Why is '--no-gc-sections' added in build with MAIN_MODULE/SIDE_MODULE=1?

[alexander-veselov]

emscripten/tools/building.py

Line 298 in a62c811

cmd.append('--no-gc-sections')

--no-gc-sections
Disable garbage collection of unused sections.

'--gc-sections' greatly decreases function count (which helps us not to run into the limit of functions). But we can't use it because of '--no-gc-sections'. Why was it added? Is there a way to bypass this and will it be correct?

[sbc100]

Because with -sMAIN_MODULE=1 we explicitly want to include everything. If you want to DCE your main module you need to build with MAIN_MODULE=2. That is the main difference between =1 and =2. One does DCE and the other conservatively keeps everything alive in case it is needed at runtime.

[Maksim-Bozhko]

Yes, I think so, let me double check function counts and compare. I will tell you for sure soon.

[Maksim-Bozhko]

So I counted functions in side module and in main module, for 3 cases:
SIDE_MODULE=1 && MAIN_MODULE=1
main : [funcs]: 25049
side: [funcs]: 3157

SIDE_MODULE=2 && MAIN_MODULE=2
main : [funcs] : 10760
side: [funcs] : 1774

SIDE_MODULE=1 && MAIN_MODULE=1 with cmd.append('--gc-sections')
main : [funcs] : 21451
side: [funcs] : 2183

So =2 seems to be more agressive still, but it is close

[Maksim-Bozhko]

@sbc100 so to answer your initial question: Yes we would need an option to turn on --gc-sections. If we remove --no-gc-sections and add --gc-sections isntead, everything seems to work with SIDE_MODULE/MAIN_MODULE=1 and we didn't see cases of removing anything that we actually needed, I assume because it respects exports and assumes all exports are needed. It also reduces size of wasm_apply_data_reloc function dramatically which allows us to be under function size limit(in combination with -O1)

[sbc100]

OK sounds like we can maybe just remove --no-gc-sections completely! I'll give it a try. Thanks for investigating.

[sbc100]

This was fixed in #17056