Linking a staticlib into shared C library or executable with gcc causes considerbly larger binary than linking into a cdylib or executable with rustc

[sdroege]

See https://github.com/sdroege/rust-staticlib-bloat-test for the code. This doesn't use cargo to make the actual compiler invocations clear (the same behaviour can be observed with cargo).

With Rust 1.69 and gcc 13.1.1 on Linux x86-64 the following can be observed:

	unstripped	stripped (--strip-all)
Rust executable (my_test)	4.1MB	315kB
Rust cdylib (libstaticlib.so)	4.0MB	290kB
gcc executable (my_staticlib_test)	5.2MB (+ 27%)	1.1MB (+ 355%)
gcc shared library (libmy_staticlib_test_dylib.so)	5.6MB (+ 40%)	1.5MB (+ 517%)

The actual size comes from the enabled overflow checks (-C debug-assertions=on) , and thus all the panic and string formatting machinery that is included. However, this should be approximately the same code in either case so that doesn't explain the difference. (Disabling the overflow checks gives a more expected result because the function from the library compiles down to a few instructions) Also note that this compiles with -C panic=abort so having the assertion in an extern "C" function is not UB.

The question here now is what exactly the difference between both cases is, and how the staticlib case can be improved. Maybe this requires changes in the compiler, maybe this just requires invoking the linker differently (in which case this would be useful to add to the documentation).

[bjorn3]

For cdylib we use a version script to limit the exported symbols in combination with --gc-sections to remove unused functions. Did you pass --gc-sections when linking yourself?

[sdroege]

That makes a lot of sense, sorry for missing that. Passing -Wl,--gc-sections solves it for the executable. I guess for the shared library a version script to not re-export random Rust symbols would also fix it (will test later and update this comment it does).

Do you think this would be useful to add to e.g. https://doc.rust-lang.org/reference/linkage.html ?

[bjorn3]

I guess for the shared library a version script to not re-export random Rust symbols would also fix it (will test later and update this comment).

You need both. Without the version script the linker needs to export all symbols, but without --gc-sections the linker won't omit unused functions even if the version script says they aren't exported.

Do you think this would be useful to add to e.g. https://doc.rust-lang.org/reference/linkage.html ?

Yeah, could be useful to add a recommendation to pass -Wl,--gc-sections (and -dead_strip on macOS) there.

[sdroege]

You need both. Without the version script the linker needs to export all symbols, but without --gc-sections the linker won't omit unused functions even if the version script says they aren't exported.

Yeah, with that it's also the same in both cases. Thanks!

I'll prepare a PR for adding a sentence about that to the docs some time later :)