[PATCH] fix PR opt/2391 (mostly ARM)

[Richard Henderson]

On Fri, Jan 10, 2003 at 10:56:43AM -0800, Adam Nemet wrote:
> +static unsigned HOST_WIDE_INT
> +nonzero_bits1 (x, mode, known_x, known_x_ret)
> +     rtx x;
> +     enum machine_mode mode;
> +     rtx known_x;
> +     unsigned HOST_WIDE_INT known_x_ret;
> +{
> +  if (x == known_x)
> +    return known_x_ret;

I believe you should be using rtx_equal_p here instead of pointer
equality.  Similarly with the other comparisons below.

I also believe that you should be checking MODE against the mode
that was used for KNOWN_X as well.  This means passing around a
third cache value as well.

I considered whether it might be useful to have a more general
lookaside cache for this.  I'm guessing not, since the memory
allocation overhead would tend to dominate the common case.

> -nonzero_bits (x, mode)
> +nonzero_bits2 (x, mode, known_x, known_x_ret)
>       rtx x;
>       enum machine_mode mode;
> +     rtx known_x;
> +     unsigned HOST_WIDE_INT known_x_ret;
>  {
>    unsigned HOST_WIDE_INT nonzero = GET_MODE_MASK (mode);
>    unsigned HOST_WIDE_INT inner_nz;
> @@ -8156,7 +8218,7 @@ nonzero_bits (x, mode)
>        && GET_MODE_BITSIZE (GET_MODE (x)) <= HOST_BITS_PER_WIDE_INT
>        && GET_MODE_BITSIZE (mode) > GET_MODE_BITSIZE (GET_MODE (x)))
>      {
> -      nonzero &= nonzero_bits (x, GET_MODE (x));
> +      nonzero &= nonzero_bits1 (x, GET_MODE (x), known_x, known_x_ret);

You might consider redefining nonzero_bits for the duration of this
function.  It would avoid cluttering the text, and allow you to
easily add the third cache argument.

> @@ -11325,7 +11459,45 @@ update_table_tick (x)
> +	    /* If x0 and x1 are identical then there is no need to
> +	       process x0.  */
> +	    if (x0 == x1)
> +	      break;
> +
> +	    /* If x0 is identical to a subexpression of x1 then while
> +	       processing x1, x0 has already been processed.  Thus we
> +	       are done with x.  */
> +	    if ((GET_RTX_CLASS (GET_CODE (x1)) == '2'
> +		 || GET_RTX_CLASS (GET_CODE (x1)) == 'c')
> +		&& (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1)))
> +	      break;
> +
> +	    /* If x1 is identical to a subexpression of x0 then we
> +	       still have to process the rest of x0.  */
> +	    if ((GET_RTX_CLASS (GET_CODE (x0)) == '2'
> +		 || GET_RTX_CLASS (GET_CODE (x0)) == 'c')
> +		&& (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1)))

Um, how often do these trigger?


r~