Re: Unused header file inclusion

Hi,

On 2019-07-31 11:23:22 -0400, Alvaro Herrera wrote:
> I think removing unnecessary include lines from header files is much
> more useful than from .c files.  However, nowadays even I am not very
> convinced that that is a very fruitful use of time, since many/most
> developers use ccache which will reduce the compile times anyway in many
> cases; and development machines are typically much faster than ten years
> ago.

IDK, I find the compilation times annoying. And it's gotten quite
noticably worse with all the speculative execution mitigations. Although
to some degree that's not really the fault of individual compilations,
but our buildsystem being pretty slow.

I think there's also just modularity reasons for removing includes from
headers. We've some pretty oddly interlinked systems, often without good
reason (*). Cleaning those up imo is well worth the time - but hardly
can be automated.

If one really wanted to automate removal of header files, it'd need to
be a lot smarter than just checking whether a file fails to compile if
one header is removed. In the general case we'd have to test if the .c
file itself uses any of the symbols from the possibly-to-be-removed
header. That's hard to do without using something like llvm's
libclang. The one case that's perhaps a bit easier to automate, and
possibly worthwhile: If a header is not indirectly included (possibly
testable via #ifndef HEADER_NAME_H #error 'already included' etc), and
compilation doesn't fail with it removed, *then* it's actually likely
useless (except for portability cases).

* I think a lot of the interlinking stems from the bad idea to use
typedef's everywhere. In contrast to structs they cannot be forward
declared portably in our version of C. We should use a lot more struct
forward declarations, and just not use the typedef.

Greetings,

Andres Freund