Messages in this thread | | | Date | Thu, 17 Nov 2022 13:55:07 +0000 (UTC) | From | Richard Biener <> | Subject | Re: [PATCH 00/46] gcc-LTO support for the kernel |
| |
On Thu, 17 Nov 2022, Ard Biesheuvel wrote:
> On Thu, 17 Nov 2022 at 12:43, Peter Zijlstra <peterz@infradead.org> wrote: > > > > On Thu, Nov 17, 2022 at 08:50:59AM +0000, Richard Biener wrote: > > > On Thu, 17 Nov 2022, Peter Zijlstra wrote: > > > > > > > On Mon, Nov 14, 2022 at 08:40:50PM +0100, Ard Biesheuvel wrote: > > > > > On Mon, 14 Nov 2022 at 12:44, Jiri Slaby (SUSE) <jirislaby@kernel.org> wrote: > > > > > > > > > > > > Hi, > > > > > > > > > > > > this is the first call for comments (and kbuild complaints) for this > > > > > > support of gcc (full) LTO in the kernel. Most of the patches come from > > > > > > Andi. Me and Martin rebased them to new kernels and fixed the to-use > > > > > > known issues. Also I updated most of the commit logs and reordered the > > > > > > patches to groups of patches with similar intent. > > > > > > > > > > > > The very first patch comes from Alexander and is pending on some x86 > > > > > > queue already (I believe). I am attaching it only for completeness. > > > > > > Without that, the kernel does not boot (LTO reorders a lot). > > > > > > > > > > > > In our measurements, the performance differences are negligible. > > > > > > > > > > > > The kernel is bigger with gcc LTO due to more inlining. > > > > > > > > > > OK, so if I understand this correctly: > > > > > - the performance is the same > > > > > - the resulting image is bigger > > > > > - we need a whole lot of ugly hacks to placate the linker. > > > > > > > > > > Pardon my cynicism, but this cover letter does not mention any > > > > > advantages of LTO, so what is the point of all of this? > > > > > > > > Seconded; I really hate all the ugly required for the GCC-LTO > > > > 'solution'. There not actually being any benefit just makes it a very > > > > simple decision to drop all these patches on the floor. > > > > > > I'd say that instead a prerequesite for the series would be to actually > > > enforce hidden visibility for everything not part of the kernel module > > > API so the compiler can throw away unused functions. Currently it has > > > to keep everything because with a shared object there might be external > > > references to everything exported from individual TUs. > > > > I'm not sure what you're on about; only symbols annotated with > > EXPORT_SYMBOL*() are accessible from modules (aka DSOs) and those will > > have their address taken. You can feely eliminate any unused symbol.
But IIRC that's not reflected on the ELF level by making EXPORT_SYMBOL*() symbols public and the rest hidden - instead all symbols global in the C TUs will become public and the module dynamic loader details are hidden from GCCs view of the kernel image as ELF relocatable object.
> > > There was a size benefit mentioned for module-less monolithic kernels > > > as likely used in embedded setups, not sure if that's enough motivation > > > to properly annotate symbols with visibility - and as far as I understand > > > all these 'required' are actually such fixes. > > > > I'm not seeing how littering __visible is useful or desired, doubly so > > for that static hack, that's just a crude work around for GCC LTO being > > inferior for not being able to read inline asm. > > We have an __ADDRESSABLE() macro and asmlinkage modifier to annotate > symbols that may appear to the compiler as though they are never > referenced. > > Would it be possible to repurpose those so that the LTO code knows > which symbols it must not remove?
I find
/* * Force the compiler to emit 'sym' as a symbol, so that we can reference * it from inline assembler. Necessary in case 'sym' could be inlined * otherwise, or eliminated entirely due to lack of references that are * visible to the compiler. */ #define ___ADDRESSABLE(sym, __attrs) \ static void * __used __attrs \ __UNIQUE_ID(__PASTE(__addressable_,sym)) = (void *)&sym; #define __ADDRESSABLE(sym) \ ___ADDRESSABLE(sym, __section(".discard.addressable"))
that should be enough to force LTO keeping 'sym' - unless there's a linker script that discards .discard.addressable which I fear LTO will notice, losing the effect. A more direct way would be to attach __used to 'sym' directly. __ADDRESSABLE doesn't seem to be used directly but instead I see cases like
#define __define_initcall_stub(__stub, fn) \ int __init __stub(void); \ int __init __stub(void) \ { \ return fn(); \ } \ __ADDRESSABLE(__stub)
where one could have added __used to the __stub prototypes instead?
The folks who worked on LTO enablement of the kernel should know the real issue better - I understand asm()s are a pain because GCC refuses to parse the assembler string heuristically for used symbols (but it can never be more than heuristics). The issue with asm()s is not so much elimination (__used solves that) but that GCC can end up moving the asm() and the refered to symbols to different link-time units causing unresolved symbols for non-global symbols. -fno-toplevel-reorder should fix that at some cost.
Richard.
-- Richard Biener <rguenther@suse.de> SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg, Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman; HRB 36809 (AG Nuernberg)
| |