Re: [PATCH v2 02/41] Makefile: Allow LTO to be disabled for a build

From: Tom Rini <trini@konsulko.com>
To: Simon Glass <sjg@chromium.org>
Cc: "Heinrich Schuchardt" <heinrich.schuchardt@canonical.com>,
	"U-Boot Mailing List" <u-boot@lists.denx.de>,
	"Michal Simek" <michal.simek@xilinx.com>,
	"Daniel Schwierzeck" <daniel.schwierzeck@gmail.com>,
	"Steffen Jaeckel" <jaeckel-floss@eyet-services.de>,
	"Marek Behún" <marek.behun@nic.cz>,
	"Lukas Auer" <lukas.auer@aisec.fraunhofer.de>,
	"Dennis Gilmore" <dennis@ausil.us>,
	"Masahiro Yamada" <yamada.masahiro@socionext.com>,
	"Ilias Apalodimas" <ilias.apalodimas@linaro.org>,
	"Heinrich Schuchardt" <xypron.glpk@gmx.de>
Subject: Re: [PATCH v2 02/41] Makefile: Allow LTO to be disabled for a build
Date: Thu, 4 Nov 2021 10:55:25 -0400	[thread overview]
Message-ID: <20211104145525.GM24579@bill-the-cat> (raw)
In-Reply-To: <CAPnjgZ2UJgKP38_1p1ZCBsSMaRuWvxcHrFGd2OdF2uVv57dQJQ@mail.gmail.com>

[-- Attachment #1: Type: text/plain, Size: 4960 bytes --]

On Wed, Nov 03, 2021 at 08:49:01PM -0600, Simon Glass wrote:
> Hi Tom,
> 
> On Mon, 1 Nov 2021 at 15:58, Tom Rini <trini@konsulko.com> wrote:
> >
> > On Sun, Oct 31, 2021 at 05:46:43PM -0600, Simon Glass wrote:
> > > Hi Tom,
> > >
> > > On Wed, 27 Oct 2021 at 07:13, Tom Rini <trini@konsulko.com> wrote:
> > > >
> > > > On Wed, Oct 27, 2021 at 02:21:17PM +0200, Heinrich Schuchardt wrote:
> > > > >
> > > > >
> > > > > On 10/27/21 10:50, Ilias Apalodimas wrote:
> > > > > > Hi Simon
> > > > > >
> > > > > > How does this patch related to the standard boot series? Shouldn't
> > > > > > this be a completely separate patch?
> > > > > >
> > > > > > Thanks
> > > > > > /Ilias
> > > > > >
> > > > > > On Sun, 24 Oct 2021 at 02:26, Simon Glass <sjg@chromium.org> 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@chromium.org>
> > > > > > > ---
> > > > > > >
> > > > > > > (no changes since v1)
> > > > > > >
> > > > > > >   Makefile                           | 18 +++++++++++++-----
> > > > > > >   arch/arm/config.mk                 |  4 ++--
> > > > > > >   arch/arm/include/asm/global_data.h |  2 +-
> > > > > > >   doc/build/gcc.rst                  | 17 +++++++++++++++++
> > > > > > >   4 files changed, 33 insertions(+), 8 deletions(-)
> > > > > > >
> > > > > > > diff --git a/Makefile b/Makefile
> > > > > > > index b79b2319ff6..7057723e046 100644
> > > > > > > --- a/Makefile
> > > > > > > +++ b/Makefile
> > > > > > > @@ -434,6 +434,9 @@ KBUILD_CFLAGS       += -fshort-wchar -fno-strict-aliasing
> > > > > > >   KBUILD_AFLAGS   := -D__ASSEMBLY__
> > > > > > >   KBUILD_LDFLAGS  :=
> > > > > > >
> > > > > > > +# Set this to "n" use of LTO for this build, e.g. LTO_BUILD=n
> > > > > > > +LTO_BUILD      ?= y
> > > > >
> > > > > This does not allow LTO_BUILD=y to enable LTO for CONFIG_LTO=n.
> > > >
> > > > I don't understand why we need this patch at all.  If you want to
> > > > disable LTO, disable LTO.  Yes, LTO makes linking take longer which can
> > > > be annoying on iterative development.  I have a few different "HACK: DO
> > > > NOT PUSH:  ..." things I git am at the start of a branch, depending on
> > > > needs.  You can just do that to drop "imply LTO" from the SANDBOX stanza
> > > > in arch/Kconfig.  We do not need a whole thing around a CONFIG option
> > > > that can be disabled in the defconfig, or local .config file even.
> > > >
> > >
> > > Cranky time.
> > >
> > > Of course we don't *need* it. I could just buy a slower build machine
> > > and type with two fingers. There are lots of ways to slow things down
> > > and LTO is one of them. I change branches at least a dozen times a day
> > > and am always trying things out from patchwork. I am sure others do
> > > too. LTO dramatically slows down builds. Having a way to easily do
> > > this from the build system saves time.
> >
> > Maybe the answer is that LTO just isn't appropriate for sandbox.  We're
> > not doing any specific tests for LTO anywhere (nor does that seem
> > appropriate), and we do have platforms in CI that run tests other than
> > building, with LTO.
> 
> It has value as a test, I presume, and a demo of how it works. Also it
> runs most of the tests.
> 
> But I'm happy to disable it if that helps.
> 
> Still, it doesn't really solve the issue. The same thing happens when
> building real boards.

Well, a big part of the problem here is I strongly disagree with a
make-line flag to override a CONFIG option.  I also hear your use case
of "I build this platform so frequently per development session it's a
noticeable slowdown and a 'LOCAL:' commit will also screw up my
workflow".  So where do we go?  LTO is a size versus speed trade-off
(and -ffunction-sections/-fdata-sections/--gc-unused is a much smaller
speed trade-off) that's more important on real hardware (and also not
used as often as it might be, at least so far).  On real boards, there's
less of a "just turn it off" option because it's required to be small
enough to use on the hardware.  Since that's not the case on sandbox,
maybe we just need to turn it on in more QEMU platforms (to get broader
test coverage in CI) and off in sandbox.

-- 
Tom

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 659 bytes --]