[lno] [patch] vectorizer update - support constants.

[Dorit Naishlos]

> This is a problem in the ppc machine description.  According to
> ...
> that instruction is legal.  To fix this, you need a more complex
> test instead of "TARGET_ALTIVEC".

Yes, that does the trick.
I'll send a patch shortly.

thanks very much,
dorit



                                                                                                                              
                      Richard Henderson                                                                                       
                      <rth@redhat.com>         To:       Dorit Naishlos/Haifa/IBM@IBMIL                                       
                                               cc:       Daniel Jacobowitz <drow@mvista.com>, gcc-patches@gcc.gnu.org         
                      23/01/2004 00:11         Subject:  Re: [lno] [patch] vectorizer update - support constants.             
                                                                                                                              




On Thu, Jan 22, 2004 at 11:58:45PM +0200, Dorit Naishlos wrote:
> The vectorizer actually generates the 0 initialization out of the loop,
but
> it is propagated into the store insn inside the loop during gcse:
>
> (insn 17 16 18 1 (set (mem:V8HI (plus:SI (reg:SI 124)
>                 (reg:SI 127)) [3 S16 A128])
>         (const_vector:V8HI [
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>             ])) 561 {*movv8hi_internal1} (nil)
>     (nil))
>
> ...and unfortunately remains in the loop from that point.

This is a problem in the ppc machine description.  According to

  [(set (match_operand:V8HI 0 "nonimmediate_operand" "=m,v,v,o,r,r,v")
        (match_operand:V8HI 1 "input_operand" "v,m,v,r,o,r,W"))]
  "TARGET_ALTIVEC"

that instruction is legal.  To fix this, you need a more complex
test instead of "TARGET_ALTIVEC".  Notice that movsi_internal1 uses

  "gpc_reg_operand (operands[0], SImode)
   || gpc_reg_operand (operands[1], SImode)"

This is the correct type of test, though of course that exact test
would fail for vxx registers.


r~