Re: [PATCH] doc: clarify semantics of vector bitwise shifts

[Alexander Monakov <amonakov@ispras.ru>]

[-- Attachment #1: Type: text/plain, Size: 3322 bytes --]


On Wed, 24 May 2023, Richard Biener wrote:

> On Wed, May 24, 2023 at 2:54 PM Alexander Monakov via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
> >
> > Explicitly say that bitwise shifts for narrow types work similar to
> > element-wise C shifts with integer promotions, which coincides with
> > OpenCL semantics.
> 
> Do we need to clarify that v << w with v being a vector of shorts
> still yields a vector of shorts and not a vector of ints?

I don't think so, but if necessary we could add "and the result was
truncated back to the base type":

    When the base type is narrower than @code{int}, element-wise shifts
    are performed as if operands underwent C integer promotions, and
    the result was truncated back to the base type, like in OpenCL. 

> Btw, I don't see this promotion reflected in the IL.  For
> 
> typedef short v8hi __attribute__((vector_size(16)));
> 
> v8hi foo (v8hi a, v8hi b)
> {
>   return a << b;
> }
> 
> I get no masking of 'b' and vector lowering if the target doens't handle it
> yields
> 
>   short int _5;
>   short int _6;
> 
>   _5 = BIT_FIELD_REF <a_1(D), 16, 0>;
>   _6 = BIT_FIELD_REF <b_2(D), 16, 0>;
>   _7 = _5 << _6;
> 
> which we could derive ranges from for _6 (apparantly we don't yet).

Here it depends on how we define the GIMPLE-level semantics of bit-shift
operators for narrow types. To avoid changing lowering we could say that
shifting by up to 31 bits is well-defined for narrow types.

RTL-level semantics are also undocumented, unfortunately.

> Even
> 
> typedef int v8hi __attribute__((vector_size(16)));
> 
> v8hi x;
> int foo (v8hi a, v8hi b)
> {
>   x = a << b;
>   return (b[0] > 33);
> }
> 
> isn't optimized currently (but could - note I've used 'int' elements here).

Yeah. But let's constrain the optimizations first.

> So, I don't see us making sure the hardware does the right thing for
> out-of bound values.

I think in practice it worked out even if GCC did not pay attention to it,
because SIMD instructions had to facilitate autovectorization for C with
corresponding shift semantics.

Alexander

> 
> Richard.
> 
> > gcc/ChangeLog:
> >
> >         * doc/extend.texi (Vector Extensions): Clarify bitwise shift
> >         semantics.
> > ---
> >  gcc/doc/extend.texi | 7 ++++++-
> >  1 file changed, 6 insertions(+), 1 deletion(-)
> >
> > diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi
> > index e426a2eb7d..6b4e94b6a1 100644
> > --- a/gcc/doc/extend.texi
> > +++ b/gcc/doc/extend.texi
> > @@ -12026,7 +12026,12 @@ elements in the operand.
> >  It is possible to use shifting operators @code{<<}, @code{>>} on
> >  integer-type vectors. The operation is defined as following: @code{@{a0,
> >  a1, @dots{}, an@} >> @{b0, b1, @dots{}, bn@} == @{a0 >> b0, a1 >> b1,
> > -@dots{}, an >> bn@}}@. Vector operands must have the same number of
> > +@dots{}, an >> bn@}}@.  When the base type is narrower than @code{int},
> > +element-wise shifts are performed as if operands underwent C integer
> > +promotions, like in OpenCL.  This makes vector shifts by up to 31 bits
> > +well-defined for vectors with @code{char} and @code{short} base types.
> > +
> > +Operands of binary vector operations must have the same number of
> >  elements.
> >
> >  For convenience, it is allowed to use a binary vector operation
> > --
> > 2.39.2
> >
>