Re: [PATCH] improve float to double on PPC a little

[Dale Johannesen <dalej at apple dot com>]

On May 27, 2004, at 5:01 PM, Andrew Pinski wrote:

> Since on rs6000, the floating point registers are always in double 
> precision,
> the need for conversion is not needed at all.  This improves the 
> conversion by
> make the cost for the case where the registers are different.

It would be good to teach hard register copyprop that SF->DF 
conversions are
really copies, too; they can be removed freely.   I've had the 
following patch
in Apple's branch for some time, but don't know how to make it 
FSF-compliant.
(Might need adjustment now that we have 128-bit long doubles, haven't 
exercised it.)

--- ./regrename.c       Tue Mar 30 19:53:12 2004
*************** static bool
*** 1268,1273 ****
--- 1267,1280 ----
  mode_change_ok (enum machine_mode orig_mode, enum machine_mode 
new_mode,
                unsigned int regno ATTRIBUTE_UNUSED)
  {
+   /* APPLE LOCAL begin add mode change case */
+ #ifdef TARGET_POWERPC
+   /* This arises from FLOAT_EXTEND which is really a NOP.  */
+   if (orig_mode == SFmode && new_mode == DFmode)
+     return true;
+ #endif
+   /* APPLE LOCAL end add mode change case */
+
    if (GET_MODE_SIZE (orig_mode) < GET_MODE_SIZE (new_mode))
      return false;

*************** copyprop_hardreg_forward_1 (basic_block
*** 1713,1718 ****
--- 1720,1734 ----
        /* Notice copies.  */
        if (set && REG_P (SET_DEST (set)) && REG_P (SET_SRC (set)))
        copy_value (SET_DEST (set), SET_SRC (set), vd);
+       /* APPLE LOCAL begin record that float extend is a copy */
+ #ifdef TARGET_POWERPC
+       /* FLOAT_EXTEND is actually a copy; record that too.  */
+       if (set && REG_P (SET_DEST (set))
+         && GET_CODE (SET_SRC (set)) == FLOAT_EXTEND
+         && REG_P (XEXP (SET_SRC (set), 0)))
+         copy_value (SET_DEST (set), XEXP (SET_SRC (set), 0), vd);
+ #endif
+       /* APPLE LOCAL end record that float extend is a copy */

        if (insn == BB_END (bb))
        break;