perf: Enable link-time-optimization for performance critical libraries

[mrousavy]

Summary:

Enables Link Time Optimization for performance critical libraries to optimize the native code across modules.

This might increase release build-time a tiny bit (not yet measured), but should improve runtime performance or startup performance a bit (60ms in the Discord app, not yet measured in other apps)

Comparisons

Native Libraries Size

• Without -flto (current main): 19,4 MB for rn-tester apk (assembleRelease, arm64-v8a)
• With -flto (current PR): 19,4 MB for rn-tester apk (assembleRelease, arm64-v8a)

..so no difference.

Release apk build time

• Without -flto (current main): 8:40.24
• With -flto (current PR): 10:19.57

...so 1min 20s increase in build time 😢

Release apk performance

Without -flto (current main)	With -flto (current PR)

I'm not sure how accurate the flashlight measurements are, but according to the average of 10 tests in this CI, enabling -flto makes the app start ~70ms (or ~10%) faster... 👀

Release apk performance

• Without -flto (current main): ...
• With -flto (current PR): ...

Changelog:

[GENERAL] [ADDED] - Enabled link-time-optimization to improve performance of some parts of the C++ codebase

Test Plan:

1. Build once with, and once without -flto (release builds!)
2. Measure build time difference between clean builds
3. Measure startup time difference
4. Measure perceived performance when stress-testing something (maybe the Turbo Modules vs Nitro Modules benchmark got a bit tighter with this)

[NickGerleman]

I think @cortinico had tried to enable this before, globally in the build, by setting INTERPROCEDURAL_OPTIMIZATION to true?

Are there places where this is not getting set? Or is this doing something like setting flto=thin maybe? Some of these should definitely already have some form of LTO enabled.

[cortinico]

So in theory this should not be needed, because the Android NDK toolchain already should be setting those flags for us.

Specifically the release versions of React Native are already built with:

LANGUAGE_COMPILE_FLAGS = -g -DANDROID -fdata-sections -ffunction-sections -funwind-tables -fstack-protector-strong -no-canonical-prefixes -D__BIONIC_NO_PAGE_SIZE_MACRO -D_FORTIFY_SOURCE=2 -Wformat -Werror=format-security   -O3 -DNDEBUG  -flto=thin

[cortinico]

On top of this, this would set the compilation flag for both Debug and Release, which is something we don't really want

[cortinico]

I think @cortinico had tried to enable this before, globally in the build, by setting INTERPROCEDURAL_OPTIMIZATION to true?

Are there places where this is not getting set? Or is this doing something like setting flto=thin maybe? Some of these should definitely already have some form of LTO enabled.

Yeah see my other comment. We already set -flto=thin on all the targets

[NickGerleman]

@mrousavy how reproducible are the wins from using -flto instead of -flto=thin?

If there is a good win there, I could imagine it making sense to enable for anything prebuilt, since end-users won't often hit the build time impact. But we would want to reconcile the different ways this is enabled.

[mrousavy]

let me get back to you on this after some more testing. :)

[cortinico]

let me get back to you on this after some more testing. :)

Eager to hear your findings @mrousavy :)