[C++] Avoid using dynamic_cast where possible by using hand rolled RTTI #3584
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
I should note that a similar performance boost was seen by an internal google team. @jcking's implementation is a more general version of that |
jcking
force-pushed
the
cpp-dynamic-cast-alt
branch
from
2da8925
to
31e81f6
Compare
|
@parrt I'm going to upload my draft PR for posterity, and include the performance numbers. |
|
I'm waiting on Google to approve the DCO in this repo, but here's my version https://github.com/antlr/antlr4/compare/dev...QuantamHD:surelog_perf?expand=1 It results in about a 25% performance uplift in walltime (808s vs 557s on Google benchmark task), and 11% reduction in profile cycles. Before Changes After Changes After the changes strcmp and dynamic_cast are nowhere near the top of the profile. I think @jcking's changes are great. I made mine more targeted by implementing specific is* functions for hot paths whereas Justin's is far more general. Either way removing RTTI in hot paths is a must. |
hzeller
reviewed
Mar 15, 2022
| LOOP_END = 12, | ||
| }; | ||
|
|
||
| std::string atnStateTypeName(ATNStateType atnStateType); |
There was a problem hiding this comment.
Not necessarily now (not sure what the current baseline requirements are, is antlr already compiling with c++17?), but longer term, returning std::string_view here is probably a good idea.
There was a problem hiding this comment.
We could use string_view, but then we cannot ever dynamically allocate. For example, if we wanted to do "INVALID(%d)".
There was a problem hiding this comment.
Given that ATNStateType is an enum class it would be very hard to get an invalid value in there. So I'd prefer performance over new string allocations to be prepared for an unlikely force-typecast any time.
There was a problem hiding this comment.
atnStateTypeName and transitionTypeName are only used by toString methods, which are not called anywhere in the ANTLR runtime except for debugging, which is off by default. So performance is not an issue in this case, unless the user is calling toString in which cause I would suggest not unless debugging because a bunch of them are not performant.
jcking
force-pushed
the
cpp-dynamic-cast-alt
branch
from
31e81f6
to
4ac9a74
Compare
hzeller
reviewed
Mar 15, 2022
| LOOP_END = 12, | ||
| }; | ||
|
|
||
| std::string atnStateTypeName(ATNStateType atnStateType); |
There was a problem hiding this comment.
Given that ATNStateType is an enum class it would be very hard to get an invalid value in there. So I'd prefer performance over new string allocations to be prepared for an unlikely force-typecast any time.
mike-lischke
approved these changes
Mar 16, 2022
There was a problem hiding this comment.
From my standpoint the patch is fine, but @hzeller has a few good points here, which might be worth to be considered.
jcking
force-pushed
the
cpp-dynamic-cast-alt
branch
from
4ac9a74
to
f4985e0
Compare
jcking
force-pushed
the
cpp-dynamic-cast-alt
branch
from
f4985e0
to
4202f4a
Compare
|
Ok, great! I'll address the php failing issue later. |
parrt
added a commit
that referenced
this pull request
Apr 10, 2022
* Get rid of reflection in CodeGenerator * Rename TargetType -> Language * Remove TargetType enum, use String instead as it was before Create CodeGenerator only one time during grammar processing, refactor code * Add default branch to appendEscapedCodePoint for unofficial targets (Kotlin) * Remove getVersion() overrides from Targets since they return the same value * Remove getLanguage() overrides from Targets since common implementation returns correct value * [again] don't use "quiet" option for mvn tests...hard to figure out what's wrong when failed. * normalize targets to 80 char strings for ATN serialization, except Java which needs big strings for efficiency. * Update actions.md fixed a small typo * Rename `CodeGenerator.createCodeGenerator` to `CodeGenerator.create` * Replace constants on string literals in `appendEscapedCodePoint` * Restore API of Target getLanguage(): protected -> public as it was before appendUnicodeEscapedCodePoint(int codePoint, StringBuilder sb, boolean escape): protected -> private (it's a new helper method, no need for API now) Added comment for appendUnicodeEscapedCodePoint * Introduce caseInsensitive lexer rule option, fixes #3436 * don't ahead of time compile for DART. See 8ca8804#commitcomment-62642779 * Simplify test rig related to timeouts (#3445) * remove all -q quiet mvn options to see output on CI servers. * run the various unit test classes in parallel rather than each individual test method, all except for Swift at the moment: `-Dparallel=classes -DthreadCount=4` * use bigger machine at circleci * No more test groups like parser1, parser2. * simplify Swift like the other tests * fix whitespace issues * use 4.10 not 4.9.4 * improve releasing antlr doc * Add Support For Swift Package Manager (#3132) * Add Swift Package Manager Support * Swift Package Dynamic * 【fix】【test】Fix run process path Co-authored-by: Terence Parr <parrt@cs.usfca.edu> * use src 11 for tool, but 8 for plugin/runtime (#3450) * use src 11 for tool, but 8 for plugin/runtime/runtime-tests. * use 11 in CI builds * cpp/cmake: Fix library install directories (#3447) This installs DLLs in bin directory instead of lib. * Python local import fixes (#3232) * Fixed pygrun relative import issue * Added name to contributors.txt Co-authored-by: Terence Parr <parrt@cs.usfca.edu> * Update javadoc to 8 and 11 (#3454) * no need for plugin in runtime, always gen svg from dot for javadoc, gen 1.8 not 1.7 doc for runtime. Gen 11 for tool. * tweak doc for 1.8 runtime. Test rig should gen 1.8 not 1.7 * [Go] Fix (*BitSet).equals (#3455) * set tool version for testing * oops reversion tool version as it's not sync'd with runtime and not time to release yet. * Remove unused variable from generated code (#3459) * [C++] Fix bugs in UnbufferedCharStream (#3420) * Escape bad words during grammar generation (#3451) * Escape reserved words during grammar generation, fixes #1070 (for -> for_ but RULE_for) Deprecate USE_OF_BAD_WORD * Make name and escapedName consistent across tool and codegen classes Fix other pull request notes * Rename NamedActionChunk to SymbolRefChunk * try out windows runners * rename workflow * Update windows.yml Fix cmd line issue * fix maven issue on windows * use jdk 11 * remove arch arg * display Github status for windows * try testing python3 on windows * try new run for python3 windows * try new run for python3 windows (again) * try new run for python3 windows (again2) * try new run for python3 windows (again3) * try new run for python3 windows (again4) * try new run for python3 windows (again5) * try new run for python3 windows * try new run for python3 windows * try new run for python3 windows * ugh i give up. python won't install on github actions. * Update windows.yml try python 3 * Update windows.yml * Update run-tests-python3.cmd * Update run-tests-python3.cmd * Create run-tests-python2.cmd * Update windows.yml * Update run-tests-python2.cmd * Update windows.yml * Update windows.yml * Update windows.yml * Create run-tests-javascript.cmd * Update run-tests-javascript.cmd * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Create run-tests-csharp.cmd * Update windows.yml * fix warnings in C# CI * Update windows.yml * Update windows.yml * Create run-tests-dart.cmd * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update run-tests-dart.cmd * Update run-tests-dart.cmd * Update run-tests-dart.cmd * Update run-tests-dart.cmd * Update windows.yml * Update windows.yml * Update windows.yml * Create run-tests-go.cmd * Update windows.yml * Update windows.yml * Update windows.yml * GitHub action php (#3474) * Update windows.yml * Create run-tests-php.cmd * Update run-tests-php.cmd * Update run-tests-php.cmd * Update run-tests-php.cmd * Update run-tests-php.cmd * Update windows.yml * Update windows.yml * Update windows.yml * Update run-tests-php.cmd * Update windows.yml * Cleanup ci (#3476) * Delete .appveyor directory * Delete .travis directory * Improve CI concurrency (#3477) * Update windows.yml * Update windows.yml * Update windows.yml * Optimize toArray replace toArray(new T[size]) with toArray(new T[0]) for better performance https://shipilev.net/blog/2016/arrays-wisdom-ancients/#_conclusion * add contributor * resolve conflicts * fix-maven-concurrency (#3479) * fix-maven-concurrency * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update run-tests-python2.cmd * Update run-tests-python3.cmd * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update windows.yml * Update run-tests-php.cmd * Update windows.yml * Update run-tests-dart.cmd * Update run-tests-csharp.cmd * Update run-tests-go.cmd * Update run-tests-java.cmd * Update run-tests-javascript.cmd * Update run-tests-php.cmd * Update run-tests-python2.cmd * Update run-tests-python3.cmd * increase Windows CI concurrency for all targets except Dart * Preserve line separators for input runtime tests data (#3483) * Preserve line separators for input data in runtime tests, fix test data Refactor and improve performance of BaseRuntimeTest * Add LineSeparator (\n, \r\n) tests * Set up .gitattributes for LineSeparator_LF.txt (eol=lf) and LineSeparator_CRLF.txt (eol=crlf) * Restore `\n` for all input in runtime tests, add extra LexerExec tests (LineSeparatorLf, LineSeparatorCrLf) * Add generated LargeLexer test, remove LargeLexer.txt descriptor * tweak name to be GeneratedLexerDescriptors * [JavaScript] Migrate from jest to jasmine * [C++] Fix Windows min/max macro collision * [C++] Update cmake README.md to C++17 * remove unnecessary comparisons. * Add useful function writeSerializedATNIntegerHistogram for writing out information concerning how many of each integer value appear in a serialized ATN. * fix comment indicating what goes in the serialized ATN. * move writeSerializedATNIntegerHistogram out of runtime. * follow guidelines * Fix .interp file parsing test for the Java runtime. Also includes separating the generation of the .interp file from writing it out so that we can use both independently. * Delete files no longer needed. Should have been part of #3520 * [C++] Optimizations and cleanups and const correctness, oh my * [C++] Optimize LL1Analyzer * [C++] Fix missing virtual destructors * Remove not used PROTECTED, PUBLIC, PRIVATE tokens from ANTLRLexer.g * Remove ANTLR 3 stuff from ANTLR grammars, deprecate ANTLR 3 errors * Remove not used imaginary tokens from ANTLRParser.g * Fix misprints in grammars * ATN serialized data: remove shifting by 2, remove UUID; fix #3515 Regenerate XPathLexer files * Disable native runtime tests (see #3521) * Implement Java-specific ATN data optimization (+-2 shift) * [C++] Remove now unused antlrcpp::Guid * pull new branch diagram from master * use dev not master branch for CI github * update doc from master * add back missing author * [C++] Fix const correctness in ATN and DFA * keep getSerializedATNSegmentLimit at max int * Fixes #3259 make InErrorRecoveryMode public for go * Change code gen template to capitalize InErrorRecoveryMode * [C++] Improve multithreaded performance, fix TSAN error, and fix profiling ATN simulator setup bug * Get rid of unnecessary allocations and calculations in SerializedATN * Get rid of excess char escaping in generated files, decrease size of output files Fix creation of excess fragments for Dart, Cpp, PHP runtimes * Swift: fix binary serialization and use instead of JSON * Fix targetCharValueEscape, make them final and static * [C++] Cleanup ATNDeserializer and remove related deprecated methods from ATNSimulator * Fix for #3557 (getting "go test" to work again). * Convert Python2/3 to use int arrays not strings for ATN encodings (#3561) * Convert Python2/3 to use int arrays not strings for ATN encodings. Also make target indicate int vs string. * rename and reverse ATNSerializedAsInts * add override * remove unneeded method * [C++] Drastically improve multi-threaded performance (#3550) Thanks guys. A major advancement. * [C++] Remove duplicate includes and remove unused includes (#3563) * [C++] Lazily deserialize ATN in generated code (#3562) * [Docs] Update Swift Docs (#3458) * Add Swift Package Manager Support * Swift Package Dynamic * 【fix】【test】Fix run process path * [Docs] [Swift] update link, remove expired descriptions Co-authored-by: Terence Parr <parrt@cs.usfca.edu> * Ascii only ATN serialization (#3566) * go back to generating pure ascii ATN serializations to avoid issues where target compilers might assume ascii vs utf-8. * forgot I had to change php on previous ATN serialization tweak. * change how we escapeChar() per target. * oops; gotta use escapeChar method * rm unneeded case * add @OverRide * use ints not chars for C# (#3567) * use ints not chars for C# * oops. remove 'quotes' * regen from XPathLexer.g4 * simplify ATN with bypass alts mechanism in Java. * Change string to int[] for serialized ATN for C#; removed unneeded `use System` from XPathLexer.g4; regen that grammar. * [C++] Use camel case name in generated lexers and parsers (#3565) * Change string to int array for serialized ATN for JavaScript (#3568) * perf: Add default implementation for Visit in ParseTreeVisitor. (#3569) * perf: Add default implementation for Visit in ParseTreeVisitor. Reference: https://github.com/antlr/antlr4/blob/ad29539cd2e94b2599e0281515f6cbb420d29f38/runtime/Java/src/org/antlr/v4/runtime/tree/AbstractParseTreeVisitor.java#L18 * doc: add contributor * Don't use utf decoding...these are just ints (#3573) * [Go] Cleanup and fix ATN deserialization verification (#3574) * [C++] Force generated static data type name to titlecase (#3572) * Use int array not string for ATN in Swift (#3575) * [C++] Fix generated Lexer static data constructor (#3576) * Use int array not string for ATN in Dart (#3578) * Fix PHP codegen to support int ATN serialization (#3579) * Update listener documentation to satisfy the discussion about improving exception handling: #3162 * tweak * [C++] Remove unused LexerATNSimulator::match_calls (#3570) * [C++] Remove unused LexerATNSimulator::match_calls * Remove match_calls from other targets * [Java] Preserve serialized ATN version 3 compatibility (#3583) * add jcking to the contributors list * Update releasing-antlr.md * [C++] Avoid using dynamic_cast where possible by using hand rolled RTTI (#3584) * Revert "[Java] Preserve serialized ATN version 3 compatibility (#3583)" This reverts commit 01bc811. * [C++] Add ANTLR4CPP_PUBLIC attributes to various symbols (#3588) * Update editorconfig for c++ (#3586) * Make it easier to contribute: Add c++ configuration for .editorconfig. Using the observed style with 2 indentation spaces. Signed-off-by: Henner Zeller <hzeller@google.com> * Add hzeller to contributors.txt Signed-off-by: Henner Zeller <hzeller@google.com> * Fix code style and typing to support PHP 8 (#3582) * [Go] Port locking algorithm from C++ to Go (#3571) * Use linux DCO not our old contributors certificate of origin * [C++] Fix bugs in SemanticContext (#3595) * [Go] Do not export Array2DHashSet which is an implementation detail (#3597) * Revert "Use linux DCO not our old contributors certificate of origin" This reverts commit b0f8551. * Use signed ints for ATN serialization not uint16, except for java (#3591) * refactor serialize so we don't need comments * more cleanup during refactor * store language in serializer obj * A lexer rule token type should never be -1 (EOF). 0 is fragment but then must be > 0. * Go uses int not uint16 for ATN now. java/go/python3 pass * remove checks for 0xFFFF in Go. * C++ uint16_t to int for ATN. * add mac php dir; fix type on accept() for generated code to be mixed. * Add test from @KvanTTT. This PR fixes #3555 for non-Java targets. * cleanup and add big lexer from #3546 * increase mvn mem size to 2G * increase mvn mem size to 8G * turn off the big ATN lexer test as we have memory issues during testing. * Fixes #3592 * Revert "C++ uint16_t to int for ATN." This reverts commit 4d2ebbf. # Conflicts: # runtime/Cpp/runtime/src/atn/ATNSerializer.cpp # runtime/Cpp/runtime/src/tree/xpath/XPathLexer.cpp * C++ uint16_t to int32_t for ATN. * rm unnecessary include file, updating project file. get rid of the 0xFFFF does in the C++ deserialization * rm refs to 0xFFFF in swift * javascript tests were running as Node...added to ignore list. * don't distinguish between 16 and 32 bit char sets in serialization; Python2/3 updated to work with this change. * update C++ to deserialize only 32-bit sets * 0xFFFF -> -1 for C++ target. * get other targets to use 32-bit sets in serialization. tests pass locally. * refactor to reduce code size * add comment * oops. comment out call to writeSerializedATNIntegerHistogram(). I wonder if this is why it ran out of memory during testing? * all but Java, Node, PHP, Go work now for the huge lexer file; I have set them to ignore. note that the swift target takes over a minute to lex it. I've turned off Node but it does not seem to terminate but it could terminate eventually. * all but Java, Node, PHP, Go work now for the huge lexer file; I have set them to ignore. note that the swift target takes over a minute to lex it. I've turned off Node but it does not seem to terminate but it could terminate eventually. * Turn off this big lexer because we get memory errors during continuous integration * Intermediate commit where I have shuffled around all of the -1 flipping and bumping by two. work still needs to be done because the token stream rewriter stuff fails. and I assume the other decoding for human readability testing if doesn't work * convert decode to use int[]; remove dead code. don't use serializeAsChar stuff. more tests pass. * more tests passing. simplify. When copying atn, must run ATN through serializer to set some state flags. * 0xFFFD+ are not valid char * clean up. tests passing now * huge clean up. Got Java working with 32-bit ATNs!Still working on cleanup but I want to run the tests * Cleanup the hack I did earlier; everything still seems to work * Use linux DCO not our old contributors certificate of origin * remove bump-by-2 code * clean up per @KvanTTT. Can't test locally on this box. Will see what CI says. * tweak comment * Revert "Use linux DCO not our old contributors certificate of origin" This reverts commit b0f8551. * see if C++ works in CI for huge ATN * Use linux DCO not our old contributors certificate of origin (#3598) * Use linux DCO not our old contributors certificate of origin * Revert "Use linux DCO not our old contributors certificate of origin" This reverts commit b0f8551. * use linux DCO * use linux DCO * Use linux DCO not our old contributors certificate of origin * update release documentation Signed-off-by: Terence Parr <parrt@antlr.org> * Equivalent of #3537 * clean up setup * clean up doc version * [Swift] improvements to equality functions (#3302) * fix default equality * equality cases * optional unwrapping * [Swift] Use for in loops (#3303) * common for in loops * reversed loop * drop first loop * for in with default BitSet * [Go] Fix symbol collision in generated lexers and parsers (#3603) * [C++] Refactor and optimize SemanticContext (#3594) * [C++] Devirtualize hand rolled RTTI for performance (#3609) * [C++] Add T::is for type hierarchy checks and remove some dynamic_cast (#3612) * [C++] Avoid copying statically generated serialized ATNs (#3613) * [C++] Refactor PredictionContext and yet more performance improvements (#3608) * [C++] Cleanup DFA, DFAState, LexerAction, and yet more performance improvements (#3615) * fix dependabot issues * [Swift] use stdlib (single pass) (#3602) * this was added to the stdlib in Swift 5 * &>> is defined as lhs >> (rhs % lhs.bitwidth) * the stdlib has these * reduce loops * use indices * append(contentsOf:) * Array literal init works for sets too! * inline and remove bit query functions * more optional handling (#3605) * [C++] Minor improvements to PredictionContext (#3616) * use php runtime dev branch to test dev * update doc to be more explicit about the interaction between lexer actions and semantic predicates; Fixes #3611. Fixes #3606. Signed-off-by: Terence Parr <parrt@antlr.org> * Refactor js runtime in preparation of future improvements * refactor, 1 file per class, use import, use module semantics, use webpack 5, use eslint * all tests pass * simplifications and alignment with standard js idioms * simplifications and alignment with standard js idioms * support reading legacy ATN * support both module and non-module imports * fix failing tests * fix failing tests * No longer necessary too generate sets or single atom transit that are bigger than 16bits. (#3620) * Updated getting started with Cpp documentation. (#3628) Included specific examples of using ANTLR4_TAG and ANTLR4_ZIP_REPOSITORY in the sample CMakeLists file. * [C++] Free ATNConfig lookup set in readonly ATNConfigSet (#3630) * [C++] Implement configurable PredictionContextMergeCache (#3627) * Allow to choose to switch off building tests in C++ (#3624) The new option to cmake ANTLR_BUILD_CPP_TESTS is default on (so the behavior is as before), but it provides a way to switch off if not needed. The C++ tests pull in an external dependency (googletests), which might conflict if ANTLR is used as a subproject in another cmake project. Signed-off-by: Henner Zeller <h.zeller@acm.org> * Fix NPE for undefined label, fix #2788 * An interval ought to be a value Interval was a pointer to 2 Ints it ought to be just 2 Ints, which is smaller and more semantically correct, with no need for a cache. However, this technically breaks metadata and AnyObject conformance but people shouldn't be relying on those for an Interval. * [C++] Remove more dynamic_cast usage * [C++] Introduce version macros * add license prefix * Prep 4.10 (#3599) * Tweak doc * Swift was referring to hardcoded version * Start version update script. * add files to update * clean up setup * clean up setup * clean up setup * don't need file * don't need file * Fixes #3600. add instructions and associated code necessary to build the xpath lexers. * clean up version nums * php8 * php8 * php8 * php8 * php8 * php8 * php8 * php8 * tweak doc * ok, i give up. php won't bump up too v8 * tweak doc * version number bumped to 4.10 in runtime. * Change the doc for releasing and update to use latest ST 4.3.2 * fix dart version to 4.10.0 * cmd files Cannot use export bash command. * try fixing php ci again * working on deploy Signed-off-by: Terence Parr <parrt@antlr.org> * php8 always install. * set js to 4.10.0 not 4.10 * turn off apt update for php circleci * try w/o cimg/php * try setting branch * ok i give up * tweak * update docs for release. * php8 circleci * use 3.5.3 antlr * use 3.5.3-SNAPSHOT antlr * use full 3.5.3 antlr * [Swift] reduce Optionals in APIs (#3621) * ParserRuleContext.children see comment in removeLastChild * TokenStream.getText * Parser._parseListeners this might require changes to the code templates? * ATN {various} * make computeReachSet return empty, not nil * overrides refine optionality * BufferedTokenStream getHiddenTokensTo{Left, Right} return empty not nil * Update Swift.stg * avoid breakage by adding overload of `getText` in extension * tweak to kick off build Signed-off-by: Terence Parr <parrt@antlr.org> * try parallelism: 4 circleci * Revert "[Swift] reduce Optionals in APIs (#3621)" This reverts commit b5ccba0. * tweaks to doc * Improve the deploy script and tweak the released doc. * use 4.10 not Snapshot for scripts Co-authored-by: Ivan Kochurkin <kvanttt@gmail.com> Co-authored-by: Alexandr <60813335+Alex-Andrv@users.noreply.github.com> Co-authored-by: 100mango <100mango@users.noreply.github.com> Co-authored-by: Biswapriyo Nath <nathbappai@gmail.com> Co-authored-by: Benjamin Spiegel <bspiegel11@gmail.com> Co-authored-by: Justin King <jcking@google.com> Co-authored-by: Eric Vergnaud <eric.vergnaud@wanadoo.fr> Co-authored-by: Harry Chan <harry.chan@codersatlas.com> Co-authored-by: Ken Domino <kenneth.domino@domemtech.com> Co-authored-by: chenquan <chenquan.dev@gmail.com> Co-authored-by: Marcos Passos <marcospassos@users.noreply.github.com> Co-authored-by: Henner Zeller <h.zeller@acm.org> Co-authored-by: Dante Broggi <34220985+Dante-Broggi@users.noreply.github.com> Co-authored-by: chris-miner <94078897+chris-miner@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
dynamic_castin hot code paths is much slower than using hand rolled RTTI using an enum. CPU profiling shows thatdynamic_castis one of the single most expensive calls. This change switches to using a hand rolled RTTI-like implementation using an enum.