Improve zstd-sys/build.rs: Fallback to -Os or -O2 if -Oz is not supported

[NobodyXu]

Signed-off-by: Jiahao XU [email protected]

[NobodyXu]

@gyscos Can you review this PR please?

[NobodyXu]

@gyscos Pinging

[gyscos]

Unfortunately I'm not sure is_flag_supported does its job: it seems that for some versions of gcc, it will accept a flag, even though this flag leads to a non-working compilation.

I'm still trying to figure out how to safely enable any form of lto on GCC that will work on both recent gcc version, older gcc version used in cross (gcc 9), and the gcc used in the windows-gnu target...

[NobodyXu]

Unfortunately I'm not sure is_flag_supported does its job: it seems that for some versions of gcc, it will accept a flag, even though this flag leads to a non-working compilation.

I'm still trying to figure out how to safely enable any form of lto on GCC that will work on both recent gcc version, older gcc version used in cross (gcc 9), and the gcc used in the windows-gnu target...

Perhaps we should just compile a simple hello-world with lto to see if that works?
If it can be compiled, then it's likely it would work.

I can try implementing that in a new PR and meanwhile I will reduce the scope of this PR

[gyscos]

I think a simpler option to start would be letting the end user override this "guesswork" and explicitly enable LTO or other options.

Some magic to detect supported features would also be valuable, but it's not clear if it should go here or in cc itself.

[gyscos]

Actually, do we need to turn on O2 manually? What's the advantage vs letting cc pick the optimization level?

I realize it's only when none of Oz or Os works (not sure when that happens), but that still looks like a different decision.

[NobodyXu]

Actually, do we need to turn on O2 manually? What's the advantage vs letting cc pick the optimization level?

It's only turned on if feature "thin" is enabled.
Using -O2 probably generates smaller code than -O3 (or even -O1 sometimes).

[NobodyXu]

I realize it's only when none of Oz or Os works (not sure when that happens),

I remember that gcc doesn't support -Oz and only -Os?

[NobodyXu]

Unfortunately I'm not sure is_flag_supported does its job: it seems that for some versions of gcc, it will accept a flag, even though this flag leads to a non-working compilation.

I'm still trying to figure out how to safely enable any form of lto on GCC that will work on both recent gcc version, older gcc version used in cross (gcc 9), and the gcc used in the windows-gnu target...

I checked Build::is_flag_supported and it actually attempts to compile the file with the flag provided.

I think the problem lies in this line, where it does not check the exit status of the command and assumes that if the stderr is empty, everything works as expected.

[NobodyXu]

@gyscos I've opend rust-lang/cc-rs#757 which might help fix the LTO issue in zstd.