Re: [PATCH] riscv: pass machine size to sparse

[Palmer Dabbelt]

On Thu, 31 May 2018 08:35:54 PDT (-0700), luc.vanoostenryck@xxxxxxxxx wrote:
> On Thu, May 31, 2018 at 05:09:21AM -0700, Palmer Dabbelt wrote:
> > On Mon, 28 May 2018 23:14:20 PDT (-0700), yamada.masahiro@xxxxxxxxxxxxx wrote:
> > > 2018-05-29 15:11 GMT+09:00 Christoph Hellwig <hch@xxxxxxxxxxxxx>:
> > > > On Mon, May 28, 2018 at 06:35:05PM +0200, Luc Van Oostenryck wrote:
> > > > > By default, sparse assumes a 64bit machine when compiled on x86-64
> > > > > and 32bit when compiled on anything else.
> > > > >
> > > > > This can of course create all sort of problems when this doesn't
> > > > > correspond to the target's machine size, like issuing false
> > > > > warnings like: 'shift too big (32) for type unsigned long' or
> > > > > is 64bit while sparse was compiled on a 32bit machine, or worse,
> > > > > to not emit legitimate warnings.
> > > > >
> > > > > Fix this by passing the appropriate -m32/-m64 flag to sparse.
> > > >
> > > > Can we please move this to the common Kbuild code using the
> > > > CONFIG_64BIT syombol?  This really should not need boiler plate in
> > > > every architecture.
> > >
> > >
> > > I agree.
> > >
> > > Luc did so for -mbig/little-endian:
> > > https://patchwork.kernel.org/patch/10433957/
> > >
> > > We should do likewise for -m32/64.
> >
> > Sorry for being a bit slow here, but I like the idea of making the
> > 32/64-bit issue generic as it seems like it'll be necessary for
> > every port.
>
> Sure, Christophe asked it too.
> I've sent a new version doing it once and for all for all archs
> but I forgot to CC you:
> 	https://lkml.org/lkml/2018/5/30/948
>
> > Looking through the patch for big/little-endian I did
> > notice:
> >
> > * RISC-V compilers set "__riscv_xlen" to the length of an X
> > (integer) register  in bits.
> > * RISC-V compilers define "__riscv", and it doesn't appear we inform
> > sparse  about that.
> >
> > These two might not be that interesting, but we do already have some
> > cases where we're checking for __riscv_xlen in Linux.  I've yet to
> > successfully use sparse, but adding at least
> >
> >    CHECKFLAGS += -D__riscv
> >
> > seems reasonable,
>
> Sure (but I don't see a dependency in the kernel (yet)).
>
> > and possibly also some sort of
> >
> >    ifeq ($(CONFIG_ARCH_RV64I),y)
> >            CHECKFLAGS += -D__riscv_xlen=64
> >    else
> >            CHECKFLAGS += -D__riscv_xlen=32
> >    fi
> >
> > might be necessary.
>
> Yes, this one is really needed.
> I'll send a patch for this one and __riscv.
>
> > We strive to follow the generic rules for
> > ABI-related stuff like __LP64__ but I don't think there's any
> > generic mapping for XLEN.  Similarly there's "__riscv_flen" and
> > "__riscv_float_abi_*", but those are less likely to be used by the
> > kernel so they're probably not worth worrying about for now.
>
> Yes, I agree.
> Note that sparse will define __LP64__ (and _LP64) when in -m64 mode.
>
> > There's also a bunch of other RISC-V macros, the only one of which
> > we're currently using is "__riscv_muldiv" (and that's in a manner
> > that's unlikely to trigger any sort of static analysis).  Between a
> > lack of Kconfig options and a glibc port we're essentially mandating
> > IMA right now, so these probably don't matter.
>
> Yes, I just saw. I think also it's better to leave it so for now.
> And if it becomes more used, then better to infer it from the compiler
> than harcoding it.

Makes sense.  Thanks for the work!