This is used to help the user catch cases where some value gets
ovewritten by a new one. This should not happen, as this will
probably leak.
Unfortunately this uncovered a bug in the C++ parser itself: the
lookahead value was not destroyed between two calls to yylex. For
instance if the previous lookahead was a std::string, and then an int,
then the value of the std::string was correctly taken (i.e., the
lookahead was now an empty string), but std::string structure itself
was not reclaimed.
This is now done in variant::build(other&) (which is used to take the
value of the lookahead): other is not only stolen from its value, it
is also destroyed. This incurs a new performance penalty of a few
percent, and union becomes faster again.
* data/lalr1-fusion.cc (variant::build(other&)): Destroy other.
(b4_variant_if): New.
(variant::built): New.
Use it whereever the status of the variant changes.
* etc/bench.pl.in: Check the penalty of %define assert.
* data/lalr1-fusion.cc (b4_symbol_variant): Adjust additional
arguments.
(variant::build): New overload for
copy-construction-that-destroys.
(variant::swap): New.
(parser::yypush_): Use it in variant mode.
* data/bison.m4 (b4_copyright): Fix the indentation of the
copyright year paragraph.
Use b4_copyright_years when no years are given.
* data/lalr1.cc, data/lalr1-fusion.cc, data/location.cc
(b4_copyright_years): New.
Use it.
In order to make it easy to perform benchmarks to ensure that there are no
performance loss, lalr1.cc is forked into lalr1-fusion.cc. Eventually,
lalr1-fusion.cc will replace lalr1.cc.
Meanwhile, to make sure that lalr1-fusion.cc is correctly exercized by the
test suite, the user must install a symbolic link from lalr1.cc to it.
Instead of having three stacks (state, value, location), use a stack
of triples. This considerably simplifies the code (and it will be
easier not to require locations as currently does the C++ parser),
and also gives a 10% speedup according to etc/bench (probably mainly since
memory allocation is done once instead of three times).
Another motivation is to make it easier to destruct properly
semantic values: now that they are bound to their state (hence
symbol type) it will be easier to call the appropriate destructor.
These changes should probably benefit the C parser too.
* data/lalr1.cc: Copy as... * data/lalr1-fusion.cc: this new
file.
(b4_rhs_value, b4_rhs_location): New definitions overriding those
from c++.m4.
(state_stack_type, semantic_stack_type, location_stack_type)
(yystate_stack_, yysemantic_stack_, yylocation_stack_): Remove.
(data_type, stack_type, yystack_): New.
(YYLLOC_DEFAULT, yypush_): Adjust.
(yyerror_range): Now based on data_type, not location_type.
