[Bug rtl-optimization/37490] New: [4.4 Regression] IRA causes FP code to have more spills

["pinskia at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

I had quickly looked at PR 37488 and decided to try it out on PPC and noticed
that IRA generates worse code than what was being generated before.
Here is a simple testcase which does not have any indirect gotos in it:
void g(void);
const void* test(int action, void* ptr, int a, int *b)
{
   if (action == 0) {
      return 0;
   } else {
      char* ip = ptr;
      double reg_f_a;
      double reg_f[1];
      reg_f_a = 0.0;
      reg_f[0] = 0.0;
      if(a) goto L2;
L1: {
      int t1 = *(int*)(++ip);
      reg_f_a = reg_f_a + reg_f[t1];
      g();
      if(*(++b)) goto L1;
   }
L2:
      g();
      *(double*)ptr = reg_f_a;
   }
   return 0;
}

--- CUT ---
-       fmr 31,0
+       std 0,128(1)
+       lfd 31,128(1)

-       fmr 31,0
-       mr 31,4
-       addi 29,1,112
+       std 0,128(1)
+       mr 29,4
+       lfd 31,128(1)
+       addi 28,1,112

For some reason IRA decides it is better to use the GPRs for 0.0 than it is to
use the FPRs.


-- 
           Summary: [4.4 Regression] IRA causes FP code to have more spills
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: pinskia at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37490