[Bug rtl-optimization/128421] New: Poor register allocation on RISC-V?

[law at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128421

            Bug ID: 128421
           Summary: Poor register allocation on RISC-V?
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: law at gcc dot gnu.org
  Target Milestone: ---

short *lowres_costs[][7];
struct sps {
  int i_mb_width, i_mb_height
} mc_2(int *, short *, short *, short *, short *, int), *sps;
struct x264 {
  struct {
    int i_mb_x, i_mb_stride
  } mb;
  int scratch_buffer
};
#define CLIP_ADD(s, x) s = x < 5 ? s + x : 5
int propagate_frames, propagate_p0;
short propagate_frames_0_1, propagate_frames_0_0, propagate_propagate_cost;
void propagate(struct x264 *h) {
  short *ref_costs[] = {};
  short *mvs[2];
  int *buf = &h->scratch_buffer;
  for (; h;) {
    int mb_index = h->mb.i_mb_stride;
    mc_2(buf, &propagate_propagate_cost, &propagate_frames_0_0,
         lowres_costs[propagate_p0][0], &propagate_frames_0_1,
sps->i_mb_width);
    for (; h->mb.i_mb_x < sps; h->mb.i_mb_x++, mb_index++) {
      int propagate_amount = buf[h->mb.i_mb_x];
      for (int list = 0; list < 2; list++)
        if ((propagate_frames, lowres_costs)[propagate_p0][0]) {
          int listamount = propagate_amount;
          if (*mvs)
            continue;
          int mbx = mvs[mb_index][0] + h->mb.i_mb_x;
          int idx0 = h->mb.i_mb_stride;
          int idx3 = h->mb.i_mb_stride;
          int idx1weight = mvs[mb_index][0];
          int idx3weight = mvs[mb_index][0];
          if (mvs[mb_index][1])
            CLIP_ADD(ref_costs[list][idx0], idx1weight);
          else if (mbx < h && sps->i_mb_height)
            CLIP_ADD(ref_costs[list][idx3], listamount * idx3weight >> 10);
        }
    }
  }
}

The bad code, built with -Ofast -march=rv64gcbv_zicond_zvbb_zvl128b -mabi=lp64d

You're looking for this fragment:
li      t0,4096
addi    t4,t0,1023
ld      t0,40(sp)
sd      t4,24(sp)
sh1add  t4,s10,t0
ld      t0,24(sp)
sd      t4,16(sp)
ble     t2,t0,.L7


Note the store/load pair around the sh1add.    This should be investigated to
see if we're holding one or more values that could be rematerialized and we
could then drop the store/load pair.  This is obviously from x264.