[PATCH v2 02/13] Makefile: Allow LTO to be disabled for a build

[Simon Glass]

Hi Tom,

On Mon, 14 Mar 2022 at 13:45, Tom Rini <trini at konsulko.com> wrote:
>
> On Mon, Mar 14, 2022 at 01:21:02PM -0600, Simon Glass wrote:
> > Hi Tom,
> >
> > On Mon, 14 Mar 2022 at 12:29, Tom Rini <trini at konsulko.com> wrote:
> > >
> > > On Mon, Mar 14, 2022 at 12:24:42PM -0600, Simon Glass wrote:
> > > > Hi Tom,
> > > >
> > > > On Mon, 14 Mar 2022 at 06:49, Tom Rini <trini at konsulko.com> wrote:
> > > > >
> > > > > On Sat, Mar 12, 2022 at 10:58:44AM -0700, Simon Glass wrote:
> > > > > > Hi Tom,
> > > > > >
> > > > > > On Mon, 7 Mar 2022 at 07:33, Tom Rini <trini at konsulko.com> wrote:
> > > > > > >
> > > > > > > On Fri, Mar 04, 2022 at 08:42:57AM -0700, Simon Glass wrote:
> > > > > > >
> > > > > > > > LTO (Link-Time Optimisation) is an very useful feature which can
> > > > > > > > significantly reduce the size of U-Boot binaries. So far it has been
> > > > > > > > made available for selected ARM boards and sandbox.
> > > > > > > >
> > > > > > > > However, incremental builds are much slower when LTO is used. For example,
> > > > > > > > an incremental build of sandbox takes 2.1 seconds on my machine, but 6.7
> > > > > > > > seconds with LTO enabled.
> > > > > > > >
> > > > > > > > Add a LTO_BUILD=n parameter to the build, so it can be disabled during
> > > > > > > > development if needed, for faster builds.
> > > > > > > >
> > > > > > > > Add some documentation about LTO while we are here.
> > > > > > > >
> > > > > > > > Signed-off-by: Simon Glass <sjg at chromium.org>
> > > > > > >
> > > > > > > We don't need this since you can do:
> > > > > > > make EXTRA_CFLAGS="-fno-lto" EXTRA_LDFLAGS="-fno-lto"
> > > > > > > to pass -fno-lto to compile/linking and disable lto and per
> > > > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=46905 this has been working
> > > > > > > for some time.
> > > > > >
> > > > > > Thanks for that, it is a big pain point for me, picking up this patch
> > > > > > for every series I write. The incremental build time for sandbox goes
> > > > > > from 3 seconds to 27 seconds on my laptop with LTO, which is
> > > > > > intolerable.
> > > > >
> > > > > Yeah, I noticed it's visible on my laptop, but not at all on my desktop
> > > > > (i7 vs Ryzen 7).
> > > > >
> > > > > > The EXTRA_CFLAGS says it is for 'Backward compatibility' and it still
> > > > > > does the various LTO things (i.e. it changes the build logic). It
> > > > >
> > > > > We're unlikely to move to newer Linux kernel kbuild logic so this isn't
> > > > > going away, and there's not much in the way of logic that's changed for
> > > > > LTO that I see.
> > > > >
> > > > > > seems odd to me to enable the option and then disable it later in the
> > > > > > command line. It is therefore not quite equivalent. But it seems to
> > > > > > work well enough for me fom a small amount of testing. If you are
> > > > > > really set on not having a special option for it, I can live with it
> > > > > > for now. I'm also not convinced that my patch entirely removes the LTO
> > > > > > stuff in a consistent way.
> > > > >
> > > > > Yeah, I really don't want to go down the path of overriding CONFIG
> > > > > options via make/environment logic.  I'm also open to turning off LTO on
> > > > > sandbox and on with qemu-* so it gets wider CI testing.
> > > >
> > > > Yes you did mention that, but the problem is that LTO is very handy
> > > > with sandbox, to test the strange things that happen. For example, I
> > > > found the bug where LTO was dropping a linker-list item, using
> > > > sandbox. We could perhaps make one of the sandbox builds not use LTO,
> > > > e.g. sandbox_flattree ?
> > >
> > > Well, the big issue with LTO+sandbox is that it slows down your
> > > workflow, so I would think you want the inverse, one platform does
> > > enable it?
> >
> > It slows down all boards that use it, actually, so I normally don't
> > want it enabled for my IDE. This is not sandbox-specific.
>
> The degree of slowdown depends on what you're building on.  I see
> several seconds of link time on my laptop and nothing on my desktop
> (which isn't that high end either, it's not an EPYC or anything) when
> LTO is/isn't enabled.

For me, on 16-core AMD desktop, incremental build

LTO: real 0m7.744s
No LTO: real 0m3.172s

On 4-core laptop:

No LTO: real 0m2.429s
LTO: real 0m15.650s

Perhaps the disconnect here is that you just don't see any difference?
What times are you seeing?

Regards,
SImon