Re: [PATCH] rs6000: Accept const pointer operands for MMA builtins [PR109073]

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

Hi!

On Thu, Mar 09, 2023 at 05:30:53PM +0800, Kewen.Lin wrote:
> on 2023/3/9 07:01, Peter Bergner via Gcc-patches wrote:
> > PR109073 shows a problem where GCC 11 and GCC 10 do not accept a const
> > __vector_pair pointer operand to some MMA builtins, which GCC 12 and later
> > correctly accept.  Fixed here by initializing the builtins to accept const
> > pointers.

"Pointers to const" is the more correct.  A "const pointer" is e.g.
  int *const p;
not the same thing at all, and sometimes this is useful to have ;-)

> > This patch was tested in both GCC 11 and GCC 10 on powerpc64le-linux and
> > showed no regressions.  Ok for backports?

It isn't truly a backport. You can put it on 11 and 10 at the same time,
there is no benefit doing it on 11 only first.

> >  	{
> >  	  op[nopnds++] = build_pointer_type (void_type_node);
> >  	  if (d->code == MMA_BUILTIN_DISASSEMBLE_ACC)
> > -	    op[nopnds++] = build_pointer_type (vector_quad_type_node);
> > +	    op[nopnds++] = build_pointer_type (build_qualified_type
> > +						 (vector_quad_type_node,
> > +						  TYPE_QUAL_CONST));
> 
> Nit: Maybe we can build them out of the loop once and then just use the
> built one in the loop.

Or as globals even.  Currently we have X and pointer to X, but no
pointer to const X (and no const X either, but that isn't so useful).

The generic code doesn't have this either, hrm.

(snip)

> Simply testing __builtin_mma_xxmtacc and __builtin_mma_xxmfacc as below:
> 
> $ cat test.C
> void foo0(const __vector_quad *acc) {
>   __builtin_mma_xxmtacc(acc);
>   __builtin_mma_xxmfacc(acc);
> }
> 
> test.C:2:25: error: invalid conversion from ‘const __vector_quad*’ to ‘__vector_quad*’ [-fpermissive]
>     2 |   __builtin_mma_xxmtacc(acc);
> 
> test.C:3:25: error: invalid conversion from ‘const __vector_quad*’ to ‘__vector_quad*’ [-fpermissive]
>     3 |   __builtin_mma_xxmfacc(acc);
> 
> They also suffered the same error on gcc11 branch but not on trunk.

Yeah, there is more to be done here.

> Besides, I'm not sure if the existing bif declarations using ptr_vector_pair_type_node
> and ptr_vector_quad_type_node are all intentional, at least it looks weird to me that
> we declare const __vector_pair* for this __builtin_vsx_stxvp, which is meant to store 32
> bytes into the memory provided by the pointer biasing the sizetype offset, but the "const"
> qualifier seems to tell that this bif doesn't modify the memory pointed by the given pointer.

That looks like a bug.  Well it is one even.  Is it fixed on trunk?

Since the patch is a strict improvement already, it is okay for 11 and
10.  But you (Peter) may want to flesh it out a bit first?  Or first
commit only this if that works better for you.


Segher