https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79663
--- Comment #1 from amker at gcc dot gnu.org ---
Root cause understood. After patch, combine_chains looks like:
/* Process in reverse order so dominance point is ready when it comes
to the root ref. */
for (i = ch1->refs.length (); i > 0; i--)
{
r1 = ch1->refs[i - 1];
r2 = ch2->refs[i - 1];
nw = XCNEW (struct dref_d);
nw->distance = r1->distance;
nw->stmt = stmt_combining_refs (r1, r2, i == 1 ? insert : NULL);
/* Record dominance point where root combined stmt should be inserted
for chains with 0 length. Though all root refs dominate following
refs, it's possible the combined stmt doesn't. See PR70754. */
if (ch1->length == 0
&& (insert == NULL || stmt_dominates_stmt_p (nw->stmt, insert)))
insert = nw->stmt;
tmp_refs.safe_push (nw);
}
//...
root_stmt = get_chain_root (new_chain)->stmt;
for (i = 1; new_chain->refs.iterate (i, &nw); i++)
{
if (nw->distance == new_chain->length
&& !stmt_dominates_stmt_p (nw->stmt, root_stmt))
{
new_chain->has_max_use_after = true;
break;
}
}
Note the original PR70754 only happens for ZERO length chains, so I skipped
non-ZERO chains in the first for loop. Problem with non-ZERO chains is
insertion point for refs of the chain could be the same statement, so
originally combined statements of the chain are inserted as:
combined_stmt_for_ref0
combined_stmt_for_ref1
...
combined_stmt_for_refn
insert_point
But now the first for-loop visits all refs in reversing order, the inserted
statements are like:
combined_stmt_for_refn
...
combined_stmt_for_ref1
combined_stmt_for_ref0
insert_point
Thus in the second for-loop, condition "!stmt_dominates_stmt_p" evaluates to
true, we set new_chain->has_max_use_after to true in this case.
In determine_unroll_factor, we have:
/* The best unroll factor for this chain is equal to the number of
temporary variables that we create for it. */
af = chain->length;
if (chain->has_max_use_after)
af++;
Causing unnecessary unrolling and the regression.
There are two possible fix.
1) Separate fix of PR70754 for ZERO/non-ZERO chains. kind of duplicate the
first loop for the two cases.
2) Simply remove (ch1->length == 0) condition in the first loop, thus
root_stmt will be inserted in dominant point. I am testing this one.