Re: [PATCH] powerpc: e500: Fix compilation with gcc e500 compiler

["Pali Rohár" <pali@kernel.org>]

On Saturday 09 July 2022 09:16:13 Christophe Leroy wrote:
> Le 08/07/2022 à 19:14, Pali Rohár a écrit :
> > On Monday 04 July 2022 15:13:58 Pali Rohár wrote:
> >> On Monday 04 July 2022 14:07:10 Arnd Bergmann wrote:
> >>> On Mon, Jul 4, 2022 at 12:39 PM Pali Rohár <pali@kernel.org> wrote:
> >>>> On Monday 04 July 2022 20:23:29 Michael Ellerman wrote:
> >>>>> On 2 July 2022 7:44:05 pm AEST, "Pali Rohár" <pali@kernel.org> wrote:
> >>>>>> On Tuesday 24 May 2022 11:39:39 Pali Rohár wrote:
> >>>>>>> gcc e500 compiler does not support -mcpu=powerpc option. When it is
> >>>>>>> specified then gcc throws compile error:
> >>>>>>>
> >>>>>>>    gcc: error: unrecognized argument in option ‘-mcpu=powerpc’
> >>>>>>>    gcc: note: valid arguments to ‘-mcpu=’ are: 8540 8548 native
> >>>>>>>
> >>>>>>> So do not set -mcpu=powerpc option when CONFIG_E500 is set. Correct option
> >>>>>>> -mcpu=8540 for CONFIG_E500 is set few lines below in that Makefile.
> >>>>>>>
> >>>>>>> Signed-off-by: Pali Rohár <pali@kernel.org>
> >>>>>>> Cc: stable@vger.kernel.org
> >>>>>>
> >>>>>> Michael, do you have any objections about this patch?
> >>>>>
> >>>>> I don't particularly like it :)
> >>>>>
> >>>>>  From the discussion with Segher, it sounds like this is a problem with a specific build of gcc that you're using, not a general problem with gcc built with e500 support.
> >>>>
> >>>> Well, the "full" build of gcc for e500 cores with SPE does not support
> >>>> -mcpu=powerpc option. So I think this is a general problem. I do not
> >>>> think that this is "specific build" as this is the correct build of gcc
> >>>> for these processors with e500 cores.
> >>>>
> >>>> "stripped". build of gcc without SPE support for e500 cores does not
> >>>> have this problem...
> >>>
> >>> I can see a couple of problems with the CPU selection, but I don't think
> >>> this is a major one, as nobody should be using those SPE compilers for
> >>> building the kernel. Just use a modern powerpc-gcc build.
> >>
> >> The point is to use same compiler for building kernel as for the all
> >> other parts of the system.
> >>
> >> I just do not see reason why for kernel it is needed to build completely
> >> different toolchain and compiler.
> >>
> >>>>> Keying it off CONFIG_E500 means it will fix your problem, but not anyone else who has a different non-e500 compiler that also doesn't support -mcpu=powerpc (for whatever reason).
> >>>>>
> >>>>> So I wonder if a better fix is to use cc-option when setting -mcpu=powerpc.
> >>>>>
> >>>>
> >>>> Comment for that code which adds -mpcu=powerpc says:
> >>>>
> >>>>    they are needed to set a sane 32-bit cpu target for the 64-bit cross
> >>>>    compiler which may default to the wrong ISA.
> >>>>
> >>>> So I'm not sure how to handle this in other way. GCC uses -mpcu=8540
> >>>> option for specifying to compile code for e500 cores and seems that
> >>>> -mcpu=8540 is supported by all e500 compilers...
> >>>>
> >>>> Few lines below is code
> >>>>
> >>>>    CFLAGS-$(CONFIG_E500) += $(call cc-option,-mcpu=8540 -msoft-float,-mcpu=powerpc)
> >>>>
> >>>> which for e500 kernel builds user either -mcpu=8540 or -mcpu=powerpc
> >>>> (probably as a fallback if -mcpu=8540 is not supported).
> >>>
> >>> The -mcpu=powerpc fallback can probably be skipped here, that must have been
> >>> for compilers predating the addition of -mcpu=8540, and even the oldest ones
> >>> support that now.
> >>
> >> Ok, makes sense.
> >>
> >>>> So for me it looks like that problematic code
> >>>>
> >>>>    KBUILD_CFLAGS         += -mcpu=powerpc
> >>>>    KBUILD_AFLAGS         += -mcpu=powerpc
> >>>>
> >>>> needs to be somehow skipped when compiling for CONFIG_E500.
> >>>>> My change which skips that code base on ifndef CONFIG_E500 should be
> >>>> fine as when CONFIG_E500 is disabled it does nothing and when it is
> >>>> enabled then code
> >>>>
> >>>>    CFLAGS-$(CONFIG_E500) += $(call cc-option,-mcpu=8540 -msoft-float,-mcpu=powerpc)
> >>>>
> >>>> is called which sets -mcpu option suitable for e500.
> >>>
> >>> I think this part is indeed fishy, but adding another special case for E500
> >>> seems to take it in the wrong direction.
> >>>
> >>> Nick added this in 4bf4f42a2feb ("powerpc/kbuild: Set default generic
> >>> machine type
> >>> for 32-bit compile") as a compile-time fix to prevent the default target from
> >>> getting used when the compiler supports both 64-bit and 32-bit. This is the
> >>> right idea, but it's inconsistent to pass different flags depending on the type
> >>> of toolchain, and it loses the more specific options.
> >>>
> >>> Another problem I see is that a kernel that is built for both E500 and E500MC
> >>> uses -mcpu=e500mc and may not actually work on the older ones either
> >>> (even with your patch).
> >>
> >> That is probably truth, -mcpu=8540 should have been chosen. (Anyway it
> >> should have been called -mcpu=e500, no idea why gcc still name it 8540.)
> >>
> >>> I think what you actually want is to set one option for each of the
> >>> possible CPU types:
> >>>
> >>> CFLAGS_CPU-$(CONFIG_PPC_BOOK3S_32) := -mcpu=powerpc
> >>> CFLAGS_CPU-$(CONFIG_PPC_85xx) := -mcpu=8540
> >>> CFLAGS_CPU-$(CONFIG_PPC8xx) := -mcpu=860
> >>> CFLAGS_CPU-$(CONFIG_PPC44x) := -mcpu=440
> >>> CFLAGS_CPU-$(CONFIG_PPC40x) := -mcpu=405
> >>> ifdef CONFIG_CPU_LITTLE_ENDIAN
> >>> CFLAGS_CPU-$(CONFIG_BOOK3S_64) := -mcpu=power8
> >>> else
> >>> CFLAGS_CPU-$(CONFIG_BOOK3S_64) := -mcpu=power5
> >>> endif
> >>> CFLAGS_CPU-$(CONFIG_BOOK3E_64) := -mcpu=powerpc64
> >>
> >> Yes, this is something I would expect that in Makefile should be.
> > 
> > So what about this change?
> > 
> > diff --git a/arch/powerpc/Makefile b/arch/powerpc/Makefile
> > index a0cd70712061..74a608b5796a 100644
> > --- a/arch/powerpc/Makefile
> > +++ b/arch/powerpc/Makefile
> > @@ -15,22 +15,7 @@ HAS_BIARCH	:= $(call cc-option-yn, -m32)
> >   # Set default 32 bits cross compilers for vdso and boot wrapper
> >   CROSS32_COMPILE ?=
> >   
> > -ifeq ($(HAS_BIARCH),y)
> > -ifeq ($(CROSS32_COMPILE),)
> > -ifdef CONFIG_PPC32
> > -# These options will be overridden by any -mcpu option that the CPU
> > -# or platform code sets later on the command line, but they are needed
> > -# to set a sane 32-bit cpu target for the 64-bit cross compiler which
> > -# may default to the wrong ISA.
> > -KBUILD_CFLAGS		+= -mcpu=powerpc
> > -KBUILD_AFLAGS		+= -mcpu=powerpc
> > -endif
> > -endif
> > -endif
> > -
> > -ifdef CONFIG_PPC_BOOK3S_32
> > -KBUILD_CFLAGS		+= -mcpu=powerpc
> > -endif
> > +CFLAGS-$(CONFIG_PPC_BOOK3S_32) += -mcpu=powerpc
> 
> This comes too early, it is overriden by later CFLAGS-$(CONFIG_PPC32) := 
> something
> 
> >   
> >   # If we're on a ppc/ppc64/ppc64le machine use that defconfig, otherwise just use
> >   # ppc64_defconfig because we have nothing better to go on.
> > @@ -163,17 +148,14 @@ CFLAGS-$(CONFIG_PPC32)	+= $(call cc-option, $(MULTIPLEWORD))
> >   
> >   CFLAGS-$(CONFIG_PPC32)	+= $(call cc-option,-mno-readonly-in-sdata)
> >   
> > -ifdef CONFIG_PPC_BOOK3S_64
> >   ifdef CONFIG_CPU_LITTLE_ENDIAN
> > -CFLAGS-$(CONFIG_GENERIC_CPU) += -mcpu=power8
> > -CFLAGS-$(CONFIG_GENERIC_CPU) += $(call cc-option,-mtune=power9,-mtune=power8)
> > +CFLAGS-$(CONFIG_PPC_BOOK3S_64) += -mcpu=power8
> > +CFLAGS-$(CONFIG_PPC_BOOK3S_64) += $(call cc-option,-mtune=power9,-mtune=power8)
> >   else
> > -CFLAGS-$(CONFIG_GENERIC_CPU) += $(call cc-option,-mtune=power7,$(call cc-option,-mtune=power5))
> > -CFLAGS-$(CONFIG_GENERIC_CPU) += $(call cc-option,-mcpu=power5,-mcpu=power4)
> > -endif
> > -else ifdef CONFIG_PPC_BOOK3E_64
> > -CFLAGS-$(CONFIG_GENERIC_CPU) += -mcpu=powerpc64
> > +CFLAGS-$(CONFIG_PPC_BOOK3S_64) += $(call cc-option,-mtune=power7,$(call cc-option,-mtune=power5))
> > +CFLAGS-$(CONFIG_PPC_BOOK3S_64) += $(call cc-option,-mcpu=power5,-mcpu=power4)
> 
> So before that change I got -mcpu=power9
> 
> Now I get -mtune=power7 -mcpu=power5 -mcpu=power9

I did it like Arnd wrote.

And seems that it does not work and now is fully out of the scope of the
original issue. Now I'm really lost here.

So I nobody comes with better solution, I would prefer to stick with my
original version which targets _only_ e500 cores.

Any other suggestion?

> 
> 
> >   endif
> > +CFLAGS-$(CONFIG_PPC_BOOK3E_64) += -mcpu=powerpc64
> >   
> >   ifdef CONFIG_FUNCTION_TRACER
> >   CC_FLAGS_FTRACE := -pg
> > @@ -193,13 +175,8 @@ endif
> >   CFLAGS-$(CONFIG_E5500_CPU) += $(E5500_CPU)
> >   CFLAGS-$(CONFIG_E6500_CPU) += $(call cc-option,-mcpu=e6500,$(E5500_CPU))
> >   
> > -ifdef CONFIG_PPC32
> > -ifdef CONFIG_PPC_E500MC
> > -CFLAGS-y += $(call cc-option,-mcpu=e500mc,-mcpu=powerpc)
> > -else
> > +CFLAGS-$(CONFIG_PPC_E500MC) += $(call cc-option,-mcpu=e500mc,-mcpu=powerpc)
> 
> Before I got -mcpu=e6500
> 
> Now I get  -mcpu=powerpc64 -mcpu=e6500 -mcpu=e500mc -mcpu=8540
> 
> >   CFLAGS-$(CONFIG_E500) += $(call cc-option,-mcpu=8540 -msoft-float,-mcpu=powerpc)
> > -endif
> > -endif
> >   
> >   asinstr := $(call as-instr,lis 9$(comma)foo@high,-DHAVE_AS_ATHIGH=1)
> >   
> > 
> > 
> >> But what to do with fallback value?
> >>
> >>> For the non-generic CPU types, there is also CONFIG_TARGET_CPU,
> >>> and the list above could just get folded into that instead.
> >>>
> >>>         Arnd
> 
> 
> Christophe