Re: Ping: [patch] Fix PR22372

[Richard Henderson <rth at redhat dot com>]

On Wed, Aug 10, 2005 at 03:05:19PM +0300, Ira Rosen wrote:
> http://gcc.gnu.org/ml/gcc-patches/2005-07/msg01877.html

Not ok.

First, 

!   if (TREE_TYPE (vec_oprnd1) != vectype)

Never use direct pointer comparison.  Always use the 
lang_hook.type_compatible_p function.

Second, 

!                         build1 (NOP_EXPR, vectype, vec_oprnd1));

Never build NOP_EXPRs yourself.  Use fold_convert.

Third, none of this would have been necessary if you'd had
formed the address correctly in the first place.  From the
test case, the current source creates

  char * vect_pib.35;
  vector char * vect_pib.32;
  vect_pib.35_45 = &ib;
  vect_pib.32_51 = (vector char *) vect_pib.35_45;

your patched source creates

  char[16] * vect_pib.35;
  vector char * vect_pib.32;
  vect_pib.35_45 = &ib;
  vect_pib.32_51 = (vector char *) vect_pib.35_45;

and an ideal result would be either

  vector char * vect_pib.32;
  vect_pib.32_51 = (vector char *) &ib;
or
  vect_pib.32_51 = (vector char *) &ib[0];

eliminating the vect_pib.35 variable entirely.

If this is too difficult, I'd prefer emitting 

  char * vect_pib.35;
  vector char * vect_pib.32;
  vect_pib.35_45 = &ib[0];
  vect_pib.32_51 = (vector char *) vect_pib.35_45;

over your current patch.


r~