C++ runtime: Make Parser::exitRule() non-virtual #2162
Conversation
Add missing override markers to the following functions of the C++ runtime: - TokensStartState::getStateType() - TagChunk::toString() - TextChunk::toString() The missing markers made builds against the API with -Wsuggest-override choke.
Parser::exitRule() in the C++ runtime is a virtual function which is not reimplemented anywhere. OTOH, it is invoked during the execution of every rule, which can cause a noticeable performance hit. This commit removes its virtual qualifier. It should make a difference particularly for large grammars, because the number of rules corresponds to the number of the Parser object's virtual functions, and, consequently, its vtable lookup time. Tested with a VHDL grammar of 436 rules, where it brings down parsing time from 75 to 44 seconds on unoptimized compilation, i.e. a 40% speed gain. Still a lot slower than an equivalent java parser, though, which takes 2.64 seconds for the same input.
|
This is a very interesting outcome. VMT lookup shouldn't have such an impact and indeed I cannot reproduce this result with my (also fairly large grammar, > 500 parser rules, > 800 lexer rules). I used deeply nested expression to produce heavy recursion. But as I said, I can't reproduce that. Could you attach your grammar and sample input here? I then can try with your input to see if that makes a difference. |
:-(. Did you compile with -O3 by chance? Which compiler do you use? I measured with GCC 7.2.1 from -O0 up to -O2, with -O3 the gap (nearly) closed, 44 seconds vs 40 or so, i.e. 10%.
I've attached it to an email to you. I'm iffy about publishing at this point, because it's still part of a commercial project. |
|
I'm using clang (with XCode) and get these numbers (using your vhdl grammar): Java: C++ without Release with 'virtual': Release That's a bit different than what I get with my MySQL grammar, so it seems to be related also to the grammar structure. Comparing the release build times gives us:
hence ~45% speed increase (but 13 times slower than the Java parser, 10 times without |
|
Here are the results for my MySQL grammar (using 3 simple but highly nested queries): Java C++ (release only) w/o |
|
@mike-lischke should I merge? |
This patch aims at improving the performance of ANTLR-generated C++ parsers for large grammars.
Parser::exitRule() in the C++ runtime is a virtual function which is not reimplemented anywhere. OTOH, it is invoked during the execution of every rule, which seems to have a noticeable influence on performance. This commit removes its virtual qualifier. It should make a difference particularly for large grammars, because the number of rules corresponds to the number of the Parser object's virtual functions, and, consequently, its vtable lookup time.
Tested with a VHDL grammar of 436 rules, where it brings down parsing time from 75 to 44 seconds on unoptimized compilation, i.e. a 40% speed gain. Still a lot slower than an equivalent java parser, which takes 2.64 seconds for the same input.