CHANGES FROM 1.31

This file contains the migration of PCCTS from 1.31 in the order that
changes were made.  1.32b7 is the last beta before full 1.32.
Terence Parr, Parr Research Corporation 1995.


======================================================================
1.32b1
Added Russell Quong to banner, changed banner for output slightly
Fixed it so that you have before / after actions for C++ in class def
Fixed bug in optimizer that made it sometimes forget to set internal
        token pointers.  Only showed up when a {...} was in the "wrong spot".

======================================================================
1.32b2
Added fixes by Dave Seidel for PC compilers in 32 bit mode (config.h
and set.h).

======================================================================
1.32b3
Fixed hideous bug in code generator for wildcard and for ~token op.

from Dave Seidel

   Added pcnames.bat
   1. in antlr/main.c: change strcasecmp() to stricmp()

   2. in dlg/output.c: use DLEXER_C instead on "DLexer.C"

   3. in h/PBlackBox.h: use <iostream.h> instead of <stream.h>

======================================================================
1.32b4
When the -ft option was used, any path prefix screwed up
the gate on the .h files

Fixed yet another bug due to the optimizer.

The exception handling thing was a bit wacko:

a : ( A B )? A B
  | A C
  ;
  exception ...

caused an exception if "A C" was the input.  In other words,
it found that A C didn't match the (A B)? pred and caused
an exception rather than trying the next alt.  All I did
was to change the zzmatch_wsig() macros.

Fixed some problems in gen.c relating to the name of token
class bit sets in the output.

Added the tremendously cool generalized predicate.  For the
moment, I'll give this bried description.

a : <<predicate>>? blah
  | foo
  ;

This implies that (assuming blah and foo are syntactically
ambiguous) "predicate" indicates the semantic validity of
applying "blah".  If "predicate" is false, "foo" is attempted.

Previously, you had to say:

a : <<LA(1)==ID ? predicate : 1>>? ID
  | ID
  ;

Now, you can simply use "predicate" without the ?: operator
if you turn on ANTLR command line option: "-prc on".  This
tells ANTLR to compute that all by itself.  It computes n
tokens of lookahead where LT(n) or LATEXT(n) is the farthest
ahead you look.

If you give a predicate using "-prc on" that is followed
by a construct that can recognize more than one n-sequence,
you will get a warning from ANTLR.  For example,

a : <<isTypeName(LT(1)->getText())>>? (ID|INT)
  ;

This is wrong because the predicate will be applied to INTs
as well as ID's.  You should use this syntax to make
the predicate more specific:

a : (ID)? => <<isTypeName(LT(1)->getText())>>? (ID|INT)
  ;

which says "don't apply the predicate unless ID is the
current lookahead context".

You cannot currently have anything in the "(context)? =>"
except sequences such as:

( LPAREN ID | LPAREN SCOPE )? => <<pred>>?

I haven't tested this THAT much, but it does work for the
C++ grammar.

======================================================================
1.32b5

Added getLine() to the ANTLRTokenBase and DLGBasedToken classes
left line() for backward compatibility.
----
Removed SORCERER_TRANSFORM from the ast.h stuff.
-------
Fixed bug in code gen of ANTLR such that nested syn preds work more
efficiently now.  The ANTLRTokenBuffer was getting very large
with nested predicates.
------
Memory leak is now gone from ANTLRTokenBuf; all tokens are deleted.
For backward compatibility reasons, you have to say parser->deleteTokens()
or mytokenbuffer->deleteTokens() but later it will be the default mode.
Say this after the parser is constructed. E.g.,

    ParserBlackBox<DLGLexer, MyParser, ANTLRToken> p(stdin);
    p.parser()->deleteTokens();
    p.parser()->start_symbol();


==============================
1.32b6

Changed so that deleteTokens() will do a delete ((ANTLRTokenBase *))
on the ptr.  This gets the virtual destructor.

Fixed some weird things in the C++ header files (a few return types).

Made the AST routines correspond to the book and SORCERER stuff.

New token stuff:  See testcpp/14/test.g

ANTLR accepts a #pragma gc_tokens which says
[1]     Generate label = copy(LT(1)) instead of label=LT(1) for
        all labeled token references.
[2]     User now has to define ANTLRTokenPtr (as a class or a typedef
        to just a pointer) as well as the ANTLRToken class itself.
		See the example.

To delete tokens in token buffer, use deleteTokens() message on parser.

        All tokens that fall off the ANTLRTokenBuffer get deleted
        which is what currently happens when deleteTokens() message
        has been sent to token buffer.

We always generate ANTLRTokenPtr instead of 'ANTLRToken *' now.
Then if no pragma set, ANTLR generates a

        class ANTLRToken;
        typedef ANTLRToken *ANTLRTokenPtr;

in each file.

Made a warning for x:rule_ref <<$x>>; still no warning for $i's, however.
class BB {

a : x:b y:A <<$x
$y>>
  ;

b : B;

}
generates
Antlr parser generator   Version 1.32b6   1989-1995
test.g, line 3: error: There are no token ptrs for rule references: '$x'

===================
1.32b7:

[With respect to token object garbage collection (GC), 1.32b7
 backtracks from 1.32b6, but results in better and less intrusive GC.
 This is the last beta version before full 1.32.]

BIGGEST CHANGES:

o	The "#pragma gc_tokens" is no longer used.

o	.C files are now .cpp files (hence, makefiles will have to
	be changed; or you can rerun genmk).  This is a good move,
	but causes some backward incompatibility problems.  You can
	avoid this by changing CPP_FILE_SUFFIX to ".C" in pccts/h/config.h.

o	The token object class hierarchy has been flattened to include
	only three classes: ANTLRAbstractToken, ANTLRCommonToken, and
	ANTLRCommonNoRefCountToken.  The common token now does garbage
	collection via ref counting.

o	"Smart" pointers are now used for garbage collection.  That is,
	ANTLRTokenPtr is used instead of "ANTLRToken *".

o	The antlr.1 man page has been cleaned up slightly.

o	The SUN C++ compiler now complains less about C++ support code.

o	Grammars which subclass ANTLRCommonToken must wrap all token
	pointer references in mytoken(token_ptr).  This is the only
	serious backward incompatibility.  See below.


MINOR CHANGES:

--------------------------------------------------------
1	deleteTokens()

The deleteTokens() message to the parser or token buffer has been changed
to one of:

    void noGarbageCollectTokens()   { inputTokens->noGarbageCollectTokens(); }
    void garbageCollectTokens()     { inputTokens->garbageCollectTokens(); }

The token buffer deletes all non-referenced tokens by default now.

--------------------------------------------------------
2	makeToken()

The makeToken() message returns a new type.  The function should look
like:

    virtual ANTLRAbstractToken *makeToken(ANTLRTokenType tt,
                                          ANTLRChar *txt,
                                          int line)
    {
        ANTLRAbstractToken *t = new ANTLRCommonToken(tt,txt);
        t->setLine(line);
        return t;
    }

--------------------------------------------------------
3	TokenType

Changed TokenType-> ANTLRTokenType  (often forces changes in AST defs due
to #[] constructor called to AST(tokentype, string)).

--------------------------------------------------------
4	AST()

You must define AST(ANTLRTokenPtr t) now in your AST class definition.
You might also have to include ATokPtr.h above the definition; e.g.,
if AST is defined in a separate file, such as AST.h, it's a good idea
to include ATOKPTR_H (ATokPtr.h).  For example,

	#include ATOKPTR_H
	class AST : public ASTBase {
	protected:
	    ANTLRTokenPtr token;
	public:
	    AST(ANTLRTokenPtr t) { token = t; }
	    void preorder_action() {
	        char *s = token->getText();
	        printf(" %s", s);
	    }
	};

Note the use of smart pointers rather than "ANTLRToken *".

--------------------------------------------------------
5	SUN C++

From robertb oakhill.sps.mot.com Bob Bailey. Changed ANTLR C++ output
to avoid an error in Sun C++ 3.0.1.  Made "public" return value
structs created to hold multiple return values public.

--------------------------------------------------------
6	genmk

Fixed genmk so that target List.* is not included anymore.  It's
called SList.* anyway.

--------------------------------------------------------
7	\r vs \n

Scott Vorthmann <vorth cmu.edu> fixed antlr.g in ANTLR so that \r
is allowed as the return character as well as \n.

--------------------------------------------------------
8	Exceptions

Bug in exceptions attached to labeled token/tokclass references.  Didn't gen
code for exceptions.  This didn't work:

a : "help" x:ID
  ;
        exception[x]
        catch MismatchedToken : <<printf("eh?\n");>>

Now ANTLR generates (which is kinda big, but necessary):

        if ( !_match_wsig(ID) ) {
                if ( guessing ) goto fail;
                _signal=MismatchedToken;
                switch ( _signal ) {
                case MismatchedToken :
                        printf("eh?\n");
                        _signal = NoSignal;
                        break;
                default :
                        goto _handler;
                }
        }

which implies that you can recover and continue parsing after a missing/bad
token reference.

--------------------------------------------------------
9	genmk

genmk now correctly uses config file for CPP_FILE_SUFFIX stuff.

--------------------------------------------------------
10	general cleanup / PURIFY

Anthony Green <green vizbiz.com> suggested a bunch of good general
clean up things for the code; he also suggested a few things to
help out the "PURIFY" memory allocation checker.

--------------------------------------------------------
11	$-variable references.

Manuel ORNATO indicated that a $-variable outside of a rule caused
ANTLR to crash.  I fixed this.

12	Tom Moog suggestion

Fail action of semantic predicate needs "{}" envelope.  FIXED.

13	references to LT(1).

I have enclosed all assignments such as:

             _t22 = (ANTLRTokenPtr)LT(1);

in "if ( !guessing )" so that during backtracking the reference count
for token objects is not increased.


TOKEN OBJECT GARBAGE COLLECTION

1	INTRODUCTION

The class ANTLRCommonToken is now garbaged collected through a "smart"
pointer called ANTLRTokenPtr using reference counting.  Any token
object not referenced by your grammar actions is destroyed by the
ANTLRTokenBuffer when it must make room for more token objects.
Referenced tokens are then destroyed in your parser when local
ANTLRTokenPtr objects are deleted.  For example,

a : label:ID ;

would be converted to something like:

void yourclass::a(void)
{
	zzRULE;
	ANTLRTokenPtr label=NULL;	// used to be ANTLRToken *label;
        zzmatch(ID);
        label = (ANTLRTokenPtr)LT(1);
	consume();
	...
}

When the "label" object is destroyed (it's just a pointer to your
input token object LT(1)), it decrements the reference count on the
object created for the ID.  If the count goes to zero, the object
pointed by label is deleted.

To correctly manage the garbage collection, you should use
ANTLRTokenPtr instead of "ANTLRToken *".  Most ANTLR support code
(visible to the user) has been modified to use the smart pointers.

***************************************************************
Remember that any local objects that you create are not deleted when a
lonjmp() is executed.  Unfortunately, the syntactic predicates (...)?
use setjmp()/longjmp().  There are some situations when a few tokens
will "leak".
***************************************************************

2	DETAILS

o	The default is to perform token object garbage collection.
	You may use parser->noGarbageCollectTokens() to turn off
	garbage collection.


o	The type ANTLRTokenPtr is always defined now (automatically).
	If you do not wish to use smart pointers, you will have to
	redefined ANTLRTokenPtr by subclassing, changing the header
	file or changing ANTLR's code generation (easy enough to
	do in gen.c).

o	If you don't use ParserBlackBox, the new initialization sequence is:

	    ANTLRTokenPtr aToken = new ANTLRToken;
	    scan.setToken(mytoken(aToken));

	where mytoken(aToken) gets an ANTLRToken * from the smart pointer.

o	Define C++ preprocessor symbol DBG_REFCOUNTTOKEN to see a bunch of
	debugging stuff for reference counting if you suspect something.


3	WHY DO I HAVE TO TYPECAST ALL MY TOKEN POINTERS NOW??????

If you subclass ANTLRCommonToken and then attempt to refer to one of
your token members via a token pointer in your grammar actions, the
C++ compiler will complain that your token object does not have that
member.  For example, if you used to do this

<<
class ANTLRToken : public ANTLRCommonToken {
        int muck;
	...
};
>>

class Foo {
a : t:ID << t->muck = ...; >> ;
}

Now, you must do change the t->muck reference to:

a : t:ID << mytoken(t)->muck = ...; >> ;

in order to downcast 't' to be an "ANTLRToken *" not the
"ANTLRAbstractToken *" resulting from ANTLRTokenPtr::operator->().
The macro is defined as:

/*
 * Since you cannot redefine operator->() to return one of the user's
 * token object types, we must down cast.  This is a drag.  Here's
 * a macro that helps.  template: "mytoken(a-smart-ptr)->myfield".
 */
#define mytoken(tp) ((ANTLRToken *)(tp.operator->()))

You have to use macro mytoken(grammar-label) now because smart
pointers are not specific to a parser's token objects.  In other
words, the ANTLRTokenPtr class has a pointer to a generic
ANTLRAbstractToken not your ANTLRToken; the ANTLR support code must
use smart pointers too, but be able to work with any kind of
ANTLRToken.  Sorry about this, but it's C++'s fault not mine.  Some
nebulous future version of the C++ compilers should obviate the need
to downcast smart pointers with runtime type checking (and by allowing
different return type of overridden functions).

A way to have backward compatible code is to shut off the token object
garbage collection; i.e., use parser->noGarbageCollectTokens() and
change the definition of ANTLRTokenPtr (that's why you get source code
<wink>).


PARSER EXCEPTION HANDLING

I've noticed some weird stuff with the exception handling.  I intend
to give this top priority for the "book release" of ANTLR.

==========
1.32 Full Release

o	Changed Token class hierarchy to be (Thanks to Tom Moog):

        ANTLRAbstractToken
          ANTLRRefCountToken
             ANTLRCommonToken
          ANTLRNoRefCountCommonToken

o	Added virtual panic() to ANTLRAbstractToken.  Made ANTLRParser::panic()
	virtual also.

o	Cleaned up the dup() stuff in AST hierarchy to use shallowCopy() to
	make node copies.  John Farr at Medtronic suggested this.  I.e.,
	if you want to use dup() with either ANTLR or SORCERER or -transform
	mode with SORCERER, you must defined shallowCopy() as:

	virtual PCCTS_AST *shallowCopy()
	{
	    return new AST;
	    p->setDown(NULL);
	    p->setRight(NULL);
	    return p;
	}

	or

	virtual PCCTS_AST *shallowCopy()
	{
	    return new AST(*this);
	}
	
	if you have defined a copy constructor such as

	AST(const AST &t)	// shallow copy constructor
	{
		token = t.token;
		iconst = t.iconst;
		setDown(NULL);
		setRight(NULL);
	}

o	Added a warning with -CC and -gk are used together.  This is broken,
	hence a warning is appropriate.

o	Added warning when #-stuff is used w/o -gt option.

o	Updated MPW installation.

o	"Miller, Philip W." <MILLERPW f1groups.fsd.jhuapl.edu> suggested
	that genmk be use RENAME_OBJ_FLAG RENAME_EXE_FLAG instead of
	hardcoding "-o" in genmk.c.

o	made all exit() calls use EXIT_SUCCESS or EXIT_FAILURE.

===========================================================================
1.33

EXIT_FAILURE and EXIT_SUCCESS were not always defined.  I had to modify
a bunch of files to use PCCTS_EXIT_XXX, which forces a new version.  Sorry
about that.