ORC-2013: [C++] Bump to CMake 3.25+ to use FetchContent

[wgtmac]

What changes were proposed in this pull request?

• Rename builtin FindXXX.cmake to FindXXXAlt.cmake to not interfere with cannonical package name. These are now only used when XXX_HOME or its friends are set. This includes:
  • FindGTest.cmake -> FindGTestAlt.cmake
  • FindLZ4.cmake -> FindLZ4Alt.cmake
  • FindProtobuf.cmake -> FindProtobufAlt.cmake
  • FindSnappy.cmake -> FindSnappyAlt.cmake
  • FindZLIB.cmake -> FindZLIBAlt.cmake
  • FindZSTD.cmake -> FindZSTDAlt.cmake
• Bump CMake minimum requirement to 3.25 to leverage FetchContent with FIND_PACKAGE_ARGS. For 3rd party dependencies now, we try to find system installed one via CONFIG mode or fallback to vendor it. This includes:
  • GTest
  • LZ4
  • Protobuf
  • Snappy
  • ZSTD
  • ZLIB: note that we use an unreleased version (1.4.0) to work around its misuse of build and install interface
  • Sparsehash: always vendor it with workaround to its CMake misuse

Why are the changes needed?

• Orc unable to find dependencies, even when they exist #2408
• Search for CMake and/or pkgconfig package data by default #2147
• Allow using external lib instead of trying to build bundled duplicates #2016
• GH-46375: [C++] Add adapters/orc directory to Meson arrow#46906

How was this patch tested?

• Pass all CIs
• Test it manually

Was this patch authored or co-authored using generative AI tooling?

No

[wgtmac]

This PR is ready for review.

The two failed CIs are due to outdated CMake versions from amazonlinux23 and debian11. Should we remove them? @dongjoon-hyun

For now XXX_HOME variables are still supported and take higher priority than FetchContent call, so Apache Arrow C++ should not be affected. Is it possible for Apache Arrow to leverage this improvement? @kou @raulcd

Can conda leverage improvement from this PR, or is there any additional change that is required to add? @h-vetinari @WillAyd

[kou]

If we call this in top-level, changed variables in this context are visible after this clause.

How about using function() or block()?

function(build_protobuf)
  prepare_fetchcontent()
  ...
endfunction()
build_protobuf()

https://cmake.org/cmake/help/latest/command/block.html

block()
  prepare_fetchcontent()
  ...
endblock()

[wgtmac]

Good suggestion! I've switched to use blocks.

[kou]

We may be able to remove SnappyAlt entirely by setting Snappy_ROOT:

Suggested change

	elseif (NOT "${SNAPPY_HOME}" STREQUAL "")
	find_package (Snappy REQUIRED)
	find_package (SnappyAlt REQUIRED)
	if (ORC_PREFER_STATIC_SNAPPY AND SNAPPY_STATIC_LIB)
	orc_add_resolved_library (orc_snappy ${SNAPPY_STATIC_LIB} ${SNAPPY_INCLUDE_DIR})
	else ()
	orc_add_resolved_library (orc_snappy ${SNAPPY_LIBRARY} ${SNAPPY_INCLUDE_DIR})
	endif ()
	list (APPEND ORC_SYSTEM_DEPENDENCIES Snappy)
	list (APPEND ORC_SYSTEM_DEPENDENCIES SnappyAlt)
	list (APPEND ORC_INSTALL_INTERFACE_TARGETS "$<INSTALL_INTERFACE:Snappy::snappy>")
	orc_provide_find_module (Snappy)
	orc_provide_find_module (SnappyAlt)
	else ()
	else ()
	if (NOT "${SNAPPY_HOME}" STREQUAL "" AND "${Snappy_ROOT}" STREQUAL "")
	set(Snappy_ROOT "${SNAPPY_HOME}")
	endif()

See also: https://cmake.org/cmake/help/latest/variable/PackageName_ROOT.html

[wgtmac]

Doesn't it break arrow-cpp (where SNAPPY_LIBRARY and SNAPPY_INCLUDE_DIR are explicitly provided)?

Update: I've tried to apply this change with or without CONFIG in the FIND_PACKAGE_ARGS. In both cases, system snappy is used instead of arrow vendored one.

-- Apache ORC: Found the Protobuf headers: /Users/gangwu/Projects/arrow/cpp/orc_build/protobuf_ep-install/include
-- Apache ORC: Found the Protobuf library: arrow::protobuf::libprotobuf
-- Apache ORC: Found the Protoc library: arrow::protobuf::libprotoc
-- Apache ORC: Found the Protoc executable: arrow::protobuf::protoc
-- Apache ORC: Found the Protobuf static library: /Users/gangwu/Projects/arrow/cpp/orc_build/protobuf_ep-install/lib/libprotobuf.a
-- Apache ORC: Found the Protoc static library: /Users/gangwu/Projects/arrow/cpp/orc_build/protobuf_ep-install/lib/libprotoc.a
-- Apache ORC: Providing CMake module for FindProtobufAlt as part of CMake package
-- Apache ORC: Using system snappy                     <==================
-- Apache ORC: ZLIB_HOME: /Users/gangwu/Projects/arrow/cpp/orc_build/zlib_ep/src/zlib_ep-install
-- Apache ORC: Found the ZLIB header: /Users/gangwu/Projects/arrow/cpp/orc_build/zlib_ep/src/zlib_ep-install/include/zlib.h
-- Apache ORC: Found the ZLIB library: ZLIB::ZLIB
-- Apache ORC: Found the ZLIB static library: /Users/gangwu/Projects/arrow/cpp/orc_build/zlib_ep/src/zlib_ep-install/lib/libz.a
-- Apache ORC: Providing CMake module for FindZLIBAlt as part of CMake package

[wgtmac]

I'm now using CONFIG mode in the FIND_PACKAGE_ARGS so it is not compatible if we remove XXXAlt.

[kou]

We may be able to remove FindZLIBAlt.cmake. CMake's FindZLIB provides ZLIB_ROOT hint: https://cmake.org/cmake/help/latest/module/FindZLIB.html#hints

[kou]

We may not need to install header files.

ORC doesn't use Snappy in public headers, right?

[wgtmac]

We need to install it, otherwise downstream project may fail like below:

CMake Error at CMakeLists.txt:9 (target_link_libraries):
  Cannot find source file:

    /tmp/orc/include/snappy-c.h

The reason is that Snappy adds these 4 header files as install interface: https://github.com/google/snappy/blob/1.2.2/CMakeLists.txt#L266