[PATCH] Fix vector rotate regression (PR tree-optimization/57233)

[Marc Glisse]

On Thu, 26 Jun 2014, Jakub Jelinek wrote:

> So like this?  I've also changed get_compute_type so that it will DTRT
> even for -mavx and V4DImode vectors, so e.g. f5/f6/f8 routines in
> avx-pr57233.c improve.  Also, even for shifts by scalar, if e.g. target
> doesn't have shifts by scalar at all, and only has narrower vector by vector
> shifts, it should handle this case too.

All that? Cool!

> @@ -1455,11 +1507,83 @@ expand_vector_operations_1 (gimple_stmt_
> 	{
>           op = optab_for_tree_code (code, type, optab_scalar);
>
> +	  compute_type = get_compute_type (code, op, type);
> +	  if (compute_type == type)
> +	    return;
> 	  /* The rtl expander will expand vector/scalar as vector/vector
> -	     if necessary.  Don't bother converting the stmt here.  */
> -	  if (optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing
> -	      && optab_handler (opv, TYPE_MODE (type)) != CODE_FOR_nothing)
> +	     if necessary.  Pick one with wider vector type.  */
> +	  tree compute_vtype = get_compute_type (code, opv, type);
> +	  if (count_type_subparts (compute_vtype)
> +	      > count_type_subparts (compute_type))
> +	    {
> +	      compute_type = compute_vtype;
> +	      op = opv;
> +	    }
> +	}
> +
> +      if (code == LROTATE_EXPR || code == RROTATE_EXPR)
> +	{
> +	  if (compute_type == NULL_TREE)
> +	    compute_type = get_compute_type (code, op, type);
> +	  if (compute_type == type)
> 	    return;
> +	  /* Before splitting vector rotates into scalar rotates,
> +	     see if we can't use vector shifts and BIT_IOR_EXPR
> +	     instead.  For vector by vector rotates we'd also
> +	     need to check BIT_AND_EXPR and NEGATE_EXPR, punt there
> +	     for now, fold doesn't seem to create such rotates anyway.  */
> +	  if (compute_type == TREE_TYPE (type)
> +	      && !VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2)))
> +	    {
> +	      optab oplv, opl, oprv, opr, opo;
> +	      oplv = optab_for_tree_code (LSHIFT_EXPR, type, optab_vector);
> +	      /* Right shift always has to be logical, no matter what
> +		 signedness type has.  */
> +	      oprv = vlshr_optab;
> +	      opo = optab_for_tree_code (BIT_IOR_EXPR, type, optab_default);
> +	      opl = optab_for_tree_code (LSHIFT_EXPR, type, optab_scalar);
> +	      oprv = lshr_optab;
> +	      opr = optab_for_tree_code (RSHIFT_EXPR, type, optab_scalar);

Looks like there are some typos in there, you are assigning to oprv twice.

-- 
Marc Glisse