CHANGES_FROM_133.txt revision 4fd606d1f5abe38e1f42c38de1d2e895166bd0f4
=======================================================================
List of Implemented Fixes and Changes for Maintenance Releases of PCCTS
=======================================================================
DISCLAIMER
The software and these notes are provided "as is". They may include
typographical or technical errors and their authors disclaims all
liability of any kind or nature for damages due to error, fault,
defect, or deficiency regardless of cause. All warranties of any
kind, either express or implied, including, but not limited to, the
implied warranties of merchantability and fitness for a particular
purpose are disclaimed.
-------------------------------------------------------
Note: Items #153 to #1 are now in a separate file named
-------------------------------------------------------
#312. (Changed in MR33) Bug caused by change #299.
In change #299 a warning message was suppressed when there was
no LT(1) in a semantic predicate and max(k,ck) was 1. The
changed caused the code which set a default predicate depth for
the semantic predicate to be left as 0 rather than set to 1.
This manifested as an error at line #1559 of mrhost.c
Reported by Peter Dulimov.
#311. (Changed in MR33) Added sorcer/lib to Makefile.
Reported by Dale Martin.
#310. (Changed in MR32) In C mode zzastPush was spelled zzastpush in one case.
Reported by Jean-Claude Durand
#309. (Changed in MR32) Renamed baseName because of VMS name conflict
Renamed baseName to pcctsBaseName to avoid library name conflict with
VMS library routine. Reported by Jean-Fran�ois PI�RONNE.
#308. (Changed in MR32) Used "template" as name of formal in C routine
In astlib.h routine ast_scan a formal was named "template". This caused
problems when the C code was compiled with a C++ compiler. Reported by
Sabyasachi Dey.
#307. (Changed in MR31) Compiler dependent bug in function prototype generation
The code which generated function prototypes contained a bug which
was compiler/optimization dependent. Under some circumstance an
extra character would be included in portions of a function prototype.
Reported by David Cook.
#306. (Changed in MR30) Validating predicate following a token
A validating predicate which immediately followed a token match
consumed the token after the predicate rather than before. Prior
to this fix (in the following example) isValidTimeScaleValue() in
the predicate would test the text for TIMESCALE rather than for
NUMBER:
time_scale :
TIMESCALE
<<isValidTimeScaleValue(LT(1)->getText())>>?
ts:NUMBER
( us:MICROSECOND << tVal = ...>>
| ns:NANOSECOND << tVal = ... >>
)
Reported by Adalbert Perbandt.
#305. (Changed in MR30) Alternatives with guess blocks inside (...)* blocks.
In MR14 change #175 fixed a bug in the prediction expressions for guess
blocks which were of the form (alpha)? beta. Unfortunately, this
resulted in a new bug as exemplified by the example below, which computed
the first set for r as {B} rather than {B C}:
r : ( (A)? B
| C
)*
This example doesn't make any sense as A is not a prefix of B, but it
illustrates the problem. This bug did not appear for:
r : ( (A)?
| C
)*
because it does not use the (alpha)? beta form.
Item #175 fixed an asymmetry in ambiguity messages for the following
constructs which appear to have identical ambiguities (between repeating
the loop vs. exiting the loop). MR30 retains this fix, but the implementation
is slightly different.
r_star : ( (A B)? )* A ;
r_plus : ( (A B)? )+ A ;
Reported by Arpad Beszedes (beszedes inf.u-szeged.hu).
#304. (Changed in MR30) Crash when mismatch between output value counts.
For a rule such as:
r1 : r2>[i,j];
r2 >[int i, int j] : A;
If there were extra actuals for the reference to rule r2 from rule r1
there antlr would crash. This bug was introduced by change #276.
Reported by Sinan Karasu.
#303. (Changed in MR30) DLGLexerBase::replchar
DLGLexerBase::replchar and the C mode routine zzreplchar did not work
properly when the new character was 0.
Reported with fix by Philippe Laporte
#302. (Changed in MR28) Fix significant problems in initial release of MR27.
#301. (Changed in MR27) Default tab stops set to 2 spaces.
To have antlr generate true tabs rather than spaces, use "antlr -tab 0".
To generate 4 spaces per tab stop use "antlr -tab 4"
#300. (Changed in MR27)
Consider the following methods of constructing an AST from ID:
rule1!
: id:ID << #0 = #[id]; >> ;
rule2!
: id:ID << #0 = #id; >> ;
rule3
: ID ;
rule4
: id:ID << #0 = #id; >> ;
For rule_2, the AST corresponding to id would always be NULL. This
is because the user explicitly suppressed AST construction using the
"!" operator on the rule. In MR27 the use of an AST expression
such as #id overrides the "!" operator and forces construction of
the AST.
This fix does not apply to C mode ASTs when the ASTs are referenced
using numbers rather than symbols.
For C mode, this requires that the (optional) function/macro zzmk_ast
be defined. This functions copies information from an attribute into
a previously allocated AST.
Reported by Jan Langer (jan langernetz.de)
#299. (Changed in MR27) Don't warn if k=1 and semantic predicate missing LT(i)
If a semantic does not have a reference to LT(i) or (C mode LATEXT(i))
then pccts doesn't know how many lookahead tokens to use for context.
However, if max(k,ck) is 1 then there is really only one choice and
the warning is unnecessary.
#298. (Changed in MR27) Removed "register" for lastpos in dlgauto.c zzgettok
#297. (Changed in MR27) Incorrect prototypes when used with classic C
There were a number of errors in function headers when antlr was
built with compilers that do not have __STDC__ or __cplusplus set.
The functions which have variable length argument lists now use
PCCTS_USE_STDARG rather than __USE_PROTOTYPES__ to determine
whether to use stdargs or varargs.
#296. (Changed in MR27) Complex return types in rules.
The following return type was not properly handled when
unpacking a struct with containing multiple return values:
rule > [int i, IIR_Bool (IIR_Decl::*constraint)()] : ...
Instead of using "constraint", the program got lost and used
an empty string.
Reported by P.A. Wilsey.
#295. (Changed in MR27) Extra ";" following zzGUESS_DONE sometimes.
Certain constructs with guess blocks in MR23 led to extra ";"
preceding the "else" clause of an "if".
Reported by P.A. Wilsey.
#294. (Changed in MR27) Infinite loop in antlr for nested blocks
An oversight in detecting an empty alternative sometimes led
to an infinite loop in antlr when it encountered a rule with
nested blocks and guess blocks.
Reported by P.A. Wilsey.
#293. (Changed in MR27) Sorcerer optimization of _t->type()
Sorcerer generated code may contain many calls to _t->type() in a
single statement. This change introduces a temporary variable
to eliminate unnnecesary function calls.
Change implemented by Tom Molteno (tim videoscript.com).
#292. (Changed in MR27)
WARNING: Item #267 changes the signature of methods in the AST class.
**** Be sure to revise your AST functions of the same name ***
#291. (Changed in MR24)
Fix to serious code generation error in MR23 for (...)+ block.
#290. (Changed in MR23)
Item #247 describes a change in the way {...} blocks handled
an error. Consider:
r1 : {A} b ;
b : B;
with input "C".
Prior to change #247, the error would resemble "expected B -
found C". This is correct but incomplete, and therefore
misleading. In #247 it was changed to "expected A, B - found
C". This was fine, except for users of parser exception
handling because the exception was generated in the epilogue
for {...} block rather than in rule b. This made it difficult
for users of parser exception handling because B was not
expected in that context. Those not using parser exception
handling didn't notice the difference.
The current change restores the behavior prior to #247 when
parser exceptions are present, but retains the revised behavior
otherwise. This change should be visible only when exceptions
are in use and only for {...} blocks and sub-blocks of the form
(something|something | something | epsilon) where epsilon represents
an empty production and it is the last alternative of a sub-block.
In contrast, (something | epsilon | something) should generate the
same code as before, even when exceptions are used.
Reported by Philippe Laporte (philippe at transvirtual.com).
#289. (Changed in MR23) Bug in matching complement of a #tokclass
Prior to MR23 when a #tokclass was matched in both its complemented form
and uncomplemented form, the bit set generated for its first use was used
for both cases. However, the prediction expression was correctly computed
in both cases. This meant that the second case would never be matched
because, for the second appearance, the prediction expression and the
set to be matched would be complements of each other.
Consider:
#token A "a"
#token B "b"
#token C "c"
#tokclass AB {A B}
r1 : AB /* alt 1x */
| ~AB /* alt 1y */
;
Prior to MR23, this resulted in alternative 1y being unreachable. Had it
been written:
r2 : ~AB /* alt 2x */
: AB /* alt 2y */
then alternative 2y would have become unreachable.
This bug was only for the case of complemented #tokclass. For complemented
#token the proper code was generated.
#288. (Changed in MR23) #errclass not restricted to choice points
The #errclass directive is supposed to allow a programmer to define
print strings which should appear in syntax error messages as a replacement
for some combinations of tokens. For instance:
#errclass Operator {PLUS MINUS TIMES DIVIDE}
If a syntax message includes all four of these tokens, and there is no
"better" choice of error class, the word "Operator" will be used rather
than a list of the four token names.
Prior to MR23 the #errclass definitions were used only at choice points
(which call the FAIL macro). In other cases where there was no choice
(e.g. where a single token or token class were matched) the #errclass
information was not used.
With MR23 the #errclass declarations are used for syntax error messages
when matching a #tokclass, a wildcard (i.e. "*"), or the complement of a
#token or #tokclass (e.g. ~Operator).
Please note that #errclass may now be defined using #tokclass names
(see Item #284).
Reported by Philip A. Wilsey.
#287. (Changed in MR23) Print name for #tokclass
Item #148 describes how to give a print name to a #token so that,for
example, #token ID could have the expression "identifier" in syntax
error messages. This has been extended to #tokclass:
#token ID("identifier") "[a-zA-Z]+"
#tokclass Primitive("primitive type")
{INT, FLOAT, CHAR, FLOAT, DOUBLE, BOOL}
This is really a cosmetic change, since #tokclass names do not appear
in any error messages.
#286. (Changed in MR23) Makefile change to use of cd
In cases where a pccts subdirectory name matched a directory identified
in a $CDPATH environment variable the build would fail. All makefile
cd commands have been changed from "cd xyz" to "cd ./xyz" in order
to avoid this problem.
#285. (Changed in MR23) Check for null pointers in some dlg structures
An invalid regular expression can cause dlg to build an invalid
structure to represent the regular expression even while it issues
error messages. Additional pointer checks were added.
Reported by Robert Sherry.
#284. (Changed in MR23) Allow #tokclass in #errclass definitions
Previously, a #tokclass reference in the definition of an
#errclass was not handled properly. Instead of being expanded
into the set of tokens represented by the #tokclass it was
treated somewhat like an #errclass. However, in a later phase
when all #errclass were expanded into the corresponding tokens
the #tokclass reference was not expanded (because it wasn't an
#errclass). In effect the reference was ignored.
This has been fixed.
Problem reported by Mike Dimmick (mike dimmick.demon.co.uk).
#283. (Changed in MR23) Option -tmake invoke's parser's tmake
When the string #(...) appears in an action antlr replaces it with
a call to ASTBase::tmake(...) to construct an AST. It is sometimes
useful to change the tmake routine so that it has access to information
in the parser - something which is not possible with a static method
in an application where they may be multiple parsers active.
The antlr option -tmake replaces the call to ASTBase::tmake with a call
to a user supplied tmake routine.
#282. (Changed in MR23) Initialization error for DBG_REFCOUNTTOKEN
When the pre-processor symbol DBG_REFCOUNTTOKEN is defined
incorrect code is generated to initialize ANTLRRefCountToken::ctor and
dtor.
Fix reported by Sven Kuehn (sven sevenkuehn.de).
#281. (Changed in MR23) Addition of -noctor option for Sorcerer
Added a -noctor option to suppress generation of the blank ctor
for users who wish to define their own ctor.
Contributed by Jan Langer (jan langernetz.de).
#280. (Changed in MR23) Syntax error message for EOF token
The EOF token now receives special treatment in syntax error messages
because there is no text matched by the eof token. The token name
of the eof token is used unless it is "@" - in which case the string
"<eof>" is used.
Problem reported by Erwin Achermann (erwin.achermann switzerland.org).
#279. (Changed in MR23) Exception groups
There was a bug in the way that exception groups were attached to
alternatives which caused problems when there was a block contained
in an alternative. For instance, in the following rule;
statement : IF S { ELSE S }
exception ....
;
the exception would be attached to the {...} block instead of the
entire alternative because it was attached, in error, to the last
alternative instead of the last OPEN alternative.
Reported by Ty Mordane (tymordane hotmail.com).
#278. (Changed in MR23) makefile changes
Contributed by Tomasz Babczynski (faster lab05-7.ict.pwr.wroc.pl).
The -cfile option is not absolutely needed: when extension of
source file is one of the well-known C/C++ extensions it is
treated as C/C++ source
The gnu make defines the CXX variable as the default C++ compiler
name, so I added a line to copy this (if defined) to the CCC var.
Added a -sor option: after it any -class command defines the class
name for sorcerer, not for ANTLR. A file extended with .sor is
treated as sorcerer input. Because sorcerer can be called multiple
times, -sor option can be repeated. Any files and classes (one class
per group) after each -sor makes one tree parser.
Not implemented:
1. Generate dependences for user c/c++ files.
2. Support for -sor in c mode not.
I have left the old genmk program in the directory as genmk_old.c.
#277. (Changed in MR23) Change in macro for failed semantic predicates
In the past, a semantic predicate that failed generated a call to
the macro zzfailed_pred:
#ifndef zzfailed_pred
#define zzfailed_pred(_p) \
if (guessing) { \
zzGUESS_FAIL; \
} else { \
something(_p)
}
#endif
If a user wished to use the failed action option for semantic predicates:
rule : <<my_predicate>>? [my_fail_action] A
| ...
the code for my_fail_action would have to contain logic for handling
the guess part of the zzfailed_pred macro. The user should not have
to be aware of the guess logic in writing the fail action.
The zzfailed_pred has been rewritten to have three arguments:
arg 1: the stringized predicate of the semantic predicate
arg 2: 0 => there is no user-defined fail action
1 => there is a user-defined fail action
arg 3: the user-defined fail action (if defined)
otherwise a no-operation
The zzfailed_pred macro is now defined as:
#ifndef zzfailed_pred
#define zzfailed_pred(_p,_hasuseraction,_useraction) \
if (guessing) { \
zzGUESS_FAIL; \
} else { \
zzfailed_pred_action(_p,_hasuseraction,_useraction) \
}
#endif
With zzfailed_pred_action defined as:
#ifndef zzfailed_pred_action
#define zzfailed_pred_action(_p,_hasuseraction,_useraction) \
if (_hasUserAction) { _useraction } else { failedSemanticPredicate(_p); }
#endif
In C++ mode failedSemanticPredicate() is a virtual function.
In C mode the default action is a fprintf statement.
Suggested by Erwin Achermann (erwin.achermann switzerland.org).
#276. (Changed in MR23) Addition of return value initialization syntax
In an attempt to reduce the problems caused by the PURIFY macro I have
added new syntax for initializing the return value of rules and the
antlr option "-nopurify".
A rule with a single return argument:
r1 > [Foo f = expr] :
now generates code that resembles:
Foo r1(void) {
Foo _retv = expr;
...
}
A rule with more than one return argument:
r2 > [Foo f = expr1, Bar b = expr2 ] :
generates code that resembles:
struct _rv1 {
Foo f;
Bar b;
}
_rv1 r2(void) {
struct _rv1 _retv;
_retv.f = expr1;
_retv.b = expr2;
...
}
C++ style comments appearing in the initialization list may cause problems.
#275. (Changed in MR23) Addition of -nopurify option to antlr
A long time ago the PURIFY macro was introduced to initialize
return value arguments and get rid of annying messages from program
that checked for unitialized variables.
This has caused significant annoyance for C++ users that had
classes with virtual functions or non-trivial contructors because
it would zero the object, including the pointer to the virtual
function table. This could be defeated by redefining
the PURIFY macro to be empty, but it was a constant surprise to
new C++ users of pccts.
I would like to remove it, but I fear that some existing programs
depend on it and would break. My temporary solution is to add
an antlr option -nopurify which disables generation of the PURIFY
macro call.
The PURIFY macro should be avoided in favor of the new syntax
for initializing return arguments described in item #275.
To avoid name clash, the PURIFY macro has been renamed PCCTS_PURIFY.
#274. (Changed in MR23) DLexer.cpp renamed to DLexer.h
(Changed in MR23) ATokPtr.cpp renamed to ATokPtrImpl.h
These two files had .cpp extensions but acted like .h files because
there were included in other files. This caused problems for many IDE.
I have renamed them. The ATokPtrImpl.h was necessary because there was
already an ATokPtr.h.
#273. (Changed in MR23) Default win32 library changed to multi-threaded DLL
The model used for building the Win32 debug and release libraries has changed
to multi-threaded DLL.
To make this change in your MSVC 6 project:
Project -> Settings
Select the C++ tab in the right pane of the dialog box
Select "Category: Code Generation"
Under "Use run-time library" select one of the following:
Multi-threaded DLL
Debug Multi-threaded DLL
Suggested by Bill Menees (bill.menees gogallagher.com)
#272. (Changed in MR23) Failed semantic predicate reported via virtual function
In the past, a failed semantic predicated reported the problem via a
macro which used fprintf(). The macro now expands into a call on
the virtual function ANTLRParser::failedSemanticPredicate().
#271. (Changed in MR23) Warning for LT(i), LATEXT(i) in token match actions
An bug (or at least an oddity) is that a reference to LT(1), LA(1),
or LATEXT(1) in an action which immediately follows a token match
in a rule refers to the token matched, not the token which is in
the lookahead buffer. Consider:
r : abc <<action alpha>> D <<action beta>> E;
In this case LT(1) in action alpha will refer to the next token in
the lookahead buffer ("D"), but LT(1) in action beta will refer to
the token matched by D - the preceding token.
A warning has been added for users about this when an action
following a token match contains a reference to LT(1), LA(1), or LATEXT(1).
This behavior should be changed, but it appears in too many programs
now. Another problem, perhaps more significant, is that the obvious
fix (moving the consume() call to before the action) could change the
order in which input is requested and output appears in existing programs.
This problem was reported, along with a fix by Benjamin Mandel
(beny sd.co.il). However, I felt that changing the behavior was too
dangerous for existing code.
#270. (Changed in MR23) Removed static objects from PCCTSAST.cpp
There were some statically allocated objects in PCCTSAST.cpp
These were changed to non-static.
#269. (Changed in MR23) dlg output for initializing static array
The output from dlg contains a construct similar to the
following:
struct XXX {
static const int size;
static int array1[5];
};
const int XXX::size = 4;
int XXX::array1[size+1];
The problem is that although the expression "size+1" used in
the definition of array1 is equal to 5 (the expression used to
declare array), it is not considered equivalent by some compilers.
Reported with fix by Volker H. Simonis (simonis informatik.uni-tuebingen.de)
#268. (Changed in MR23) syn() routine output when k > 1
The syn() routine is supposed to print out the text of the
token causing the syntax error. It appears that it always
used the text from the first lookahead token rather than the
appropriate one. The appropriate one is computed by comparing
the token codes of lookahead token i (for i = 1 to k) with
the FIRST(i) set.
This has been corrected in ANTLRParser::syn().
Reported by Bill Menees (bill.menees gogallagher.com)
#267. (Changed in MR23) AST traversal functions client data argument
The AST traversal functions now take an extra (optional) parameter
which can point to client data:
preorder_action(void* pData = NULL)
preorder_before_action(void* pData = NULL)
preorder_after_action(void* pData = NULL)
**** Warning: this changes the AST signature. ***
**** Be sure to revise your AST functions of the same name ***
Bill Menees (bill.menees gogallagher.com)
#266. (Changed in MR23) virtual function printMessage()
Bill Menees (bill.menees gogallagher.com) has completed the
tedious taks of replacing all calls to fprintf() with calls
to the virtual function printMessage(). For classes which
have a pointer to the parser it forwards the printMessage()
call to the parser's printMessage() routine.
This should make it significanly easier to redirect pccts
error and warning messages.
#265. (Changed in MR23) Remove "labase++" in C++ mode
In C++ mode labase++ is called when a token is matched.
It appears that labase is not used in C++ mode at all, so
this code has been commented out.
#264. (Changed in MR23) Complete rewrite of ParserBlackBox.h
The parser black box (PBlackBox.h) was completely rewritten
by Chris Uzdavinis (chris atdesk.com) to improve its robustness.
#263. (Changed in MR23) -preamble and -preamble_first rescinded
Changes for item #253 have been rescinded.
#262. (Changed in MR23) Crash with -alpha option during traceback
Under some circumstances a -alpha traceback was started at the
"wrong" time. As a result, internal data structures were not
initialized.
Reported by Arpad Beszedes (beszedes inf.u-szeged.hu).
#261. (Changed in MR23) Defer token fetch for C++ mode
Item #216 has been revised to indicate that use of the defer fetch
option (ZZDEFER_FETCH) requires dlg option -i.
#260. (MR22) Raise default lex buffer size from 8,000 to 32,000 bytes.
ZZLEXBUFSIZE is the size (in bytes) of the buffer used by dlg
generated lexers. The default value has been raised to 32,000 and
the value used by antlr, dlg, and sorcerer has also been raised to
32,000.
#259. (MR22) Default function arguments in C++ mode.
If a rule is declared:
rr [int i = 0] : ....
then the declaration generated by pccts resembles:
void rr(int i = 0);
however, the definition must omit the default argument:
void rr(int i) {...}
In the past the default value was not omitted. In MR22
the generated code resembles:
void rr(int i /* = 0 */ ) {...}
Implemented by Volker H. Simonis (simonis informatik.uni-tuebingen.de)
Note: In MR23 this was changed so that nested C style comments
("/* ... */") would not cause problems.
#258. (MR22) Using a base class for your parser
In item #102 (MR10) the class statement was extended to allow one
to specify a base class other than ANTLRParser for the generated
parser. It turned out that this was less than useful because
the constructor still specified ANTLRParser as the base class.
The class statement now uses the first identifier appearing after
the ":" as the name of the base class. For example:
class MyParser : public FooParser {
Generates in MyParser.h:
class MyParser : public FooParser {
Generates in MyParser.cpp something that resembles:
MyParser::MyParser(ANTLRTokenBuffer *input) :
FooParser(input,1,0,0,4)
{
token_tbl = _token_tbl;
traceOptionValueDefault=1; // MR10 turn trace ON
}
The base class constructor must have a signature similar to
that of ANTLRParser.
#257. (MR21a) Removed dlg statement that -i has no effect in C++ mode.
This was incorrect.
#256. (MR21a) Malformed syntax graph causes crash after error message.
In the past, certain kinds of errors in the very first grammar
element could cause the construction of a malformed graph
representing the grammar. This would eventually result in a
fatal internal error. The code has been changed to be more
resistant to this particular error.
#255. (MR21a) ParserBlackBox(FILE* f)
This constructor set openByBlackBox to the wrong value.
Reported by Kees Bakker (kees_bakker tasking.nl).
#254. (MR21a) Reporting syntax error at end-of-file
When there was a syntax error at the end-of-file the syntax
error routine would substitute "<eof>" for the programmer's
end-of-file symbol. This substitution is now done only when
the programmer does not define his own end-of-file symbol
or the symbol begins with the character "@".
Reported by Kees Bakker (kees_bakker tasking.nl).
#253. (MR21) Generation of block preamble (-preamble and -preamble_first)
*** This change was rescinded by item #263 ***
The antlr option -preamble causes antlr to insert the code
BLOCK_PREAMBLE at the start of each rule and block. It does
not insert code before rules references, token references, or
actions. By properly defining the macro BLOCK_PREAMBLE the
user can generate code which is specific to the start of blocks.
The antlr option -preamble_first is similar, but inserts the
code BLOCK_PREAMBLE_FIRST(PreambleFirst_123) where the symbol
PreambleFirst_123 is equivalent to the first set defined by
the #FirstSetSymbol described in Item #248.
I have not investigated how these options interact with guess
mode (syntactic predicates).
#252. (MR21) Check for null pointer in trace routine
When some trace options are used when the parser is generated
without the trace enabled, the current rule name may be a
NULL pointer. A guard was added to check for this in
restoreState.
Reported by Douglas E. Forester (dougf projtech.com).
#251. (MR21) Changes to #define zzTRACE_RULES
The macro zzTRACE_RULES was being use to pass information to
AParser.h. If this preprocessor symbol was not properly
set the first time AParser.h was #included, the declaration
of zzTRACEdata would be omitted (it is used by the -gd option).
Subsequent #includes of AParser.h would be skipped because of
the #ifdef guard, so the declaration of zzTracePrevRuleName would
never be made. The result was that proper compilation was very
order dependent.
The declaration of zzTRACEdata was made unconditional and the
problem of removing unused declarations will be left to optimizers.
Diagnosed by Douglas E. Forester (dougf projtech.com).
#250. (MR21) Option for EXPERIMENTAL change to error sets for blocks
The antlr option -mrblkerr turns on an experimental feature
which is supposed to provide more accurate syntax error messages
for k=1, ck=1 grammars. When used with k>1 or ck>1 grammars the
behavior should be no worse than the current behavior.
There is no problem with the matching of elements or the computation
of prediction expressions in pccts. The task is only one of listing
the most appropriate tokens in the error message. The error sets used
in pccts error messages are approximations of the exact error set when
optional elements in (...)* or (...)+ are involved. While entirely
correct, the error messages are sometimes not 100% accurate.
There is also a minor philosophical issue. For example, suppose the
grammar expects the token to be an optional A followed by Z, and it
is X. X, of course, is neither A nor Z, so an error message is appropriate.
Is it appropriate to say "Expected Z" ? It is correct, it is accurate,
but it is not complete.
When k>1 or ck>1 the problem of providing the exactly correct
list of tokens for the syntax error messages ends up becoming
equivalent to evaluating the prediction expression for the
alternatives twice. However, for k=1 ck=1 grammars the prediction
expression can be computed easily and evaluated cheaply, so I
decided to try implementing it to satisfy a particular application.
This application uses the error set in an interactive command language
to provide prompts which list the alternatives available at that
point in the parser. The user can then enter additional tokens to
complete the command line. To do this required more accurate error
sets then previously provided by pccts.
In some cases the default pccts behavior may lead to more robust error
recovery or clearer error messages then having the exact set of tokens.
This is because (a) features like -ge allow the use of symbolic names for
certain sets of tokens, so having extra tokens may simply obscure things
and (b) the error set is use to resynchronize the parser, so a good
choice is sometimes more important than having the exact set.
Consider the following example:
Note: All examples code has been abbreviated
to the absolute minimum in order to make the
examples concise.
star1 : (A)* Z;
The generated code resembles:
old new (with -mrblkerr)
--//----------- --------------------
for (;;) { for (;;) {
match(A); match(A);
} }
match(Z); if (! A and ! Z) then
FAIL(...{A,Z}...);
}
match(Z);
With input X
old message: Found X, expected Z
new message: Found X, expected A, Z
For the example:
star2 : (A|B)* Z;
old new (with -mrblkerr)
------------- --------------------
for (;;) { for (;;) {
if (!A and !B) break; if (!A and !B) break;
if (...) { if (...) {
<same ...> <same ...>
} }
else { else {
FAIL(...{A,B,Z}...) FAIL(...{A,B}...);
} }
} }
match(B); if (! A and ! B and !Z) then
FAIL(...{A,B,Z}...);
}
match(B);
With input X
old message: Found X, expected Z
new message: Found X, expected A, B, Z
With input A X
old message: Found X, expected Z
new message: Found X, expected A, B, Z
This includes the choice of looping back to the
star block.
The code for plus blocks:
plus1 : (A)+ Z;
The generated code resembles:
old new (with -mrblkerr)
------------- --------------------
do { do {
match(A); match(A);
} while (A) } while (A)
match(Z); if (! A and ! Z) then
FAIL(...{A,Z}...);
}
match(Z);
With input A X
old message: Found X, expected Z
new message: Found X, expected A, Z
This includes the choice of looping back to the
plus block.
For the example:
plus2 : (A|B)+ Z;
old new (with -mrblkerr)
------------- --------------------
do { do {
if (A) { <same>
match(A); <same>
} else if (B) { <same>
match(B); <same>
} else { <same>
if (cnt > 1) break; <same>
FAIL(...{A,B,Z}...) FAIL(...{A,B}...);
} }
cnt++; <same>
} }
match(Z); if (! A and ! B and !Z) then
FAIL(...{A,B,Z}...);
}
match(B);
With input X
old message: Found X, expected A, B, Z
new message: Found X, expected A, B
With input A X
old message: Found X, expected Z
new message: Found X, expected A, B, Z
This includes the choice of looping back to the
star block.
#249. (MR21) Changes for DEC/VMS systems
Jean-Fran�ois Pi�ronne (jfp altavista.net) has updated some
VMS related command files and fixed some minor problems related
users the most important differences are:
a. Revised makefile.vms
b. Revised genMMS for genrating VMS style makefiles.
#248. (MR21) Generate symbol for first set of an alternative
pccts can generate a symbol which represents the tokens which may
appear at the start of a block:
rr : #FirstSetSymbol(rr_FirstSet) ( Foo | Bar ) ;
This will generate the symbol rr_FirstSet of type SetWordType with
elements Foo and Bar set. The bits can be tested using code similar
to the following:
if (set_el(Foo, &rr_FirstSet)) { ...
This can be combined with the C array zztokens[] or the C++ routine
tokenName() to get the print name of the token in the first set.
The size of the set is given by the newly added enum SET_SIZE, a
protected member of the generated parser's class. The number of
elements in the generated set will not be exactly equal to the
value of SET_SIZE because of synthetic tokens created by #tokclass,
#errclass, the -ge option, and meta-tokens such as epsilon, and
end-of-file.
The #FirstSetSymbol must appear immediately before a block
such as (...)+, (...)*, and {...}, and (...). It may not appear
immediately before a token, a rule reference, or action. However
a token or rule reference can be enclosed in a (...) in order to
make the use of #pragma FirstSetSymbol legal.
rr_bad : #FirstSetSymbol(rr_bad_FirstSet) Foo; // Illegal
rr_ok : #FirstSetSymbol(rr_ok_FirstSet) (Foo); // Legal
Do not confuse FirstSetSymbol sets with the sets used for testing
lookahead. The sets used for FirstSetSymbol have one element per bit,
so the number of bytes is approximately the largest token number
divided by 8. The sets used for testing lookahead store 8 lookahead
sets per byte, so the length of the array is approximately the largest
token number.
If there is demand, a similar routine for follow sets can be added.
#247. (MR21) Misleading error message on syntax error for optional elements.
===================================================
The behavior has been revised when parser exception
handling is used. See Item #290
===================================================
Prior to MR21, tokens which were optional did not appear in syntax
error messages if the block which immediately followed detected a
syntax error.
Consider the following grammar which accepts Number, Word, and Other:
rr : {Number} Word;
For this rule the code resembles:
if (LA(1) == Number) {
match(Number);
consume();
}
match(Word);
Prior to MR21, the error message for input "$ a" would be:
line 1: syntax error at "$" missing Word
With MR21 the message will be:
line 1: syntax error at "$" expecting Word, Number.
The generate code resembles:
if ( (LA(1)==Number) ) {
zzmatch(Number);
consume();
}
else {
if ( (LA(1)==Word) ) {
/* nothing */
}
else {
FAIL(... message for both Number and Word ...);
}
}
match(Word);
The code generated for optional blocks in MR21 is slightly longer
than the previous versions, but it should give better error messages.
The code generated for:
{ a | b | c }
should now be *identical* to:
( a | b | c | )
which was not the case prior to MR21.
Reported by Sue Marvin (sue siara.com).
#246. (Changed in MR21) Use of $(MAKE) for calls to make
Calls to make from the makefiles were replaced with $(MAKE)
because of problems when using gmake.
Reported with fix by Sunil K.Vallamkonda (sunil siara.com).
#245. (Changed in MR21) Changes to genmk
The following command line options have been added to genmk:
-cfiles ...
To add a user's C or C++ files into makefile automatically.
The list of files must be enclosed in apostrophes. This
option may be specified multiple times.
-compiler ...
The name of the compiler to use for $(CCC) or $(CC). The
default in C++ mode is "CC". The default in C mode is "cc".
-pccts_path ...
The value for $(PCCTS), the pccts directory. The default
is /usr/local/pccts.
Contributed by Tomasz Babczynski (t.babczynski ict.pwr.wroc.pl).
#244. (Changed in MR21) Rename variable "not" in antlr.g
When antlr.g is compiled with a C++ compiler, a variable named
"not" causes problems. Reported by Sinan Karasu
#243 (Changed in MR21) Replace recursion with iteration in zzfree_ast
Another refinement to zzfree_ast in ast.c to limit recursion.
NAKAJIMA Mutsuki (muc isr.co.jp).
#242. (Changed in MR21) LineInfoFormatStr
Added an #ifndef/#endif around LineInfoFormatStr in pcctscfg.h.
#241. (Changed in MR21) Changed macro PURIFY to a no-op
***********************
*** NOT IMPLEMENTED ***
***********************
The PURIFY macro was changed to a no-op because it was causing
problems when passing C++ objects.
The old definition:
#define PURIFY(r,s) memset((char *) &(r),'\\0',(s));
The new definition:
#define PURIFY(r,s) /* nothing */
#endif
#240. (Changed in MR21) sorcerer/h/sorcerer.h _MATCH and _MATCHRANGE
Added test for NULL token pointer.
Suggested by Peter Keller (keller ebi.ac.uk)
#239. (Changed in MR21) C++ mode AParser::traceGuessFail
If tracing is turned on when the code has been generated
without trace code, a failed guess generates a trace report
even though there are no other trace reports. This
make the behavior consistent with other parts of the
trace system.
Reported by David Wigg (wiggjd sbu.ac.uk).
#238. (Changed in MR21) Namespace version #include files
Changed reference from CStdio to cstdio (and other
#include file names) in the namespace version of pccts.
Should have known better.
#237. (Changed in MR21) ParserBlackBox(FILE*)
In the past, ParserBlackBox would close the FILE in the dtor
even though it was not opened by ParserBlackBox. The problem
is that there were two constructors, one which accepted a file
name and did an fopen, the other which accepted a FILE and did
not do an fopen. There is now an extra member variable which
remembers whether ParserBlackBox did the open or not.
Suggested by Mike Percy (mpercy scires.com).
#236. (Changed in MR21) tmake now reports down pointer problem
When ASTBase::tmake attempts to update the down pointer of
an AST it checks to see if the down pointer is NULL. If it
is not NULL it does not do the update and returns NULL.
An attempt to update the down pointer is almost always a
result of a user error. This can lead to difficult to find
problems during tree construction.
With this change, the routine calls a virtual function
reportOverwriteOfDownPointer() which calls panic to
report the problem. Users who want the old behavior can
redefined the virtual function in their AST class.
Suggested by Sinan Karasu (sinan.karasu boeing.com)
#235. (Changed in MR21) Made ANTLRParser::resynch() virtual
Suggested by Jerry Evans (jerry swsl.co.uk).
#234. (Changed in MR21) Implicit int for function return value
ATokenBuffer:bufferSize() did not specify a type for the
return value.
Reported by Hai Vo-Ba (hai fc.hp.com).
#233. (Changed in MR20) Converted to MSVC 6.0
Due to external circumstances I have had to convert to MSVC 6.0
The MSVC 5.0 project files (.dsw and .dsp) have been retained as
and xxx60.dsw (where xxx is the related to the directory/project).
#232. (Changed in MR20) Make setwd bit vectors protected in parser.h
The access for the setwd array in the parser header was not
specified. As a result, it would depend on the code which
preceded it. In MR20 it will always have access "protected".
Reported by Piotr Eljasiak (eljasiak zt.gdansk.tpsa.pl).
#231. (Changed in MR20) Error in token buffer debug code.
When token buffer debugging is selected via the pre-processor
symbol DEBUG_TOKENBUFFER there is an erroneous check in
#ifdef DEBUG_TOKENBUFFER
if (i >= inputTokens->bufferSize() ||
inputTokens->minTokens() < LLk ) /* MR20 Was "<=" */
...
#endif
Reported by David Wigg (wiggjd sbu.ac.uk).
#230. (Changed in MR20) Fixed problem with #define for -gd option
There was an error in setting zzTRACE_RULES for the -gd (trace) option.
Reported by Gary Funck (gary intrepid.com).
#229. (Changed in MR20) Additional "const" for literals
"const" was added to the token name literal table.
"const" was added to some panic() and similar routine
#228. (Changed in MR20) dlg crashes on "()"
The following token defintion will cause DLG to crash.
#token "()"
When there is a syntax error in a regular expression
many of the dlg routines return a structure which has
null pointers. When this is accessed by callers it
generates the crash.
I have attempted to fix the more common cases.
Reported by Mengue Olivier (dolmen bigfoot.com).
#227. (Changed in MR20) Array overwrite
Steveh Hand (sassth unx.sas.com) reported a problem which
was traced to a temporary array which was not properly
resized for deeply nested blocks. This has been fixed.
#226. (Changed in MR20) -pedantic conformance
G. Hobbelt (i_a mbh.org) and THM made many, many minor
changes to create prototypes for all the functions and
bring antlr, dlg, and sorcerer into conformance with
the gcc -pedantic option.
This may require uses to add pccts/h/pcctscfg.h to some
files or makefiles in order to have __USE_PROTOS defined.
#225 (Changed in MR20) AST stack adjustment in C mode
The fix in #214 for AST stack adjustment in C mode missed
some cases.
Reported with fix by Ger Hobbelt (i_a mbh.org).
#224 (Changed in MR20) LL(1) and LL(2) with #pragma approx
This may take a record for the oldest, most trival, lexical
error in pccts. The regular expressions for LL(1) and LL(2)
lacked an escape for the left and right parenthesis.
Reported by Ger Hobbelt (i_a mbh.org).
#223 (Changed in MR20) Addition of IBM_VISUAL_AGE directory
Build files for antlr, dlg, and sorcerer under IBM Visual Age
have been contributed by Anton Sergeev (ags mlc.ru). They have
been placed in the pccts/IBM_VISUAL_AGE directory.
#222 (Changed in MR20) Replace __STDC__ with __USE_PROTOS
Most occurrences of __STDC__ replaced with __USE_PROTOS due to
complaints from several users.
#221 (Changed in MR20) Added #include for DLexerBase.h to PBlackBox.
Added #include for DLexerBase.h to PBlackBox.
#220 (Changed in MR19) strcat arguments reversed in #pred parse
The arguments to strcat are reversed when creating a print
name for a hash table entry for use with #pred feature.
Problem diagnosed and fix reported by Scott Harrington
(seh4 ix.netcom.com).
#219. (Changed in MR19) C Mode routine zzfree_ast
Changes to reduce use of recursion for AST trees with only right
links or only left links in the C mode routine zzfree_ast.
Implemented by SAKAI Kiyotaka (ksakai isr.co.jp).
#218. (Changed in MR19) Changes to support unsigned char in C mode
Implemented by SAKAI Kiyotaka (ksakai isr.co.jp).
#217. (Changed in MR19) Error message when dlg -i and -CC options selected
*** This change was rescinded by item #257 ***
The parsers generated by pccts in C++ mode are not able to support the
interactive lexer option (except, perhaps, when using the deferred fetch
parser option.(Item #216).
DLG now warns when both -i and -CC are selected.
This warning was suggested by David Venditti (07751870267-0001 t-online.de).
#216. (Changed in MR19) Defer token fetch for C++ mode
Implemented by Volker H. Simonis (simonis informatik.uni-tuebingen.de)
Normally, pccts keeps the lookahead token buffer completely filled.
This requires max(k,ck) tokens of lookahead. For some applications
this can cause deadlock problems. For example, there may be cases
when the parser can't tell when the input has been completely consumed
until the parse is complete, but the parse can't be completed because
the input routines are waiting for additional tokens to fill the
lookahead buffer.
When the ANTLRParser class is built with the pre-processor option
ZZDEFER_FETCH defined, the fetch of new tokens by consume() is deferred
until LA(i) or LT(i) is called.
To test whether this option has been built into the ANTLRParser class
use "isDeferFetchEnabled()".
Using the -gd trace option with the default tracein() and traceout()
routines will defeat the effort to defer the fetch because the
trace routines print out information about the lookahead token at
the start of the rule.
Because the tracein and traceout routines are virtual it is
easy to redefine them in your parser:
class MyParser {
<<
virtual void tracein(ANTLRChar * ruleName)
{ fprintf(stderr,"Entering: %s\n", ruleName); }
virtual void traceout(ANTLRChar * ruleName)
{ fprintf(stderr,"Leaving: %s\n", ruleName); }
>>
The originals for those routines are pccts/h/AParser.cpp
This requires use of the dlg option -i (interactive lexer).
This is implemented only for C++ mode.
This is experimental. The interaction with guess mode (syntactic
predicates)is not known.
#215. (Changed in MR19) Addition of reset() to DLGLexerBase
There was no obvious way to reset the lexer for reuse. The
reset() method now does this.
Suggested by David Venditti (07751870267-0001 t-online.de).
#214. (Changed in MR19) C mode: Adjust AST stack pointer at exit
In C mode the AST stack pointer needs to be reset if there will
be multiple calls to the ANTLRx macros.
Reported with fix by Paul D. Smith (psmith baynetworks.com).
#213. (Changed in MR18) Fatal error with -mrhoistk (k>1 hoisting)
When rearranging code I forgot to un-comment a critical line of
code that handles hoisting of predicates with k>1 lookahead. This
is now fixed.
Reported by Reinier van den Born (reinier vnet.ibm.com).
#212. (Changed in MR17) Mac related changes by Kenji Tanaka
Kenji Tanaka (kentar osa.att.ne.jp) has made a number of changes for
Macintosh users.
a. The following Macintosh MPW files aid in installing pccts on Mac:
They completely replace the previous Mac installation files.
b. The most significant is a change in the MAC_FILE_CREATOR symbol
in pcctscfg.h:
old: #define MAC_FILE_CREATOR 'MMCC' /* Metrowerks C/C++ Text files */
new: #define MAC_FILE_CREATOR 'CWIE' /* Metrowerks C/C++ Text files */
c. Added calls to special_fopen_actions() where necessary.
#211. (Changed in MR16a) C++ style comment in dlg
This has been fixed.
#210. (Changed in MR16a) Sor accepts \r\n, \r, or \n for end-of-line
A user requested that Sorcerer be changed to accept other forms
of end-of-line.
#209. (Changed in MR16) Name of files changed.
Old: CHANGES_FROM_1.33
New: CHANGES_FROM_133.txt
Old: KNOWN_PROBLEMS
New: KNOWN_PROBLEMS.txt
#208. (Changed in MR16) Change in use of pccts #include files
There were problems with MS DevStudio when mixing Sorcerer and
PCCTS in the same source file. The problem is caused by the
redefinition of setjmp in the MS header file setjmp.h. In
setjmp.h the pre-processor symbol setjmp was redefined to be
_setjmp. A later effort to execute #include <setjmp.h> resulted
in an effort to #include <_setjmp.h>. I'm not sure whether this
is a bug or a feature. In any case, I decided to fix it by
avoiding the use of pre-processor symbols in #include statements
altogether. This has the added benefit of making pre-compiled
headers work again.
I've replaced statements:
old: #include PCCTS_SETJMP_H
new: #include "pccts_setjmp.h"
Where pccts_setjmp.h contains:
#ifndef __PCCTS_SETJMP_H__
#define __PCCTS_SETJMP_H__
#ifdef PCCTS_USE_NAMESPACE_STD
#include <Csetjmp>
#else
#include <setjmp.h>
#endif
#endif
A similar change has been made for other standard header files
Reported by Jeff Vincent (JVincent novell.com) and Dale Davis
(DalDavis spectrace.com).
#207. (Changed in MR16) dlg reports an invalid range for: [\0x00-\0xff]
-----------------------------------------------------------------
Note from MR23: This fix does not work. I am investigating why.
-----------------------------------------------------------------
dlg will report that this is an invalid range.
Diagnosed by Piotr Eljasiak (eljasiak no-spam.zt.gdansk.tpsa.pl):
I think this problem is not specific to unsigned chars
because dlg reports no error for the range [\0x00-\0xfe].
I've found that information on range is kept in field
letter (unsigned char) of Attrib struct. Unfortunately
the letter value internally is for some reasons increased
by 1, so \0xff is represented here as 0.
That's why dlg complains about the range [\0x00-\0xff] in
if ($$.letter > $2.letter) {
error("invalid range ", zzline);
}
The fix is:
error("invalid range ", zzline);
}
#206. (Changed in MR16) Free zzFAILtext in ANTLRParser destructor
The ANTLRParser destructor now frees zzFAILtext.
Problem and fix reported by Manfred Kogler (km cast.uni-linz.ac.at).
#205. (Changed in MR16) DLGStringReset argument now const
Changed: void DLGStringReset(DLGChar *s) {...}
To: void DLGStringReset(const DLGChar *s) {...}
Suggested by Dale Davis (daldavis spectrace.com)
#204. (Changed in MR15a) Change __WATCOM__ to __WATCOMC__ in pcctscfg.h
Reported by Oleg Dashevskii (olegdash my-dejanews.com).
#203. (Changed in MR15) Addition of sorcerer to distribution kit
I have finally caved in to popular demand. The pccts 1.33mr15
kit will include sorcerer. The separate sorcerer kit will be
discontinued.
#202. (Changed) in MR15) Organization of MS Dev Studio Projects in Kit
Previously there was one workspace that contained projects for
all three parts of pccts: antlr, dlg, and sorcerer. Now each
part (and directory) has its own workspace/project and there
is an additional workspace/project to build a library from the
.cpp files in the pccts/h directory.
The library build will create pccts_debug.lib or pccts_release.lib
according to the configuration selected.
If you don't want to build pccts 1.33MR15 you can download a
ready-to-run kit for win32 from http://www.polhode.com/win32.zip.
The ready-to-run for win32 includes executables, a pre-built static
library for the .cpp files in the pccts/h directory, and a sample
application
You will need to define the environment variable PCCTS to point to
the root of the pccts directory hierarchy.
#201. (Changed in MR15) Several fixes by K.J. Cummings (cummings peritus.com)
Generation of SETJMP rather than SETJMP_H in gen.c.
(Sor B19) Declaration of ref_vars_inits for ref_var_inits in
#200. (Changed in MR15) Remove operator=() in AToken.h
User reported that WatCom couldn't handle use of
explicit operator =(). Replace with equivalent
using cast operator.
#199. (Changed in MR15) Don't allow use of empty #tokclass
Change antlr.g to disallow empty #tokclass sets.
Reported by Manfred Kogler (km cast.uni-linz.ac.at).
#198. Revised ANSI C grammar due to efforts by Manuel Kessler
Manuel Kessler (mlkessler cip.physik.uni-wuerzburg.de)
Allow trailing ... in function parameter lists.
Add bit fields.
Allow old-style function declarations.
Support cv-qualified pointers.
Better checking of combinations of type specifiers.
Release of memory for local symbols on scope exit.
Allow input file name on command line as well as by redirection.
and other miscellaneous tweaks.
This is not part of the pccts distribution kit. It must be
downloaded separately from:
#197. (Changed in MR14) Resetting the lookahead buffer of the parser
Explanation and fix by Sinan Karasu (sinan.karasu boeing.com)
Consider the code used to prime the lookahead buffer LA(i)
of the parser when init() is called:
void
ANTLRParser::
prime_lookahead()
{
int i;
for(i=1;i<=LLk; i++) consume();
dirty=0;
//lap = 0; // MR14 - Sinan Karasu (sinan.karusu boeing.com)
//labase = 0; // MR14
labase=lap; // MR14
}
When the parser is instantiated, lap=0,labase=0 is set.
The "for" loop runs LLk times. In consume(), lap = lap +1 (mod LLk) is
computed. Therefore, lap(before the loop) == lap (after the loop).
Now the only problem comes in when one does an init() of the parser
after an Eof has been seen. At that time, lap could be non zero.
Assume it was lap==1. Now we do a prime_lookahead(). If LLk is 2,
then
consume()
{
NLA = inputTokens->getToken()->getType();
dirty--;
lap = (lap+1)&(LLk-1);
}
or expanding NLA,
token_type[lap&(LLk-1)]) = inputTokens->getToken()->getType();
dirty--;
lap = (lap+1)&(LLk-1);
so now we prime locations 1 and 2. In prime_lookahead it used to set
lap=0 and labase=0. Now, the next token will be read from location 0,
NOT 1 as it should have been.
This was never caught before, because if a parser is just instantiated,
then lap and labase are 0, the offending assignment lines are
basically no-ops, since the for loop wraps around back to 0.
#196. (Changed in MR14) Problems with "(alpha)? beta" guess
Consider the following syntactic predicate in a grammar
with 2 tokens of lookahead (k=2 or ck=2):
rule : ( alpha )? beta ;
alpha : S t ;
t : T U
| T
;
beta : S t Z ;
When antlr computes the prediction expression with one token
of lookahead for alts 1 and 2 of rule t it finds an ambiguity.
Because the grammar has a lookahead of 2 it tries to compute
two tokens of lookahead for alts 1 and 2 of t. Alt 1 clearly
has a lookahead of (T U). Alt 2 is one token long so antlr
tries to compute the follow set of alt 2, which means finding
the things which can follow rule t in the context of (alpha)?.
This cannot be computed, because alpha is only part of a rule,
and antlr can't tell what part of beta is matched by alpha and
what part remains to be matched. Thus it impossible for antlr
to properly determine the follow set of rule t.
Prior to 1.33MR14, the follow of (alpha)? was computed as
FIRST(beta) as a result of the internal representation of
guess blocks.
With MR14 the follow set will be the empty set for that context.
Normally, one expects a rule appearing in a guess block to also
appear elsewhere. When the follow context for this other use
is "ored" with the empty set, the context from the other use
results, and a reasonable follow context results. However if
there is *no* other use of the rule, or it is used in a different
manner then the follow context will be inaccurate - it was
inaccurate even before MR14, but it will be inaccurate in a
different way.
For the example given earlier, a reasonable way to rewrite the
grammar:
rule : ( alpha )? beta
alpha : S t ;
t : T U
| T
;
beta : alpha Z ;
If there are no other uses of the rule appearing in the guess
block it will generate a test for EOF - a workaround for
representing a null set in the lookahead tests.
If you encounter such a problem you can use the -alpha option
to get additional information:
line 2: error: not possible to compute follow set for alpha
in an "(alpha)? beta" block.
With the antlr -alpha command line option the following information
is inserted into the generated file:
#if 0
Trace of references leading to attempt to compute the follow set of
alpha in an "(alpha)? beta" block. It is not possible for antlr to
compute this follow set because it is not known what part of beta has
already been matched by alpha and what part remains to be matched.
Rules which make use of the incorrect follow set will also be incorrect
1 #token T alpha/2 line 7 brief.g
2 end alpha alpha/3 line 8 brief.g
2 end (...)? block at start/1 line 2 brief.g
#endif
At the moment, with the -alpha option selected the program marks
any rules which appear in the trace back chain (above) as rules with
possible problems computing follow set.
Reported by Greg Knapen (gregory.knapen bell.ca).
#195. (Changed in MR14) #line directive not at column 1
Under certain circunstances a predicate test could generate
a #line directive which was not at column 1.
Reported with fix by David K�gedal (davidk lysator.liu.se)
#194. (Changed in MR14) (C Mode only) Demand lookahead with #tokclass
In C mode with the demand lookahead option there is a bug in the
code which handles matches for #tokclass (zzsetmatch and
zzsetmatch_wsig).
The bug causes the lookahead pointer to get out of synchronization
with the current token pointer.
The problem was reported with a fix by Ger Hobbelt (hobbelt axa.nl).
#193. (Changed in MR14) Use of PCCTS_USE_NAMESPACE_STD
The pcctscfg.h now contains the following definitions:
#ifdef PCCTS_USE_NAMESPACE_STD
#define PCCTS_STDIO_H <Cstdio>
#define PCCTS_STDLIB_H <Cstdlib>
#define PCCTS_STDARG_H <Cstdarg>
#define PCCTS_SETJMP_H <Csetjmp>
#define PCCTS_STRING_H <Cstring>
#define PCCTS_ASSERT_H <Cassert>
#define PCCTS_ISTREAM_H <istream>
#define PCCTS_IOSTREAM_H <iostream>
#define PCCTS_NAMESPACE_STD namespace std {}; using namespace std;
#else
#define PCCTS_STDIO_H <stdio.h>
#define PCCTS_STDLIB_H <stdlib.h>
#define PCCTS_STDARG_H <stdarg.h>
#define PCCTS_SETJMP_H <setjmp.h>
#define PCCTS_STRING_H <string.h>
#define PCCTS_ASSERT_H <assert.h>
#define PCCTS_ISTREAM_H <istream.h>
#define PCCTS_IOSTREAM_H <iostream.h>
#define PCCTS_NAMESPACE_STD
#endif
The runtime support in pccts/h uses these pre-processor symbols
consistently.
Also, antlr and dlg have been changed to generate code which uses
these pre-processor symbols rather than having the names of the
#include files hard-coded in the generated code.
This required the addition of "#include pcctscfg.h" to a number of
files in pccts/h.
It appears that this sometimes causes problems for MSVC 5 in
combination with the "automatic" option for pre-compiled headers.
In such cases disable the "automatic" pre-compiled headers option.
Suggested by Hubert Holin (Hubert.Holin Bigfoot.com).
#192. (Changed in MR14) Change setText() to accept "const ANTLRChar *"
Changed ANTLRToken::setText(ANTLRChar *) to setText(const ANTLRChar *).
This allows literal strings to be used to initialize tokens. Since
the usual token implementation (ANTLRCommonToken) makes a copy of the
input string, this was an unnecessary limitation.
Suggested by Bob McWhirter (bob netwrench.com).
#191. (Changed in MR14) HP/UX aCC compiler compatibility problem
Needed to explicitly declare zzINF_DEF_TOKEN_BUFFER_SIZE and
zzINF_BUFFER_TOKEN_CHUNK_SIZE as ints in pccts/h/AParser.cpp.
Reported by David Cook (dcook bmc.com).
#190. (Changed in MR14) IBM OS/2 CSet compiler compatibility problem
Name conflict with "_cs" in pccts/h/ATokenBuffer.cpp
Reported by David Cook (dcook bmc.com).
#189. (Changed in MR14) -gxt switch in C mode
The -gxt switch in C mode didn't work because of incorrect
initialization.
Reported by Sinan Karasu (sinan boeing.com).
#188. (Changed in MR14) Added pccts/h/DLG_stream_input.h
This is a DLG stream class based on C++ istreams.
Contributed by Hubert Holin (Hubert.Holin Bigfoot.com).
#187. (Changed in MR14) Rename config.h to pcctscfg.h
The PCCTS configuration file has been renamed from config.h to
pcctscfg.h. The problem with the original name is that it led
to name collisions when pccts parsers were combined with other
software.
All of the runtime support routines in pccts/h/* have been
changed to use the new name. Existing software can continue
to use pccts/h/config.h. The contents of pccts/h/config.h is
now just "#include "pcctscfg.h".
I don't have a record of the user who suggested this.
#186. (Changed in MR14) Pre-processor symbol DllExportPCCTS class modifier
Classes in the C++ runtime support routines are now declared:
class DllExportPCCTS className ....
By default, the pre-processor symbol is defined as the empty
string. This if for use by MSVC++ users to create DLL classes.
Suggested by Manfred Kogler (km cast.uni-linz.ac.at).
#185. (Changed in MR14) Option to not use PCCTS_AST base class for ASTBase
Normally, the ASTBase class is derived from PCCTS_AST which contains
functions useful to Sorcerer. If these are not necessary then the
user can define the pre-processor symbol "PCCTS_NOT_USING_SOR" which
will cause the ASTBase class to replace references to PCCTS_AST with
references to ASTBase where necessary.
The class ASTDoublyLinkedBase will contain a pure virtual function
shallowCopy() that was formerly defined in class PCCTS_AST.
Suggested by Bob McWhirter (bob netwrench.com).
#184. (Changed in MR14) Grammars with no tokens generate invalid tokens.h
Reported by Hubert Holin (Hubert.Holin bigfoot.com).
#183. (Changed in MR14) -f to specify file with names of grammar files
In DEC/VMS it is difficult to specify very long command lines.
The -f option allows one to place the names of the grammar files
in a data file in order to bypass limitations of the DEC/VMS
command language interpreter.
Addition supplied by Bernard Giroud (b_giroud decus.ch).
#182. (Changed in MR14) Output directory option for DEC/VMS
Fix some problems with the -o option under DEC/VMS.
Fix supplied by Bernard Giroud (b_giroud decus.ch).
#181. (Changed in MR14) Allow chars > 127 in DLGStringInput::nextChar()
Changed DLGStringInput to cast the character using (unsigned char)
so that languages with character codes greater than 127 work
without changes.
Suggested by Manfred Kogler (km cast.uni-linz.ac.at).
#180. (Added in MR14) ANTLRParser::getEofToken()
Added "ANTLRToken ANTLRParser::getEofToken() const" to match the
setEofToken routine.
Requested by Manfred Kogler (km cast.uni-linz.ac.at).
#179. (Fixed in MR14) Memory leak for BufFileInput subclass of DLGInputStream
The BufFileInput class described in Item #142 neglected to release
the allocated buffer when an instance was destroyed.
Reported by Manfred Kogler (km cast.uni-linz.ac.at).
#178. (Fixed in MR14) Bug in "(alpha)? beta" guess blocks first sets
In 1.33 vanilla, and all maintenance releases prior to MR14
there is a bug in the handling of guess blocks which use the
"long" form:
(alpha)? beta
inside a (...)*, (...)+, or {...} block.
This problem does *not* apply to the case where beta is omitted
or when the syntactic predicate is on the leading edge of an
alternative.
The problem is that both alpha and beta are stored in the
syntax diagram, and that some analysis routines would fail
to skip the alpha portion when it was not on the leading edge.
Consider the following grammar with -ck 2:
r : ( (A)? B )* C D
| A B /* forces -ck 2 computation for old antlr */
/* reports ambig for alts 1 & 2 */
| B C /* forces -ck 2 computation for new antlr */
/* reports ambig for alts 1 & 3 */
;
The prediction expression for the first alternative should be
LA(1)={B C} LA(2)={B C D}, but previous versions of antlr
would compute the prediction expression as LA(1)={A C} LA(2)={B D}
Reported by Arpad Beszedes (beszedes inf.u-szeged.hu) who provided
a very clear example of the problem and identified the probable cause.
#177. (Changed in MR14) #tokdefs and #token with regular expression
In MR13 the change described by Item #162 caused an existing
feature of antlr to fail. Prior to the change it was possible
to give regular expression definitions and actions to tokens
which were defined via the #tokdefs directive.
This now works again.
Reported by Manfred Kogler (km cast.uni-linz.ac.at).
#176. (Changed in MR14) Support for #line in antlr source code
Note: this was implemented by Arpad Beszedes (beszedes inf.u-szeged.hu).
In 1.33MR14 it is possible for a pre-processor to generate #line
directives in the antlr source and have those line numbers and file
names used in antlr error messages and in the #line directives
generated by antlr.
The #line directive may appear in the following forms:
#line ll "sss" xx xx ...
where ll represents a line number, "sss" represents the name of a file
enclosed in quotation marks, and xxx are arbitrary integers.
The following form (without "line") is not supported at the moment:
# ll "sss" xx xx ...
The result:
zzline
is replaced with ll from the # or #line directive
FileStr[CurFile]
is updated with the contents of the string (if any)
following the line number
Note
----
The file-name string following the line number can be a complete
name with a directory-path. Antlr generates the output files from
the input file name (by replacing the extension from the file-name
with .c or .cpp).
If the input file (or the file-name from the line-info) contains
a path:
"../grammar.g"
the generated source code will be placed in "../grammar.cpp" (i.e.
in the parent directory). This is inconvenient in some cases
(even the -o switch can not be used) so the path information is
removed from the #line directive. Thus, if the line-info was
#line 2 "../grammar.g"
then the current file-name will become "grammar.g"
In this way, the generated source code according to the grammar file
will always be in the current directory, except when the -o switch
is used.
#175. (Changed in MR14) Bug when guess block appears at start of (...)*
In 1.33 vanilla and all maintenance releases prior to 1.33MR14
there is a bug when a guess block appears at the start of a (...)+.
Consider the following k=1 (ck=1) grammar:
rule :
( (STAR)? ZIP )* ID ;
Prior to 1.33MR14, the generated code resembled:
...
zzGUESS_BLOCK
while ( 1 ) {
if ( ! LA(1)==STAR) break;
zzGUESS
if ( !zzrv ) {
zzmatch(STAR);
zzCONSUME;
zzGUESS_DONE
zzmatch(ZIP);
zzCONSUME;
...
Note that the routine uses STAR for the prediction expression
rather than ZIP. With 1.33MR14 the generated code resembles:
...
while ( 1 ) {
if ( ! LA(1)==ZIP) break;
...
This problem existed only with (...)* blocks and was caused
by the slightly more complicated graph which represents (...)*
blocks. This caused the analysis routine to compute the first
set for the alpha part of the "(alpha)? beta" rather than the
beta part.
Both (...)+ and {...} blocks handled the guess block correctly.
Reported by Arpad Beszedes (beszedes inf.u-szeged.hu) who provided
a very clear example of the problem and identified the probable cause.
#174. (Changed in MR14) Bug when action precedes syntactic predicate
In 1.33 vanilla, and all maintenance releases prior to MR14,
there was a bug when a syntactic predicate was immediately
preceded by an action. Consider the following -ck 2 grammar:
rule :
<<int i;>>
(alpha)? beta C
| A B
;
alpha : A ;
beta : A B;
Prior to MR14, the code generated for the first alternative
resembled:
...
zzGUESS
if ( !zzrv && LA(1)==A && LA(2)==A) {
alpha();
zzGUESS_DONE
beta();
zzmatch(C);
zzCONSUME;
} else {
...
The prediction expression (i.e. LA(1)==A && LA(2)==A) is clearly
wrong because LA(2) should be matched to B (first[2] of beta is {B}).
With 1.33MR14 the prediction expression is:
...
if ( !zzrv && LA(1)==A && LA(2)==B) {
alpha();
zzGUESS_DONE
beta();
zzmatch(C);
zzCONSUME;
} else {
...
This will only affect users in which alpha is shorter than
than max(k,ck) and there is an action immediately preceding
the syntactic predicate.
This problem was reported by reported by Arpad Beszedes
(beszedes inf.u-szeged.hu) who provided a very clear example
of the problem and identified the presence of the init-action
as the likely culprit.
#173. (Changed in MR13a) -glms for Microsoft style filenames with -gl
With the -gl option antlr generates #line directives using the
exact name of the input files specified on the command line.
An oddity of the Microsoft C and C++ compilers is that they
don't accept file names in #line directives containing "\"
even though these are names from the native file system.
With -glms option, the "\" in file names appearing in #line
directives is replaced with a "/" in order to conform to
Microsoft compiler requirements.
Reported by Erwin Achermann (erwin.achermann switzerland.org).
#172. (Changed in MR13) \r\n in antlr source counted as one line
Some MS software uses \r\n to indicate a new line. Antlr
now recognizes this in counting lines.
Reported by Edward L. Hepler (elh ece.vill.edu).
#171. (Changed in MR13) #tokclass L..U now allowed
The following is now allowed:
#tokclass ABC { A..B C }
Reported by Dave Watola (dwatola amtsun.jpl.nasa.gov)
#170. (Changed in MR13) Suppression for predicates with lookahead depth >1
In MR12 the capability for suppression of predicates with lookahead
depth=1 was introduced. With MR13 this had been extended to
predicates with lookahead depth > 1 and released for use by users
on an experimental basis.
Consider the following grammar with -ck 2 and the predicate in rule
"a" with depth 2:
r1 : (ab)* "@"
;
ab : a
| b
;
a : (A B)? => <<p(LATEXT(2))>>? A B C
;
b : A B C
;
Normally, the predicate would be hoisted into rule r1 in order to
determine whether to call rule "ab". However it should *not* be
hoisted because, even if p is false, there is a valid alternative
in rule b. With "-mrhoistk on" the predicate will be suppressed.
If "-info p" command line option is present the following information
will appear in the generated code:
while ( (LA(1)==A)
#if 0
Part (or all) of predicate with depth > 1 suppressed by alternative
without predicate
pred << p(LATEXT(2))>>?
depth=k=2 ("=>" guard) rule a line 8 t1.g
tree context:
(root = A
B
)
The token sequence which is suppressed: ( A B )
The sequence of references which generate that sequence of tokens:
1 to ab r1/1 line 1 t1.g
2 ab ab/1 line 4 t1.g
3 to b ab/2 line 5 t1.g
4 b b/1 line 11 t1.g
5 #token A b/1 line 11 t1.g
6 #token B b/1 line 11 t1.g
#endif
A slightly more complicated example:
r1 : (ab)* "@"
;
ab : a
| b
;
a : (A B)? => <<p(LATEXT(2))>>? (A B | D E)
;
b : <<q(LATEXT(2))>>? D E
;
In this case, the sequence (D E) in rule "a" which lies behind
the guard is used to suppress the predicate with context (D E)
in rule b.
while ( (LA(1)==A || LA(1)==D)
#if 0
Part (or all) of predicate with depth > 1 suppressed by alternative
without predicate
pred << q(LATEXT(2))>>?
depth=k=2 rule b line 11 t2.g
tree context:
(root = D
E
)
The token sequence which is suppressed: ( D E )
The sequence of references which generate that sequence of tokens:
1 to ab r1/1 line 1 t2.g
2 ab ab/1 line 4 t2.g
3 to a ab/1 line 4 t2.g
4 a a/1 line 8 t2.g
5 #token D a/1 line 8 t2.g
6 #token E a/1 line 8 t2.g
#endif
&&
#if 0
pred << p(LATEXT(2))>>?
depth=k=2 ("=>" guard) rule a line 8 t2.g
tree context:
(root = A
B
)
#endif
(! ( LA(1)==A && LA(2)==B ) || p(LATEXT(2)) ) {
ab();
...
#169. (Changed in MR13) Predicate test optimization for depth=1 predicates
When the MR12 generated a test of a predicate which had depth 1
it would use the depth >1 routines, resulting in correct but
inefficient behavior. In MR13, a bit test is used.
#168. (Changed in MR13) Token expressions in context guards
The token expressions appearing in context guards such as:
(A B)? => <<test(LT(1))>>? someRule
are computed during an early phase of antlr processing. As
a result, prior to MR13, complex expressions such as:
~B
~L..U
TokClassName
~TokClassName
were not computed properly. This resulted in incorrect
context being computed for such expressions.
In MR13 these context guards are verified for proper semantics
in the initial phase and then re-evaluated after complex token
expressions have been computed in order to produce the correct
behavior.
Reported by Arpad Beszedes (beszedes inf.u-szeged.hu).
#167. (Changed in MR13) ~L..U
Prior to MR13, the complement of a token range was
not properly computed.
#166. (Changed in MR13) token expression L..U
The token U was represented as an unsigned char, restricting
the use of L..U to cases where U was assigned a token number
less than 256. This is corrected in MR13.
#165. (Changed in MR13) option -newAST
To create ASTs from an ANTLRTokenPtr antlr usually calls
"new AST(ANTLRTokenPtr)". This option generates a call
to "newAST(ANTLRTokenPtr)" instead. This allows a user
to define a parser member function to create an AST object.
Similar changes for ASTBase::tmake and ASTBase::link were not
thought necessary since they do not create AST objects, only
use existing ones.
#164. (Changed in MR13) Unused variable _astp
For many compilations, we have lived with warnings about
the unused variable _astp. It turns out that this varible
can *never* be used because the code which references it was
commented out.
This investigation was sparked by a note from Erwin Achermann
#163. (Changed in MR13) Incorrect makefiles for testcpp examples
All the examples in pccts/testcpp/* had incorrect definitions
in the makefiles for the symbol "CCC". Instead of CCC=CC they
had CC=$(CCC).
There was an additional problem in testcpp/1/test.g due to the
change in ANTLRToken::getText() to a const member function
(Item #137).
Reported by Maurice Mass (maas cuci.nl).
#162. (Changed in MR13) Combining #token with #tokdefs
When it became possible to change the print-name of a
#token (Item #148) it became useful to give a #token
statement whose only purpose was to giving a print name
to the #token. Prior to this change this could not be
combined with the #tokdefs feature.
#161. (Changed in MR13) Switch -gxt inhibits generation of tokens.h
#160. (Changed in MR13) Omissions in list of names for remap.h
When a user selects the -gp option antlr creates a list
of macros in remap.h to rename some of the standard
antlr routines from zzXXX to userprefixXXX.
There were number of omissions from the remap.h name
list related to the new trace facility. This was reported,
along with a fix, by Bernie Solomon (bernard ug.eds.com).
#159. (Changed in MR13) Violations of classic C rules
There were a number of violations of classic C style in
the distribution kit. This was reported, along with fixes,
by Bernie Solomon (bernard ug.eds.com).
#158. (Changed in MR13) #header causes problem for pre-processors
A user who runs the C pre-processor on antlr source suggested
that another syntax be allowed. With MR13 such directives
such as #header, #pragma, etc. may be written as "\#header",
"\#pragma", etc. For escaping pre-processor directives inside
a #header use something like the following:
\#header
<<
\#include <stdio.h>
>>
#157. (Fixed in MR13) empty error sets for rules with infinite recursion
When the first set for a rule cannot be computed due to infinite
left recursion and it is the only alternative for a block then
the error set for the block would be empty. This would result
in a fatal error.
Reported by Darin Creason (creason genedax.com)
#156. (Changed in MR13) DLGLexerBase::getToken() now public
#155. (Changed in MR13) Context behind predicates can suppress
With -mrhoist enabled the context behind a guarded predicate can
be used to suppress other predicates. Consider the following grammar:
r0 : (r1)+;
r1 : rp
| rq
;
rp : <<p LATEXT(1)>>? B ;
rq : (A)? => <<q LATEXT(1)>>? (A|B);
In earlier versions both predicates "p" and "q" would be hoisted into
rule r0. With MR12c predicate p is suppressed because the context which
follows predicate q includes "B" which can "cover" predicate "p". In
other words, in trying to decide in r0 whether to call r1, it doesn't
really matter whether p is false or true because, either way, there is
a valid choice within r1.
#154. (Changed in MR13) Making hoist suppression explicit using <<nohoist>>
A common error, even among experienced pccts users, is to code
an init-action to inhibit hoisting rather than a leading action.
An init-action does not inhibit hoisting.
This was coded:
rule1 : <<;>> rule2
This is what was meant:
rule1 : <<;>> <<;>> rule2
With MR13, the user can code:
rule1 : <<;>> <<nohoist>> rule2
The following will give an error message:
rule1 : <<nohoist>> rule2
If the <<nohoist>> appears as an init-action rather than a leading
action an error message is issued. The meaning of an init-action
containing "nohoist" is unclear: does it apply to just one
alternative or to all alternatives ?
-------------------------------------------------------
Note: Items #153 to #1 are now in a separate file named
-------------------------------------------------------