Re: patch e500: unspec spe_evmra

Re: patch e500: unspec spe_evmra

[David Edelsohn]

	* config/rs6000/spe.md ("spe_evmra"): Change to unspec.

Sure.

David

patch e500: unspec spe_evmra

[Aldy Hernandez]

The spe_evmra instruction is causing grief because gcc will insist in
using the accumulator as a source operand.  Since we have no direct way
of coping from the accumulator, it's just easier to declare this
unspec and keep the scheduler from messing us up.

Ok?

2002-07-29  Aldy Hernandez  <aldyh@redhat.com>

	* config/rs6000/spe.md ("spe_evmra"): Change to unspec.


Index: config/rs6000/spe.md
===================================================================
RCS file: /cvs/uberbaum/gcc/config/rs6000/spe.md,v
retrieving revision 1.1
diff -c -p -r1.1 spe.md
*** config/rs6000/spe.md	25 Jul 2002 02:22:47 -0000	1.1
--- config/rs6000/spe.md	29 Jul 2002 20:00:57 -0000
***************
*** 2028,2034 ****
  (define_insn "spe_evmra"
    [(set (match_operand:V2SI 0 "gpc_reg_operand" "=r")
          (match_operand:V2SI 1 "gpc_reg_operand" "r"))
!    (set (reg:V2SI SPE_ACC_REGNO) (match_dup 1))]
    "TARGET_SPE"
    "evmra %0,%1"
    [(set_attr "type" "veccomplex")
--- 2028,2035 ----
  (define_insn "spe_evmra"
    [(set (match_operand:V2SI 0 "gpc_reg_operand" "=r")
          (match_operand:V2SI 1 "gpc_reg_operand" "r"))
!    (set (reg:V2SI SPE_ACC_REGNO)
! 	(unspec:V2SI [(match_dup 1)] 726))]
    "TARGET_SPE"
    "evmra %0,%1"
    [(set_attr "type" "veccomplex")

Re: patch e500: unspec spe_evmra

[Graham Stott]

Aldy,

Aldy Hernandez wrote:
>
> --- 2028,2035 ----
>   (define_insn "spe_evmra"
>     [(set (match_operand:V2SI 0 "gpc_reg_operand" "=r")
>           (match_operand:V2SI 1 "gpc_reg_operand" "r"))
> !    (set (reg:V2SI SPE_ACC_REGNO)
> ! 	(unspec:V2SI [(match_dup 1)] 726))]
>     "TARGET_SPE"
>     "evmra %0,%1"
>     [(set_attr "type" "veccomplex")
> (define_constants


Don't much like the magic 726 unspec number might I suggest now
would be a good time for the rs6000 backend to start using the
define_contants md construct and banish the magic numbers.

Cheers
Graham

Re: patch e500: unspec spe_evmra

[David Edelsohn]

>>>>> Graham Stott writes:

Graham> Don't much like the magic 726 unspec number might I suggest now
Graham> would be a good time for the rs6000 backend to start using the
Graham> define_contants md construct and banish the magic numbers.

	Aldy already is using that in spe.md.  He is suppose to convert
the entire backend (mainly spe.md and altivec.md) to that method one of
these days.

David

Re: patch e500: unspec spe_evmra

[Aldy Hernandez]

On Mon, Jul 29, 2002 at 04:20:52PM -0400, David Edelsohn wrote:
> >>>>> Graham Stott writes:
> 
> Graham> Don't much like the magic 726 unspec number might I suggest now
> Graham> would be a good time for the rs6000 backend to start using the
> Graham> define_contants md construct and banish the magic numbers.
> 
> 	Aldy already is using that in spe.md.  He is suppose to convert
> the entire backend (mainly spe.md and altivec.md) to that method one of
> these days.

Doh!  Yes, that would be me, Mr. Procrastinator.  I'm avoiding it
because they're soooo many.  Will do though.

I'll start with moving the altivec builtins into altivec.md.

Aldy