On Thu, Nov 14, 2019 at 05:28:06PM +0100, Jan Kiszka wrote:
On 26.08.19 08:43, Fabrice Fontaine wrote:
Le lun. 26 août 2019 à 07:57, Jan Kiszka jan.kiszka@web.de a écrit :
On 25.08.19 21:11, Fabrice Fontaine wrote:
Le dim. 25 août 2019 à 17:49, Jan Kiszka jan.kiszka@web.de a écrit :
On 25.08.19 17:13, Fabrice Fontaine wrote:
Le dim. 25 août 2019 à 16:11, Jan Kiszka jan.kiszka@web.de a écrit : > > On 25.08.19 15:43, Fabrice Fontaine wrote: > > Hello, > > Le dim. 25 août 2019 à 13:44, Jan Kiszka jan.kiszka@web.de a écrit : > > > > > > On 01.05.19 15:08, Fabrice Fontaine wrote: > > > > When CROSS_BUILD_TOOLS is set, set HOSTCFLAGS to CFLAGS otherwise CC > > > > will be used with HOSTCFLAGS which seems wrong > > > > > > > > Signed-off-by: Fabrice Fontaine fontaine.fabrice@gmail.com > > > > --- > > > > tools/Makefile | 1 + > > > > 1 file changed, 1 insertion(+) > > > > > > > > diff --git a/tools/Makefile b/tools/Makefile > > > > index 12a3027e23..eadeba417d 100644 > > > > --- a/tools/Makefile > > > > +++ b/tools/Makefile > > > > @@ -272,6 +272,7 @@ subdir- += env > > > > > > > > ifneq ($(CROSS_BUILD_TOOLS),) > > > > override HOSTCC = $(CC) > > > > +override HOSTCFLAGS = $(CFLAGS) > > > > > > > > quiet_cmd_crosstools_strip = STRIP $^ > > > > cmd_crosstools_strip = $(STRIP) $^; touch $@ > > > > > > > > > > This eats - among other things - -O2, normally set in /Makefile. And that breaks > > > CROSS_BUILD_TOOLS=y with CONFIG_FIT but without CONFIG_FIT_SIGNATURE because "if > > > (!IMAGE_ENABLE_SIGN)" is no longer optimized. I tend to believe this should > > > simply be reverted (which is what I'm doing locally in order to fix my builds). > > I don't think this patch should be reverted, I sent it to fix a > > cross-compilation build issue with host-openssl. > > > > Indeed, without this patch, with CROSS_BUILD_TOOLS=y, tools/Makefile > > sets HOSTCC=$(CC) but it forgets > > to override HOSTCFLAGS, though, so the tools are built with target > > compiler but host CFLAGS. > > > > This raises a build failure if host-openssl is built before uboot-tools because > > uboot-tools links with openssl headers from host which depends on pthread.h > > > > More information can be found on buildroot's patchwork here: > > - https://patchwork.ozlabs.org/patch/1092227 > > - https://patchwork.ozlabs.org/patch/1093385 > > > > > > Jan > > > > Best Regards, > > > > Fabrice > > > > So what is your suggestion to fix the O2-regression best? I'm far from being an expert in uboot so please excuse me if my answer contains some mistakes. However, IMHO, we should not pass -O2 or any other optimization flags. From my understanding, CFLAGS (and so HOSTCFLAGS when building tools for the target through cross-tools) will (or should?) contain -O2 or -Os depending on CONFIG_CC_OPTIMIZE_FOR_SIZE value. However, I do understand that this is a change in the former behavior that was always passing -O2 in HOSTCFLAGS. So, if you really want to pass -O2, I would add it to HOSTCFLAGS in tools/Makefile (after the override).
This is a hack. U-boot relies on >= -O2 for proper build, so it must inject this setting on its own - which it did prior to your change. I think top-level HOSTCFLAGS belongs to the build, irrespective of native vs. cross. You may allow appending (and, thus, overwriting) settings but you cannot remove mandatory ones.
However, I have to confess that I don't understand why the build breaks "if (!IMAGE_ENABLE_SIGN)" is no longer optimized."
IMAGE_ENABLE_SIGN is 0 when CONFIG_FIT_SIGNATURE is not set. That will remove all code from tools/image-host.c which is not reachable then. When it's not removed because of default optimization settings, we get linker errors because common/image.sig.c is not built without CONFIG_FIT_SIGNATURE. I assume this way was preferred over #ifdef'ery, but it's broken now.
Thanks a lot for your explanation. I was able to reproduce the issue and I think that https://github.com/ffontaine/u-boot/commit/61a162e5c0343fc55304902042f8c9adf... should fixed it. Can you review it? I can also directly send a PR.
I still think that dropping the content of HOSTCFLAGS is the actual bug. There is more than -O2 that you lose. If you want your CFLAGS to be respected, append them.
I add Arnout to the list to get his feedback as he is the one that suggested this solution during review of https://patchwork.ozlabs.org/patch/1092227.
Nothing happened, bug is still present in master. Any suggestions how to resolve it?
OK, I've looked over the thread again. Where and how are you hitting a problem here again? This area is indeed a bit funky and I see OpenEmbedded has long been doing "shove everything we need in via CC, do not use CFLAGS/etc". Thanks!