Re: [PATCH] arm: add an option for erratum 657417

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday, August 24, 2016 8:56:46 PM CEST Nicholas Piggin wrote:
> On Wed, 24 Aug 2016 19:05:30 +1000
> Nicholas Piggin <npiggin@gmail.com> wrote:
> > Thanks, I'll take a look.
> 
> Okay, I can't reproduce your bad linking times even with gc-sections. It's
> possible I'm doing something wrong, but with my patches + your patch and
> standard arm allyesconfig:
> 
>   20:33:56  AR      built-in.o
>   20:33:57  LD      vmlinux.o
>   MODPOST vmlinux.o
>   20:34:12  GEN     .version
>   CHK     include/generated/compile.h
>   UPD     include/generated/compile.h
>   CC      init/version.o
>   AR      init/built-in.o
>   20:34:24  KSYM    .tmp_kallsyms1.o
>   20:34:45  KSYM    .tmp_kallsyms2.o
>   20:34:54  LD      vmlinux
>   20:35:07  SORTEX  vmlinux
>   20:35:07  SYSMAP  System.map
> 
> I have about 71 seconds for the final link phase.


I've tracked down my remaining build time regression to
a bad binutils snapshot (2.26.51) I had been using, and upgraded
to the 2.27 release now, which is roughly the same as what
you have:

14:45:41   LINK    vmlinux
14:45:41  AR      built-in.o
14:45:42  LD      vmlinux.o
14:51:49   MODPOST vmlinux.o
14:51:51  GEN     .version
14:51:51   CHK     include/generated/compile.h
  UPD     include/generated/compile.h
14:51:51   CC      init/version.o
14:51:52   AR      init/built-in.o
14:52:04  KSYM    .tmp_kallsyms1.o
14:52:31  KSYM    .tmp_kallsyms2.o
14:52:43  LD      vmlinux
14:52:55  SORTEX  vmlinux
14:52:55  SYSMAP  System.map
14:52:56   OBJCOPY arch/arm/boot/Image

The long minutes that were spent in "arm-linux-gnueabi-ld 
-r -o vmlinux.o --whole-archive built-in.o" are all gone now.

I still see a problem with big-endian builds failing with
thinarc/gc-sections, I'll investigate that some other day,
or you could have a look at that if you want to make sure
it's an ARM specific problem, not something with your
patches in general.

The patch that I sent for enabling the two on ARM blocks
out CONFIG_CPU_BIG_ENDIAN, so just revert that hunk to see
the problem. It's possible that it only breaks when doing
a big-endian build after a little-endian build without
a "make clean" inbetween.

	Arnd