readme

yacc源码

	     BTYACC -- backtracking yacc
	     ===========================

BTYACC was created by Chris Dodd using ideas from many
places and lots of code from the Berkeley Yacc
distribution, which is a public domain yacc clone put
together by the good folks at Berkeley.  This code is
distributed with NO WARRANTEE and is public domain.  It is
certain to contain bugs, which you should report to:
	[e-mail]

Vadim Maslov of Siber Systems <[e-mail]> modified
BTYACC to make it suitable for production environment. 
Specifically, preprocessing commands %define, %ifdef,
%endif, %include were added and many bugs were fixed.

See the README.BYACC, NEW_FEATURES, and ACKNOWLEDGEMENTS
files for more about Berkeley Yacc and other sources of
info.

		Version 2.1 changes
		-------------------
		  by Vadim Maslov

** Added preprocessor statements:

%define <variable> 
%ifdef <variable>
%endif

These statements have the same meaning as in C preprocessor.  
They allow to introduce modularity and versioning 
into grammar building.  

Preprocessor statement %include was introduced earlier,
in ver. 2.0.
				

		Version 2.0 changes
		-------------------
		  by Vadim Maslov

** Added 

%include file_name

command.

It acts just like #include in C.

Helps a lot when the list of tokens used by this grammar 
is generated by another program.


** Changed 16-bit short numbers to 32-bit int numbers in
grammar tables, so that huge grammar tables (tables that
are larger than 32768 elements) resulting from huge
grammars (Cobol grammar, for instance) can work correctly.
You need to have 32-bit integer to index table bigger than
32768 elements, 16-bit integer is not enough.

The original BtYacc just generated non-working tables
larger than 32768 elements without even notifying about
the table overflow.


** Make error recovery work correctly when error happens
while processing nested conflicts. Original BtYacc could
infinitely cycle in certain situations that involved error
recovery while in nested conflict.

More detailed explanation: when we have nested conflicts
(conflict that happens while trial-processing another
conflict), it leads btyacc into NP-complete searching of
conflict tree. The ultimate goal is YYVALID operator that
selects a particular branch of that tree as a valid one.

If no YYVALID is found on the tree, then error recovery
takes over.  The problem with this is that error recovery
is started in the same state context that exists on the
last surveyed branch of the conflict tree.  Sometimes this
last branch may be of zero length and it results in
recovering to exactly the same state as existed before
entering the conflict. BtYacc cycles then.

We solved this problem by memorizing the longest path in
the conflict tree while browsing it. If we ever get into
error recovery, we restore state that existed on the
longest path.  Effectively we say: if we have an error,
let us move forward as far as we possibly could while we
were browsing the conflict tree.


** Introduce YYVALID_NESTED operation in addition to
simply YYVALID.  When we have a nested conflict (conflict
while processing in trial mode for another conflict), we
want to relate YYVALID to a particular level of conflict
being in trial.

Since we mostly anticipate only 2-level nested conflicts
YYVALID_NESTED tells the parser to satisfy only the
internal conflict.  Therefore, in 1-level conflict
situation YYVALID_NESTED acts like a regular YYVALID, but
in 2-level conflict it is a no-op and the other YYVALID
for outer conflict will be searched for.


** Improved grammar error diagnostics.


** Improved handling of situation where /tmp directory is
missing.  Original btyacc just died quietly when /tmp
directory was missing.  We added code that states the
problem explicitly. While on UNIX /tmp directory is always
present, it may be missing on WIN32 systems, therefore
diagnosing this situation is important.


	BackTracking changes
        --------------------

	  by Chriss Dodd


BTYACC is a modified version of yacc that supports
automatic backtracking and semantic disambiguation to
parse ambiguous grammars, as well as syntactic sugar for
inherited attributes (which tend to introduce conflicts).
Whenever a btyacc generated parser runs into a
shift-reduce or reduce-reduce error in the parse table, it
remembers the current parse point (yacc stack and input
stream state), and goes into trial parse mode.  It then
continues parsing, ignoring most rule actions.  If it runs
into an error (either through the parse table or through
an action calling YYERROR), it backtracks to the most
recent conflict point and tries a different alternative.
If it finds a successful parse (reaches the end of the
input or an action calls YYVALID), it backtracks to the
point where it first entered trial parse mode, and
continues with a full parse (executing all actions),
following the path of the successful trial.

Actions in btyacc come in two flavors -- {}-actions, which
are only executed when not in trial mode, and []-actions
which are executed regardless of mode.  There are also
inherited attributes, which look like arguments (they are
enclosed in "()") and act like []-actions.

What this buys you:

* No more lexer feedback hack.  In yacc grammars for C, a
standard hack, know as the "lexer feedback hack" is used
to find typedef names.  The lexer uses semantic
information to decide if any given identifier is a
typedef-name or not and returns a special token.  With
btyacc, you no longer need to do this; the lexer should
just always return an identifier.  The btyacc grammar then
needs a rule of the form:

typename: ID [ if (!IsTypeName(LookupId($1))) YYERROR; ]

While the hack works adequately well for parsing C, it
becomes a nightmare when you try to parse something like
C++, where treating an ID as a typedef becomes heavily
dependent on context.

* Easy disambiguation via simple ordering.  Btyacc runs
its trials via the rule "try shifting first, then try
reducing by the order that the conflicting rules appear in
the input file".  This means you can deal with semantic a
disambiguation rule like:
    [1] If it looks like a declaration it is, otherwise
    [2] If it looks like an expression it is, otherwise
    [3] it is a syntax error
	[Ellis & Stroustrup, Annotated C++ Reference Manual, p93]

To deal with this, you need only put all the rules for
declarations before the rules for expressions in the
grammar file.

* No extra cost if you do not use it.  Backtracking is
only triggered when the parse hits a shift/reduce or
reduce/reduce conflict in the table.  If you have no
conflicts in your grammar, there is no extra cost, other
than some extra code which will never be invoked.

* C++ and ANSI C compatible parsers.  The parsers produced
by btyacc can be compiled with C++ correctly.  If you
"#define" YYSTYPE to be some C++ type with constructor and
destructor, everything will work fine.  My favorite is
"#define YYSTYPE SmartPointer", where SmartPointer is a
smart pointer type that does garbage collection on the
pointed to objects.

BTYACC was originally written to make it easy to write a
C++ parser (my goal was to be able to use the grammar out
of the back of the ARM with as few modifications as
possible).  Anyone who has ever looked at Jim Roskind
public domain C++ yacc grammar, or the yacc-based grammar
used in g++ knows how difficult this is.  BTYACC is very
useful for parsing any ambiguous grammar, particularly
ones that come from trying to merge two (or more) complete
grammars.

Limitations of the backtracking: Currently, the generated
parser does NO pruning of alternate parsing paths.  To
avoid an exponential explosion of possible paths (and
parsing time), you need to manually tell the parser when
it can throw away saved paths using YYVALID.  In pratice,
this turns out to be fairly easy to do.  A C++ parser (for
example) can juts put a [YYVALID;] after every complete
declaration and statement rule, corresponding to pruning
the backtracking state after seeing a ';' or '}' -- there
will never be a situation in which it is useful to
backtrack past either of these.

Inherited attributes in btyacc:

Inherited attributes look a lot like function arguments to
non-terminals, which is what they end up being in a
recursive descent parser, but NOT how they are implemented
in btyacc.  Basically they are just syntactic sugar for
embedded semantic actions and $0, $-1, ... in normal yacc.
btyacc gives you two big advantages besides just the
syntax:
    1. it does type checking on the inherited attributes, 
       so you do not have to specify $<type>0 and makes sure 
       you give the correct number of arguments (inherited 
       attributes) to every use of a non-terminal.
    2. It "collapses" identical actions from that are produced 
       from inherited attributes.  This eliminates many 
       potential reduce-reduce conflicts arising from 
       the inherited attributes.

You use inherited attributes by declaring the types of the
attributes in the preamble with a type declaration and
declaring names of the attributes on the lhs of the yacc
rule.  You can of course have more than one rule with the
same lhs, and you can even give them different names in
each, but the type and number must be the same.

Here is a small example:
%type <t1> lhs(<t1>, <t2>)  /* lhs takes two inherited attributes */
	   stuff(<t1>, <t2>)
%%
lhs($i1, $i2) : { $$ = $i1 }
	      | lhs($i1, $i2) stuff($1,$i2) { $$ = $2; }

This is roughly equivalent to the following yacc code:
lhs : { $$ = $<t1>-1; }
    | lhs [ $<t1>$ = $-1; ] [ $<t2>$ = $<t2>0; ] stuff { $$ = $4; }
    ;

See the file "test/t2.y" for a longer and more complete
example.  At the current time, the start symbol cannot
have any arguments.

Variant parsers:

Btyacc supports the -S flag to use a different parser
skeleton, changing the way that the parser is called and
used.  The skeleton "push.skel" is included to produce a
"passive" parser that you feed tokens to (rather than
having the parser call a seperate yylex routine).  With
push.skel, yyparse is defined as follows:

int yyparse(int token, YYSTYPE yylval)

You should call yyparse repeatedly with successive tokens
of input.  It returns 0 if more input is needed, 1 for a
successful parse, and -1 for an unrecoverable parse error.