[C++] Fix Meson Dependency Declarations for WrapDB

[WillAyd]

Describe the bug, including details regarding any error messages, version, and platform.

As discovered in mesonbuild/wrapdb#2099 (comment), we have a few issues that need to be corrected for the next release to make the WrapDB entry for Arrow easier to use. Specifically, we need to:

• Ensure the pkg-config entry for acero includes the Acero lib itself
• Create a dependency to match the arrow-csv pkgconfig
• Create a dependency to match the arrow-filesystem pkgconfig
• Create a dependency to match the arrow-json pkgconfig
• Create a dependency to match the arrow-tensorflow pkgconfig

Additionally, it is recommended that we call meson.override_dependency for our exposed dependencies (including arrow and arrow-compute) so that the wrap entry does not need to depend on the internal variable name (see mesonbuild/wrapdb#2206 (comment)). We may also want to explore lowering the Meson version requirement, unless Apache Arrow truly needs version 1.3.0 as a minimum

Component(s)

C++

[bgilbert]

You can set the Meson version to >=0, configure, and see what constructs Meson warns you about. That approach isn't perfect (Meson doesn't warn about every possible bug or consideration) but will help narrow down the minimum.

[WillAyd]

Nice idea - thanks @bgilbert !

[WillAyd]

From a quick glance I don't think we can relax the min version of 1.3.0 here, at least according to the following warnings:

src/arrow/meson.build:423: WARNING: Project targets '>=0' but uses feature introduced in '1.3.0': cpp_shared_args arg in library.
src/arrow/util/meson.build:62: WARNING: Project targets '>=0' but uses feature introduced in '1.3.0': compiler.has_define.

cpp_shared_args in particular is used heavily to define macros for symbol export visbiility. We may be able to always define that macro even when static builds are used, although I think that makes the intent far less clear

[bgilbert]

compiler.has_header_symbol() might be a sufficient substitute for compiler.has_define(), and if you don't care about the get_option('default_library') == 'both' case, you could condition the macros on the value of default_library. But it also seems reasonable to stick with 1.3.0 for now.

[kou]

Issue resolved by pull request 47541
#47541