Re: [PATCH v2] powerpc: slightly improve cache helpers

[Nick Desaulniers]

On Mon, Jul 22, 2019 at 10:58 AM Segher Boessenkool
<segher@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Mon, Jul 22, 2019 at 10:21:07AM -0700, Nick Desaulniers wrote:
> > On Sun, Jul 21, 2019 at 11:19 PM Segher Boessenkool
> > <segher@xxxxxxxxxxxxxxxxxxx> wrote:
> > > On Sun, Jul 21, 2019 at 07:41:40PM -0700, Nathan Chancellor wrote:
> > > > On Sun, Jul 21, 2019 at 01:01:50PM -0500, Segher Boessenkool wrote:
> > > > > On Sun, Jul 21, 2019 at 12:58:46AM -0700, Nathan Chancellor wrote:
> > > > > > 0000017c clear_user_page:
> > > > > >      17c: 94 21 ff f0                     stwu 1, -16(1)
> > > > > >      180: 38 80 00 80                     li 4, 128
> > > > > >      184: 38 63 ff e0                     addi 3, 3, -32
> > > > > >      188: 7c 89 03 a6                     mtctr 4
> > > > > >      18c: 38 81 00 0f                     addi 4, 1, 15
> > > > > >      190: 8c c3 00 20                     lbzu 6, 32(3)
> > > > > >      194: 98 c1 00 0f                     stb 6, 15(1)
> > > > > >      198: 7c 00 27 ec                     dcbz 0, 4
> > > > > >      19c: 42 00 ff f4                     bdnz .+65524
> > > > >
> > > > > Uh, yeah, well, I have no idea what clang tried here, but that won't
> > > > > work.  It's copying a byte from each target cache line to the stack,
> > > > > and then does clears the cache line containing that byte on the stack.
> > > > >
> > > > > I *guess* this is about "Z" and not about "%y", but you'll have to ask
> > > > > the clang people.
> > > > >
> > > > > Or it may be that they do not treat inline asm operands as lvalues
> > > > > properly?  That rings some bells.  Yeah that looks like it.
> > >
> > > The code is
> > >   __asm__ __volatile__ ("dcbz %y0" : : "Z"(*(u8 *)addr) : "memory");
> > >
> > > so yeah it looks like clang took that  *(u8 *)addr  as rvalue, and
> > > stored that in stack, and then used *that* as memory.
> >
> > What's the %y modifier supposed to mean here?
>
> It prints a memory address for an indexed operand.
>
> If you write just "%0" it prints addresses that are a single register
> as "0(r3)" instead of "0,r3".  Some instructions do not allow offset
> form.
>
> > addr is in the list of
> > inputs, so what's wrong with using it as an rvalue?
>
> It seems to use *(u8 *)addr as rvalue.  Asm operands are lvalues.  It
> matters a lot for memory operands.

Hmm...not sure that's specified behavior.  Anyways, I've filed:
https://bugs.llvm.org/show_bug.cgi?id=42762
to see if folks more familiar with LLVM's ppc backend have some more thoughts.

I recommend considering reverting commit 6c5875843b87 ("powerpc:
slightly improve cache helpers") until the issue is resolved in clang,
otherwise I'll probably just turn off our CI builds of PPC32 for the
time being.
-- 
Thanks,
~Nick Desaulniers