cex: update NEWS for 3.7

[bison.git] / TODO

* Bison 3.7
** Cex
*** Improve gnulib
Don't do this (counterexample.c):

// This is the fastest way to get the tail node from the gl_list API.
gl_list_node_t
list_get_end (gl_list_t list)
{
  gl_list_node_t sentinel = gl_list_add_last (list, NULL);
  gl_list_node_t res = gl_list_previous_node (list, sentinel);
  gl_list_remove_node (list, sentinel);
  return res;
}

*** Ambiguous rewriting
If the user is stupid enough to have equal rules, then the derivations are
harder to read:

    Reduce/reduce conflict on tokens $end, "+", "⊕":
        2 exp: exp "+" exp .
        3 exp: exp "+" exp .
      Example                  exp "+" exp •
      First derivation         exp ::=[ exp "+" exp • ]
      Example                  exp "+" exp •
      Second derivation        exp ::=[ exp "+" exp • ]

Do we care about this?  In color, we use twice the same color here, but we
could try to use the same color for the same rule.

*** XML reports
Show the counterexamples.  This is going to be really hard and/or painful.
Unless we play it dumb (little structure).

** Bistromathic
- How about not evaluating incomplete lines when the text is not finished
  (as shells do).

** Questions
*** Java
- Should i18n be part of the Lexer?  Currently it's a static method of
  Lexer.

- is there a migration path that would allow to use TokenKinds in
  yylex?

- define the tokens as an enum too.

- promote YYEOF rather than EOF.

*** D
- is there a way to attach yysymbol_name to the enum itself?  As we did
  in Java.

- It would be better to have TokenKind as return value.  Can we use
  reflection to support both output types?

** YYerror
https://git.savannah.gnu.org/gitweb/?p=gettext.git;a=blob;f=gettext-runtime/intl/plural.y;h=a712255af4f2f739c93336d4ff6556d932a426a5;hb=HEAD

should be updated to not use YYERRCODE.  Returning an undef token is good
enough.

** Java
*** calc.at
Stop hard-coding "Calc".  Adjust local.at (look for FIXME).

** A dev warning for b4_
Maybe we should check for m4_ and b4_ leaking out of the m4 processing, as
Autoconf does.  It would have caught overquotation issues.

** doc
I feel it's ugly to use the GNU style to declare functions in the doc.  It
generates tons of white space in the page, and may contribute to bad page
breaks.

** consistency
token vs terminal.

** api.token.raw
The YYUNDEFTOK could be assigned a semantic value so that yyerror could be
used to report invalid lexemes.

** push parsers
Consider deprecating impure push parsers.  They add a lot of complexity, for
a bad feature.  On the other hand, that would make it much harder to sit
push parsers on top of pull parser.  Which is currently not relevant, since
push parsers are measurably slower.

** %define parse.error formatted
How about pushing bistromathics' yyreport_syntax_error as another standard
way to generate the error message, and leave to the user the task of
providing the message formats?  Currently in bistro, it reads:

    const char *
    error_format_string (int argc)
    {
      switch (argc)
        {
        default: /* Avoid compiler warnings. */
        case 0: return _("%@: syntax error");
        case 1: return _("%@: syntax error: unexpected %u");
          // TRANSLATORS: '%@' is a location in a file, '%u' is an
          // "unexpected token", and '%0e', '%1e'... are expected tokens
          // at this point.
          //
          // For instance on the expression "1 + * 2", you'd get
          //
          // 1.5: syntax error: expected - or ( or number or function or variable before *
        case 2: return _("%@: syntax error: expected %0e before %u");
        case 3: return _("%@: syntax error: expected %0e or %1e before %u");
        case 4: return _("%@: syntax error: expected %0e or %1e or %2e before %u");
        case 5: return _("%@: syntax error: expected %0e or %1e or %2e or %3e before %u");
        case 6: return _("%@: syntax error: expected %0e or %1e or %2e or %3e or %4e before %u");
        case 7: return _("%@: syntax error: expected %0e or %1e or %2e or %3e or %4e or %5e before %u");
        case 8: return _("%@: syntax error: expected %0e or %1e or %2e or %3e or %4e or %5e or %6e before %u");
        }
    }

The message would have to be generated in a string, and pushed to yyerror.
Which will be a pain in the neck in yacc.c.

If we want to do that, we should think very carefully about the syntax of
the format string.

** yyclearin does not invoke the lookahead token's %destructor
https://lists.gnu.org/r/bug-bison/2018-02/msg00000.html
Rici:

> Modifying yyclearin so that it calls yydestruct seems like the simplest
> solution to this issue, but it is conceivable that such a change would
> break programs which already perform some kind of workaround in order to
> destruct the lookahead symbol. So it might be necessary to use some kind of
> compatibility %define, or to create a new replacement macro with a
> different name such as yydiscardin.
>
> At a minimum, the fact that yyclearin does not invoke the %destructor
> should be highlighted in the documentation, since it is not at all obvious.

** Issues in i18n

Les catégories d'avertissements incluent :
  conflicts-sr      conflits S/R (activé par défaut)
  conflicts-rr      conflits R/R (activé par défaut)
  dangling-alias    l'alias chaîne n'est pas attaché à un symbole
  deprecated        construction obsolète
  empty-rule        règle vide sans %empty
  midrule-values    valeurs de règle intermédiaire non définies ou inutilisées
  precedence        priorité et associativité inutiles
  yacc              incompatibilités avec POSIX Yacc
  other             tous les autres avertissements (activé par défaut)
  all               tous les avertissements sauf « dangling-alias » et « yacc »
  no-CATEGORY       désactiver les avertissements dans CATEGORIE
  none              désactiver tous les avertissements
  error[=CATEGORY]  traiter les avertissements comme des erreurs

Line -1 and -3 should mention CATEGORIE, not CATEGORY.

* Bison 3.8
** Rewrite glr.cc
Get rid of scaffolding in glr.c.

** Unit rules / Injection rules (Akim Demaille)
Maybe we could expand unit rules (or "injections", see
https://homepages.cwi.nl/~daybuild/daily-books/syntax/2-sdf/sdf.html), i.e.,
transform

        exp: arith | bool;
        arith: exp '+' exp;
        bool: exp '&' exp;

into

        exp: exp '+' exp | exp '&' exp;

when there are no actions.  This can significantly speed up some grammars.
I can't find the papers.  In particular the book 'LR parsing: Theory and
Practice' is impossible to find, but according to 'Parsing Techniques: a
Practical Guide', it includes information about this issue.  Does anybody
have it?

** clean up (Akim Demaille)
Do not work on these items now, as I (Akim) have branches with a lot of
changes in this area (hitting several files), and no desire to have to fix
conflicts.  Addressing these items will happen after my branches have been
merged.

*** lalr.c
Introduce a goto struct, and use it in place of from_state/to_state.
Rename states1 as path, length as pathlen.
Introduce inline functions for things such as nullable[*rp - ntokens]
where we need to map from symbol number to nterm number.

There are probably a significant part of the relations management that
should be migrated on top of a bitsetv.

*** closure
It should probably take a "state*" instead of two arguments.

*** traces
The "automaton" and "set" categories are not so useful.  We should probably
introduce lr(0) and lalr, just the way we have ielr categories.  The
"closure" function is too verbose, it should probably have its own category.

"set" can still be used for summariring the important sets.  That would make
tests easy to maintain.

*** complain.*
Rename these guys as "diagnostics.*" (or "diagnose.*"), since that's the
name they have in gcc, clang, etc.  Likewise for the complain_* series of
functions.

*** ritem
states/nstates, rules/nrules, ..., ritem/nritems
Fix the latter.

* Better error messages
The users are not provided with enough tools to forge their error messages.
See for instance "Is there an option to change the message produced by
YYERROR_VERBOSE?" by Simon Sobisch, on bison-help.

See also
https://www.cs.tufts.edu/~nr/cs257/archive/clinton-jefferey/lr-error-messages.pdf
https://research.swtch.com/yyerror
http://gallium.inria.fr/~fpottier/publis/fpottier-reachability-cc2016.pdf

* Modernization
Fix data/skeletons/yacc.c so that it defines YYPTRDIFF_T properly for modern
and older C++ compilers.  Currently the code defaults to defining it to
'long' for non-GCC compilers, but it should use the proper C++ magic to
define it to the same type as the C ptrdiff_t type.

* Completion
Several features are not available in all the backends.

- lac: D, Java (easy)
- push parsers: glr.c, glr.cc, lalr1.cc (not very difficult)
- token constructors: Java, C, D (a bit difficult)
- glr: D, Java (super difficult)

* Bugs
** Autotest has quotation issues
tests/input.at:1730:AT_SETUP([%define errors])

->

$ ./tests/testsuite -l | grep errors | sed q
  38: input.at:1730      errors

* Short term
** Get rid of YYPRINT and b4_toknum
Besides yytoknum is wrong when api.token.raw is defined.

** Better design for diagnostics
The current implementation of diagnostics is ad hoc, it grew organically.
It works as a series of calls to several functions, with dependency of the
latter calls on the former.  For instance:

      complain (&sym->location,
                sym->content->status == needed ? complaint : Wother,
                _("symbol %s is used, but is not defined as a token"
                  " and has no rules; did you mean %s?"),
                quote_n (0, sym->tag),
                quote_n (1, best->tag));
      if (feature_flag & feature_caret)
        location_caret_suggestion (sym->location, best->tag, stderr);

We should rewrite this in a more FP way:

1. build a rich structure that denotes the (complete) diagnostic.
   "Complete" in the sense that it also contains the suggestions, the list
   of possible matches, etc.

2. send this to the pretty-printing routine.  The diagnostic structure
   should be sufficient so that we can generate all the 'format' of
   diagnostics, including the fixits.

If properly done, this diagnostic module can be detached from Bison and be
put in gnulib.  It could be used, for instance, for errors caught by
xgettext.

There's certainly already something alike in GCC.  At least that's the
impression I get from reading the "-fdiagnostics-format=FORMAT" part of this
page:

https://gcc.gnu.org/onlinedocs/gcc/Diagnostic-Message-Formatting-Options.html

** Graphviz display code thoughts
The code for the --graph option is over two files: print_graph, and
graphviz. This is because Bison used to also produce VCG graphs, but since
this is no longer true, maybe we could consider these files for fusion.

An other consideration worth noting is that print_graph.c (correct me if I
am wrong) should contain generic functions, whereas graphviz.c and other
potential files should contain just the specific code for that output
format. It will probably prove difficult to tell if the implementation is
actually generic whilst only having support for a single format, but it
would be nice to keep stuff a bit tidier: right now, the construction of the
bitset used to show reductions is in the graphviz-specific code, and on the
opposite side we have some use of \l, which is graphviz-specific, in what
should be generic code.

Little effort seems to have been given to factoring these files and their
rint{,-xml} counterpart. We would very much like to re-use the pretty format
of states from .output for the graphs, etc.

Since graphviz dies on medium-to-big grammars, maybe consider an other tool?

** push-parser
Check it too when checking the different kinds of parsers.  And be
sure to check that the initial-action is performed once per parsing.

** m4 names
b4_shared_declarations is no longer what it is.  Make it
b4_parser_declaration for instance.

** yychar in lalr1.cc
There is a large difference bw maint and master on the handling of
yychar (which was removed in lalr1.cc).  See what needs to be
back-ported.


    /* User semantic actions sometimes alter yychar, and that requires
       that yytoken be updated with the new translation.  We take the
       approach of translating immediately before every use of yytoken.
       One alternative is translating here after every semantic action,
       but that translation would be missed if the semantic action
       invokes YYABORT, YYACCEPT, or YYERROR immediately after altering
       yychar.  In the case of YYABORT or YYACCEPT, an incorrect
       destructor might then be invoked immediately.  In the case of
       YYERROR, subsequent parser actions might lead to an incorrect
       destructor call or verbose syntax error message before the
       lookahead is translated.  */

    /* Make sure we have latest lookahead translation.  See comments at
       user semantic actions for why this is necessary.  */
    yytoken = yytranslate_ (yychar);


** Get rid of fake #lines [Bison: ...]
Possibly as simple as checking whether the column number is nonnegative.

I have seen messages like the following from GCC.

<built-in>:0: fatal error: opening dependency file .deps/libltdl/argz.Tpo: No such file or directory


** Discuss about %printer/%destroy in the case of C++.
It would be very nice to provide the symbol classes with an operator<<
and a destructor.  Unfortunately the syntax we have chosen for
%destroy and %printer make them hard to reuse.  For instance, the user
is invited to write something like

   %printer { debug_stream() << $$; } <my_type>;

which is hard to reuse elsewhere since it wants to use
"debug_stream()" to find the stream to use.  The same applies to
%destroy: we told the user she could use the members of the Parser
class in the printers/destructors, which is not good for an operator<<
since it is no longer bound to a particular parser, it's just a
(standalone symbol).

* Various
** Rewrite glr.cc in C++ (Valentin Tolmer)
As a matter of fact, it would be very interesting to see how much we can
share between lalr1.cc and glr.cc.  Most of the skeletons should be common.
It would be a very nice source of inspiration for the other languages.

Valentin Tolmer is working on this.

** yychar == YYEMPTY
The code in yyerrlab reads:

      if (yychar <= YYEOF)
        {
          /* Return failure if at end of input.  */
          if (yychar == YYEOF)
            YYABORT;
        }

There are only two yychar that can be <= YYEOF: YYEMPTY and YYEOF.
But I can't produce the situation where yychar is YYEMPTY here, is it
really possible?  The test suite does not exercise this case.

This shows that it would be interesting to manage to install skeleton
coverage analysis to the test suite.

* From lalr1.cc to yacc.c
** Single stack
Merging the three stacks in lalr1.cc simplified the code, prompted for
other improvements and also made it faster (probably because memory
management is performed once instead of three times).  I suggest that
we do the same in yacc.c.

(Some time later): it's also very nice to have three stacks: it's more dense
as we don't lose bits to padding.  For instance the typical stack for states
will use 8 bits, while it is likely to consume 32 bits in a struct.

We need trustworthy benchmarks for Bison, for all our backends.  Akim has a
few things scattered around; we need to put them in the repo, and make them
more useful.

* Report

** Figures
Some statistics about the grammar and the parser would be useful,
especially when asking the user to send some information about the
grammars she is working on.  We should probably also include some
information about the variables (I'm not sure for instance we even
specify what LR variant was used).

** GLR
How would Paul like to display the conflicted actions?  In particular,
what when two reductions are possible on a given lookahead token, but one is
part of $default.  Should we make the two reductions explicit, or just
keep $default?  See the following point.

** Disabled Reductions
See 'tests/conflicts.at (Defaulted Conflicted Reduction)', and decide
what we want to do.

** Documentation
Extend with error productions.  The hard part will probably be finding
the right rule so that a single state does not exhibit too many yet
undocumented ''features''.  Maybe an empty action ought to be
presented too.  Shall we try to make a single grammar with all these
features, or should we have several very small grammars?

* Extensions
** More languages?
Well, only if there is really some demand for it.

*** PHP
https://github.com/scfc/bison-php/blob/master/data/lalr1.php

*** Python
https://lists.gnu.org/r/bison-patches/2013-09/msg00000.html and following

** Multiple start symbols
Would be very useful when parsing closely related languages.  The idea is to
declare several start symbols, for instance

    %start stmt expr
    %%
    stmt: ...
    expr: ...

and to generate parse(), parse_stmt() and parse_expr().  Technically, the
above grammar would be transformed into

   %start yy_start
   %token YY_START_STMT YY_START_EXPR
   %%
   yy_start: YY_START_STMT stmt | YY_START_EXPR expr

so that there are no new conflicts in the grammar (as would undoubtedly
happen with yy_start: stmt | expr).  Then adjust the skeletons so that this
initial token (YY_START_STMT, YY_START_EXPR) be shifted first in the
corresponding parse function.

** %include
This is a popular demand.  We already made many changes in the parser that
should make this reasonably easy to implement.

Bruce Mardle <marblypup@yahoo.co.uk>
https://lists.gnu.org/archive/html/bison-patches/2015-09/msg00000.html

However, there are many other things to do before having such a feature,
because I don't want a % equivalent to #include (which we all learned to
hate).  I want something that builds "modules" of grammars, and assembles
them together, paying attention to keep separate bits separated, in pseudo
name spaces.

** Push parsers
There is demand for push parsers in Java and C++.  And GLR I guess.

** Generate code instead of tables
This is certainly quite a lot of work.  See
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.50.4539.

** $-1
We should find a means to provide an access to values deep in the
stack.  For instance, instead of

        baz: qux { $$ = $<foo>-1 + $<bar>0 + $1; }

we should be able to have:

  foo($foo) bar($bar) baz($bar): qux($qux) { $baz = $foo + $bar + $qux; }

Or something like this.

** %if and the like
It should be possible to have %if/%else/%endif.  The implementation is
not clear: should it be lexical or syntactic.  Vadim Maslow thinks it
must be in the scanner: we must not parse what is in a switched off
part of %if.  Akim Demaille thinks it should be in the parser, so as
to avoid falling into another CPP mistake.

(Later): I'm sure there's actually good case for this.  People who need that
feature can use m4/cpp on top of Bison.  I don't think it is worth the
trouble in Bison itself.

** XML Output
There are couple of available extensions of Bison targeting some XML
output.  Some day we should consider including them.  One issue is
that they seem to be quite orthogonal to the parsing technique, and
seem to depend mostly on the possibility to have some code triggered
for each reduction.  As a matter of fact, such hooks could also be
used to generate the yydebug traces.  Some generic scheme probably
exists in there.

XML output for GNU Bison and gcc
   http://www.cs.may.ie/~jpower/Research/bisonXML/

XML output for GNU Bison
   http://yaxx.sourceforge.net/

* Coding system independence
Paul notes:

        Currently Bison assumes 8-bit bytes (i.e. that UCHAR_MAX is
        255).  It also assumes that the 8-bit character encoding is
        the same for the invocation of 'bison' as it is for the
        invocation of 'cc', but this is not necessarily true when
        people run bison on an ASCII host and then use cc on an EBCDIC
        host.  I don't think these topics are worth our time
        addressing (unless we find a gung-ho volunteer for EBCDIC or
        PDP-10 ports :-) but they should probably be documented
        somewhere.

        More importantly, Bison does not currently allow NUL bytes in
        tokens, either via escapes (e.g., "x\0y") or via a NUL byte in
        the source code.  This should get fixed.

* Broken options?
** %token-table
** Skeleton strategy
Must we keep %token-table?

* Precedence

** Partial order
It is unfortunate that there is a total order for precedence.  It
makes it impossible to have modular precedence information.  We should
move to partial orders (sounds like series/parallel orders to me).

This is a prerequisite for modules.

* Pre and post actions.
From: Florian Krohm <florian@edamail.fishkill.ibm.com>
Subject: YYACT_EPILOGUE
To: bug-bison@gnu.org
X-Sent: 1 week, 4 days, 14 hours, 38 minutes, 11 seconds ago

The other day I had the need for explicitly building the parse tree. I
used %locations for that and defined YYLLOC_DEFAULT to call a function
that returns the tree node for the production. Easy. But I also needed
to assign the S-attribute to the tree node. That cannot be done in
YYLLOC_DEFAULT, because it is invoked before the action is executed.
The way I solved this was to define a macro YYACT_EPILOGUE that would
be invoked after the action. For reasons of symmetry I also added
YYACT_PROLOGUE. Although I had no use for that I can envision how it
might come in handy for debugging purposes.
All is needed is to add

#if YYLSP_NEEDED
    YYACT_EPILOGUE (yyval, (yyvsp - yylen), yylen, yyloc, (yylsp - yylen));
#else
    YYACT_EPILOGUE (yyval, (yyvsp - yylen), yylen);
#endif

at the proper place to bison.simple. Ditto for YYACT_PROLOGUE.

I was wondering what you think about adding YYACT_PROLOGUE/EPILOGUE
to bison. If you're interested, I'll work on a patch.

* Better graphics
Equip the parser with a means to create the (visual) parse tree.


Local Variables:
mode: outline
coding: utf-8
fill-column: 76
End:

-----

Copyright (C) 2001-2004, 2006, 2008-2015, 2018-2020 Free Software
Foundation, Inc.

This file is part of Bison, the GNU Compiler Compiler.

Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.3 or
any later version published by the Free Software Foundation; with no
Invariant Sections, with no Front-Cover Texts, and with no Back-Cover
Texts.  A copy of the license is included in the "GNU Free
Documentation License" file as part of this distribution.