[Bug tree-optimization/77283] New: Revision 238005 disables loop unrolling

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77283

            Bug ID: 77283
           Summary: Revision 238005 disables loop unrolling
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: bergner at gcc dot gnu.org
  Target Milestone: ---

Revision 238005 (richi's fix to path splitting to handle empty else blocks)
disables loop unrolling for the test case below (extracted from one of our
benchmarks) leading to a performance regression.

bergner@genoa:~/gcc/BUGS/$ cat foo.c 
void
foo (double *x, double *a, double *b, long n, double limit)
{
  long i;
  for (i=0; i < n; i++)
    if (a[i] < limit)
      x[i] = b[i];
}
bergner@genoa:~/gcc/BUGS/$
/home/bergner/gcc/build/gcc-fsf-mainline-r238005/gcc/xgcc
-B/home/bergner/gcc/build/gcc-fsf-mainline-r238005/gcc -O3 -funroll-loops -S
foo.c 
bergner@genoa:~/gcc/BUGS/$ cat foo.s 

foo:
        cmpdi 0,6,0
        ble 0,.L1
        sldi 6,6,3
        li 9,0
        .p2align 4,,15
.L3:
        lfdx 0,4,9
        fcmpu 7,0,1
        bnl 7,.L8
        lfdx 2,5,9
        stfdx 2,3,9
        addi 9,9,8
        cmpld 5,9,6
        bne 5,.L3
.L1:
        blr
        .p2align 4,,15
.L8:
        addi 9,9,8
        cmpld 1,9,6
        bne 1,.L3
        blr


bergner@genoa:~/gcc/BUGS/$
/home/bergner/gcc/build/gcc-fsf-mainline-r238004/gcc/xgcc
-B/home/bergner/gcc/build/gcc-fsf-mainline-r238004/gcc -O3 -funroll-loops -S
foo.c 
bergner@genoa:~/gcc/BUGS/$ cat foo.s

foo:
        cmpdi 0,6,0
        ble 0,.L1
        lfd 0,0(4)
        sldi 6,6,3
        addi 8,6,-8
        srdi 0,8,3
        rldicl 10,0,0,61
        fcmpu 7,0,1
        blt 7,.L8
.L59:
        li 9,8
        cmpld 1,9,6
        beq 1,.L1
        cmpdi 5,10,0
        beq 5,.L30
        cmpdi 6,10,1
        beq 6,.L46
        cmpdi 0,10,2
        beq 0,.L47
        cmpdi 7,10,3
        beq 7,.L48
        cmpdi 1,10,4
        beq 1,.L49
        cmpdi 5,10,5
        beq 5,.L50
        cmpdi 6,10,6
        beq 6,.L51
        lfdx 3,4,9
        fcmpu 0,3,1
        bnl 0,.L61
        lfdx 4,5,9
        stfdx 4,3,9
.L61:
        addi 9,9,8
.L51:
        lfdx 5,4,9
        fcmpu 7,5,1
        bnl 7,.L62
        lfdx 6,5,9
        stfdx 6,3,9
.L62:
        addi 9,9,8
.L50:
        lfdx 7,4,9
        fcmpu 1,7,1
        bnl 1,.L63
        lfdx 8,5,9
        stfdx 8,3,9
[snip]


An executable version of the test case above is:

bergner@genoa:~/gcc/BUGS/LTC144447$ cat loop.c 
#define SIZE 1024*1000

void
__attribute__ ((noinline))
foo (double *x, double *a, double *b, long n, double limit)
{
  long i;
  for (i=0; i < n; i++)
    if (a[i] < limit)
      x[i] = b[i];
}

double x[SIZE], a[SIZE], b[SIZE];

int
main (void)
{
  long i;
  for (i=0; i < 3000; i++)
    foo (x, a, b, SIZE, 1.0);
  return 0;
}

bergner@genoa:~/gcc/BUGS/$
/home/bergner/gcc/build/gcc-fsf-mainline-r238004/gcc/xgcc
-B/home/bergner/gcc/build/gcc-fsf-mainline-r238004/gcc -O3 -funroll-loops
loop.c 
bergner@genoa:~/gcc/BUGS/$ time ./a.out 

real    0m3.729s
user    0m3.690s
sys     0m0.021s

bergner@genoa:~/gcc/BUGS/$
/home/bergner/gcc/build/gcc-fsf-mainline-r238005/gcc/xgcc
-B/home/bergner/gcc/build/gcc-fsf-mainline-r238005/gcc -O3 -funroll-loops
loop.c 
bergner@genoa:~/gcc/BUGS/$ time ./a.out 

real    0m6.939s
user    0m6.851s
sys     0m0.040s