Functions

src/Functions/ is the registry of scalar functions. There are hundreds of them; the directory has more .cpp files than any other under src/.

Architecture

Every function is a class that implements IFunction (src/Functions/IFunction.h). It declares its name, argument signature, and a vectorized executeImpl(...) that takes a ColumnsWithTypeAndName and returns a ColumnPtr. Most functions are templated over column types and have specialized integer/float/string paths. The FunctionFactory (src/Functions/FunctionFactory.cpp) maps the name to a constructor.

Each function lives in its own .cpp file. They self-register with an anonymous-namespace void registerFunction<Name>() invoked from src/Functions/registerFunctions.cpp.

Big families

Family	Examples	File / dir
Arithmetic	+, -, *, /, intDiv, modulo, abs, negate, least, greatest	FunctionBinaryArithmetic.h (~160 KB header), plus.cpp, minus.cpp, multiply.cpp, divide.cpp
Comparison	equals, less, greater, …, like, match, multiMatchAny	equals.cpp, less.cpp, …
Logical	and, or, not, xor	FunctionsLogical.cpp
String	concat, substring, lower, upper, replaceAll, regex*, lpad, rpad, position, extract*, splitByChar	concat.cpp, substring.cpp, Functions/StringHelpers.cpp, Functions/extract*
Date/Time	toDate, toDateTime, toStartOfMonth, formatDateTime, addDays, subtractMinutes, dateDiff, now, today	DateTimeTransforms.h (~94 KB), makeDate.cpp, dateDiff.cpp, formatDateTime.cpp
Conversion	toUInt32, toString, CAST, parseDateTimeBestEffort, toType*	CastOverloadResolver.cpp, toString.cpp, parseDateTimeBestEffort.cpp
JSON	JSONExtract*, JSONHas, JSONLength, simdjson	FunctionsJSON.cpp (~90 KB)
Hashing	cityHash64, sipHash64, MD5, SHA*, xxHash*, farmHash*	FunctionsHashing.h (~70 KB), FunctionsHashingMisc.cpp, FunctionsHashingMurmur.cpp, CRC.cpp
Encoding	base64Encode, hex, unhex, base58Encode	FunctionBase64Conversion.h, FunctionBase58Conversion.h, FunctionBaseAI.h, hex.cpp
Encryption	encrypt, decrypt, tryDecrypt, aes_*	encrypt.cpp, decrypt.cpp, tryDecrypt.cpp
Random	rand, randNormal, randCanonical, generateUUIDv4	FunctionsRandom.cpp
Arrays	arrayMap, arrayFilter, arrayJoin, arrayReduce, arrayConcat, arrayResize, arraySort, groupArray*	a long tail in array*.cpp
Higher-order	arrayMap(x -> ...), arrayFilter(x -> ...)	implemented via IFunction::isSuitableForShortCircuitArgumentsExecution and lambdas
Tuple/Map	tupleElement, mapKeys, mapValues, mapContains, mapApply	tupleElement.cpp, map*.cpp
Geo	geohashEncode, geohashDecode, pointInPolygon, geoToS2, s2*	geohash*.cpp, geo*.cpp, pointInPolygon.cpp
Vectors	cosineDistance, L1Norm, L2Norm, dotProduct	vectorFunctions.cpp (~130 KB)
URL/IP	domain, topLevelDomain, IPv4StringToNum, cutQueryString	URL/, IP/ (subdirs)
Bitmap	bitmapAnd, bitmapOr, bitmapHasAll	bitmap*.cpp
External dictionaries	dictGet, dictGetOrDefault, dictHas, dictGetHierarchy	FunctionsExternalDictionaries.cpp
Sketches/HLL	uniqHLL12*, groupBitmap*, quantilesTDigest (most in aggregates)	shares code with src/AggregateFunctions/
AI	chatbotAnswer, embedText, aiGenerateText	Functions/AI/, builds on FunctionBaseAI.h and contrib/ai-sdk-cpp
Misc	version, hostname, errorCodeToName, getMacro, currentRoles, currentDatabase	version.cpp, getMacro.cpp, errorCodeToName.cpp, …

src/Functions/registerFunctions*.cpp carries the giant list of register* calls grouped by category.

Vectorized execution

A function receives ColumnPtrs (typed wrappers around a contiguous buffer) and produces a ColumnPtr. Hot paths are templated over physical types (e.g. ColumnVector<UInt32>) so the compiler unrolls the inner loop. For variable-length data (String, Array), functions operate on the offsets buffer plus the values buffer.

The generic flow:

1. Resolve overload via IFunctionOverloadResolver and castColumn to align argument types if needed.
2. Bind a concrete IExecutableFunction.
3. The executor calls executeImpl(arguments, result_type, input_rows_count) once per Chunk.
4. IFunction::isSuitableForConstantFolding, useDefaultImplementationForConstants, useDefaultImplementationForNulls, etc. control which fast paths are taken.

Lambdas / higher-order

arrayMap, arrayFilter, arrayCount, etc. accept a lambda. Internally the function compiles the lambda body via ActionsDAG, evaluates it against an offset-flattened view of the array, and assembles results.

JIT compilation

src/Interpreters/ExpressionJIT.cpp (using vendored LLVM under contrib/llvm-project) compiles inner expression loops to native code when compile_expressions=1. This kicks in for filter and projection DAGs that are evaluated many times per query.

CAST

src/Functions/CastOverloadResolver.cpp is a meta-function that picks the right typed conversion. It handles all the awkward cases (Nullable → not, LowCardinality ↔ regular, Decimal rounding, DateTime time-zone reinterpretation, Variant projection, Dynamic dispatch).

Adding a function

1. Write a .cpp under src/Functions/.
2. Subclass IFunction (or one of the helpers like FunctionFactory::register<NameOf>).
3. Add a void registerFunctionFoo(FunctionFactory & factory); and call it from registerFunctions.cpp.
4. Add tests under tests/queries/0_stateless/.

Per AGENTS.md: refer to the function as f, not f(), in commit messages and prose.

Related pages

• Aggregate functions
• Data types and columns
• Interpreters — ActionsDAG, JIT, expression analyzer.