[Bug target/35045] [4.3 Regression] gcc-4.3 generates wrong code on i386 with -O3

["rguenth at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #16 from rguenth at gcc dot gnu dot org  2008-02-01 14:08 -------
The problem is that we retain

(insn 38 32 61 4 test-cacoshf.c:16 (clobber (reg/i:SC 0 ax)) -1 (nil))

after postreload and gcse rightfully assumes such clobbers are removed
(and hoists over them).  They are supposed to be removed by reload1.c:reload()
which does

  /* Make a pass over all the insns and delete all USEs which we inserted
     only to tag a REG_EQUAL note on them.  Remove all REG_DEAD and REG_UNUSED
     notes.  Delete all CLOBBER insns, except those that refer to the return
     value and the special mem:BLK CLOBBERs added to prevent the scheduler
     from misarranging variable-array code, and simplify (subreg (reg))
     operands.  Also remove all REG_RETVAL and REG_LIBCALL notes since they
     are no longer useful or accurate.  Strip and regenerate REG_INC notes
     that may have been moved around.  */

...
                && (!REG_P (XEXP (PATTERN (insn), 0))
                    || ! REG_FUNCTION_VALUE_P (XEXP (PATTERN (insn), 0)))))
          {
            delete_insn (insn);
            continue;
          }

where the exception for return value clobbers was added by rth with

r30401 | rth | 1999-11-05 01:35:10 +0100 (Fri, 05 Nov 1999) | 9 lines

        * function.c (diddle_return_value): New.
        (expand_function_end): Use it.
        * stmt.c (expand_null_return): Likewise.
        (expand_value_return): Likewise.

        * reg-stack.c (subst_stack_regs_pat): Handle clobbers at top-level.

        * reload1.c (reload): Don't remove return value clobbers.

seemingly as an optimization(?) with stmt.c:expand_null_return():

void
expand_null_return (void)
{
  /* If this function was declared to return a value, but we
     didn't, clobber the return registers so that they are not
     propagated live to the rest of the function.  */
  clobber_return_register ();


which doesn't make sense to me (_which_ rest of the function?).  Maybe
this refers to missed optimizations with the RTL inliner?

Removing this restriction from reload fixes this problem.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35045