any info for latest kernel LTO?

[honeymak]

i would like to know any steps for gentoo-sources kernel to do LTO or -flto=auto in any place

btw, is it stable or latest tech for kernel LTO?

thx

[Goverp]

IIRC it's possible to build an LTO kernel using gcc, but the normal path is to use clang/LLVM.
You have to select LTO in the kernel configuration, usually

Code: Select all

General architecture-dependent options
  Link Time Optimization (LTO)
    Clang ThinLTO (experimental)

AND you must pass some environment options to the kernel make:

Code: Select all

CC='clang'
LLVM='1'

I doubt it's a good idea to try Clan Full LTO, but I've not tried it.

I've been compiling my kernels this way for ages.

[honeymak]

i am a gcc guy

i am actually editing kernel makefiles for my cflags
while -flto=auto does not work that way

that's why i am trying to ask here if any extra info
clang/llvm way sorta not my intention

[krumpf]

To compile kernel with GCC and LTO, read here : https://wiki.gentoo.org/wiki/Kernel/Opt ... timization
The gentoo-sources package itself doesn't care about flto, it's just decompressing the archive containing source files with some gentoo specific patches. As Goverp said, it's easier to LTOize the kernel when building it with Clang.

[honeymak]

hmmmm.........clang lto is not my way........clang is kinda.....not very mature to me......and has to change so many stuff

btw......LTO stuff is kinda latest tech, right?
i guess i will just wait for some time......

i do edit/change kernel Makefiles for my current 'cflags' stuff
but it's small scale and even just several lines of flags

looks like the gcc pgo kernel info page is kinda lengthy

it's kinda hard for me to trace/track all the changes for every kernel i use....
so....just wait for some more centralise easy changes if any....later.....

i will settle here for now.....hmm...

[sam_]

it's a work-in-progress.

GCC 16 (not yet released) will have some work towards making top-level asm work via heuristics, avoiding the need to patch many places in the kernel with explicit annotations (which the kernel developers previously rejected).

Some patches are still needed for the kernel: https://gitlab.com/mixal_iirec/linux_gcc_lto_patches

Even with these patches (some will be dropped in the end), module support is not yet there. i don't recommend trying it yet, but there is progress on this front.

[Juippisi]

btw......LTO stuff is kinda latest tech, right?

No we're in https://wiki.gentoo.org/wiki/Kernel/Opt ... ng_AutoFDO and https://wiki.gentoo.org/wiki/Kernel/Opt ... _Propeller era already. LTO is old tech. Unfortunately these are all clang-related tech.

i guess i will just wait for some time......
looks like the gcc pgo kernel info page is kinda lengthy
so....just wait for some more centralise easy changes if any....later.....

I believe you missed the train for "easy gcc lto'ing". I think cachyos used to maintain a patch to enable lto+gcc for kernels, but they've since moved to compile kernels with clang as well. To my knowledge there's no longer an easy way to get gcc+lto work with kernels.

[Zucca]

Another point of view:

• One should also weigh the pros and cons of LTOing kernel.
  
  I'm certainly not 100% sure, but normally kernel tries to "stay away" as much as possible so that user space programs could run as fast as possible.
  So unless you've been "bottlenecked" by say network load, filesystem load or something else that's kernel's job... you may only benefit a minuscule amount of LTOing the kernel.
  
  So if you're experiencing lag/under-performance on some area, do performance checks with and without LTO.
  
  When it comes to kernel... I wouldn't do anything too experimental. Kernel silently misbehaving can cause all kinds of issues. So if you're going for it, I'd highly suggest using clang.