o For some unknown reason, the complexity of the benchmark platform
prevented some C++ compilers from generating optimal code, properly
reflective of the real performance in actual deployment.
o Added the json_benchmarks_simple target, which performs the same
suite of tests as json_benchmarks.
o Simplified the benchmark platform, and emit an "Average" TPS
(Transactions Per Second) value reflective of aggregate parse/output
performance.
o We assume the same character int_type as the unerlying std::istream
o There are no assumptions on the value of eof(), other than that it
will not be a valid unsigned char value.
o To retain performance, we do not allow swapping out the underlying
std::streambuf during our use of the std::istream for parsing.
o We can retain -Weffc++ and specify default initializers by using
initializer lists. The risks are low (of additional non-conformat
compilers), because there is already one other such initialization
used in the code-base.
o An (-'ve valued, typically -1) EOF must never be allowed in
token_string, as it be converted to 255 -- a legitimate value.
o Comparing against a specific eof() (-1, typically) is more costly than
detecting +'ve/-'ve. Since EOF is the only non-positive value allowed
we can use the simpler test.
o Removed unnecessary test for token_string size, as it is already
tested in the method, and must never occur in correct code; used an
assert instead.
o Return its contents when necessary. In many cases, this avoids
construction of multiple copies of the yytext token. Exceeds
performance of current develop branch.
o Use std::streambuf I/O instead of std::istream; does not maintain
(unused) istream flags.
o Further simplify get/unget handling.
o Restore original handling of NUL in input stream; ignored during
token_string escaping.
Perry Kundert