Re: [RFA/alpha] Add handling of FP control insn in software-single step

[Richard Henderson <rth@redhat.com>]

On Tue, May 17, 2005 at 12:34:00PM +1000, Joel Brobecker wrote:
> +fp_register_zero_p (char *buf)
> +{
> +  return ((buf[1] & 0x0f) == 0 && buf[2] == 0 && buf[3] == 0
> +          && buf[4] == 0 && buf[5] == 0 && buf[6] == 0 && buf[7] == 0);
> +}
...
> +        case 0x31:              /* FBEQ */
> +          if (fp_register_zero_p (reg))
> +            goto branch_taken;
> +          break;
> +        case 0x36:              /* FBGE */
> +          if (fp_register_sign_bit (reg) == 0 || fp_register_zero_p (reg))
> +            goto branch_taken;
> +          break;
> +        case 0x37:              /* FBGT */
> +          if (fp_register_sign_bit (reg) == 0 && ! fp_register_zero_p (reg))
> +            goto branch_taken;
> +          break;
> +        case 0x33:              /* FBLE */
> +          if (fp_register_sign_bit (reg) == 1 || fp_register_zero_p (reg))
> +            goto branch_taken;
> +          break;
> +        case 0x32:              /* FBLT */
> +          if (fp_register_sign_bit (reg) == 1 && ! fp_register_zero_p (reg))
> +            goto branch_taken;
> +          break;
> +        case 0x35:              /* FBNE */
> +          if (! fp_register_zero_p (reg))
> +            goto branch_taken;
> +          break;

NACK.  Why are you ignoring the entire exponent?  These are not the
operations that fp branches perform.  The tests are:

  zero		(reg & 0x7fff_ffff_ffff_ffff) == 0
  sign		(reg & 0x8000_0000_0000_0000) != 0

  fbeq		zero
  fbne		!zero
  fbge		!sign || zero
  fbgt		!sign && !zero
  fble		sign || zero
  fblt		sign && !zero

This is a nearly normal integer test, except for the special case
of 0x8000_0000_0000_0000, i.e. -0.0, which is treated as zero.


r~