Build with LTO in release mode by default

[kbieganski]

Enables link time optimizations by default in release mode, yielding 3-12% better performance for the entire flow.

nangate45/ibex run	Mean [s]	Min [s]	Max [s]	Relative
No LTO	256.417 ± 1.089	252.747	258.500	1.12 ± 0.01
With LTO	228.625 ± 1.979	223.222	230.904	1.00

nangate45/black_parrot run	Mean [s]	Min [s]	Max [s]	Relative
No LTO	2961.542 ± 14.714	2937.151	2987.307	1.06 ± 0.01
With LTO	2795.283 ± 26.873	2725.632	2822.192	1.00

nangate45/ariane133 run	Mean [s]	Min [s]	Max [s]	Relative
No LTO	5807.242 ± 17.377	5791.852	5832.0591	1.03 ± 0.02
With LTO	5651.473 ± 71.611	5550.869	5739.271	1.00

[github-actions[bot]]

clang-tidy review says "All clean, LGTM! :+1:"

[QuantamHD]

I've tried enabling this before, but the GUI fails to run under LTO. Have you tried running the GUI?

[QuantamHD]

https://github.com/InBetweenNames/gentooLTO/issues/444

[kbieganski]

Ah, no, I've only been building without the GUI. I'll try it, and if it doesn't work, I'll add another condition that checks if BUILD_GUI=off.

[QuantamHD]

It also might be possible to compile the GUI without LTO and the other parts with, but I never looked into how that might work.

[QuantamHD]

If LTO doesn't work out AutoFDO might be a more stable option https://gcc.gnu.org/wiki/AutoFDO/Tutorial

[maliberty]

Please also give a summary of the impact on link time of using whatever option you settle on.

[kbieganski]

	Full build	Linking only
No LTO	8 min 23 s	5 s
With LTO	9 min 23 s	1 min 30 s

On GCC 13.2.1, i7-9700 (8 cores)

The GUI actually works for me with LTO (Arch Linux, Qt 5.15.12). I'll try with Clang. Also, seems like the CI is failing on Clang.

[maliberty]

The linking only hit is too substantial to be on by default imo. I expect the full build on a bigger machine will also look worse. Let's make this an option. It could be on for the pre-built binary releases (@vvbandeira )

[rovinski]

If it's not possible to build only the GUI without LTO, then maybe CMake can just error if both LTO and the GUI are enabled.

[kbieganski]

If we're not going to enable it by default, there's no point in this PR. Closing it.

For future reference:

• If you want to build with LTO, use -DCMAKE_INTERPROCEDURAL_OPTIMIZATION=ON (or -DLINK_TIME_OPTIMIZATION=ON) when configuring
• The GUI worked for me with LTO (Qt 5.15.12), so this issue may have been resolved. (Also, for CI I recommend -DBUILD_GUI=no)
• Use a new-ish compiler and avoid the BFD linker (the default GNU linker). For me, BFD segfaulted with Clang. Use gold, lld, or mold. They're much faster anyway: LTO link time with Clang+any of these is about twice as fast as GCC+BFD.

[maliberty]

We could make it the default but there needs to be an option to turn it off for developers in an edit/compile/run loop.

[kbieganski]

For sure, you don't want LTO while developing. I'll get back to this then.

[github-actions[bot]]

clang-tidy review says "All clean, LGTM! :+1:"

[github-actions[bot]]

clang-tidy review says "All clean, LGTM! :+1:"

[maliberty]

@kbieganski are you still planning to get back to this?

[kbieganski]

I'm busy with other stuff at the moment, so if anyone is willing to take over, go ahead.