[Bug tree-optimization/122755] New: Missed optimization opportunity for loop-split

[dizhao at os dot amperecomputing.com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122755

            Bug ID: 122755
           Summary: Missed optimization opportunity for loop-split
           Product: gcc
           Version: 16.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dizhao at os dot amperecomputing.com
  Target Milestone: ---

This tiny code is simplified from gcc/tree-ssa-sccvn.cc. After
inling dominated_by_p into can_track_predicate_on_edge, the
for-loop can be optimized by splitting it upon conditional
statement "if (dom_computed[dir_index] == 2)". Because once
the expression is TRUE, the loop won't change its value, so
it is (at least) a semi-invariant for the second kind of 
transformation in pass lsplit. (Compile options is "-O3".)


#include <vector>

struct et_node
{
  void *data;
  int dfs_num_in, dfs_num_out;
};

struct basic_block {
  struct et_node * dom[2];
  int index;
};

class edge
{
public:
  basic_block* src;
  basic_block* dest;
  int flags;
};

extern bool
et_below (struct et_node *down, struct et_node *up);

int dom_computed[2];

bool
dominated_by_p (unsigned int dir_index, basic_block* bb1, basic_block* bb2)
{
  struct et_node *n1 = bb1->dom[dir_index], *n2 = bb2->dom[dir_index];

  if (dom_computed[dir_index] == 2)
    return (n1->dfs_num_in >= n2->dfs_num_in
            && n1->dfs_num_out <= n2->dfs_num_out);

  return et_below (n1, n2);
}

bool
can_track_predicate_on_edge (std::vector<edge *> preds)
{
  for (unsigned i = 0; i < preds.size (); ++i)
    {
      edge *e = preds[i];
      if (!dominated_by_p (0, e->src, e->dest))
        return false;
    }

  return true;
}


Currently the check in tree-ssa-loop-split.cc:control_dep_semi_invariant_p
fails around "i < preds.size ()", because it is not semi-invariant.
Since its false edge is loop-exit and won't change the value of 
"(dom_computed[dir_index] == 2", I think we can refine the checks in
control_dep_semi_invariant_p to optimize such cases.