scanning example.
 -In table based output split up transitions into targets and actions. This
  allows actions to be omitted.
 -Broke the components of the state array into separate arrays. Requires
  adding some fields where they could previously be omitted, however allows
  finer grained control over the sizes of items and an overal size reduction.
  Also means that state numbers are not an offset into the state array but
  instead a sequence of numbers, meaning the context array does not have any
  wasted bits.
 -Action lists and transition also have their types chosen to be the smallest
  possible for accomodating the contained values.
 -Changed curs state stored in fsm struct from _cs to curs. Keep fsm->curs ==
  -1 while in machine. Added tests curs1 and curs2.
 -Implemented the notion of context. Context can be embedded in states using
  >:, $:, @: and %: operators. These embed a named context into start states,
  all states, non-start/non-final and final states. If the context is declared
  using a context statment
    context name;
  then the context can be quered for any state using fsm_name_ctx_name(state)
  in C code and fsm_name::ctx_name(state) in C++ code. This feature makes it
  possible to determine what "part" of the machine is currently active.
 -Fixed crash on machine generation of graphs with no final state. If there
  is no reference to a final state in a join operation, don't generate one. 
 -Updated Vim sytax: added labels to inline code, added various C++ keywords.
  Don't highlight name separations as labels. Added switch labels, improved
  alphtype, element and getkey.
 -Fixed line info in error reporting of bad epsilon trans. 
 -Fixed fstate() for tab code gen.
 -Removed references to malloc.h.

Ragel 3.5 - May 29, 2004
========================
 -When parse errors occur, the partially generated output file is deleted and
  an non-zero exit status is returned.
 -Updated Vim syntax file.
 -Implemented the setting of the element type that is passed to the execute
  routine as well as method for specifying how ragel should retrive the key
  from the element type. This lets ragel process arbitrary structures inside
  of which is the key that is parsed.
      element struct Element;
      getkey fpc->character;
 -The current state is now implemented with an int across all machines. This
  simplifies working with current state variables. For example this allows a
  call stack to be implemented in user code.
 -Implemented a method for retrieving the current state, the target state, and
  any named states. 
      fcurs           -retrieve the current state
      ftargs          -retrieve the target state
      fstate(name)    -retrieve a named state.
 -Implemented a mechanism for jumping to and calling to a state stored in a
  variable.
      fgoto *<expr>;  -goto the state returned by the C/C++ expression.
      fcall *<expr>;  -call the state returned by the C/C++ expression.
 -Implemented a mechanism for specifying the next state without immediately
  transfering control there (any code following statement is executed).
      fnext label;    -set the state pointed to by label as the next state.
      fnext *<expr>;  -set the state returned by the C/C++ expression as the
                       next.
 -Action references are determined from the final machine instead of during
  the parse tree walk. Some actions can be referenced in the parse tree but not
  show up in the final machine. Machine analysis is now done based on this new
  computation.
 -Named state lookup now employs a breadth-first search in the lookup and
  allows the user to fully qualify names, making it possible to specify
  jumps/calls into parts of the machine deep in the name hierarchy. Each part
  of name (separated by ::) employs a breadth first search from it's starting
  point.
 -Name references now must always refer to a single state. Since references to
  multiple states is not normally intended, it no longer happens
  automatically. This frees the programmer from thinking about whether or not
  a state reference is unique. It also avoids the added complexity of
  determining when to merge the targets of multiple references. The effect of
  references to multiple states can be explicitly created using the join
  operator and epsilon transitions.
 -M option was split into -S and -M. -S specifies the machine spec to generate
  for graphviz output and dumping. -M specifies the machine definition or
  instantiation.
 -Machine function parameters are now prefixed with and underscore to
  avoid the hiding of class members.

Ragel 3.4 - May 8, 2004
=======================
 -Added the longest match kleene star operator **, which is synonymous 
  with ( ( <machine> ) $0 %1 ) *.
 -Epsilon operators distinguish between leaving transitions (going to an
  another expression in a comma separated list) and non-leaving transitions.
  Leaving actions and priorities are appropriately transferred.
 -Relative priority of following ops changed to:
      1. Action/Priority 
      2. Epsilon 
      3. Label
  If label is done first then the isolation of the start state in > operators
  will cause the label to point to the old start state that doesn't have the
  new action/priority.
 -Merged >! and >~, @! and @~, %! and %~, and $! and $~ operators to have one
  set of global error action operators (>!, @!, %! and $!) that are invoked on
  error by unexpected characters as well as by unexepected EOF.
 -Added the fpc keyword for use in action code. This is a pointer to the
  current character. *fpc == fc. If an action is invoked on EOF then fpc == 0.
 -Added >^, @^, %^, and $^ local error operators. Global error operators (>!,
  @!, $!, and %!) cause actions to be invoked if the final machine fails.
  Local error actions cause actions to be invoked if if the current machine
  fails.
 -Changed error operators to mean embed global/local error actions in:
     >! and !^  -the start state.
     @! and @^  -states that are not the start state and are not final.
     %! and %^  -final states.
     $! and $^  -all states.
 -Added >@! which is synonymous >! then @!
 -Added >@^ which is synonymous >^ then @^
 -Added @%! which is synonymous @! then %!
 -Added @%^ which is synonymous >^ then @^
 -FsmGraph representation of transition lists was changed from a mapping of
  alphabet key -> transition objects using a BST to simply a list of
  transition objects. Since the transitions are no longer divided by
  single/range, the fast finding of transition objects by key is no longer
  required functionality and can be eliminated. This new implementation uses
  the same amount of memory however causes less allocations. It also make more
  sense for supporting error transitions with actions. Previously an error
  transition was represented by a null value in the BST.
 -Regular expression ranges are checked to ensure that lower <= upper.
 -Added printf-like example.
 -Added atoi2, erract2, and gotcallret to the test suite.
 -Improved build test to support make -jN and simplified the compiling and
  running of tests.

Ragel 3.3 - Mar 7, 2004
=======================
 -Portability bug fixes were made. Minimum and maximum integer values are
  now taken from the system. An alignment problem on 64bit systems
  was fixed.

Ragel 3.2 - Feb 28, 2004
========================
 -Added a Vim syntax file.
 -Eliminated length var from generated execute code in favour of an end
  pointer. Using length requires two variables be read and written. Using an
  end pointer requires one variable read and written and one read. Results in
  more optimizable code.
 -Minimization is now on by default.
 -States are ordered in output by depth first search.
 -Bug in minimization fixed. States were not being distinguished based on
  error actions. 
 -Added null and empty builtin machines.
 -Added EOF error action operators. These are >~, >@, $~, and %~. EOF error
  operators embed actions to take if the EOF is seen and interpreted as an
  error. The operators correspond to the following states:
    -the start state
    -any state with a transition to a final state
    -any state with a transiion out
    -a final state
 -Fixed bug in generation of unreference machine vars using -M. Unreferenced
  vars don't have a name tree built underneath when starting from
  instantiations. Need to instead build the name tree starting at the var.
 -Calls, returns, holds and references to fc in out action code are now
  handled for ipgoto output.
 -Only actions referenced by an instantiated machine expression are put into
  the action index and written out.
 -Added rlscan, an example that lexes Ragel input.

Ragel 3.1 - Feb 18, 2004
========================
 -Duplicates in OR literals are removed and no longer cause an assertion
  failure.
 -Duplicate entry points used in goto and call statements are made into
  deterministic entry points.
 -Base FsmGraph code moved from aapl into ragel, as an increasing amount
  of specialization is required. Too much time was spent attempting to
  keep it as a general purpose template.
 -FsmGraph code de-templatized and heirarchy squashed to a single class.
 -Single transitions taken out of FsmGraph code. In the machine construction
  stage, transitions are now implemented only with ranges and default
  transtions. This reduces memory consumption, simplifies code and prevents
  covered transitions. However it requires the automated selection of single
  transitions to keep goto-driven code lean.
 -Machine reduction completely rewritten to be in-place. As duplicate
  transitions and actions are found and the machine is converted to a format
  suitable for writing as C code or as GraphViz input, the memory allocated
  for states and transitions is reused, instead of newly allocated.
 -New reduction code consolodates ranges, selects a default transition, and
  selects single transitions with the goal of joining ranges that are split by
  any number of single characters.
 -Line directive changed from "# <num> <file>" to the more common format
  "#line <num> <file>".
 -Operator :! changed to @!. This should have happened in last release.
 -Added params example.

Ragel 3.0 - Jan 22, 2004
========================
 -Ragel now parses the contents of struct statements and action code. 
 -The keyword fc replaces the use of *p to reference the current character in
  action code.
 -Machine instantiations other than main are allowed.
 -Call, jump and return statements are now available in action code. This
  facility makes it possible to jump to an error handling machine, call a
  sub-machine for parsing a field or to follow paths through a machine as
  determined by arbitrary C code.
 -Added labels to the language. Labels can be used anywhere in a machine
  expression to define an entry point. Also references to machine definitions
  cause the implicit creation of a label.
 -Added epsilon transitions to the language. Epsilon operators may reference
  labels in the current name scope resolved when join operators are evaluated
  and at the root of the expression tree of machine assignment/instantiation.
 -Added the comma operator, which joins machines together without drawing any
  transitions between them. This operator is useful in combination with
  labels, the epsilon operator and user code transitions for defining machines
  using the named state and transition list paradigm. It is also useful for
  invoking transitions based on some analysis of the input or on the
  environment.
 -Added >!, :!, $!, %! operators for specifying actions to take should the
  machine fail. These operators embed actions to execute if the machine
  fails in
    -the start state
    -any state with a transition to a final state
    -any state with a transiion out
    -a final state
  The general rule is that if an action embedding operator embeds an action
  into a set of transitions T, then the error-counterpart with a !  embeds an
  action into the error transition taken when any transition T is a candidate,
  but does not match the input.
 -The finishing augmentation operator ':' has been changed to '@'. This
  frees the ':' symbol for machine labels and avoids hacks to the parser to
  allow the use of ':' for both labels and finishing augmentations. The best
  hack required that label names be distinct from machine definition names as
  in main := word : word; This restriction is not good because labels are
  local to the machine that they are used in whereas machine names are global
  entities. Label name choices should not be restricted by the set of names
  that are in use for machines.
 -Named priority syntax now requires parenthesis surrounding the name and
  value pair. This avoids grammar ambiguities now that the ',' operator has
  been introduced and makes it more clear that the name and value are an
  asscociated pair.
 -Backslashes are escaped in line directive paths.

Ragel 2.2 - Oct 6, 2003
=======================
 -Added {n}, {,n}, {n,} {n,m} repetition operators.
    <expr> {n}    -- exactly n repetitions
    <expr> {,n}   -- zero to n repetitions
    <expr> {n,}   -- n or more repetitions
    <expr> {n,m}  -- n to m repetitions
 -Bug in binary search table in Aapl fixed. Fixes crashing on machines that
  add to action tables that are implicitly shared among transitions.
 -Tests using obsolete minimization algorithms are no longer built and run by
  default.
 -Added atoi and concurrent from examples to the test suite.

Ragel 2.1 - Sep 22, 2003
========================
 -Bug in priority comparison code fixed. Segfaulted on some input with many
  embedded priorities.
 -Added two new examples.

Ragel 2.0 - Sep 7, 2003
=======================
 -Optional (?), One or More (+) and Kleene Star (*) operators changed from
  prefix to postfix. Rationale is that postfix version is far more common in
  regular expression implementations and will be more readily understood.
 -All priority values attached to transitions are now accompanied by a name.
  Transitions no longer have default priority values of zero assigned
  to them. Only transitions that have different priority values assigned
  to the same name influence the NFA-DFA conversion. This scheme reduces
  side-effects of priorities.
 -Removed the %! statement for unsetting pending out priorities. With
  named priorities, it is not necessary to clear the priorities of a
  machine with $0 %! because non-colliding names can be used to avoid
  side-effects.
 -Removed the clear keyword, which was for removing actions from a machine.
  Not required functionality and it is non-intuitive to have a language
  feature that undoes previous definitions.
 -Removed the ^ modifier to repetition and concatenation operators. This
  undocumented feature prevented out transitions and out priorities from being
  transfered from final states to transitions leaving machines. Not required
  functionality and complicates the language unnecessarily.
 -Keyword 'func' changed to 'action' as a part of the phasing out of the term
  'function' in favour of 'action'. Rationale is that the term 'function'
  implies that the code is called like a C function, which is not necessarily
  the case. The term 'action' is far more common in state machine compiler
  implementations.
 -Added the instantiation statement, which looks like a standard variable
  assignment except := is used instead of =. Instantiations go into the
  same graph dictionary as definitions. In the the future, instantiations
  will be used as the target for gotos and calls in action code.
 -The main graph should now be explicitly instantiated. If it is not,
  a warning is issued.
 -Or literal basic machines ([] outside of regular expressions) now support
  negation and ranges. 
 -C and C++ interfaces lowercased. In the C interface an underscore now
  separates the fsm machine and the function name. Rationale is that lowercased
  library and generated routines are more common.
    C output:
      int fsm_init( struct clang *fsm );
      int fsm_execute( struct clang *fsm, char *data, int dlen );
      int fsm_finish( struct clang *fsm );
    C++ output:
      int fsm::init( );
      int fsm::execute( char *data, int dlen );
      int fsm::finish( );
 -Init, execute and finish all return -1 if the machine is in the error state
  and can never accept, 0 if the machine is in a non-accepting state that has a
  path to a final state and 1 if the machine is in an accepting state.
 -Accept routine eliminated. Determining whether or not the machine accepts is
  done by examining the return value of the finish routine.
 -In C output, fsm structure is no longer a typedef, so referencing requires
  the struct keyword. This is to stay in line with C language conventions.
 -In C++ output, constructor is no longer written by ragel. As a consequence,
  init routine is not called automatically. Allows constructor to be supplied
  by user as well as the return value of init to be examined without calling it
  twice.
 -Static start state and private structures are taken out of C++ classes.

Ragel 1.5.4 - Jul 14, 2003
==========================
 -Workaround for building with bison 1.875, which produces an
  optimization that doesn't build with newer version gcc.

Ragel 1.5.3 - Jul 10, 2003
==========================
 -Fixed building with versions of flex that recognize YY_NO_UNPUT.
 -Fixed version numbers in ragel.spec file.

Ragel 1.5.2 - Jul 7, 2003
=========================
 -Transition actions and out actions displayed in the graphviz output.
 -Transitions on negative numbers handled in graphviz output.
 -Warning generated when using bison 1.875 now squashed.
 
Ragel 1.5.1 - Jun 21, 2003
==========================
 -Bugs fixed: Don't delete the output objects when writing to standard out.
  Copy mem into parser buffer with memcpy, not strcpy. Fixes buffer mem errror.
 -Fixes for compiling with Sun WorkShop 6 compilers.

Ragel 1.5.0 - Jun 10, 2003
==========================
 -Line directives written to the output so that errors in the action code
  are properly reported in the ragel input file.
 -Simple graphviz dot file output format is supported. Shows states and
  transitions. Does not yet show actions.
 -Options -p and -f dropped in favour of -d output format.
 -Added option -M for specifying the machine to dump with -d or the graph to
  generate with -V.
 -Error recovery implemented.
 -Proper line and column number tracking implemented in the scanner.
 -All action/function code is now embedded in the main Execute routine. Avoids
  duplication of action code in the Finish routine and the need to call
  ExecFuncs which resulted in huge code bloat. Will also allow actions to
  modify cs when fsm goto, call and return is supported in action code.
 -Fsm spec can have no statements, nothing will be generated.
 -Bug fix: Don't accept ] as the opening of a .-. range a reg exp.
 -Regular expression or set ranges (ie /[0-9]/) are now handled by the parser
  and consequently must be well-formed. The following now generates a parser
  error: /[+-]/ and must be rewritten as /[+\-]/. Also fixes a bug whereby ]
  might be accepted as the opening of a .-. range causing /[0-9]-[0-9]/ to
  parse incorrectly.
 -\v, \f, and \r are now treated as whitespace in an fsm spec.

Ragel 1.4.1 - Nov 19, 2002
==========================
 -Compile fixes. The last release (integer alphabets) was so exciting
  that usual portability checks got bypassed.

Ragel 1.4.0 - Nov 19, 2002
==========================
 -Arbitrary integer alphabets are now fully supported! A new language
  construct:
  'alphtype <type>' added for specifying the type of the alphabet. Default
  is 'char'. Possible alphabet types are:
       char, unsigned char, short, unsigned short, int, unsigned int
 -Literal machines specified in decimal format can now be negative when the
  alphabet is a signed type.
 -Literal machines (strings, decimal and hex) have their values checked for
  overflow/underflow against the size of the alphabet type.
 -Table driven and goto driven output redesigned to support ranges. Table
  driven uses a binary search for locating single characters and ranges. Goto
  driven uses a switch statement for single characters and nested if blocks for
  ranges.
 -Switch driven output removed due to a lack of consistent advantages. Most of
  the time the switch driven FSM is of no use because the goto FSM makes
  smaller and faster code. Under certain circumstances it can produce smaller
  code than a goto driven fsm and be almost as fast, but some sporadic case
  does not warrant maintaining it.
 -Many warnings changed to errors.
 -Added option -p for printing the final fsm before minimization. This lets
  priorities be seen. Priorties are all reset to 0 before minimization. The
  exiting option -f prints the final fsm after minimization.
 -Fixed a bug in the clang test and example that resulted in redundant actions
  being executed.

Ragel 1.3.4 - Nov 6, 2002
=========================
 -Fixes to Chapter 1 of the guide.
 -Brought back the examples and made them current.
 -MSVC is no longer supported for compiling windows binaries because its
  support for the C++ standard is frustratingly inadequate, it will cost money
  to upgrade if it ever gets better, and MinGW is a much better alternative.
 -The build system now supports the --host= option for building ragel
  for another system (used for cross compiling a windows binary with MinGW).
 -Various design changes and fixes towards the goal of arbitrary integer
  alphabets and the handling of larger state machines were made.
 -The new shared vector class is now used for action lists in transitions and
  states to reduce memory allocations.
 -An avl tree is now used for the reduction of transitions and functions of an
  fsm graph before making the final machine. The tree allows better scalability
  and performance by not requiring consecutively larger heap allocations.
 -Final stages in the separation of fsm graph code from action embedding and
  priority assignment is complete. Makes the base graph leaner and easier to reuse
  in other projects (like Keller).

Ragel 1.3.3 - Oct 22, 2002
==========================
 -More diagrams were added to section 1.7.1 of the user guide.
 -FSM Graph code was reworked to spearate the regex/nfa/minimizaion graph
  algorithms from the manipulation of state and transition properties.
 -An rpm spec file from Cris Bailiff was added. This allows an rpm for ragel
  to be built with the command 'rpm -ta ragel-x.x.x.tar.gz'
 -Fixes to the build system and corresponding doc updates in the README.
 -Removed autil and included the one needed source file directly in the top
  level ragel directory.
 -Fixed a bug that nullified the 20 times speedup in large compilations
  claimed by the last version.
 -Removed awk from the doc build (it was added with the last release -- though
  not mentioned in the changelog).
 -Install of man page was moved to the doc dir. The install also installs the
  user guide to $(PREFIX)/share/doc/ragel/

Ragel 1.3.2 - Oct 16, 2002
==========================
 -Added option -v (or --version) to show version information.
 -The subtract operator no longer removes transition data from the machine
  being subtracted. This is left up to the user for the purpose of making it
  possible to transfer transitions using subtract and also for speeding up the
  subtract routine. Note that it is possible to explicitly clear transition
  data before a doing a subtract.
 -Rather severe typo bug fixed. Bug was related to transitions with higher
  priorities taking precedence. A wrong ptr was being returned. It appears to
  have worked most of the time becuase the old ptr was deleted and the new one
  allocated immediatly after so the old ptr often pointed to the same space.
  Just luck though.
 -Bug in the removing of dead end paths was fixed. If the start state
  has in transitions then those paths were not followed when finding states to
  keep. Would result in non-dead end states being removed from the graph.
 -In lists and in ranges are no longer maintained as a bst with the key as the
  alphabet character and the value as a list of transitions coming in on that
  char. There is one list for each of inList, inRange and inDefault. Now that
  the required functionality of the graph is well known it is safe to remove
  these lists to gain in speed and footprint. They shouldn't be needed.
 -IsolateStartState() runs on modification of start data only if the start
  state is not already isolated, which is now possible with the new in list
  representation.
 -Concat, Or and Star operators now use an approximation to
  removeUnreachableStates that does not require a traversal of the entire
  graph. This combined with an 'on-the-fly' management of final bits and final
  state status results is a dramatic speed increase when compiling machines
  that use those operators heavily. The strings2 test goes 20 times faster.
 -Before the final minimization, after all fsm operations are complete,
  priority data is reset which enables better minimization in cases where
  priorities would otherwise separate similar states.
 
Ragel 1.3.1 - Oct 2, 2002
=========================
 -Range transitions are now used to implement machines made with /[a-z]/ and
  the .. operator as well as most of the builtin machines. The ranges are not
  yet reflected in the output code, they are expanded as if they came from the
  regular single transitions. This is one step closer to arbitrary integer
  output.
 -The builtin machine 'any' was added. It is equiv to the builtin extend,
  matching any characters.
 -The builtin machine 'cntrl' now includes newline.
 -The builtin machine 'space' now includes newline.
 -The builtin machine 'ascii' is now the range 0-127, not all characters.
 -A man page was written.
 -A proper user guide was started. Chapter 1: Specifying Ragel Programs
  was written. It even has some diagrams :)

Ragel 1.3.0 - Sep 4, 2002
=========================
 -NULL keyword no longer used in table output.
 -Though not yet in use, underlying graph structure changed to support range
  transitions. As a result, most of the code that walks transition lists is now
  implemented with an iterator that hides the complexity of the transition
  lists and ranges. Range transitions will be used to implement /[a-z]/ style
  machines and machines made with the .. operator. Previously a single
  transition would be used for each char in the range, which is very costly.
  Ranges eliminate much of the space complexity and allow for the .. operator
  to be used with very large (integer) alphabets.
 -New minimization similar to Hopcroft's alg. It does not require n^2 space and
  runs close to O(n*log(n)) (an exact analysis of the alg is very hard). It is
  much better than the stable and approx minimization and obsoletes them both.
  An exact implementation of Hopcroft's alg is desirable but not possible
  because the ragel implementation does not assume a finite alphabet, which
  Hopcroft's requires. Ragel will support arbitrary integer alphabets which
  must be treated as an infinite set for implementation considerations.
 -New option -m using above described minimization to replace all previous
  minimization options. Old options sill work but are obsolete and not
  advertised with -h.
 -Bug fixed in goto style output. The error exit set the current state to 0,
  which is actually a valid state. If the machine was entered again it would go
  into the first state, very wrong. If the first state happened to be final then
  an immediate finish would accept when in fact it should fail.
 -Slightly better fsm minimization now capable due to clearing of the
  transition ordering numbers just prior to minimization. 
 
Ragel 1.2.2 - May 25, 2002
==========================
 -Configuration option --prefix now works when installing.
 -cc file extension changed to cpp for better portability.
 -Unlink of output file upon error no longer happens, removes dependency on
  unlink system command.
 -All multiline strings removed: not standard c++.
 -Awk build dependency removed.
 -MSVC 6.0 added to the list of supported compilers (with some tweaking of
  bison and flex output).

Ragel 1.2.1 - May 13, 2002
==========================
 -Automatic dependencies were fixed, they were not working correctly.
 -Updated AUTHORS file to reflect contributors.
 -Code is more C++ standards compliant: compiles with g++ 3.0
 -Fixed bugs that only showed up in g++ 3.0
 -Latest (unreleased) Aapl.
 -Configuration script bails out if bison++ is installed. Ragel will not
  compile with bison++ because it is coded in c++ and bison++ automatically
  generates a c++ parser. Ragel uses a c-style bison parser.

Ragel 1.2.0 - May 3, 2002
=========================
 -Underlying graph structure now supports default transitions. The result is
  that a transition does not need to be made for each char of the alphabet
  when making 'extend' or '/./' machines. Ragel compiles machines that
  use the aforementioned primitives WAY faster.
 -The ugly hacks needed to pick default transitions now go away due to
  the graph supporting default transitions directly.
 -If -e is given, but minimization is not turned on, print a warning.
 -Makefiles use automatic dependencies.

Ragel 1.1.0 - Apr 15, 2002
==========================
 -Added goto fsm: much faster than any other fsm style.
 -Default operator (if two machines are side by side with no operator
  between them) is concatenation. First showed up in 1.0.4.
 -The fsm machine no longer auotmatically builds the flat table for
  transition indicies. Instead it keeps the key,ptr pair. In tabcodegen
  the flat table is produced. This way very large alphabets with sparse
  transitions will not consume large amounts of mem. This is also in prep
  for fsm graph getting a default transition.
 -Generated code contains a statement explicitly stating that ragel fsms
  are NOT covered by the GPL. Technically, Ragel copies part of itself
  to the output to make the generic fsm execution routine (for table driven
  fsms only) and so the output could be considered under the GPL. But this
  code is very trivial and could easlily be rewritten. The actual fsm data
  is subject to the copyright of the source. To promote the use of Ragel,
  a special exception is made for the part of the output copied from Ragel:
  it may be used without restriction.
 -Much more elegant code generation scheme is employed. Code generation
  class members need only put the 'codegen' keyword after their 'void' type
  in order to be automatically registerd to handle macros of the same name.
  An awk script recognises this keyword and generates an appropriate driver.
 -Ragel gets a test suite.
 -Postfunc and prefunc go away because they are not supported by non
  loop-driven fsms (goto, switch) and present duplicate functionality. 
  Universal funcs can be implemented by using $ operator.
 -Automatic dependencies used in build system, no more make depend target.
 -Code generation section in docs.
 -Uses the latests aapl.

Ragel 1.0.5 - Mar 3, 2002
=========================
 -Bugfix in SetErrorState that caused an assertion failure when compiling
  simple machines that did not have full transition tables (and thus did
  not show up on any example machines). Assertion failure did not occur
  when using the switch statement code as ragel does not call SetErrorState
  in that case.
 -Fixed some missing includes, now compiles on redhat.
 -Moved the FsmMachTrans Compare class out of FsmMachTrans. Some compilers
  don't deal with nested classes in templates too well.
 -Removed old unused BASEREF in fsmgraph and ragel now compiles using
  egcs-2.91.66 and presumably SUNWspro. The baseref is no longer needed
  because states do not support being elements in multiple lists. I would
  rather be able to support more compilers than have this feature.
 -Started a README with compilation notes. Started an AUTHORS file.
 -Started the user documentation. Describes basic machines and operators.

Ragel 1.0.4 - Mar 1, 2002
=========================
 -Ported to the version of Aapl just after 2.2.0 release. See 
  http://www.ragel.ca/aapl/ for details on aapl.
 -Fixed a bug in the clang example: the newline machine was not stared.
 -Added explanations to the clang and mailbox examples. This should
  help people that want to learn the lanuage as the manual is far from
  complete.

Ragel 1.0.3 - Feb 2, 2002
=========================
 -Added aapl to the ragel tree. No longer requires you to download
  and build aapl separately. Should avoid discouraging impatient users
  from compiling ragel.
 -Added the examples to the ragel tree.
 -Added configure script checks for bison and flex.
 -Fixed makefile so as not to die with newer versions of bison that
  write the header of the parser to a .hh file.
 -Started ChangeLog file.

Ragel 1.0.2 - Jan 30, 2002
==========================
 -Bug fix in calculating highIndex for table based code. Was using
  the length of out tranisition table rather than the value at the
  end.
 -If high/low index are at the limits, output a define in their place,
  not the high/low values themselves so as not to cause compiler warnings.
 -If the resulting machines don't have any indicies or functions, then
  omit the empty unrefereced static arrays so as not to cause compiler
  warnings about unused static vars.
 -Fixed variable sized indicies support. The header cannot have any
  reference to INDEX_TYPE as that info is not known at the time the header
  data is written. Forces us to use a void * for pointers to indicies. In
  the c++ versions we are forced to make much of the data non-member
  static data in the code portion for the same reason.

Ragel 1.0.1 - Jan 28, 2002
==========================
 -Exe name change from reglang to ragel.
 -Added ftabcodegen output code style which uses a table for states and
  transitions but uses a switch statement for the function execution.
 -Reformatted options in usage dump to look better.
 -Support escape sequences in [] sections of regular expressions.

Ragel 1.0 - Jan 25, 2002
========================
 -Initial release.