Re: [PATCH 1/7] RISC-V: Top-Level Makefile for riscv{32,64}

[Palmer Dabbelt]

On Tue, 06 Jun 2017 10:39:04 PDT (-0700), merker@xxxxxxxxxx wrote:
> On Mon, Jun 05, 2017 at 09:56:33PM -0700, Palmer Dabbelt wrote:
>> On Mon, 29 May 2017 03:50:47 PDT (-0700), Arnd Bergmann wrote:
>> > On Sat, May 27, 2017 at 2:57 AM, Palmer Dabbelt <palmer@xxxxxxxxxxx> wrote:
>> >> On Tue, 23 May 2017 04:30:50 PDT (-0700), Arnd Bergmann wrote:
>> >>> On Tue, May 23, 2017 at 2:41 AM, Palmer Dabbelt <palmer@xxxxxxxxxxx> wrote:
>> >>>> RISC-V has both 32-bit and 64-bit base ISAs, but they are very similar.
>> >>>> Like some other platforms, we'd like to share one arch directory between
>> >>>> the two of them.
>> >>>
>> >>> I think we mainly do the others for backwards-compatibility with ancient
>> >>> build scripts, and we don't need that here. Instead, you could add one more
>> >>> line to the 'SUBARCH:=' statement that interprets the uname output.
>> >>
>> >> I don't think that does the same thing.  The desired effect of this diff is:
>> >>
>> >>  * "uname -m" when running on a RISC-V machine returns either riscv32 or
>> >>    riscv64, as that's what tools like autoconf expect when trying to find
>> >>    tuples.
>> >>
>> >>  * I can cross compile for riscv32 and riscv64.  That's currently controlled by
>> >>    a Kconfig setting, but ARCH=riscv32 vs ARCH=riscv64 controlls what defconfig
>> >>    sets.
>> >>
>> >>  * I can natively compile for riscv32 and riscv64.  That uses the same Kconfig
>> >>    setting, and the same ARCH=riscv32 vs ARCH=riscv64 switch for defconfig.
>> >
>> > Right, but my point is that a new architecture should not rely on 'ARCH='
>> > to pick the defconfig, we only do that on a couple of architectures for
>> > backwards compatibility with old scripts.
>> >
>> >> Neither of the two Kconfig issues is a big deal, but we de need "uname -m" to
>> >> return "riscv64" or "riscv32" not "riscv".  I think the only way to do that is
>> >> to set SRCARCH, but I'd be happy to change it if there's a better way.  I think
>> >> if I just do this
>> >>
>> >> diff --git a/Makefile b/Makefile
>> >> index 0606f28..4adc609 100644
>> >> --- a/Makefile
>> >> +++ b/Makefile
>> >> @@ -232,7 +232,8 @@ SUBARCH := $(shell uname -m | sed -e s/i.86/x86/ -e s/x86_64/x86/ \
>> >>                                   -e s/arm.*/arm/ -e s/sa110/arm/ \
>> >>                                   -e s/s390x/s390/ -e s/parisc64/parisc/ \
>> >>                                   -e s/ppc.*/powerpc/ -e s/mips.*/mips/ \
>> >> -                                 -e s/sh[234].*/sh/ -e s/aarch64.*/arm64/ )
>> >> +                                 -e s/sh[234].*/sh/ -e s/aarch64.*/arm64/ \
>> >> +                                 -e s/riscv.*/riscv/ )
>> >>
>> >>  # Cross compiling and selecting different set of gcc/bin-utils
>> >>  # ---------------------------------------------------------------------------
>> >> @@ -269,14 +270,6 @@ ifeq ($(ARCH),x86_64)
>> >>          SRCARCH := x86
>> >>  endif
>> >>
>> >> -# Additional ARCH settings for RISC-V
>> >> -ifeq ($(ARCH),riscv32)
>> >> -       SRCARCH := riscv
>> >> -endif
>> >> -ifeq ($(ARCH),riscv64)
>> >> -       SRCARCH := riscv
>> >> -endif
>> >> -
>> >>  # Additional ARCH settings for sparc
>> >>  ifeq ($(ARCH),sparc32)
>> >>         SRCARCH := sparc
>> >>
>> >> then I'll end up with "uname -m" as "riscv" -- I haven't tried it, but that's
>> >> why we ended up with this diff in the first place.
>> >
>> > Do you mean the "uname -m" output comes from "${SRCARCH}" at
>> > the time of the kernel build? That would be easy enough to change
>> > by simply hardcoding it depending on CONFIG_64BIT.
>>
>> OK, I didn't know about COMPAT_UTS_MACHINE.  That's a much better solution,
>> I'll use that.
>
> Hello Palmer,
>
> I suppose the commit:
>
>   https://github.com/riscv/riscv-linux/commit/8c826930d2a19ecd4f1036f10a380dc4fddd0da5
>
> aims to address this, but it appears to be incomplete. It lacks
> the first fragment of the patch above, i.e. the conversion from
> the "uname -m" output (i.e. "riscv{64,32}") to the canonical
> arch string (i.e. "riscv").  As a result, a native build (which
> normally doesn't explicitly pass ARCH=riscv to make and therefore
> relies on the output of "uname -m") would fail.

Sorry about that, I dropped the commit.  I'm in the middle of bisecting to find
a regression, so the patch might now show up for a bit, but it'll be fixed.