[Bug rtl-optimization/58033] counterproductive bb-reorder

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58033

--- Comment #6 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Another example on SH:

const char* test (const char* s0, int c, int* rout)
{
  int r = 0;

  for (int i = 0; i < c; ++i)
    r += s0[i];

  *rout = r;
  return s0;
}


compiled with -O2:
_test:
        cmp/pl  r5
        bf      .L4
        mov     #0,r1
        mov     r4,r2
        .align 2
.L3:
        mov.b   @r2+,r3
        dt      r5
        bf/s    .L3
        add     r3,r1
.L2:
        mov.l   r1,@r6
        rts
        mov     r4,r0
    .align 1
.L4:
        bra     .L2
        mov     #0,r1


compiled with -O2 -fno-reorder-blocks:

_test:
        cmp/pl  r5
        bf/s    .L4
        mov     #0,r1
        mov     r4,r2
    .align 2
.L3:
        mov.b   @r2+,r3
        dt      r5
        bf/s    .L3
        add     r3,r1
        bra     .L7
        mov.l   r1,@r6
        .align 1
.L4:
        mov.l   r1,@r6
.L7:
        rts
        mov     r4,r0

.. which is better, except for the redundant stores.  Folding the two stores
gives the minimal code:
_test:
        cmp/pl  r5
        bf/s    .L4
        mov     #0,r1
        mov     r4,r2
    .align 2
.L3:
        mov.b   @r2+,r3
        dt      r5
        bf/s    .L3
        add     r3,r1
.L4:
        mov.l   r1,@r6
        rts
        mov     r4,r0