Re: [PATCH] powerpc: workaround clang codegen bug in dcbz

[Segher Boessenkool <segher@kernel.crashing.org>]

On Tue, Jul 30, 2019 at 04:30:29PM +0200, Arnd Bergmann wrote:
> On Tue, Jul 30, 2019 at 3:49 PM Segher Boessenkool
> <segher@kernel.crashing.org> wrote:
> >
> > On Tue, Jul 30, 2019 at 09:34:28AM +0200, Arnd Bergmann wrote:
> > > Upon a second look, I think the issue is that the "Z" is an input argument
> > > when it should be an output. clang decides that it can make a copy of the
> > > input and pass that into the inline asm. This is not the most efficient
> > > way, but it seems entirely correct according to the constraints.
> >
> > Most dcb* (and all icb*) do not change the memory pointed to.  The
> > memory is an input here, logically as well, and that is obvious.
> 
> Ah, right. I had only thought of dcbz here, but you are right that using
> an output makes little sense for the others.
> 
> readl() is another example where powerpc currently uses "Z" for an
> input, which illustrates this even better.

in_le32 and friends?  Yeah, huh.  If LLVM copies that to the stack as
well, its (not byte reversing) read will be atomic just fine, so things
will still work correctly.

The things defined with DEF_MMIO_IN_D (instead of DEF_MMIO_IN_X) do not
look like they will work correctly if an update form address is chosen,
but that won't happen because the constraint is "m" instead of "m<>",
making the %Un pretty useless (it will always be the empty string).

> > As I said many times already, LLVM does not seem to treat all asm
> > operands as lvalues.  That is a bug.  And it is critical for memory
> > operands for example, as should be obvious if you look at at for a few
> > seconds (you pass *that* memory, not a copy of it).  The thing you pass
> > has an identity.  It's an lvalue.  This is true for *all* inline asm
> > operands, not just output operands and memory operands, but it is most
> > obvious there.
> 
> >From experimentation, I would guess that llvm handles "m" correctly, but
> not "Z". See https://godbolt.org/z/uqfDx_ for another example.

Yeah, it does not treat "Z" as a memory constraint apparently, and it
special cases output operands and memory operands to be lvalues, but
does not do that for everything else as it should.

> > Or, LLVM might have a bug elsewhere.
> >
> > Either way, the asm is fine, and it has worked fine in GCC since
> > forever.  Changing this constraint to be an output constraint would
> > just be obfuscation (we could change *all* operands to *everything* to
> > be inout ("+") constraints, and it won't affect correctness, just the
> > reader's sanity).
> 
> I would still argue that for dcbz specifically, an output makes more
> sense than an input, but as you say that does not solve the others.

An output would be somewhat misleading.  dcbz zeroes the whole aligned
cache block sized region of memory its operand points into.  The kernel
dcbz functions do not easily know the cache block size I think, and
besides, you want a "memory" clobber anyway, also for the other dcb*,
so it won't help anything.  Also, the compiler can almost never use the
extra info ("affects the aligned 32B or 128B block this points into")
usefully anyway; it will usually see it as "can alias pretty much
anything".  Just use a "memory" clobber :-/


Segher