Re: [PATCH], Add PowerPC ISA 3.0 IEEE 128-bit floating point round to odd built-in functions

[Michael Meissner <meissner at linux dot vnet dot ibm dot com>]

On Fri, Sep 29, 2017 at 12:10:07PM -0500, Segher Boessenkool wrote:
> Hi Mike,
> 
> On Thu, Sep 28, 2017 at 06:34:23PM -0400, Michael Meissner wrote:
> > This patch addss built-in functions on PowerPC ISA 3.0 (power9) that allow the
> > user to access the round to odd IEEE 128-bit floating point instructions.
> 
> > --- gcc/config/rs6000/rs6000.md	(revision 253267)
> > +++ gcc/config/rs6000/rs6000.md	(working copy)
> > @@ -14505,7 +14505,9 @@ (define_insn_and_split "trunc<mode>sf2_h
> >    "#"
> >    "&& 1"
> >    [(set (match_dup 2)
> > -	(unspec:DF [(match_dup 1)] UNSPEC_ROUND_TO_ODD))
> > +	(unspec:DF [(float_truncate:DF
> > +		     (match_dup 1))]
> > +		   UNSPEC_ROUND_TO_ODD))
> >     (set (match_dup 0)
> >  	(float_truncate:SF (match_dup 2)))]
> >  {
> 
> I don't think this is correct.  It says to first truncate the f128 to DF,
> and then round it to odd; I think you want to do the truncation with
> round-to-odd rounding mode already.
> 
> > +(define_insn "mul<mode>3_odd"
> > +  [(set (match_operand:IEEE128 0 "altivec_register_operand" "=v")
> > +	(unspec:IEEE128
> > +	 [(mult:IEEE128
> > +	   (match_operand:IEEE128 1 "altivec_register_operand" "v")
> > +	   (match_operand:IEEE128 2 "altivec_register_operand" "v"))]
> > +	 UNSPEC_ROUND_TO_ODD))]
> > +  "TARGET_FLOAT128_HW && FLOAT128_IEEE_P (<MODE>mode)"
> > +  "xsmulqpo %0,%1,%2"
> > +  [(set_attr "type" "vecfloat")
> > +   (set_attr "size" "128")])
> 
> Similar here (and everywhere else): it does an f128 mul, so rounding
> with whatever rounding mode is current, and *then* it rounds to odd.
> 
> > +(define_insn "sqrt<mode>2_odd"
> > +  [(set (match_operand:IEEE128 0 "altivec_register_operand" "=v")
> > +	(unspec:IEEE128
> > +	 [(sqrt:IEEE128
> > +	   (match_operand:IEEE128 1 "altivec_register_operand" "v"))]
> > +	 UNSPEC_ROUND_TO_ODD))]
> > +  "TARGET_FLOAT128_HW && FLOAT128_IEEE_P (<MODE>mode)"
> > +   "xssqrtqpo %0,%1"
> 
> (One space too many here).
> 
> Everything else looks fine, but that unspec thing needs fixing.  Can be
> later, things will likely work for now, so okay for trunk.  Thanks.
> 
> How do other ports deal with this?  Insns with a specific rounding mode?
> Have a separate unspec for every operation?  Not very nice either :-(

I just want to do the minimal work so the glibc people can use this
instruction, but I am not that motivated to go back and modify it further.  So,
I would prefer to do it right before committing it, unless you want to refine
it.

As I see it, there are 2 ways to encode the RTL:

    1)	Use one UNSPEC name, and add the operation inside of the spec;

    2)	Add separate UNSPEC's for each operation.

Either one is fine with me.  If you would prefer separate names, I can do it.

In theory we could do something like define_subst, but that is a large time
waster to learn how to do it for 8 specialized instructions.

Similarly, I don't see the need for doing more general support, until we have
more data types and more rounding modes.

-- 
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meissner@linux.vnet.ibm.com, phone: +1 (978) 899-4797