------------------------------------------------------------ This is the second part of a two part file. This is a list of changes to pccts 1.33 prior to MR13 For more recent information see CHANGES_FROM_133.txt ------------------------------------------------------------ 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. #153. (Changed in MR12b) Bug in computation of -mrhoist suppression set Consider the following grammar with k=1 and "-mrhoist on": r1 : (A)? => ((p>>? x /* l1 */ | r2 /* l2 */ ; r2 : A /* l4 */ | (B)? => <>? y /* l5 */ ; In earlier versions the mrhoist routine would see that both l1 and l2 contained predicates and would assume that this prevented either from acting to suppress the other predicate. In the example above it didn't realize the A at line l4 is capable of suppressing the predicate at l1 even though alt l2 contains (indirectly) a predicate. This is fixed in MR12b. Reported by Reinier van den Born (reinier@vnet.ibm.com) #153. (Changed in MR12a) Bug in computation of -mrhoist suppression set An oversight similar to that described in Item #152 appeared in the computation of the set that "covered" a predicate. If a predicate expression included a term such as p=AND(q,r) the context of p was taken to be context(q) & context(r), when it should have been context(q) | context(r). This is fixed in MR12a. #152. (Changed in MR12) Bug in generation of predicate expressions The primary purpose for MR12 is to make quite clear that MR11 is obsolete and to fix the bug related to predicate expressions. In MR10 code was added to optimize the code generated for predicate expression tests. Unfortunately, there was a significant oversight in the code which resulted in a bug in the generation of code for predicate expression tests which contained predicates combined using AND: r0 : (r1)* "@" ; r1 : (AAA)? => <

>? r2 ; r2 : (BBB)? => <>? Q | (BBB)? => <>? Q ; In MR11 (and MR10 when using "-mrhoist on") the code generated for r0 to predict r1 would be equivalent to: if ( LA(1)==Q && (LA(1)==AAA && LA(1)==BBB) && ( p && ( q || r )) ) { This is incorrect because it expresses the idea that LA(1) *must* be AAA in order to attempt r1, and *must* be BBB to attempt r2. The result was that r1 became unreachable since both condition can not be simultaneously true. The general philosophy of code generation for predicates can be summarized as follows: a. If the context is true don't enter an alt for which the corresponding predicate is false. If the context is false then it is okay to enter the alt without evaluating the predicate at all. b. A predicate created by ORing of predicates has context which is the OR of their individual contexts. c. A predicate created by ANDing of predicates has (surprise) context which is the OR of their individual contexts. d. Apply these rules recursively. e. Remember rule (a) The correct code should express the idea that *if* LA(1) is AAA then p must be true to attempt r1, but if LA(1) is *not* AAA then it is okay to attempt r1, provided that *if* LA(1) is BBB then one of q or r must be true. if ( LA(1)==Q && ( !(LA(1)==AAA || LA(1)==BBB) || ( ! LA(1) == AAA || p) && ( ! LA(1) == BBB || q || r ) ) ) { I believe this is fixed in MR12. Reported by Reinier van den Born (reinier@vnet.ibm.com) #151a. (Changed in MR12) ANTLRParser::getLexer() As a result of several requests, I have added public methods to get a pointer to the lexer belonging to a parser. ANTLRTokenStream *ANTLRParser::getLexer() const Returns a pointer to the lexer being used by the parser. ANTLRTokenStream is the base class of DLGLexer ANTLRTokenStream *ANTLRTokenBuffer::getLexer() const Returns a pointer to the lexer being used by the ANTLRTokenBuffer. ANTLRTokenStream is the base class of DLGLexer You must manually cast the ANTLRTokenStream to your program's lexer class. Because the name of the lexer's class is not fixed. Thus it is impossible to incorporate it into the DLGLexerBase class. #151b.(Changed in MR12) ParserBlackBox member getLexer() The template class ParserBlackBox now has a member getLexer() which returns a pointer to the lexer. #150. (Changed in MR12) syntaxErrCount and lexErrCount now public See Item #127 for more information. #149. (Changed in MR12) antlr option -info o (letter o for orphan) If there is more than one rule which is not referenced by any other rule then all such rules are listed. This is useful for alerting one to rules which are not used, but which can still contribute to ambiguity. For example: start : a Z ; unused: a A ; a : (A)+ ; will cause an ambiguity report for rule "a" which will be difficult to understand if the user forgets about rule "unused" simply because it is not used in the grammar. #148. (Changed in MR11) #token names appearing in zztokens,token_tbl In a #token statement like the following: #token Plus "\+" the string "Plus" appears in the zztokens array (C mode) and token_tbl (C++ mode). This string is used in most error messages. In MR11 one has the option of using some other string, (e.g. "+") in those tables. In MR11 one can write: #token Plus ("+") "\+" #token RP ("(") "\(" #token COM ("comment begin") "/\*" A #token statement is allowed to appear in more than one #lexclass with different regular expressions. However, the token name appears only once in the zztokens/token_tbl array. This means that only one substitute can be specified for a given #token name. The second attempt to define a substitute name (different from the first) will result in an error message. #147. (Changed in MR11) Bug in follow set computation There is a bug in 1.33 vanilla and all maintenance releases prior to MR11 in the computation of the follow set. The bug is different than that described in Item #82 and probably more common. It was discovered in the ansi.g grammar while testing the "ambiguity aid" (Item #119). The search for a bug started when the ambiguity aid was unable to discover the actual source of an ambiguity reported by antlr. The problem appears when an optimization of the follow set computation is used inappropriately. The result is that the follow set used is the "worst case". In other words, the error can lead to false reports of ambiguity. The good news is that if you have a grammar in which you have addressed all reported ambiguities you are ok. The bad news is that you may have spent time fixing ambiguities that were not real, or used k=2 when ck=2 might have been sufficient, and so on. The following grammar demonstrates the problem: ------------------------------------------------------------ expr : ID ; start : stmt SEMI ; stmt : CASE expr COLON | expr SEMI | plain_stmt ; plain_stmt : ID COLON ; ------------------------------------------------------------ When compiled with k=1 and ck=2 it will report: warning: alts 2 and 3 of the rule itself ambiguous upon { IDENTIFIER }, { COLON } When antlr analyzes "stmt" it computes the first[1] set of all alternatives. It finds an ambiguity between alts 2 and 3 for ID. It then computes the first[2] set for alternatives 2 and 3 to resolve the ambiguity. In computing the first[2] set of "expr" (which is only one token long) it needs to determine what could follow "expr". Under a certain combination of circumstances antlr forgets that it is trying to analyze "stmt" which can only be followed by SEMI and adds to the first[2] set of "expr" the "global" follow set (including "COLON") which could follow "expr" (under other conditions) in the phrase "CASE expr COLON". #146. (Changed in MR11) Option -treport for locating "difficult" alts It can be difficult to determine which alternatives are causing pccts to work hard to resolve an ambiguity. In some cases the ambiguity is successfully resolved after much CPU time so there is no message at all. A rough measure of the amount of work being peformed which is independent of the CPU speed and system load is the number of tnodes created. Using "-info t" gives information about the total number of tnodes created and the peak number of tnodes. Tree Nodes: peak 1300k created 1416k lost 0 It also puts in the generated C or C++ file the number of tnodes created for a rule (at the end of the rule). However this information is not sufficient to locate the alternatives within a rule which are causing the creation of tnodes. Using: antlr -treport 100000 .... causes antlr to list on stdout any alternatives which require the creation of more than 100,000 tnodes, along with the lookahead sets for those alternatives. The following is a trivial case from the ansi.g grammar which shows the format of the report. This report might be of more interest in cases where 1,000,000 tuples were created to resolve the ambiguity. ------------------------------------------------------------------------- There were 0 tuples whose ambiguity could not be resolved by full lookahead There were 157 tnodes created to resolve ambiguity between: Choice 1: statement/2 line 475 file ansi.g Choice 2: statement/3 line 476 file ansi.g Intersection of lookahead[1] sets: IDENTIFIER Intersection of lookahead[2] sets: LPARENTHESIS COLON AMPERSAND MINUS STAR PLUSPLUS MINUSMINUS ONESCOMPLEMENT NOT SIZEOF OCTALINT DECIMALINT HEXADECIMALINT FLOATONE FLOATTWO IDENTIFIER STRING CHARACTER ------------------------------------------------------------------------- #145. (Documentation) Generation of Expression Trees Item #99 was misleading because it implied that the optimization for tree expressions was available only for trees created by predicate expressions and neglected to mention that it required the use of "-mrhoist on". The optimization applies to tree expressions created for grammars with k>1 and for predicates with lookahead depth >1. In MR11 the optimized version is always used so the -mrhoist on option need not be specified. #144. (Changed in MR11) Incorrect test for exception group In testing for a rule's exception group the label a pointer is compared against '\0'. The intention is "*pointer". Reported by Jeffrey C. Fried (Jeff@Fried.net). #143. (Changed in MR11) Optional ";" at end of #token statement Fixes problem of: #token X "x" << parser action >> Being confused with: #token X "x" <> #142. (Changed in MR11) class BufFileInput subclass of DLGInputStream Alexey Demakov (demakov@kazbek.ispras.ru) has supplied class BufFileInput derived from DLGInputStream which provides a function lookahead(char *string) to test characters in the input stream more than one character ahead. The default amount of lookahead is specified by the constructor and defaults to 8 characters. This does *not* include the one character of lookahead maintained internally by DLG in member "ch" and which is not available for testing via BufFileInput::lookahead(). This is a useful class for overcoming the one-character-lookahead limitation of DLG without resorting to a lexer capable of backtracking (like flex) which is not integrated with antlr as is DLG. There are no restrictions on copying or using BufFileInput.* except that the authorship and related information must be retained in the source code. The class is located in pccts/h/BufFileInput.* of the kit. #141. (Changed in MR11) ZZDEBUG_CONSUME for ANTLRParser::consume() A debug aid has been added to file ANTLRParser::consume() in file AParser.cpp: #ifdef ZZDEBUG_CONSUME_ACTION zzdebug_consume_action(); #endif Suggested by Sramji Ramanathan (ps@kumaran.com). #140. (Changed in MR11) #pred to define predicates +---------------------------------------------------+ | Note: Assume "-prc on" for this entire discussion | +---------------------------------------------------+ A problem with predicates is that each one is regarded as unique and capable of disambiguating cases where two alternatives have identical lookahead. For example: rule : <>? A | <>? A ; will not cause any error messages or warnings to be issued by earlier versions of pccts. To compare the text of the predicates is an incomplete solution. In 1.33MR11 I am introducing the #pred statement in order to solve some problems with predicates. The #pred statement allows one to give a symbolic name to a "predicate literal" or a "predicate expression" in order to refer to it in other predicate expressions or in the rules of the grammar. The predicate literal associated with a predicate symbol is C or C++ code which can be used to test the condition. A predicate expression defines a predicate symbol in terms of other predicate symbols using "!", "&&", and "||". A predicate symbol can be defined in terms of a predicate literal, a predicate expression, or *both*. When a predicate symbol is defined with both a predicate literal and a predicate expression, the predicate literal is used to generate code, but the predicate expression is used to check for two alternatives with identical predicates in both alternatives. Here are some examples of #pred statements: #pred IsLabel <>? #pred IsLocalVar <>? #pred IsGlobalVar <>? #pred IsVar <>? IsLocalVar || IsGlobalVar #pred IsScoped <>? IsLabel || IsLocalVar I hope that the use of EBNF notation to describe the syntax of the #pred statement will not cause problems for my readers (joke). predStatement : "#pred" CapitalizedName ( "<>?" | "<>?" predOrExpr | predOrExpr ) ; predOrExpr : predAndExpr ( "||" predAndExpr ) * ; predAndExpr : predPrimary ( "&&" predPrimary ) * ; predPrimary : CapitalizedName | "!" predPrimary | "(" predOrExpr ")" ; What is the purpose of this nonsense ? To understand how predicate symbols help, you need to realize that predicate symbols are used in two different ways with two different goals. a. Allow simplification of predicates which have been combined during predicate hoisting. b. Allow recognition of identical predicates which can't disambiguate alternatives with common lookahead. First we will discuss goal (a). Consider the following rule: rule0: rule1 | ID | ... ; rule1: rule2 | rule3 ; rule2: <>? ID ; rule3: <>? ID ; When the predicates in rule2 and rule3 are combined by hoisting to create a prediction expression for rule1 the result is: if ( LA(1)==ID && ( isX(LATEXT(1) || !isX(LATEXT(1) ) ) { rule1(); ... This is inefficient, but more importantly, can lead to false assumptions that the predicate expression distinguishes the rule1 alternative with some other alternative with lookahead ID. In MR11 one can write: #pred IsX <>? ... rule2: <>? ID ; rule3: <>? ID ; During hoisting MR11 recognizes this as a special case and eliminates the predicates. The result is a prediction expression like the following: if ( LA(1)==ID ) { rule1(); ... Please note that the following cases which appear to be equivalent *cannot* be simplified by MR11 during hoisting because the hoisting logic only checks for a "!" in the predicate action, not in the predicate expression for a predicate symbol. *Not* equivalent and is not simplified during hoisting: #pred IsX <>? #pred NotX <>? ... rule2: <>? ID ; rule3: <>? ID ; *Not* equivalent and is not simplified during hoisting: #pred IsX <>? #pred NotX !IsX ... rule2: <>? ID ; rule3: <>? ID ; Now we will discuss goal (b). When antlr discovers that there is a lookahead ambiguity between two alternatives it attempts to resolve the ambiguity by searching for predicates in both alternatives. In the past any predicate would do, even if the same one appeared in both alternatives: rule: <>? X | <>? X ; The #pred statement is a start towards solving this problem. During ambiguity resolution (*not* predicate hoisting) the predicates for the two alternatives are expanded and compared. Consider the following example: #pred Upper <>? #pred Lower <>? #pred Alpha <>? Upper || Lower rule0: rule1 | <>? ID ; rule1: | rule2 | rule3 ... ; rule2: <>? ID; rule3: <>? ID; The definition of #pred Alpha expresses: a. to test the predicate use the C code "isAlpha(LATEXT(1))" b. to analyze the predicate use the information that Alpha is equivalent to the union of Upper and Lower, During ambiguity resolution the definition of Alpha is expanded into "Upper || Lower" and compared with the predicate in the other alternative, which is also "Upper || Lower". Because they are identical MR11 will report a problem. ------------------------------------------------------------------------- t10.g, line 5: warning: the predicates used to disambiguate rule rule0 (file t10.g alt 1 line 5 and alt 2 line 6) are identical when compared without context and may have no resolving power for some lookahead sequences. ------------------------------------------------------------------------- If you use the "-info p" option the output file will contain: +----------------------------------------------------------------------+ |#if 0 | | | |The following predicates are identical when compared without | | lookahead context information. For some ambiguous lookahead | | sequences they may not have any power to resolve the ambiguity. | | | |Choice 1: rule0/1 alt 1 line 5 file t10.g | | | | The original predicate for choice 1 with available context | | information: | | | | OR expr | | | | pred << Upper>>? | | depth=k=1 rule rule2 line 14 t10.g | | set context: | | ID | | | | pred << Lower>>? | | depth=k=1 rule rule3 line 15 t10.g | | set context: | | ID | | | | The predicate for choice 1 after expansion (but without context | | information): | | | | OR expr | | | | pred << isUpper(LATEXT(1))>>? | | depth=k=1 rule line 1 t10.g | | | | pred << isLower(LATEXT(1))>>? | | depth=k=1 rule line 2 t10.g | | | | | |Choice 2: rule0/2 alt 2 line 6 file t10.g | | | | The original predicate for choice 2 with available context | | information: | | | | pred << Alpha>>? | | depth=k=1 rule rule0 line 6 t10.g | | set context: | | ID | | | | The predicate for choice 2 after expansion (but without context | | information): | | | | OR expr | | | | pred << isUpper(LATEXT(1))>>? | | depth=k=1 rule line 1 t10.g | | | | pred << isLower(LATEXT(1))>>? | | depth=k=1 rule line 2 t10.g | | | | | |#endif | +----------------------------------------------------------------------+ The comparison of the predicates for the two alternatives takes place without context information, which means that in some cases the predicates will be considered identical even though they operate on disjoint lookahead sets. Consider: #pred Alpha rule1: <>? ID | <>? Label ; Because the comparison of predicates takes place without context these will be considered identical. The reason for comparing without context is that otherwise it would be necessary to re-evaluate the entire predicate expression for each possible lookahead sequence. This would require more code to be written and more CPU time during grammar analysis, and it is not yet clear whether anyone will even make use of the new #pred facility. A temporary workaround might be to use different #pred statements for predicates you know have different context. This would avoid extraneous warnings. The above example might be termed a "false positive". Comparison without context will also lead to "false negatives". Consider the following example: #pred Alpha #pred Beta rule1: <>? A | rule2 ; rule2: <>? A | <>? B ; The predicate used for alt 2 of rule1 is (Alpha || Beta). This appears to be different than the predicate Alpha used for alt1. However, the context of Beta is B. Thus when the lookahead is A Beta will have no resolving power and Alpha will be used for both alternatives. Using the same predicate for both alternatives isn't very helpful, but this will not be detected with 1.33MR11. To properly handle this the predicate expression would have to be evaluated for each distinct lookahead context. To determine whether two predicate expressions are identical is difficult. The routine may fail to identify identical predicates. The #pred feature also compares predicates to see if a choice between alternatives which is resolved by a predicate which makes the second choice unreachable. Consider the following example: #pred A <>? #pred B <>? #pred A_or_B A || B r : s | t ; s : <>? ID ; t : <>? ID ; ---------------------------------------------------------------------------- t11.g, line 5: warning: the predicate used to disambiguate the first choice of rule r (file t11.g alt 1 line 5 and alt 2 line 6) appears to "cover" the second predicate when compared without context. The second predicate may have no resolving power for some lookahead sequences. ---------------------------------------------------------------------------- #139. (Changed in MR11) Problem with -gp in C++ mode The -gp option to add a prefix to rule names did not work in C++ mode. This has been fixed. Reported by Alexey Demakov (demakov@kazbek.ispras.ru). #138. (Changed in MR11) Additional makefiles for non-MSVC++ MS systems Sramji Ramanathan (ps@kumaran.com) has supplied makefiles for building antlr and dlg with Win95/NT development tools that are not based on MSVC5. They are pccts/antlr/AntlrMS.mak and pccts/dlg/DlgMS.mak. The first line of the makefiles require a definition of PCCTS_HOME. These are in additiion to the AntlrMSVC50.* and DlgMSVC50.* supplied by Jeff Vincent (JVincent@novell.com). #137. (Changed in MR11) Token getType(), getText(), getLine() const members -------------------------------------------------------------------- If you use ANTLRCommonToken this change probably does not affect you. -------------------------------------------------------------------- For a long time it has bothered me that these accessor functions in ANTLRAbstractToken were not const member functions. I have refrained from changing them because it require users to modify existing token class definitions which are derived directly from ANTLRAbstractToken. I think it is now time. For those who are not used to C++, a "const member function" is a member function which does not modify its own object - the thing to which "this" points. This is quite different from a function which does not modify its arguments Most token definitions based on ANTLRAbstractToken have something like the following in order to create concrete definitions of the pure virtual methods in ANTLRAbstractToken: class MyToken : public ANTLRAbstractToken { ... ANTLRTokenType getType() {return _type; } int getLine() {return _line; } ANTLRChar * getText() {return _text; } ... } The required change is simply to put "const" following the function prototype in the header (.h file) and the definition file (.cpp if it is not inline): class MyToken : public ANTLRAbstractToken { ... ANTLRTokenType getType() const {return _type; } int getLine() const {return _line; } ANTLRChar * getText() const {return _text; } ... } This was originally proposed a long time ago by Bruce Guenter (bruceg@qcc.sk.ca). #136. (Changed in MR11) Added getLength() to ANTLRCommonToken Classes ANTLRCommonToken and ANTLRCommonTokenNoRefCountToken now have a member function: int getLength() const { return strlen(getText()) } Suggested by Sramji Ramanathan (ps@kumaran.com). #135. (Changed in MR11) Raised antlr's own default ZZLEXBUFSIZE to 8k #134a. (ansi_mr10.zip) T.J. Parr's ANSI C grammar made 1.33MR11 compatible There is a typographical error in the definition of BITWISEOREQ: #token BITWISEOREQ "!=" should be "\|=" When this change is combined with the bugfix to the follow set cache problem (Item #147) and a minor rearrangement of the grammar (Item #134b) it becomes a k=1 ck=2 grammar. #134b. (ansi_mr10.zip) T.J. Parr's ANSI C grammar made 1.33MR11 compatible The following changes were made in the ansi.g grammar (along with using -mrhoist on): ansi.g ====== void tracein(char *) ====> void tracein(const char *) void traceout(char *) ====> void traceout(const char *) getType()==IDENTIFIER ? isTypeName(LT(1)->getText()) : 1>>? ====> <getText())>>? <<(LT(1)->getType()==LPARENTHESIS && LT(2)->getType()==IDENTIFIER) ? \ isTypeName(LT(2)->getText()) : 1>>? ====> (LPARENTHESIS IDENTIFIER)? => <getText())>>? <<(LT(1)->getType()==LPARENTHESIS && LT(2)->getType()==IDENTIFIER) ? \ isTypeName(LT(2)->getText()) : 1>>? ====> (LPARENTHESIS IDENTIFIER)? => <getText())>>? added to init(): traceOptionValueDefault=0; added to init(): traceOption(-1); change rule "statement": statement : plain_label_statement | case_label_statement | <<;>> expression SEMICOLON | compound_statement | selection_statement | iteration_statement | jump_statement | SEMICOLON ; plain_label_statement : IDENTIFIER COLON statement ; case_label_statement : CASE constant_expression COLON statement | DEFAULT COLON statement ; support.cpp =========== void tracein(char *) ====> void tracein(const char *) void traceout(char *) ====> void traceout(const char *) added to tracein(): ANTLRParser::tracein(r); // call superclass method added to traceout(): ANTLRParser::traceout(r); // call superclass method Makefile ======== added to AFLAGS: -mrhoist on -prc on #133. (Changed in 1.33MR11) Make trace options public in ANTLRParser In checking T.J. Parr's ANSI C grammar for compatibility with 1.33MR11 discovered that it was inconvenient to have the trace facilities with protected access. #132. (Changed in 1.33MR11) Recognition of identical predicates in alts Prior to 1.33MR11, there would be no ambiguity warning when the very same predicate was used to disambiguate both alternatives: test: ref B | ref C ; ref : <>? A In 1.33MR11 this will cause the warning: warning: the predicates used to disambiguate rule test (file v98.g alt 1 line 1 and alt 2 line 2) are identical and have no resolving power ----------------- Note ----------------- This is different than the following case test: <>? A B | <>? A C ; In this case there are two distinct predicates which have exactly the same text. In the first example there are two references to the same predicate. The problem represented by this grammar will be addressed later. #131. (Changed in 1.33MR11) Case insensitive command line options Command line switches like "-CC" and keywords like "on", "off", and "stdin" are no longer case sensitive in antlr, dlg, and sorcerer. #130. (Changed in 1.33MR11) Changed ANTLR_VERSION to int from string The ANTLR_VERSION was not an integer, making it difficult to perform conditional compilation based on the antlr version. Henceforth, ANTLR_VERSION will be: (base_version * 10000) + release number thus 1.33MR11 will be: 133*100+11 = 13311 Suggested by Rainer Janssen (Rainer.Janssen@Informatik.Uni-Oldenburg.DE). #129. (Changed in 1.33MR11) Addition of ANTLR_VERSION to .h The following code is now inserted into .h amd stdpccts.h: #ifndef ANTLR_VERSION #define ANTLR_VERSION 13311 #endif Suggested by Rainer Janssen (Rainer.Janssen@Informatik.Uni-Oldenburg.DE) #128. (Changed in 1.33MR11) Redundant predicate code in (<>? ...)+ Prior to 1.33MR11, the following grammar would generate redundant tests for the "while" condition. rule2 : (<>? X)+ X | B ; The code would resemble: if (LA(1)==X) { if (pred) { do { if (!pred) {zzfailed_pred(" pred");} zzmatch(X); zzCONSUME; } while (LA(1)==X && pred && pred); } else {... With 1.33MR11 the redundant predicate test is omitted. #127. (Changed in 1.33MR11) Count Syntax Errors Count DLG Errors ------------------- ---------------- C++ mode ANTLRParser:: DLGLexerBase:: syntaxErrCount lexErrCount C mode zzSyntaxErrCount zzLexErrCount The C mode variables are global and initialized to 0. They are *not* reset to 0 automatically when antlr is restarted. The C++ mode variables are public. They are initialized to 0 by the constructors. They are *not* reset to 0 by the ANTLRParser::init() method. Suggested by Reinier van den Born (reinier@vnet.ibm.com). #126. (Changed in 1.33MR11) Addition of #first <<...>> The #first <<...>> inserts the specified text in the output files before any other #include statements required by pccts. The only things before the #first text are comments and a #define ANTLR_VERSION. Requested by and Esa Pulkkinen (esap@cs.tut.fi) and Alexin Zoltan (alexin@inf.u-szeged.hu). #125. (Changed in 1.33MR11) Lookahead for (guard)? && <

>? predicates When implementing the new style of guard predicate (Item #113) in 1.33MR10 I decided to temporarily ignore the problem of computing the "narrowest" lookahead context. Consider the following k=1 grammar: start : a | b ; a : (A)? && <>? ab ; b : (B)? && <>? ab ; ab : A | B ; In MR10 the context for both "a" and "b" was {A B} because this is the first set of rule "ab". Normally, this is not a problem because the predicate which follows the guard inhibits any ambiguity report by antlr. In MR11 the first set for rule "a" is {A} and for rule "b" it is {B}. #124. A Note on the New "&&" Style Guarded Predicates I've been asked several times, "What is the difference between the old "=>" style guard predicates and the new style "&&" guard predicates, and how do you choose one over the other" ? The main difference is that the "=>" does not apply the predicate if the context guard doesn't match, whereas the && form always does. What is the significance ? If you have a predicate which is not on the "leading edge" it is cannot be hoisted. Suppose you need a predicate that looks at LA(2). You must introduce it manually. The classic example is: castExpr : LP typeName RP | .... ; typeName : <>? ID | STRUCT ID ; The problem is that isTypeName() isn't on the leading edge of typeName, so it won't be hoisted into castExpr to help make a decision on which production to choose. The *first* attempt to fix it is this: castExpr : <>? LP typeName RP | .... ; Unfortunately, this won't work because it ignores the problem of STRUCT. The solution is to apply isTypeName() in castExpr if LA(2) is an ID and don't apply it when LA(2) is STRUCT: castExpr : (LP ID)? => <>? LP typeName RP | .... ; In conclusion, the "=>" style guarded predicate is useful when: a. the tokens required for the predicate are not on the leading edge b. there are alternatives in the expression selected by the predicate for which the predicate is inappropriate If (b) were false, then one could use a simple predicate (assuming "-prc on"): castExpr : <>? LP typeName RP | .... ; typeName : <>? ID ; So, when do you use the "&&" style guarded predicate ? The new-style "&&" predicate should always be used with predicate context. The context guard is in ADDITION to the automatically computed context. Thus it useful for predicates which depend on the token type for reasons other than context. The following example is contributed by Reinier van den Born (reinier@vnet.ibm.com). +-------------------------------------------------------------------------+ | This grammar has two ways to call functions: | | | | - a "standard" call syntax with parens and comma separated args | | - a shell command like syntax (no parens and spacing separated args) | | | | The former also allows a variable to hold the name of the function, | | the latter can also be used to call external commands. | | | | The grammar (simplified) looks like this: | | | | fun_call : ID "(" { expr ("," expr)* } ")" | | /* ID is function name */ | | | "@" ID "(" { expr ("," expr)* } ")" | | /* ID is var containing fun name */ | | ; | | | | command : ID expr* /* ID is function name */ | | | path expr* /* path is external command name */ | | ; | | | | path : ID /* left out slashes and such */ | | | "@" ID /* ID is environment var */ | | ; | | | | expr : .... | | | "(" expr ")"; | | | | call : fun_call | | | command | | ; | | | | Obviously the call is wildly ambiguous. This is more or less how this | | is to be resolved: | | | | A call begins with an ID or an @ followed by an ID. | | | | If it is an ID and if it is an ext. command name -> command | | if followed by a paren -> fun_call | | otherwise -> command | | | | If it is an @ and if the ID is a var name -> fun_call | | otherwise -> command | | | | One can implement these rules quite neatly using && predicates: | | | | call : ("@" ID)? && <>? fun_call | | | (ID)? && <>? command | | | (ID "(")? fun_call | | | command | | ; | | | | This can be done better, so it is not an ideal example, but it | | conveys the principle. | +-------------------------------------------------------------------------+ #123. (Changed in 1.33MR11) Correct definition of operators in ATokPtr.h The return value of operators in ANTLRTokenPtr: changed: unsigned ... operator !=(...) to: int ... operator != (...) changed: unsigned ... operator ==(...) to: int ... operator == (...) Suggested by R.A. Nelson (cowboy@VNET.IBM.COM) #122. (Changed in 1.33MR11) Member functions to reset DLG in C++ mode void DLGFileReset(FILE *f) { input = f; found_eof = 0; } void DLGStringReset(DLGChar *s) { input = s; p = &input[0]; } Supplied by R.A. Nelson (cowboy@VNET.IBM.COM) #121. (Changed in 1.33MR11) Another attempt to fix -o (output dir) option Another attempt is made to improve the -o option of antlr, dlg, and sorcerer. This one by JVincent (JVincent@novell.com). The current rule: a. If -o is not specified than any explicit directory names are retained. b. If -o is specified than the -o directory name overrides any explicit directory names. c. The directory name of the grammar file is *not* stripped to create the main output file. However it is stil subject to override by the -o directory name. #120. (Changed in 1.33MR11) "-info f" output to stdout rather than stderr Added option 0 (e.g. "-info 0") which is a noop. #119. (Changed in 1.33MR11) Ambiguity aid for grammars The user can ask for additional information on ambiguities reported by antlr to stdout. At the moment, only one ambiguity report can be created in an antlr run. This feature is enabled using the "-aa" (Ambiguity Aid) option. The following options control the reporting of ambiguities: -aa ruleName Selects reporting by name of rule -aa lineNumber Selects reporting by line number (file name not compared) -aam Selects "multiple" reporting for a token in the intersection set of the alternatives. For instance, the token ID may appear dozens of times in various paths as the program explores the rules which are reachable from the point of an ambiguity. With option -aam every possible path the search program encounters is reported. Without -aam only the first encounter is reported. This may result in incomplete information, but the information may be sufficient and much shorter. -aad depth Selects the depth of the search. The default value is 1. The number of paths to be searched, and the size of the report can grow geometrically with the -ck value if a full search for all contributions to the source of the ambiguity is explored. The depth represents the number of tokens in the lookahead set which are matched against the set of ambiguous tokens. A depth of 1 means that the search stops when a lookahead sequence of just one token is matched. A k=1 ck=6 grammar might generate 5,000 items in a report if a full depth 6 search is made with the Ambiguity Aid. The source of the problem may be in the first token and obscured by the volume of data - I hesitate to call it information. When the user selects a depth > 1, the search is first performed at depth=1 for both alternatives, then depth=2 for both alternatives, etc. Sample output for rule grammar in antlr.g itself: +---------------------------------------------------------------------+ | Ambiguity Aid | | | | Choice 1: grammar/70 line 632 file a.g | | Choice 2: grammar/82 line 644 file a.g | | | | Intersection of lookahead[1] sets: | | | | "\}" "class" "#errclass" "#tokclass" | | | | Choice:1 Depth:1 Group:1 ("#errclass") | | 1 in (...)* block grammar/70 line 632 a.g | | 2 to error grammar/73 line 635 a.g | | 3 error error/1 line 894 a.g | | 4 #token "#errclass" error/2 line 895 a.g | | | | Choice:1 Depth:1 Group:2 ("#tokclass") | | 2 to tclass grammar/74 line 636 a.g | | 3 tclass tclass/1 line 937 a.g | | 4 #token "#tokclass" tclass/2 line 938 a.g | | | | Choice:1 Depth:1 Group:3 ("class") | | 2 to class_def grammar/75 line 637 a.g | | 3 class_def class_def/1 line 669 a.g | | 4 #token "class" class_def/3 line 671 a.g | | | | Choice:1 Depth:1 Group:4 ("\}") | | 2 #token "\}" grammar/76 line 638 a.g | | | | Choice:2 Depth:1 Group:5 ("#errclass") | | 1 in (...)* block grammar/83 line 645 a.g | | 2 to error grammar/93 line 655 a.g | | 3 error error/1 line 894 a.g | | 4 #token "#errclass" error/2 line 895 a.g | | | | Choice:2 Depth:1 Group:6 ("#tokclass") | | 2 to tclass grammar/94 line 656 a.g | | 3 tclass tclass/1 line 937 a.g | | 4 #token "#tokclass" tclass/2 line 938 a.g | | | | Choice:2 Depth:1 Group:7 ("class") | | 2 to class_def grammar/95 line 657 a.g | | 3 class_def class_def/1 line 669 a.g | | 4 #token "class" class_def/3 line 671 a.g | | | | Choice:2 Depth:1 Group:8 ("\}") | | 2 #token "\}" grammar/96 line 658 a.g | +---------------------------------------------------------------------+ For a linear lookahead set ambiguity (where k=1 or for k>1 but when all lookahead sets [i] with i>? A ; c : A ; Prior to 1.33MR10 the code generated for "start" would resemble: while { if (LA(1)==A && (!LA(1)==A || isUpper())) { a(); } }; This code is wrong because it makes rule "c" unreachable from "start". The essence of the problem is that antlr fails to recognize that there can be a valid alternative within "a" even when the predicate <>? is false. In 1.33MR10 with -mrhoist the hoisting of the predicate into "start" is suppressed because it recognizes that "c" can cover all the cases where the predicate is false: while { if (LA(1)==A) { a(); } }; With the antlr "-info p" switch the user will receive information about the predicate suppression in the generated file: -------------------------------------------------------------- #if 0 Hoisting of predicate suppressed by alternative without predicate. The alt without the predicate includes all cases where the predicate is false. WITH predicate: line 7 v1.g WITHOUT predicate: line 7 v1.g The context set for the predicate: A The lookahead set for the alt WITHOUT the semantic predicate: A The predicate: pred << isUpper(LATEXT(1))>>? depth=k=1 rule b line 9 v1.g set context: A tree context: null Chain of referenced rules: #0 in rule start (line 5 v1.g) to rule a #1 in rule a (line 7 v1.g) #endif -------------------------------------------------------------- A predicate can be suppressed by a combination of alternatives which, taken together, cover a predicate: start : (a)* "@" ; a : b | ca | cb | cc ; b : <>? ( A | B | C ) ; ca : A ; cb : B ; cc : C ; Consider a more complex example in which "c" covers only part of a predicate: start : (a)* "@" ; a : b | c ; b : <>? ( A | X ); c : A ; Prior to 1.33MR10 the code generated for "start" would resemble: while { if ( (LA(1)==A || LA(1)==X) && (! (LA(1)==A || LA(1)==X) || isUpper()) { a(); } }; With 1.33MR10 and -mrhoist the predicate context is restricted to the non-covered lookahead. The code resembles: while { if ( (LA(1)==A || LA(1)==X) && (! (LA(1)==X) || isUpper()) { a(); } }; With the antlr "-info p" switch the user will receive information about the predicate restriction in the generated file: -------------------------------------------------------------- #if 0 Restricting the context of a predicate because of overlap in the lookahead set between the alternative with the semantic predicate and one without Without this restriction the alternative without the predicate could not be reached when input matched the context of the predicate and the predicate was false. WITH predicate: line 11 v4.g WITHOUT predicate: line 12 v4.g The original context set for the predicate: A X The lookahead set for the alt WITHOUT the semantic predicate: A The intersection of the two sets A The original predicate: pred << isUpper(LATEXT(1))>>? depth=k=1 rule b line 15 v4.g set context: A X tree context: null The new (modified) form of the predicate: pred << isUpper(LATEXT(1))>>? depth=k=1 rule b line 15 v4.g set context: X tree context: null #endif -------------------------------------------------------------- The bad news about -mrhoist: (a) -mrhoist does not analyze predicates with lookahead depth > 1. (b) -mrhoist does not look past a guarded predicate to find context which might cover other predicates. For these cases you might want to use syntactic predicates. When a semantic predicate fails during guess mode the guess fails and the next alternative is tried. Limitation (a) is illustrated by the following example: start : (stmt)* EOF ; stmt : cast | expr ; cast : <>? LP ID RP ; expr : LP ID RP ; This is not much different from the first example, except that it requires two tokens of lookahead context to determine what to do. This predicate is NOT suppressed because the current version is unable to handle predicates with depth > 1. A predicate can be combined with other predicates during hoisting. In those cases the depth=1 predicates are still handled. Thus, in the following example the isUpper() predicate will be suppressed by line #4 when hoisted from "bizarre" into "start", but will still be present in "bizarre" in order to predict "stmt". start : (bizarre)* EOF ; // #1 // #2 bizarre : stmt // #3 | A // #4 ; stmt : cast | expr ; cast : <>? LP ID RP ; expr : LP ID RP ; | <>? A Limitation (b) is illustrated by the following example of a context guarded predicate: rule : (A)? <

>? // #1 (A // #2 |B // #3 ) // #4 | <> B // #5 ; Recall that this means that when the lookahead is NOT A then the predicate "p" is ignored and it attempts to match "A|B". Ideally, the "B" at line #3 should suppress predicate "q". However, the current version does not attempt to look past the guard predicate to find context which might suppress other predicates. In some cases -mrhoist will lead to the reporting of ambiguities which were not visible before: start : (a)* "@"; a : bc | d; bc : b | c ; b : <>? A; c : A ; d : A ; In this case there is a true ambiguity in "a" between "bc" and "d" which can both match "A". Without -mrhoist the predicate in "b" is hoisted into "a" and there is no ambiguity reported. However, with -mrhoist, the predicate in "b" is suppressed by "c" (as it should be) making the ambiguity in "a" apparent. The motivations for these changes were hoisting problems reported by Reinier van den Born (reinier@vnet.ibm.com) and several others. #116. (Changed in 1.33MR10) C++ mode: tracein/traceout rule name is (const char *) The prototype for C++ mode routine tracein (and traceout) has changed from "char *" to "const char *". #115. (Changed in 1.33MR10) Using guess mode with exception handlers in C mode The definition of the C mode macros zzmatch_wsig and zzsetmatch_wsig neglected to consider guess mode. When control passed to the rule's parse exception handler the routine would exit without ever closing the guess block. This would lead to unpredictable behavior. In 1.33MR10 the behavior of exceptions in C mode and C++ mode should be identical. #114. (Changed in 1.33MR10) difference in [zz]resynch() between C and C++ modes There was a slight difference in the way C and C++ mode resynchronized following a parsing error. The C routine would sometimes skip an extra token before attempting to resynchronize. The C routine was changed to match the C++ routine. #113. (Changed in 1.33MR10) new context guarded pred: (g)? && <

>? expr The existing context guarded predicate: rule : (guard)? => <

>? expr | next_alternative ; generates code which resembles: if (lookahead(expr) && (!guard || pred)) { expr() } else .... This is not suitable for some applications because it allows expr() to be invoked when the predicate is false. This is intentional because it is meant to mimic automatically computed predicate context. The new context guarded predicate uses the guard information differently because it has a different goal. Consider: rule : (guard)? && <

>? expr | next_alternative ; The new style of context guarded predicate is equivalent to: rule : <>? expr | next_alternative ; It generates code which resembles: if (lookahead(expr) && guard && pred) { expr(); } else ... Both forms of guarded predicates severely restrict the form of the context guard: it can contain no rule references, no (...)*, no (...)+, and no {...}. It may contain token and token class references, and alternation ("|"). Addition for 1.33MR11: in the token expression all tokens must be at the same height of the token tree: (A ( B | C))? && ... is ok (all height 2) (A ( B | ))? && ... is not ok (some 1, some 2) (A B C D | E F G H)? && ... is ok (all height 4) (A B C D | E )? && ... is not ok (some 4, some 1) This restriction is required in order to properly compute the lookahead set for expressions like: rule1 : (A B C)? && <>? rule2 ; rule2 : (A|X) (B|Y) (C|Z); This addition was suggested by Rienier van den Born (reinier@vnet.ibm.com) #112. (Changed in 1.33MR10) failed validation predicate in C guess mode John Lilley (jlilley@empathy.com) suggested that failed validation predicates abort a guess rather than reporting a failed error. This was installed in C++ mode (Item #4). Only now was it noticed that the fix was never installed for C mode. #111. (Changed in 1.33MR10) moved zzTRACEIN to before init action When the antlr -gd switch is present antlr generates calls to zzTRACEIN at the start of a rule and zzTRACEOUT at the exit from a rule. Prior to 1.33MR10 Tthe call to zzTRACEIN was after the init-action, which could cause confusion because the init-actions were reported with the name of the enclosing rule, rather than the active rule. #110. (Changed in 1.33MR10) antlr command line copied to generated file The antlr command line is now copied to the generated file near the start. #109. (Changed in 1.33MR10) improved trace information The quality of the trace information provided by the "-gd" switch has been improved significantly. Here is an example of the output from a test program. It shows the rule name, the first token of lookahead, the call depth, and the guess status: exit rule gusxx {"?"} depth 2 enter rule gusxx {"?"} depth 2 enter rule gus1 {"o"} depth 3 guessing guess done - returning to rule gus1 {"o"} at depth 3 (guess mode continues - an enclosing guess is still active) guess done - returning to rule gus1 {"Z"} at depth 3 (guess mode continues - an enclosing guess is still active) exit rule gus1 {"Z"} depth 3 guessing guess done - returning to rule gusxx {"o"} at depth 2 (guess mode ends) enter rule gus1 {"o"} depth 3 guess done - returning to rule gus1 {"o"} at depth 3 (guess mode ends) guess done - returning to rule gus1 {"Z"} at depth 3 (guess mode ends) exit rule gus1 {"Z"} depth 3 line 1: syntax error at "Z" missing SC ... Rule trace reporting is controlled by the value of the integer [zz]traceOptionValue: when it is positive tracing is enabled, otherwise it is disabled. Tracing during guess mode is controlled by the value of the integer [zz]traceGuessOptionValue. When it is positive AND [zz]traceOptionValue is positive rule trace is reported in guess mode. The values of [zz]traceOptionValue and [zz]traceGuessOptionValue can be adjusted by subroutine calls listed below. Depending on the presence or absence of the antlr -gd switch the variable [zz]traceOptionValueDefault is set to 0 or 1. When the parser is initialized or [zz]traceReset() is called the value of [zz]traceOptionValueDefault is copied to [zz]traceOptionValue. The value of [zz]traceGuessOptionValue is always initialzed to 1, but, as noted earlier, nothing will be reported unless [zz]traceOptionValue is also positive. When the parser state is saved/restored the value of the trace variables are also saved/restored. If a restore causes a change in reporting behavior from on to off or vice versa this will be reported. When the -gd option is selected, the macro "#define zzTRACE_RULES" is added to appropriate output files. C++ mode -------- int traceOption(int delta) int traceGuessOption(int delta) void traceReset() int traceOptionValueDefault C mode -------- int zzTraceOption(int delta) int zzTraceGuessOption(int delta) void zzTraceReset() int zzTraceOptionValueDefault The argument "delta" is added to the traceOptionValue. To turn on trace when inside a particular rule one: rule : <> ( rest-of-rule ) <> ; /* fail clause */ <> One can use the same idea to turn *off* tracing within a rule by using a delta of (-1). An improvement in the rule trace was suggested by Sramji Ramanathan (ps@kumaran.com). #108. A Note on Deallocation of Variables Allocated in Guess Mode NOTE ------------------------------------------------------ This mechanism only works for heap allocated variables ------------------------------------------------------ The rewrite of the trace provides the machinery necessary to properly free variables or undo actions following a failed guess. The macro zzUSER_GUESS_HOOK(guessSeq,zzrv) is expanded as part of the zzGUESS macro. When a guess is opened the value of zzrv is 0. When a longjmp() is executed to undo the guess, the value of zzrv will be 1. The macro zzUSER_GUESS_DONE_HOOK(guessSeq) is expanded as part of the zzGUESS_DONE macro. This is executed whether the guess succeeds or fails as part of closing the guess. The guessSeq is a sequence number which is assigned to each guess and is incremented by 1 for each guess which becomes active. It is needed by the user to associate the start of a guess with the failure and/or completion (closing) of a guess. Guesses are nested. They must be closed in the reverse of the order that they are opened. In order to free memory used by a variable during a guess a user must write a routine which can be called to register the variable along with the current guess sequence number provided by the zzUSER_GUESS_HOOK macro. If the guess fails, all variables tagged with the corresponding guess sequence number should be released. This is ugly, but it would require a major rewrite of antlr 1.33 to use some mechanism other than setjmp()/longjmp(). The order of calls for a *successful* guess would be: zzUSER_GUESS_HOOK(guessSeq,0); zzUSER_GUESS_DONE_HOOK(guessSeq); The order of calls for a *failed* guess would be: zzUSER_GUESS_HOOK(guessSeq,0); zzUSER_GUESS_HOOK(guessSeq,1); zzUSER_GUESS_DONE_HOOK(guessSeq); The default definitions of these macros are empty strings. Here is an example in C++ mode. The zzUSER_GUESS_HOOK and zzUSER_GUESS_DONE_HOOK macros and myGuessHook() routine can be used without change in both C and C++ versions. ---------------------------------------------------------------------- << #include "AToken.h" typedef ANTLRCommonToken ANTLRToken; #include "DLGLexer.h" int main() { { DLGFileInput in(stdin); DLGLexer lexer(&in,2000); ANTLRTokenBuffer pipe(&lexer,1); ANTLRCommonToken aToken; P parser(&pipe); lexer.setToken(&aToken); parser.init(); parser.start(); }; fclose(stdin); fclose(stdout); return 0; } >> << char *s=NULL; #undef zzUSER_GUESS_HOOK #define zzUSER_GUESS_HOOK(guessSeq,zzrv) myGuessHook(guessSeq,zzrv); #undef zzUSER_GUESS_DONE_HOOK #define zzUSER_GUESS_DONE_HOOK(guessSeq) myGuessHook(guessSeq,2); void myGuessHook(int guessSeq,int zzrv) { if (zzrv == 0) { fprintf(stderr,"User hook: starting guess #%d\n",guessSeq); } else if (zzrv == 1) { free (s); s=NULL; fprintf(stderr,"User hook: failed guess #%d\n",guessSeq); } else if (zzrv == 2) { free (s); s=NULL; fprintf(stderr,"User hook: ending guess #%d\n",guessSeq); }; } >> #token A "a" #token "[\t \ \n]" <> class P { start : (top)+ ; top : (which) ? <> | other <> ; <> which : which2 ; which2 : which3 ; which3 : (label)? <> | (global)? <> | (exclamation)? <> ; label : <getText());>> A ":" ; global : <getText());>> A "::" ; exclamation : <getText());>> A "!" ; other : <getText());>> "other" ; } ---------------------------------------------------------------------- This is a silly example, but illustrates the idea. For the input "a ::" with tracing enabled the output begins: ---------------------------------------------------------------------- enter rule "start" depth 1 enter rule "top" depth 2 User hook: starting guess #1 enter rule "which" depth 3 guessing enter rule "which2" depth 4 guessing enter rule "which3" depth 5 guessing User hook: starting guess #2 enter rule "label" depth 6 guessing guess failed User hook: failed guess #2 guess done - returning to rule "which3" at depth 5 (guess mode continues - an enclosing guess is still active) User hook: ending guess #2 User hook: starting guess #3 enter rule "global" depth 6 guessing exit rule "global" depth 6 guessing guess done - returning to rule "which3" at depth 5 (guess mode continues - an enclosing guess is still active) User hook: ending guess #3 enter rule "global" depth 6 guessing exit rule "global" depth 6 guessing exit rule "which3" depth 5 guessing exit rule "which2" depth 4 guessing exit rule "which" depth 3 guessing guess done - returning to rule "top" at depth 2 (guess mode ends) User hook: ending guess #1 enter rule "which" depth 3 ..... ---------------------------------------------------------------------- Remember: (a) Only init-actions are executed during guess mode. (b) A rule can be invoked multiple times during guess mode. (c) If the guess succeeds the rule will be called once more without guess mode so that normal actions will be executed. This means that the init-action might need to distinguish between guess mode and non-guess mode using the variable [zz]guessing. #107. (Changed in 1.33MR10) construction of ASTs in guess mode Prior to 1.33MR10, when using automatic AST construction in C++ mode for a rule, an AST would be constructed for elements of the rule even while in guess mode. In MR10 this no longer occurs. #106. (Changed in 1.33MR10) guess variable confusion In C++ mode a guess which failed always restored the parser state using zzGUESS_DONE as part of zzGUESS_FAIL. Prior to 1.33MR10, C mode required an explicit call to zzGUESS_DONE after the call to zzGUESS_FAIL. Consider: rule : (alpha)? beta | ... ; The generated code resembles: zzGUESS if (!zzrv && LA(1)==ID) { <==== line #1 alpha zzGUESS_DONE beta } else { if (! zzrv) zzGUESS_DONE <==== line #2a .... However, in some cases line #2 was rendered: if (guessing) zzGUESS_DONE <==== line #2b This would work for simple test cases, but would fail in some cases where there was a guess while another guess was active. One kind of failure would be to match up the zzGUESS_DONE at line #2b with the "outer" guess which was still active. The outer guess would "succeed" when only the inner guess should have succeeded. In 1.33MR10 the behavior of zzGUESS and zzGUESS_FAIL in C and and C++ mode should be identical. The same problem appears in 1.33 vanilla in some places. For example: start : { (sub)? } ; or: start : ( B | ( sub )? | C )+ ; generates incorrect code. The general principle is: (a) use [zz]guessing only when deciding between a call to zzFAIL or zzGUESS_FAIL (b) use zzrv in all other cases This problem was discovered while testing changes to item #105. I believe this is now fixed. My apologies. #105. (Changed in 1.33MR10) guess block as single alt of (...)+ Prior to 1.33MR10 the following constructs: rule_plus : ( (sub)? )+ ; rule_star : ( (sub)? )* ; generated incorrect code for the guess block (which could result in runtime errors) because of an incorrect optimization of a block with only a single alternative. The fix caused some changes to the fix described in Item #49 because there are now three code generation sequences for (...)+ blocks containing a guess block: a. single alternative which is a guess block b. multiple alternatives in which the last is a guess block c. all other cases Forms like "rule_star" can have unexpected behavior when there is a syntax error: if the subrule "sub" is not matched *exactly* then "rule_star" will consume no tokens. Reported by Esa Pulkkinen (esap@cs.tut.fi). #104. (Changed in 1.33MR10) -o option for dlg There was problem with the code added by item #74 to handle the -o option of dlg. This should fix it. #103. (Changed in 1.33MR10) ANDed semantic predicates Rescinded. The optimization was a mistake. The resulting problem is described in Item #150. #102. (Changed in 1.33MR10) allow "class parser : .... {" The syntax of the class statement ("class parser-name {") has been extended to allow for the specification of base classes. An arbirtrary number of tokens may now appear between the class name and the "{". They are output again when the class declaration is generated. For example: class Parser : public MyBaseClassANTLRparser { This was suggested by a user, but I don't have a record of who it was. #101. (Changed in 1.33MR10) antlr -info command line switch -info p - extra predicate information in generated file t - information about tnode use: at the end of each rule in generated file summary on stderr at end of program m - monitor progress prints name of each rule as it is started flushes output at start of each rule f - first/follow set information to stdout 0 - no operation (added in 1.33MR11) The options may be combined and may appear in any order. For example: antlr -info ptm -CC -gt -mrhoist on mygrammar.g #100a. (Changed in 1.33MR10) Predicate tree simplification When the same predicates can be referenced in more than one alternative of a block large predicate trees can be formed. The difference that these optimizations make is so dramatic that I have decided to use it even when -mrhoist is not selected. Consider the following grammar: start : ( all )* ; all : a | d | e | f ; a : c A B | c A C ; c : <>? ; d : <>? B C ; e : <>? B C ; f : e X Y ; In rule "a" there is a reference to rule "c" in both alternatives. The length of the predicate AAA is k=2 and it can be followed in alternative 1 only by (A B) while in alternative 2 it can be followed only by (A C). Thus they do not have identical context. In rule "all" the alternatives which refer to rules "e" and "f" allow elimination of the duplicate reference to predicate CCC. The table below summarized the kind of simplification performed by 1.33MR10. In the table, X and Y stand for single predicates (not trees). (OR X (OR Y (OR Z))) => (OR X Y Z) (AND X (AND Y (AND Z))) => (AND X Y Z) (OR X (... (OR X Y) ... )) => (OR X (... Y ... )) (AND X (... (AND X Y) ... )) => (AND X (... Y ... )) (OR X (... (AND X Y) ... )) => (OR X (... ... )) (AND X (... (OR X Y) ... )) => (AND X (... ... )) (AND X) => X (OR X) => X In a test with a complex grammar for a real application, a predicate tree with six OR nodes and 12 leaves was reduced to "(OR X Y Z)". In 1.33MR10 there is a greater effort to release memory used by predicates once they are no longer in use. #100b. (Changed in 1.33MR10) Suppression of extra predicate tests The following optimizations require that -mrhoist be selected. It is relatively easy to optimize the code generated for predicate gates when they are of the form: (AND X Y Z ...) or (OR X Y Z ...) where X, Y, Z, and "..." represent individual predicates (leaves) not predicate trees. If the predicate is an AND the contexts of the X, Y, Z, etc. are ANDed together to create a single Tree context for the group and context tests for the individual predicates are suppressed: -------------------------------------------------- Note: This was incorrect. The contexts should be ORed together. This has been fixed. A more complete description is available in item #152. --------------------------------------------------- Optimization 1: (AND X Y Z ...) Suppose the context for Xtest is LA(1)==LP and the context for Ytest is LA(1)==LP && LA(2)==ID. Without the optimization the code would resemble: if (lookaheadContext && !(LA(1)==LP && LA(1)==LP && LA(2)==ID) || ( (! LA(1)==LP || Xtest) && (! (LA(1)==LP || LA(2)==ID) || Xtest) )) {... With the -mrhoist optimization the code would resemble: if (lookaheadContext && ! (LA(1)==LP && LA(2)==ID) || (Xtest && Ytest) {... Optimization 2: (OR X Y Z ...) with identical contexts Suppose the context for Xtest is LA(1)==ID and for Ytest the context is also LA(1)==ID. Without the optimization the code would resemble: if (lookaheadContext && ! (LA(1)==ID || LA(1)==ID) || (LA(1)==ID && Xtest) || (LA(1)==ID && Ytest) {... With the -mrhoist optimization the code would resemble: if (lookaheadContext && (! LA(1)==ID) || (Xtest || Ytest) {... Optimization 3: (OR X Y Z ...) with distinct contexts Suppose the context for Xtest is LA(1)==ID and for Ytest the context is LA(1)==LP. Without the optimization the code would resemble: if (lookaheadContext && ! (LA(1)==ID || LA(1)==LP) || (LA(1)==ID && Xtest) || (LA(1)==LP && Ytest) {... With the -mrhoist optimization the code would resemble: if (lookaheadContext && (zzpf=0, (LA(1)==ID && (zzpf=1) && Xtest) || (LA(1)==LP && (zzpf=1) && Ytest) || !zzpf) { These may appear to be of similar complexity at first, but the non-optimized version contains two tests of each context while the optimized version contains only one such test, as well as eliminating some of the inverted logic (" !(...) || "). Optimization 4: Computation of predicate gate trees When generating code for the gates of predicate expressions antlr 1.33 vanilla uses a recursive procedure to generate "&&" and "||" expressions for testing the lookahead. As each layer of the predicate tree is exposed a new set of "&&" and "||" expressions on the lookahead are generated. In many cases the lookahead being tested has already been tested. With -mrhoist a lookahead tree is computed for the entire lookahead expression. This means that predicates with identical context or context which is a subset of another predicate's context disappear. This is especially important for predicates formed by rules like the following: uppperCaseVowel : <>? vowel; vowel: : <>? LETTERS; These predicates are combined using AND since both must be satisfied for rule upperCaseVowel. They have identical context which makes this optimization very effective. The affect of Items #100a and #100b together can be dramatic. In a very large (but real world) grammar one particular predicate expression was reduced from an (unreadable) 50 predicate leaves, 195 LA(1) terms, and 5500 characters to an (easily comprehensible) 3 predicate leaves (all different) and a *single* LA(1) term. #99. (Changed in 1.33MR10) Code generation for expression trees Expression trees are used for k>1 grammars and predicates with lookahead depth >1. This optimization must be enabled using "-mrhoist on". (Clarification added for 1.33MR11). In the processing of expression trees, antlr can generate long chains of token comparisons. Prior to 1.33MR10 there were many redundant parenthesis which caused problems for compilers which could handle expressions of only limited complexity. For example, to test an expression tree (root R A B C D), antlr would generate something resembling: (LA(1)==R && (LA(2)==A || (LA(2)==B || (LA(2)==C || LA(2)==D))))) If there were twenty tokens to test then there would be twenty parenthesis at the end of the expression. In 1.33MR10 the generated code for tree expressions resembles: (LA(1)==R && (LA(2)==A || LA(2)==B || LA(2)==C || LA(2)==D)) For "complex" expressions the output is indented to reflect the LA number being tested: (LA(1)==R && (LA(2)==A || LA(2)==B || LA(2)==C || LA(2)==D || LA(2)==E || LA(2)==F) || LA(1)==S && (LA(2)==G || LA(2)==H)) Suggested by S. Bochnak (S.Bochnak@@microTool.com.pl), #98. (Changed in 1.33MR10) Option "-info p" When the user selects option "-info p" the program will generate detailed information about predicates. If the user selects "-mrhoist on" additional detail will be provided explaining the promotion and suppression of predicates. The output is part of the generated file and sandwiched between #if 0/#endif statements. Consider the following k=1 grammar: start : ( all ) * ; all : ( a | b ) ; a : c B ; c : <>? | B ; b : <>? X ; Below is an excerpt of the output for rule "start" for the three predicate options (off, on, and maintenance release style hoisting). For those who do not wish to use the "-mrhoist on" option for code generation the option can be used in a "diagnostic" mode to provide valuable information: a. where one should insert null actions to inhibit hoisting b. a chain of rule references which shows where predicates are being hoisted ====================================================================== Example of "-info p" with "-mrhoist on" ====================================================================== #if 0 Hoisting of predicate suppressed by alternative without predicate. The alt without the predicate includes all cases where the predicate is false. WITH predicate: line 11 v36.g WITHOUT predicate: line 12 v36.g The context set for the predicate: B The lookahead set for alt WITHOUT the semantic predicate: B The predicate: pred << LATEXT(1)>>? depth=k=1 rule c line 11 v36.g set context: B tree context: null Chain of referenced rules: #0 in rule start (line 1 v36.g) to rule all #1 in rule all (line 3 v36.g) to rule a #2 in rule a (line 8 v36.g) to rule c #3 in rule c (line 11 v36.g) #endif && #if 0 pred << LATEXT(1)>>? depth=k=1 rule b line 15 v36.g set context: X tree context: null #endif ====================================================================== Example of "-info p" with the default -prc setting ( "-prc off") ====================================================================== #if 0 OR pred << LATEXT(1)>>? depth=k=1 rule c line 11 v36.g set context: nil tree context: null pred << LATEXT(1)>>? depth=k=1 rule b line 15 v36.g set context: nil tree context: null #endif ====================================================================== Example of "-info p" with "-prc on" and "-mrhoist off" ====================================================================== #if 0 OR pred << LATEXT(1)>>? depth=k=1 rule c line 11 v36.g set context: B tree context: null pred << LATEXT(1)>>? depth=k=1 rule b line 15 v36.g set context: X tree context: null #endif ====================================================================== #97. (Fixed in 1.33MR10) "Predicate applied for more than one ... " In 1.33 vanilla, the grammar listed below produced this message for the first alternative (only) of rule "b": warning: predicate applied for >1 lookahead 1-sequences [you may only want one lookahead 1-sequence to apply. Try using a context guard '(...)? =>' In 1.33MR10 the message is issued for both alternatives. top : (a)*; a : b | c ; b : <>? ( AAA | BBB ) | <>? ( XXX | YYY ) ; c : AAA | XXX; #96. (Fixed in 1.33MR10) Guard predicates ignored when -prc off Prior to 1.33MR10, guard predicate code was not generated unless "-prc on" was selected. This was incorrect, since "-prc off" (the default) is supposed to disable only AUTOMATIC computation of predicate context, not the programmer specified context supplied by guard predicates. #95. (Fixed in 1.33MR10) Predicate guard context length was k, not max(k,ck) Prior to 1.33MR10, predicate guards were computed to k tokens rather than max(k,ck). Consider the following grammar: a : ( A B C)? => <>? (A|X) (B|Y) (C|Z) ; The code generated by 1.33 vanilla with "-k 1 -ck 3 -prc on" for the predicate in "a" resembles: if ( (! LA(1)==A) || AAA(LATEXT(1))) {... With 1.33MR10 and the same options the code resembles: if ( (! (LA(1)==A && LA(2)==B && LA(3)==C) || AAA(LATEXT(1))) {... #94. (Fixed in 1.33MR10) Predicates followed by rule references Prior to 1.33MR10, a semantic predicate which referenced a token which was off the end of the rule caused an incomplete context to be computed (with "-prc on") for the predicate under some circum- stances. In some cases this manifested itself as illegal C code (e.g. "LA(2)==[Ep](1)" in the k=2 examples below: all : ( a ) *; a : <>? ID X | <>? Y | Z ; This might also occur when the semantic predicate was followed by a rule reference which was shorter than the length of the semantic predicate: all : ( a ) *; a : <>? ID X | <>? y | Z ; y : Y ; Depending on circumstance, the resulting context might be too generous because it was too short, or too restrictive because of missing alternatives. #93. (Changed in 1.33MR10) Definition of Purify macro Ofer Ben-Ami (gremlin@cs.huji.ac.il) has supplied a definition for the Purify macro: #define PURIFY(r, s) memset((char *) &(r), '\0', (s)); Note: This may not be the right thing to do for C++ objects that have constructors. Reported by Bonny Rais (bonny@werple.net.au). For those cases one should #define PURIFY to an empty macro in the #header or #first actions. #92. (Fixed in 1.33MR10) Guarded predicates and hoisting When a guarded predicate participates in hoisting it is linked into a predicate expression tree. Prior to 1.33MR10 this link was never cleared and the next time the guard was used to construct a new tree the link could contain a spurious reference to another element which had previosly been joined to it in the semantic predicate tree. For example: start : ( all ) *; all : ( a | b ) ; start2 : ( all2 ) *; all2 : ( a ) ; a : (A)? => <>? A ; b : (B)? => <>? B ; Prior to 1.33MR10 the code for "start2" would include a spurious reference to the BBB predicate which was left from constructing the predicate tree for rule "start" (i.e. or(AAA,BBB) ). In 1.33MR10 this problem is avoided by cloning the original guard each time it is linked into a predicate tree. #91. (Changed in 1.33MR10) Extensive changes to semantic pred hoisting ============================================ This has been rendered obsolete by Item #117 ============================================ #90. (Fixed in 1.33MR10) Semantic pred with LT(i) and i>max(k,ck) There is a bug in antlr 1.33 vanilla and all maintenance releases prior to 1.33MR10 which allows semantic predicates to reference an LT(i) or LATEXT(i) where i is larger than max(k,ck). When this occurs antlr will attempt to mark the ith element of an array in which there are only max(k,ck) elements. The result cannot be predicted. Using LT(i) or LATEXT(i) for i>max(k,ck) is reported as an error in 1.33MR10. #89. Rescinded #88. (Fixed in 1.33MR10) Tokens used in semantic predicates in guess mode Consider the behavior of a semantic predicate during guess mode: rule : a:A ( <>? b:B | c:C ); Prior to MR10 the assignment of the token or attribute to $a did not occur during guess mode, which would cause the semantic predicate to misbehave because $a would be null. In 1.33MR10 a semantic predicate with a reference to an element label (such as $a) forces the assignment to take place even in guess mode. In order to work, this fix REQUIRES use of the $label format for token pointers and attributes referenced in semantic predicates. The fix does not apply to semantic predicates using the numeric form to refer to attributes (e.g. <>?). The user will receive a warning for this case. Reported by Rob Trout (trout@mcs.cs.kent.edu). #87. (Fixed in 1.33MR10) Malformed guard predicates Context guard predicates may contain only references to tokens. They may not contain references to (...)+ and (...)* blocks. This is now checked. This replaces the fatal error message in item #78 with an appropriate (non-fatal) error messge. In theory, context guards should be allowed to reference rules. However, I have not had time to fix this. Evaluation of the guard takes place before all rules have been read, making it difficult to resolve a forward reference to rule "zzz" - it hasn't been read yet ! To postpone evaluation of the guard until all rules have been read is too much for the moment. #86. (Fixed in 1.33MR10) Unequal set size in set_sub Routine set_sub() in pccts/support/set/set.h did not work correctly when the sets were of unequal sizes. Rewrote set_equ to make it simpler and remove unnecessary and expensive calls to set_deg(). This routine was not used in 1.33 vanila. #85. (Changed in 1.33MR10) Allow redefinition of MaxNumFiles Raised the maximum number of input files to 99 from 20. Put a #ifndef/#endif around the "#define MaxNumFiles 99". #84. (Fixed in 1.33MR10) Initialize zzBadTok in macro zzRULE Initialize zzBadTok to NULL in zzRULE macro of AParser.h. in order to get rid of warning messages. #83. (Fixed in 1.33MR10) False warnings with -w2 for #tokclass When -w2 is selected antlr gives inappropriate warnings about #tokclass names not having any associated regular expressions. Since a #tokclass is not a "real" token it will never have an associated regular expression and there should be no warning. Reported by Derek Pappas (derek.pappas@eng.sun.com) #82. (Fixed in 1.33MR10) Computation of follow sets with multiple cycles Reinier van den Born (reinier@vnet.ibm.com) reported a problem in the computation of follow sets by antlr. The problem (bug) exists in 1.33 vanilla and all maintenance releases prior to 1.33MR10. The problem involves the computation of follow sets when there are cycles - rules which have mutual references. I believe the problem is restricted to cases where there is more than one cycle AND elements of those cycles have rules in common. Even when this occurs it may not affect the code generated - but it might. It might also lead to undetected ambiguities. There were no changes in antlr or dlg output from the revised version. The following fragment demonstates the problem by giving different follow sets (option -pa) for var_access when built with k=1 and ck=2 on 1.33 vanilla and 1.33MR10: echo_statement : ECHO ( echo_expr )* ; echo_expr : ( command )? | expression ; command : IDENTIFIER { concat } ; expression : operand ( OPERATOR operand )* ; operand : value | START command END ; value : concat | TYPE operand ; concat : var_access { CONCAT value } ; var_access : IDENTIFIER { INDEX } ; #81. (Changed in 1.33MR10) C mode use of attributes and ASTs Reported by Isaac Clark (irclark@mindspring.com). C mode code ignores attributes returned by rules which are referenced using element labels when ASTs are enabled (-gt option). 1. start : r:rule t:Token <<$start=$r;>> The $r refrence will not work when combined with the -gt option. 2. start : t:Token <<$start=$t;>> The $t reference works in all cases. 3. start : rule <<$0=$1;>> Numeric labels work in all cases. With MR10 the user will receive an error message for case 1 when the -gt option is used. #80. (Fixed in 1.33MR10) (...)? as last alternative of block A construct like the following: rule : a | (b)? ; does not make sense because there is no alternative when the guess block fails. This is now reported as a warning to the user. Previously, there was a code generation error for this case: the guess block was not "closed" when the guess failed. This could cause an infinite loop or other problems. This is now fixed. Example problem: #header<< #include #include "charptr.h" >> << #include "charptr.c" main () { ANTLR(start(),stdin); } >> #token "[\ \t]+" << zzskip(); >> #token "[\n]" << zzline++; zzskip(); >> #token Word "[a-z]+" #token Number "[0-9]+" start : (test1)? | (test2)? ; test1 : (Word Word Word Word)? | (Word Word Word Number)? ; test2 : (Word Word Number Word)? | (Word Word Number Number)? ; Test data which caused infinite loop: a 1 a a #79. (Changed in 1.33MR10) Use of -fh with multiple parsers Previously, antlr always used the pre-processor symbol STDPCCTS_H as a gate for the file stdpccts.h. This caused problems when there were multiple parsers defined because they used the same gate symbol. In 1.33MR10, the -fh filename is used to generate the gate file for stdpccts.h. For instance: antlr -fh std_parser1.h generates the pre-processor symbol "STDPCCTS_std_parser1_H". Reported by Ramanathan Santhanam (ps@kumaran.com). #78. (Changed in 1.33MR9) Guard predicates that refer to rules ------------------------ Please refer to Item #87 ------------------------ Guard predicates are processed during an early phase of antlr (during parsing) before all data structures are completed. There is an apparent bug in earlier versions of 1.33 which caused guard predicates which contained references to rules (rather than tokens) to reference a structure which hadn't yet been initialized. In some cases (perhaps all cases) references to rules in guard predicates resulted in the use of "garbage". #79. (Changed in 1.33MR9) Jeff Vincent (JVincent@novell.com) Previously, the maximum length file name was set arbitrarily to 300 characters in antlr, dlg, and sorcerer. The config.h file now attempts to define the maximum length filename using _MAX_PATH from stdlib.h before falling back to using the value 300. #78. (Changed in 1.33MR9) Jeff Vincent (JVincent@novell.com) Put #ifndef/#endif around definition of ZZLEXBUFSIZE in antlr. #77. (Changed in 1.33MR9) Arithmetic overflow for very large grammars In routine HandleAmbiguities() antlr attempts to compute the number of possible elements in a set that is order of number-of-tokens raised to the number-of-lookahead-tokens power. For large grammars or large lookahead (e.g. -ck 7) this can cause arithmetic overflow. With 1.33MR9, arithmetic overflow in this computation is reported the first time it happens. The program continues to run and the program branches based on the assumption that the computed value is larger than any number computed by counting actual cases because 2**31 is larger than the number of bits in most computers. Before 1.33MR9 overflow was not reported. The behavior following overflow is not predictable by anyone but the original author. NOTE In 1.33MR10 the warning message is suppressed. The code which detects the overflow allows the computation to continue without an error. The error message itself made made users worry. #76. (Changed in 1.33MR9) Jeff Vincent (JVincent@novell.com) Jeff Vincent has convinced me to make ANTLRCommonToken and ANTLRCommonNoRefCountToken use variable length strings allocated from the heap rather than fixed length strings. By suitable definition of setText(), the copy constructor, and operator =() it is possible to maintain "copy" semantics. By "copy" semantics I mean that when a token is copied from an existing token it receives its own, distinct, copy of the text allocated from the heap rather than simply a pointer to the original token's text. ============================================================ W * A * R * N * I * N * G ============================================================ It is possible that this may cause problems for some users. For those users I have included the old version of AToken.h as pccts/h/AToken_traditional.h. #75. (Changed in 1.33MR9) Bruce Guenter (bruceg@qcc.sk.ca) Make DLGStringInput const correct. Since this is infrequently subclassed, it should affect few users, I hope. #74. (Changed in 1.33MR9) -o (output directory) option Antlr does not properly handle the -o output directory option when the filename of the grammar contains a directory part. For example: antlr -o outdir pccts_src/myfile.g causes antlr create a file called "outdir/pccts_src/myfile.cpp. It SHOULD create outdir/myfile.cpp The suggested code fix has been installed in antlr, dlg, and Sorcerer. #73. (Changed in 1.33MR9) Hoisting of semantic predicates and -mrhoist ============================================ This has been rendered obsolete by Item #117 ============================================ #72. (Changed in 1.33MR9) virtual saveState()/restoreState()/guess_XXX The following methods in ANTLRParser were made virtual at the request of S. Bochnak (S.Bochnak@microTool.com.pl): saveState() and restoreState() guess(), guess_fail(), and guess_done() #71. (Changed in 1.33MR9) Access to omitted command line argument If a switch requiring arguments is the last thing on the command line, and the argument is omitted, antlr would core. antlr test.g -prc instead of antlr test.g -prc off #70. (Changed in 1.33MR9) Addition of MSVC .dsp and .mak build files The following MSVC .dsp and .mak files for pccts and sorcerer were contributed by Stanislaw Bochnak (S.Bochnak@microTool.com.pl) and Jeff Vincent (JVincent@novell.com) PCCTS Distribution Kit ---------------------- pccts/PCCTSMSVC50.dsw pccts/antlr/AntlrMSVC50.dsp pccts/antlr/AntlrMSVC50.mak pccts/dlg/DlgMSVC50.dsp pccts/dlg/DlgMSVC50.mak pccts/support/msvc.dsp Sorcerer Distribution Kit ------------------------- pccts/sorcerer/SorcererMSVC50.dsp pccts/sorcerer/SorcererMSVC50.mak pccts/sorcerer/lib/msvc.dsp #69. (Changed in 1.33MR9) Change "unsigned int" to plain "int" Declaration of max_token_num in misc.c as "unsigned int" caused comparison between signed and unsigned ints giving warning message without any special benefit. #68. (Changed in 1.33MR9) Add void return for dlg internal_error() Get rid of "no return value" message in internal_error() in file dlg/support.c and dlg/dlg.h. #67. (Changed in Sor) sor.g: lisp() has no return value Added a "void" for the return type. #66. (Added to Sor) sor.g: ZZLEXBUFSIZE enclosed in #ifndef/#endif A user needed to be able to change the ZZLEXBUFSIZE for sor. Put the definition of ZZLEXBUFSIZE inside #ifndef/#endif #65. (Changed in 1.33MR9) PCCTSAST::deepCopy() and ast_dup() bug Jeff Vincent (JVincent@novell.com) found that deepCopy() made new copies of only the direct descendents. No new copies were made of sibling nodes, Sibling pointers are set to zero by shallowCopy(). PCCTS_AST::deepCopy() has been changed to make a deep copy in the traditional sense. The deepCopy() routine depends on the behavior of shallowCopy(). In all sor examples I've found, shallowCopy() zeroes the right and down pointers. Original Tree Original deepCopy() Revised deepCopy ------------- ------------------- ---------------- a->b->c A A | | | d->e->f D D->E->F | | | g->h->i G G->H->I | | j->k J->K While comparing deepCopy() for C++ mode with ast_dup for C mode I found a problem with ast_dup(). Routine ast_dup() has been changed to make a deep copy in the traditional sense. Original Tree Original ast_dup() Revised ast_dup() ------------- ------------------- ---------------- a->b->c A->B->C A | | | d->e->f D->E->F D->E->F | | | g->h->i G->H->I G->H->I | | | j->k J->K J->K I believe this affects transform mode sorcerer programs only. #64. (Changed in 1.33MR9) anltr/hash.h prototype for killHashTable() #63. (Changed in 1.33MR8) h/charptr.h does not zero pointer after free The charptr.h routine now zeroes the pointer after free(). Reported by Jens Tingleff (jensting@imaginet.fr) #62. (Changed in 1.33MR8) ANTLRParser::resynch had static variable The static variable "consumed" in ANTLRParser::resynch was changed into an instance variable of the class with the name "resynchConsumed". Reported by S.Bochnak@microTool.com.pl #61. (Changed in 1.33MR8) Using rule>[i,j] when rule has no return values Previously, the following code would cause antlr to core when it tried to generate code for rule1 because rule2 had no return values ("upward inheritance"): rule1 : <> rule2 > [i,j] ; rule2 : Anything ; Reported by S.Bochnak@microTool.com.pl Verified correct operation of antlr MR8 when missing or extra inheritance arguments for all combinations. When there are missing or extra arguments code will still be generated even though this might cause the invocation of a subroutine with the wrong number of arguments. #60. (Changed in 1.33MR7) Major changes to exception handling There were significant problems in the handling of exceptions in 1.33 vanilla. The general problem is that it can only process one level of exception handler. For example, a named exception handler, an exception handler for an alternative, or an exception for a subrule always went to the rule's exception handler if there was no "catch" which matched the exception. In 1.33MR7 the exception handlers properly "nest". If an exception handler does not have a matching "catch" then the nextmost outer exception handler is checked for an appropriate "catch" clause, and so on until an exception handler with an appropriate "catch" is found. There are still undesirable features in the way exception handlers are implemented, but I do not have time to fix them at the moment: The exception handlers for alternatives are outside the block containing the alternative. This makes it impossible to access variables declared in a block or to resume the parse by "falling through". The parse can still be easily resumed in other ways, but not in the most natural fashion. This results in an inconsistentcy between named exception handlers and exception handlers for alternatives. When an exception handler for an alternative "falls through" it goes to the nextmost outer handler - not the "normal action". A major difference between 1.33MR7 and 1.33 vanilla is the default action after an exception is caught: 1.33 Vanilla ------------ In 1.33 vanilla the signal value is set to zero ("NoSignal") and the code drops through to the code following the exception. For named exception handlers this is the "normal action". For alternative exception handlers this is the rule's handler. 1.33MR7 ------- In 1.33MR7 the signal value is NOT automatically set to zero. There are two cases: For named exception handlers: if the signal value has been set to zero the code drops through to the "normal action". For all other cases the code branches to the nextmost outer exception handler until it reaches the handler for the rule. The following macros have been defined for convenience: C/C++ Mode Name -------------------- (zz)suppressSignal set signal & return signal arg to 0 ("NoSignal") (zz)setSignal(intValue) set signal & return signal arg to some value (zz)exportSignal copy the signal value to the return signal arg I'm not sure why PCCTS make a distinction between the local signal value and the return signal argument, but I'm loathe to change the code. The burden of copying the local signal value to the return signal argument can be given to the default signal handler, I suppose. #59. (Changed in 1.33MR7) Prototypes for some functions Added prototypes for the following functions to antlr.h zzconsumeUntil() zzconsumeUntilToken() #58. (Changed in 1.33MR7) Added defintion of zzbufsize to dlgauto.h #57. (Changed in 1.33MR7) Format of #line directive Previously, the -gl directive for line 1234 would resemble: "# 1234 filename.g". This caused problems for some compilers/pre-processors. In MR7 it generates "#line 1234 filename.g". #56. (Added in 1.33MR7) Jan Mikkelsen Move PURIFY macro invocaton to after rule's init action. #55. (Fixed in 1.33MR7) Unitialized variables in ANTLRParser Member variables inf_labase and inf_last were not initialized. (See item #50.) #54. (Fixed in 1.33MR6) Brad Schick (schick@interacess.com) Previously, the following constructs generated the same code: rule1 : (A B C)? | something-else ; rule2 : (A B C)? () | something-else ; In all versions of pccts rule1 guesses (A B C) and then consume all three tokens if the guess succeeds. In MR6 rule2 guesses (A B C) but consumes NONE of the tokens when the guess succeeds because "()" matches epsilon. #53. (Explanation for 1.33MR6) What happens after an exception is caught ? The Book is silent about what happens after an exception is caught. The following code fragment prints "Error Action" followed by "Normal Action". test : Word ex:Number <> exception[ex] catch NoViableAlt: <> ; The reason for "Normal Action" is that the normal flow of the program after a user-written exception handler is to "drop through". In the case of an exception handler for a rule this results in the exection of a "return" statement. In the case of an exception handler attached to an alternative, rule, or token this is the code that would have executed had there been no exception. The user can achieve the desired result by using a "return" statement. test : Word ex:Number <> exception[ex] catch NoViableAlt: <> ; The most powerful mechanism for recovery from parse errors in pccts is syntactic predicates because they provide backtracking. Exceptions allow "return", "break", "consumeUntil(...)", "goto _handler", "goto _fail", and changing the _signal value. #52. (Fixed in 1.33MR6) Exceptions without syntactic predicates The following generates bad code in 1.33 if no syntactic predicates are present in the grammar. test : Word ex:Number <> exception[ex] catch NoViableAlt: <> There is a reference to a guess variable. In C mode this causes a compiler error. In C++ mode it generates an extraneous check on member "guessing". In MR6 correct code is generated for both C and C++ mode. #51. (Added to 1.33MR6) Exception operator "@" used without exceptions In MR6 added a warning when the exception operator "@" is used and no exception group is defined. This is probably a case where "\@" or "@" is meant. #50. (Fixed in 1.33MR6) Gunnar Rxnning (gunnar@candleweb.no) http://www.candleweb.no/~gunnar/ Routines zzsave_antlr_state and zzrestore_antlr_state don't save and restore all the data needed when switching states. Suggested patch applied to antlr.h and err.h for MR6. #49. (Fixed in 1.33MR6) Sinan Karasu (sinan@boeing.com) Generated code failed to turn off guess mode when leaving a (...)+ block which contained a guess block. The result was an infinite loop. For example: rule : ( (x)? | y )+ Suggested code fix implemented in MR6. Replaced ... else if (zzcnt>1) break; with: C++ mode: ... else if (zzcnt>1) {if (!zzrv) zzGUESS_DONE; break;}; C mode: ... else if (zzcnt>1) {if (zzguessing) zzGUESS_DONE; break;}; #48. (Fixed in 1.33MR6) Invalid exception element causes core A label attached to an invalid construct can cause pccts to crash while processing the exception associated with the label. For example: rule : t:(B C) exception[t] catch MismatchedToken: <> Version MR6 generates the message: reference in exception handler to undefined label 't' #47. (Fixed in 1.33MR6) Manuel Ornato Under some circumstances involving a k >1 or ck >1 grammar and a loop block (i.e. (...)* ) pccts will fail to detect a syntax error and loop indefinitely. The problem did not exist in 1.20, but has existed from 1.23 to the present. Fixed in MR6. --------------------------------------------------- Complete test program --------------------------------------------------- #header<< #include #include "charptr.h" >> << #include "charptr.c" main () { ANTLR(global(),stdin); } >> #token "[\ \t]+" << zzskip(); >> #token "[\n]" << zzline++; zzskip(); >> #token B "b" #token C "c" #token D "d" #token E "e" #token LP "\(" #token RP "\)" #token ANTLREOF "@" global : ( (E liste) | liste | listed ) ANTLREOF ; listeb : LP ( B ( B | C )* ) RP ; listec : LP ( C ( B | C )* ) RP ; listed : LP ( D ( B | C )* ) RP ; liste : ( listeb | listec )* ; --------------------------------------------------- Sample data causing infinite loop --------------------------------------------------- e (d c) --------------------------------------------------- #46. (Fixed in 1.33MR6) Robert Richter (Robert.Richter@infotech.tu-chemnitz.de) This item from the list of known problems was fixed by item #18 (below). #45. (Fixed in 1.33MR6) Brad Schick (schick@interaccess.com) The dependency scanner in VC++ mistakenly sees a reference to an MPW #include file even though properly #ifdef/#endif in config.h. The suggested workaround has been implemented: #ifdef MPW ..... #define MPW_CursorCtl_Header #include MPW_CursorCtl_Header ..... #endif #44. (Fixed in 1.33MR6) cast malloc() to (char *) in charptr.c Added (char *) cast for systems where malloc returns "void *". #43. (Added to 1.33MR6) Bruce Guenter (bruceg@qcc.sk.ca) Add setLeft() and setUp methods to ASTDoublyLinkedBase for symmetry with setRight() and setDown() methods. #42. (Fixed in 1.33MR6) Jeff Katcher (jkatcher@nortel.ca) C++ style comment in antlr.c corrected. #41. (Added in 1.33MR6) antlr -stdout Using "antlr -stdout ..." forces the text that would normally go to the grammar.c or grammar.cpp file to stdout. #40. (Added in 1.33MR6) antlr -tab to change tab stops Using "antlr -tab number ..." changes the tab stops for the grammar.c or grammar.cpp file. The number must be between 0 and 8. Using 0 gives tab characters, values between 1 and 8 give the appropriate number of space characters. #39. (Fixed in 1.33MR5) Jan Mikkelsen Commas in function prototype still not correct under some circumstances. Suggested code fix installed. #38. (Fixed in 1.33MR5) ANTLRTokenBuffer constructor Have ANTLRTokenBuffer ctor initialize member "parser" to null. #37. (Fixed in 1.33MR4) Bruce Guenter (bruceg@qcc.sk.ca) In ANTLRParser::FAIL(int k,...) released memory pointed to by f[i] (as well as f itself. Should only free f itself. #36. (Fixed in 1.33MR3) Cortland D. Starrett (cort@shay.ecn.purdue.edu) Neglected to properly declare isDLGmaxToken() when fixing problem reported by Andreas Magnusson. Undo "_retv=NULL;" change which caused problems for return values from rules whose return values weren't pointers. Failed to create bin directory if it didn't exist. #35. (Fixed in 1.33MR2) Andreas Magnusson (Andreas.Magnusson@mailbox.swipnet.se) Repair bug introduced by 1.33MR1 for #tokdefs. The original fix placed "DLGmaxToken=9999" and "DLGminToken=0" in the TokenType enum in order to fix a problem with an aggresive compiler assigning an 8 bit enum which might be too narrow. This caused #tokdefs to assume that there were 9999 real tokens. The repair to the fix causes antlr to ignore TokenTypes "DLGmaxToken" and "DLGminToken" in a #tokdefs file. #34. (Added to 1.33MR1) Add public DLGLexerBase::set_line(int newValue) Previously there was no public function for changing the line number maintained by the lexer. #33. (Fixed in 1.33MR1) Franklin Chen (chen@adi.com) Accidental use of EXIT_FAILURE rather than PCCTS_EXIT_FAILURE in pccts/h/AParser.cpp. #32. (Fixed in 1.33MR1) Franklin Chen (chen@adi.com) In PCCTSAST.cpp lines 405 and 466: Change free (t) to free ( (char *)t ); to match prototype. #31. (Added to 1.33MR1) Pointer to parser in ANTLRTokenBuffer Pointer to parser in DLGLexerBase The ANTLRTokenBuffer class now contains a pointer to the parser which is using it. This is established by the ANTLRParser constructor calling ANTLRTokenBuffer:: setParser(ANTLRParser *p). When ANTLRTokenBuffer::setParser(ANTLRParser *p) is called it saves the pointer to the parser and then calls ANTLRTokenStream::setParser(ANTLRParser *p) so that the lexer can also save a pointer to the parser. There is also a function getParser() in each class with the obvious purpose. It is possible that these functions will return NULL under some circumstances (e.g. a non-DLG lexer is used). #30. (Added to 1.33MR1) function tokenName(int token) standard The generated parser class now includes the function: static const ANTLRChar * tokenName(int token) which returns a pointer to the "name" corresponding to the token. The base class (ANTLRParser) always includes the member function: const ANTLRChar * parserTokenName(int token) which can be accessed by objects which have a pointer to an ANTLRParser, but do not know the name of the parser class (e.g. ANTLRTokenBuffer and DLGLexerBase). #29. (Added to 1.33MR1) Debugging DLG lexers If the pre-processor symbol DEBUG_LEXER is defined then DLexerBase will include code for printing out key information about tokens which are recognized. The debug feature of the lexer is controlled by: int previousDebugValue=lexer.debugLexer(newValue); a value of 0 disables output a value of 1 enables output Even if the lexer debug code is compiled into DLexerBase it must be enabled before any output is generated. For example: DLGFileInput in(stdin); MyDLG lexer(&in,2000); lexer.setToken(&aToken); #if DEBUG_LEXER lexer.debugLexer(1); // enable debug information #endif #28. (Added to 1.33MR1) More control over DLG header Version 1.33MR1 adds the following directives to PCCTS for C++ mode: #lexprefix <> Adds source code to the DLGLexer.h file after the #include "DLexerBase.h" but before the start of the class definition. #lexmember <> Adds source code to the DLGLexer.h file as part of the DLGLexer class body. It appears immediately after the start of the class and a "public: statement. #27. (Fixed in 1.33MR1) Comments in DLG actions Previously, DLG would not recognize comments as a special case. Thus, ">>" in the comments would cause errors. This is fixed. #26. (Fixed in 1.33MR1) Removed static variables from error routines Previously, the existence of statically allocated variables in some of the parser's member functions posed a danger when there was more than one parser active. Replaced with dynamically allocated/freed variables in 1.33MR1. #25. (Fixed in 1.33MR1) Use of string literals in semantic predicates Previously, it was not possible to place a string literal in a semantic predicate because it was not properly "stringized" for the report of a failed predicate. #24. (Fixed in 1.33MR1) Continuation lines for semantic predicates Previously, it was not possible to continue semantic predicates across a line because it was not properly "stringized" for the report of a failed predicate. rule : <>?[ a very long statement ] #23. (Fixed in 1.33MR1) {...} envelope for failed semantic predicates Previously, there was a code generation error for failed semantic predicates: rule : <>?[ stmt1; stmt2; ] which generated code which resembled: if (! xyz()) stmt1; stmt2; It now puts the statements in a {...} envelope: if (! xyz()) { stmt1; stmt2; }; #22. (Fixed in 1.33MR1) Continuation of #token across lines using "\" Previously, it was not possible to continue a #token regular expression across a line. The trailing "\" and newline caused a newline to be inserted into the regular expression by DLG. Fixed in 1.33MR1. #21. (Fixed in 1.33MR1) Use of ">>" (right shift operator in DLG actions It is now possible to use the C++ right shift operator ">>" in DLG actions by using the normal escapes: #token "shift-right" << value=value \>\> 1;>> #20. (Version 1.33/19-Jan-97 Karl Eccleson P.A. Keller (P.A.Keller@bath.ac.uk) There is a problem due to using exceptions with the -gh option. Suggested fix now in 1.33MR1. #19. (Fixed in 1.33MR1) Tom Piscotti and John Lilley There were problems suppressing messages to stdin and stdout when running in a window environment because some functions which uses fprint were not virtual. Suggested change now in 1.33MR1. I believe all functions containing error messages (excluding those indicating internal inconsistency) have been placed in functions which are virtual. #18. (Version 1.33/ 22-Nov-96) John Bair (jbair@iftime.com) Under some combination of options a required "return _retv" is not generated. Suggested fix now in 1.33MR1. #17. (Version 1.33/3-Sep-96) Ron House (house@helios.usq.edu.au) The routine ASTBase::predorder_action omits two "tree->" prefixes, which results in the preorder_action belonging to the wrong node to be invoked. Suggested fix now in 1.33MR1. #16. (Version 1.33/7-Jun-96) Eli Sternheim Routine consumeUntilToken() does not check for end-of-file condition. Suggested fix now in 1.33MR1. #15. (Version 1.33/8 Apr 96) Asgeir Olafsson Problem with tree duplication of doubly linked ASTs in ASTBase.cpp. Suggested fix now in 1.33MR1. #14. (Version 1.33/28-Feb-96) Andreas.Magnusson@mailbox.swipnet.se Problem with definition of operator = (const ANTLRTokenPtr rhs). Suggested fix now in 1.33MR1. #13. (Version 1.33/13-Feb-96) Franklin Chen (chen@adi.com) Sun C++ Compiler 3.0.1 can't compile testcpp/1 due to goto in block with destructors. Apparently fixed. Can't locate "goto". #12. (Version 1.33/10-Nov-95) Minor problems with 1.33 code The following items have been fixed in 1.33MR1: 1. pccts/antlr/main.c line 142 "void" appears in classic C code 2. no makefile in support/genmk 3. EXIT_FAILURE/_SUCCESS instead of PCCTS_EXIT_FAILURE/_SUCCESS pccts/h/PCCTSAST.cpp pccts/h/DLexerBase.cpp pccts/testcpp/6/test.g 4. use of "signed int" isn't accepted by AT&T cfront pccts/h/PCCTSAST.h line 42 5. in call to ANTLRParser::FAIL the var arg err_k is passed as "int" but is declared "unsigned int". 6. I believe that a failed validation predicate still does not get put in a "{...}" envelope, despite the release notes. 7. The #token ">>" appearing in the DLG grammar description causes DLG to generate the string literal "\>\>" which is non-conforming and will cause some compilers to complain (scan.c function act10 line 143 of source code). #11. (Version 1.32b6) Dave Kuhlman (dkuhlman@netcom.com) Problem with file close in gen.c. Already fixed in 1.33. #10. (Version 1.32b6/29-Aug-95) pccts/antlr/main.c contains a C++ style comments on lines 149 and 176 which causes problems for most C compilers. Already fixed in 1.33. #9. (Version 1.32b4/14-Mar-95) dlgauto.h #include "config.h" The file pccts/h/dlgauto.h should probably contain a #include "config.h" as it uses the #define symbol __USE_PROTOS. Added to 1.33MR1. #8. (Version 1.32b4/6-Mar-95) Michael T. Richter (mtr@igs.net) In C++ output mode anonymous tokens from in-line regular expressions can create enum values which are too wide for the datatype of the enum assigned by the C++ compiler. Fixed in 1.33MR1. #7. (Version 1.32b4/6-Mar-95) C++ does not imply __STDC__ In err.h the combination of # directives assumes that a C++ compiler has __STDC__ defined. This is not necessarily true. This problem also appears in the use of __USE_PROTOS which is appropriate for both Standard C and C++ in antlr/gen.c and antlr/lex.c Fixed in 1.33MR1. #6. (Version 1.32 ?/15-Feb-95) Name conflict for "TokenType" Already fixed in 1.33. #5. (23-Jan-95) Douglas_Cuthbertson.JTIDS@jtids_qmail.hanscom.af.mil The fail action following a semantic predicate is not enclosed in "{...}". This can lead to problems when the fail action contains more than one statement. Fixed in 1.33MR1. #4 . (Version 1.33/31-Mar-96) jlilley@empathy.com (John Lilley) Put briefly, a semantic predicate ought to abort a guess if it fails. Correction suggested by J. Lilley has been added to 1.33MR1. #3 . (Version 1.33) P.A.Keller@bath.ac.uk Extra commas are placed in the K&R style argument list for rules when using both exceptions and ASTs. Fixed in 1.33MR1. #2. (Version 1.32b6/2-Oct-95) Brad Schick Construct #[] generates zzastnew() in C++ mode. Already fixed in 1.33. #1. (Version 1.33) Bob Bailey (robert@oakhill.sps.mot.com) Previously, config.h assumed that all PC systems required "short" file names. The user can now override that assumption with "#define LONGFILENAMES". Added to 1.33MR1.