[PATCH 5/6 ver 3] rs6000, Add vector splat builtin support

[Carl Love]

On Wed, 2020-07-01 at 12:00 -0500, Segher Boessenkool wrote:
> Hi!
> 
> On Mon, Jun 29, 2020 at 03:31:48PM -0700, Carl Love wrote:
> > On Mon, 2020-06-29 at 16:58 -0500, Segher Boessenkool wrote:
> > > On Mon, Jun 29, 2020 at 02:29:54PM -0700, Carl Love wrote:
> > > > Segher:
> > > > 
> > > > On Thu, 2020-06-25 at 17:39 -0500, Segher Boessenkool wrote:
> > > > > > +;; Return 1 if op is a constant 32-bit floating point
> > > > > > value
> > > > > > +(define_predicate "f32bit_const_operand"
> > > > > > +  (match_code "const_double")
> > > > > > +{
> > > > > > +  if (GET_MODE (op) == SFmode)
> > > > > > +    return 1;
> > > > > > +
> > > > > > +  else if ((GET_MODE (op) == DFmode) && ((UINTVAL (op) >>
> > > > > > 32)
> > > > > > ==
> > > > > > 0))
> > > > > > +   {
> > > > > > +    /* Value fits in 32-bits */
> > > > > > +    return 1;
> > > > > > +    }
> > > > > > +  else
> > > > > > +    /* Not the expected mode.  */
> > > > > > +    return 0;
> > > > > > +})
> > > > > 
> > > > > I don't think this is the correct test.  What you want to see
> > > > > is
> > > > > if
> > > > > the
> > > > > number in "op" can be converted to an IEEE single-precision
> > > > > number,
> > > > > and
> > > > > back again, losslessly.  (And subnormal SP numbers aren't
> > > > > allowed
> > > > > either, but NaNs and infinities are).
> > > > 
> > > > The predicate is used with the xxsplitw_v4sf
> > > > define_expand.  The
> > > > "user"
> > > > claims the given immediate bit pattern is the bit pattern for a
> > > > single
> > > > precision floating point number.  The immediate value is not
> > > > converted
> > > > to a float.  Rather we are splatting a bit pattern that the
> > > > "user"
> > > > already claims represents a 32-bit floating point value.  I
> > > > just
> > > > need
> > > > to make sure the immediate value actually fits into 32-bits.
> > > > 
> > > > I don't see that I need to check that the value can be
> > > > converted to
> > > > IEEE float and back.  
> > > 
> > > Ah, okay.  Can you please put that in the function comment
> > > then?  Right
> > > now it says
> > > ;; Return 1 if op is a constant 32-bit floating point value
> > > and that is quite misleading.
> > 
> > Would the following be more clear
> > 
> > ;; Return 1 if op is a constant bit pattern representing a floating
> > ;; point value that fits in 32-bits. 
> 
> Yes...  But I still don't get it.
> 
> Say you have the number  785.066650390625 .  As a single-precision
> IEEE
> float, that is x4444_4444.  But as a double-precision IEEE float, it
> is
> 0x4088_8888_8000_0000, which does not fit in 32 bits!

The value 0x4444_4444 would be mode SFmode and the test would return 1.
We would be able to splat that 32-bit wide pattern.  

The value 0x4088_8888_8000_0000 would be DFmode but when you shift it
right 32 bits it would be non-zero and 

   else if ((GET_MODE (op) == DFmode) && ((UINTVAL (op) >> 32)

would be true and the function would return 1 incorrectly.  Argh!

So yes, we have a problem.  The second part is supposed to be checking
for all zero not non zero.   The test should be:

   else if ((GET_MODE (op) == DFmode) && (((UINTVAL (op) >> 32) == 0)

                  Carl