Re: [PATCH] Fix altivec PR c/6290

[Aldy Hernandez <aldyh at redhat dot com>]

On Mon, Apr 15, 2002 at 05:33:20PM +0200, Jakub Jelinek wrote:

hi jakub.  hi bernd.

is either you or bernd going to commit this?  i haven't taken a look
at bernd's version, but this one looks ok.

regarding the other testcase that you found, i bet it gets fixed by
bernd's combine patch.

oh yeah, and you can't use CONST0_RTX like i thought, because you don't
have the mode.  i suppose we could change easy_vec_const to have a mode.

aldy

> Hi!
> 
> PR c/6290 is failing because initialize_uninitialized_subregs
> added
> (insn 70 48 4 (set (reg/v:V4SI 117)
>         (const_vector:V4SI[
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>                 (const_int 0 [0x0])
>             ] )) -1 (nil)
>     (nil))
> 
> but due to bug in easy_vector_constant rs6000_emit_move used force_const_mem
> on it which cannot work when no new pseudos can be created.
> easy_vector_constant never returned anything but 0.
> The following patch changes easy_vector_constant so that it does what
> the comment in it say.
> 
> altivec-5.c is a testcase for this, broken before this patch, fixed
> afterwards. While minimalizing the testcase, I found another testcase
> which ICEs gcc on ppc no matter whether this patch is in or not.
> 
> Ok to commit?
> 
> 2002-04-15  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR c/6290
> 	* config/rs6000/rs6000.c (easy_vector_constant): Return 1 if the
> 	CONST_VECTOR is { 0, ... 0 }.
> 
> 	* gcc.dg/altivec-5.c: New test.
> 	* gcc.dg/altivec-6.c: New test.
> 
> --- gcc/config/rs6000/rs6000.c.jj	Tue Apr  9 16:14:50 2002
> +++ gcc/config/rs6000/rs6000.c	Mon Apr 15 15:32:33 2002
> @@ -1221,18 +1221,24 @@ easy_vector_constant (op)
>  	 with CONST0_RTX for the current mode, but let's be safe
>  	 instead.  */
>  
> -      if (GET_CODE (elt) == CONST_INT && INTVAL (elt) != 0)
> -	return 0;
> -
> -      if (GET_CODE (elt) == CONST_DOUBLE
> -	       && (CONST_DOUBLE_LOW (elt) != 0
> -		   || CONST_DOUBLE_HIGH (elt) != 0))
> -	return 0;
> +      switch (GET_CODE (elt))
> +	{
> +	case CONST_INT:
> +	  if (INTVAL (elt) != 0)
> +	    return 0;
> +	  break;
> +	case CONST_DOUBLE:
> +	  if (CONST_DOUBLE_LOW (elt) != 0 || CONST_DOUBLE_HIGH (elt) != 0)
> +	    return 0;
> +	  break;
> +	default:
> +	  return 0;
> +	}
>      }
>  
>    /* We could probably generate a few other constants trivially, but
>       gcc doesn't generate them yet.  FIXME later.  */
> -  return 0;
> +  return 1;
>  }
>  
>  /* Return 1 if the operand is the constant 0.  This works for scalars
> --- gcc/testsuite/gcc.dg/altivec-5.c.jj	Mon Apr 15 17:00:40 2002
> +++ gcc/testsuite/gcc.dg/altivec-5.c	Mon Apr 15 17:02:46 2002
> @@ -0,0 +1,22 @@
> +/* PR c/6290
> +   This testcase ICEd because easy_vector_constant was broken.  */
> +/* { dg-do compile { target powerpc-*-* } } */
> +/* { dg-options "-maltivec -O2 -fno-strict-aliasing" } */
> +
> +#define vector __attribute__((vector_size(16)))
> +
> +register vector signed int reg asm ("%v31");
> +
> +vector signed int a, b;
> +
> +void
> +foo (const unsigned long x)
> +{
> +  vector signed int c;
> +  unsigned long d;
> +
> +  *(unsigned long *) &c = x;
> +  if (!__builtin_altivec_vcmpgtsw_p (0, c, b))
> +    b = __builtin_altivec_vsumsws (a, reg);
> +  __builtin_altivec_stvewx (b, 0, (int *) &d);
> +}
> --- gcc/testsuite/gcc.dg/altivec-6.c.jj	Mon Apr 15 17:03:14 2002
> +++ gcc/testsuite/gcc.dg/altivec-6.c	Mon Apr 15 17:12:19 2002
> @@ -0,0 +1,12 @@
> +/* { dg-do compile { target powerpc-*-* } } */
> +/* { dg-options "-maltivec -O2" } */
> +
> +#define vector __attribute__((vector_size(16)))
> +
> +void foo (const unsigned long x,
> +	  vector signed int a, vector signed int b)
> +{
> +  unsigned long d;
> +
> +  __builtin_altivec_stvewx (b, 0, (int *) &d);
> +}
> 
> 	Jakub