Reported by Agency for Defense Development.
https://lists.gnu.org/r/bug-bison/2020-08/msg00008.html
On an empty such as
%token FOO
BAR
FOO 0
%%
input: %empty
we crash because when we find FOO 0, we decrement ntokens (since FOO
was discovered to be EOF, which is already known to be a token, so we
increment ntokens for it, and need to cancel this). This "works well"
when EOF is properly defined in one go, but here it is first defined
and later only assign token code 0. In the meanwhile BAR was given
the token number that we just decremented.
To fix this, assign symbol numbers after parsing, not during parsing,
so that we also saw all the explicit token codes. To maintain the
current numbers (I'd like to keep no difference in the output, not
just equivalence), we need to make sure the symbols are numbered in
the same order: that of appearance in the source file. So we need the
locations to be correct, which was almost the case, except for nterms
that appeared several times as LHS (i.e., several times as "foo:
..."). Fixing the use of location_of_lhs sufficed (it appears it was
intended for this use, but its implementation was unfinished: it was
always set to "false" only).
* src/symtab.c (symbol_location_as_lhs_set): Update location_of_lhs.
(symbol_code_set): Remove broken hack that decremented ntokens.
(symbol_class_set, dummy_symbol_get): Don't set number, ntokens and
nnterms.
(symbol_check_defined): Do it.
(symbols): Don't count nsyms here.
Actually, don't count nsyms at all: let it be done in...
* src/reader.c (check_and_convert_grammar): here. Define nsyms from
ntokens and nnterms after parsing.
* tests/input.at (EOF redeclared): New.
* examples/c/bistromathic/bistromathic.test: Adjust the traces: in
"%nterm <double> exp %% input: ...", exp used to be numbered before
input.
* examples/c/bistromathic/parse.y (user_context): We need the current
line.
(yyreport_syntax_error): Quote the guilty line, with squiggles.
* examples/c/bistromathic/bistromathic.test: Adjust.
Currently autocompletion on a line with errors leaks the error
messages. It can be useful to let the user know, but GNU Readline
does not provide us with an nice way to display the error. So we
actually break into the current line of the user.
So instead, do not show these errors.
* examples/c/bistromathic/parse.y (user_context): New.
Use %param to pass it to the parser and scanner.
Keep quiet when in computing autocompletion.
That quite defeats the whole point of locations... But anyway, we
should not see these messages at all.
* examples/c/bistromathic/parse.y (expected_tokens): Fix (useless)
location tracking.
Currently when a push parser finishes its parsing (i.e., it did not
return YYPUSH_MORE), it also clears its state. It is therefore
impossible to see if it had parse errors.
In the context of autocompletion, because error recovery might have
fired, the parser is actually already in a different state. For
instance on `(1 + + <TAB>` in the bistromathic, because there's a
`exp: "(" error ")"` recovery rule, `1 + +` tokens have already been
popped, replaced by `error`, and autocompletions think we are ready
for the closing ")". So here, we would like to see if there was a
syntax error, yet `yynerrs` was cleared.
In the case of a successful parse, we still have a problem: if error
recovery succeeded, we won't know it, since, again, `yynerrs` is
clearer.
It seems much more natural to leave the parser state available for
analysis when there is a failure.
To reuse the parser, we should either:
1. provide an explicit means to reinitialize a parser state for future
parses.
2. automatically reset the parser state when it is used in a new
parse.
Option 2 requires to check whether we need to reinitialize the parser
each time we call `yypush_parse`, i.e., each time we give a new token.
This seems expensive compared to Option 1, but benchmarks revealed no
difference. Option 1 is incompatible with the documentation
("After `yypush_parse` returns a status other than `YYPUSH_MORE`, the
parser instance `yyps` may be reused for a new parse.").
So Option 2 wins, reusing the private `yynew` member to record that a
parse was finished, and therefore that the state must reset in the
next call to `yypull_parse`.
While at it, this implementation now reuses the previously enlarged
stacks from one parse to another.
* data/skeletons/yacc.c (yypstate_new): Set up the stacks in their
initial configurations (setting their bottom to the stack array), and
use yypstate_clear to reset them (moving their top to their bottom).
(yypstate_delete): Adjust.
(yypush_parse): At the beginning, clear yypstate if needed, and at the
end, record when yypstate needs to be clearer.
* examples/c/bistromathic/parse.y (expected_tokens): Do not propose
autocompletion when there are parse errors.
* examples/c/bistromathic/bistromathic.test: Check that case.
Currently, completion when there is a syntax error shows broken
locations.
* examples/c/bistromathic/parse.y (expected_tokens): Initialize the
location.
* examples/c/bistromathic/bistromathic.test: Check that.
The user gives yyexpected_tokens a limit: the max number of tokens she
wants to hear about. That's because an error message that reports a
bazillion of possible tokens is useless.
In that case yyexpected_tokens returned 0, so the user would not know
if there are too many expected tokens or none (yes, that's possible).
There are several ways to tell the user in which situation she's in:
- return some E2MANY, a negative value. Then it makes the pattern
int argsize = yypcontext_expected_tokens (ctx, arg, ARGS_MAX);
if (argsize < 0)
return argsize;
no longer valid, as for E2MANY (i) the user must generate the error
message anyway, and (ii) she should not return E2MANY
- return ARGS_MAX + 1. Then it makes it dangerous for the user, as
she has to iterate update `min (ARGS_MAX, argsize)`.
Returning 0 is definitely simpler and safer for the user, as it tells
her "this is not an error, just generate your message without a list
of expecting tokens". So let's still return 0, but set arg[0] to the
empty token when the list is really empty.
* data/skeletons/glr.c, data/skeletons/lalr1.cc, data/skeletons/lalr1.java
* data/skeletons/yacc.c (yyexpected_tokens): Put the empty symbol
first if there are no possible tokens at all.
* examples/c/bistromathic/parse.y: Demonstrate how to use that.
The shell grammar does not allow empty statements in then/else part of
an if, but examples/test failed to catch the syntax errors from the
script it ran. So exited with success anyway.
You would expect 'set -e' to suffice, but with bash 3.2 actually it
does not. As a matter of fact, I could find a way to have this behave
properly:
$ cat test.sh
set -e
cleanup ()
{
status=$?
echo "cleanup: $status"
exit $status
}
trap cleanup 0 1 2 13 15
. $1
s=$?
echo "test.sh: $s"
exit $s
$ cat bistro.test
if true; then
fi
$ /bin/sh ./test.sh ./bistro.test
./bistro.test: line 2: syntax error near unexpected token `fi'
cleanup: 0
$ echo $?
0
Remove the set -e (or the trap), and tada, it works... So we have to
deal with the error by hand.
* examples/test ($exit): Replace with...
($status): this.
Preserve the exit status of the test case.
* examples/c/bistromathic/bistromathic.test: Fix syntax error.
Readline may emit escape sequences before the prompt.
Reported by Bruno Haible.
https://lists.gnu.org/r/platform-testers/2020-05/msg00001.html.
* examples/c/bistromathic/bistromathic.test: Trust readline _only_ if
we get what we expect on some reference computation.
Don't try to build bistromathic if we don't have readline.
Reported by Bruno Haible.
https://lists.gnu.org/r/bug-bison/2020-05/msg00028.html
* configure.ac (ENABLE_BISTROMATHIC): New.
* examples/c/bistromathic/local.mk: Use it.
* examples/c/bistromathic/bistromathic.test: Exit 77 for skip.
On OpenBSD 6.5, the prompt is repeated, but not the actual command
line... Don't try to cope with that.
Reported by Bruno Haible.
https://lists.gnu.org/r/bug-bison/2020-05/msg00015.html
* examples/c/bistromathic/bistromathic.test: Skip when readline behave
this way.
* data/skeletons/yacc.c (yyparse): When the scanner returns YYERRCODE,
go directly to error recovery (yyerrlab1).
However, don't keep the error token as lookahead, that token is too
special.
* data/skeletons/lalr1.cc: Likewise.
* examples/c/bistromathic/parse.y (yylex): Use that feature to report
nicely invalid characters.
* examples/c/bistromathic/bistromathic.test: Check that.
* examples/test: Neutralize gratuitous differences such as rule
position.
* tests/calc.at: Check that case in C only.
The other case seem to be working, but that's an illusion that the
next commit will address (in fact, they can enter endless loops, and
report the error several times anyway).
* examples/c/bistromathic/parse.y: here.
* examples/c/bistromathic/bistromathic.test: Check it.
Included a stupid case where the error is actually ignored.
When the user ctrl-d the line, we left the cursor not at col 0.
Let's fix that.
This revealed a few short-comings in the testing framework.
* examples/test (run): Also display the diffs.
And support -n.
* examples/c/bistromathic/bistromathic.test
* examples/c/bistromathic/parse.y
macOS' version of readline does not repeat stdin on stdout in
non-interactive mode, contrary to the current version of GNU readline.
* examples/test: Add support for strip_prompt.
* examples/c/bistromathic/bistromathic.test (strip_prompt): Set it
when needed.
Early exit when needed.
Currently pstate_new does not set up its variables, this task is left
to yypush_parse. This was probably to share more code with usual pull
parsers, where these (local) variables are indeed initialized by
yyparse.
But as a consequence yyexpected_tokens crashes at the very beginning
of the parse, since, for instance, the stacks are not even set up.
See https://lists.gnu.org/r/bison-patches/2020-03/msg00001.html.
The fix could have very simple, but the documentation actually makes
it very clear that we can reuse a pstate for several parses:
After yypush_parse returns a status other than YYPUSH_MORE, the
parser instance yyps may be reused for a new parse.
so we need to restore the parser to its pristine state so that (i) it
is ready to run the next parse, (ii) it properly supports
yyexpected_tokens for the next run.
* data/skeletons/yacc.c (b4_initialize_parser_state_variables): New,
extracted from the top of yyparse/yypush_parse.
(yypstate_clear): New.
(yypstate_new): Use it when push parsers are enabled.
Define after the yyps macros so that we can use the same code as the
regular pull parsers.
(yyparse): Use it when push parsers are _not_ enabled.
* examples/c/bistromathic/bistromathic.test: Check the completion on
the beginning of the line.
Currently completion on "at" proposes only "atan", but does not
actually complete "at" into "atan".
* examples/c/bistromathic/parse.y (completion): Install the lcp in
matches[0].
* examples/c/bistromathic/bistromathic.test: Check that case.
Currently "(1+<TAB>" does not work as expected, because "+" is not a
word breaking character.
* examples/c/bistromathic/parse.y (init_readline): Specify our word
breaking characters.
* examples/c/bistromathic/bistromathic.test: Avoid trailing spaces.
Let's use GNU readline and its TAB autocompletion to demonstrate the
use of yyexpected_tokens.
This shows a number of weaknesses in our current approach:
- some macros (yyssp, etc.) from push parsers "leak" in user code, we
need to undefine them
- the context needed by yyexpected_tokens does not need the token,
yypstate actually suffices
- yypstate is not properly setup when first allocated, which results
in a crash of yyexpected_tokens if fired before a first token was
read. We should move initialization from yypush_parse into
yypstate_new.
* examples/c/bistromathic/parse.y (yylex): Take input as a string, not
a file.
(EXIT): New token.
(input): Adjust to work only on a line.
(line): Remove.
(symbol_count, process_line, expected_tokens, completion)
(init_readline): New.
* examples/c/bistromathic/bistromathic.test: Adjust expectations.
