static void
update_phis_for_loop_copy (struct loop *orig_loop, struct loop *new_loop)
{
tree new_ssa_name;
gimple_stmt_iterator si_new, si_orig;
edge orig_loop_latch = loop_latch_edge (orig_loop);
edge orig_entry_e = loop_preheader_edge (orig_loop);
edge new_loop_entry_e = loop_preheader_edge (new_loop);
/* Scan the phis in the headers of the old and new loops
(they are organized in exactly the same order). */
for (si_new = gsi_start_phis (new_loop->header),
si_orig = gsi_start_phis (orig_loop->header);
!gsi_end_p (si_new) && !gsi_end_p (si_orig);
gsi_next (&si_new), gsi_next (&si_orig))
{
tree def;
source_location locus;
gimple phi_new = gsi_stmt (si_new);
gimple phi_orig = gsi_stmt (si_orig);
/* Add the first phi argument for the phi in NEW_LOOP (the one
associated with the entry of NEW_LOOP) */
def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_entry_e);
locus = gimple_phi_arg_location_from_edge (phi_orig, orig_entry_e);
add_phi_arg (phi_new, def, new_loop_entry_e, locus);
/* Add the second phi argument for the phi in NEW_LOOP (the one
associated with the latch of NEW_LOOP) */
def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_loop_latch);
locus = gimple_phi_arg_location_from_edge (phi_orig, orig_loop_latch);
if (TREE_CODE (def) == SSA_NAME)
{
new_ssa_name = get_current_def (def);
if (!new_ssa_name)
/* This only happens if there are no definitions inside the
loop. Use the phi_result in this case. */
new_ssa_name = PHI_RESULT (phi_new);
}
else
/* Could be an integer. */
new_ssa_name = def;
add_phi_arg (phi_new, new_ssa_name, loop_latch_edge (new_loop), locus);
}
}
static void
unloop_loops (bitmap loop_closed_ssa_invalidated,
bool *irred_invalidated)
{
while (loops_to_unloop.length ())
{
struct loop *loop = loops_to_unloop.pop ();
int n_unroll = loops_to_unloop_nunroll.pop ();
basic_block latch = loop->latch;
edge latch_edge = loop_latch_edge (loop);
int flags = latch_edge->flags;
location_t locus = latch_edge->goto_locus;
gcall *stmt;
gimple_stmt_iterator gsi;
remove_exits_and_undefined_stmts (loop, n_unroll);
/* Unloop destroys the latch edge. */
unloop (loop, irred_invalidated, loop_closed_ssa_invalidated);
/* Create new basic block for the latch edge destination and wire
it in. */
stmt = gimple_build_call (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0);
latch_edge = make_edge (latch, create_basic_block (NULL, NULL, latch), flags);
latch_edge->probability = 0;
latch_edge->count = 0;
latch_edge->flags |= flags;
latch_edge->goto_locus = locus;
latch_edge->dest->loop_father = current_loops->tree_root;
latch_edge->dest->count = 0;
latch_edge->dest->frequency = 0;
set_immediate_dominator (CDI_DOMINATORS, latch_edge->dest, latch_edge->src);
gsi = gsi_start_bb (latch_edge->dest);
gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
}
loops_to_unloop.release ();
loops_to_unloop_nunroll.release ();
/* Remove edges in peeled copies. */
unsigned i;
edge e;
FOR_EACH_VEC_ELT (edges_to_remove, i, e)
{
bool ok = remove_path (e);
gcc_assert (ok);
}
/* Checks that LOOPS are all right:
-- sizes of loops are all right
-- results of get_loop_body really belong to the loop
-- loop header have just single entry edge and single latch edge
-- loop latches have only single successor that is header of their loop
-- irreducible loops are correctly marked
*/
void
verify_loop_structure (struct loops *loops)
{
unsigned *sizes, i, j;
sbitmap irreds;
basic_block *bbs, bb;
struct loop *loop;
int err = 0;
edge e;
/* Check sizes. */
sizes = xcalloc (loops->num, sizeof (int));
sizes[0] = 2;
FOR_EACH_BB (bb)
for (loop = bb->loop_father; loop; loop = loop->outer)
sizes[loop->num]++;
for (i = 0; i < loops->num; i++)
{
if (!loops->parray[i])
continue;
if (loops->parray[i]->num_nodes != sizes[i])
{
error ("Size of loop %d should be %d, not %d.",
i, sizes[i], loops->parray[i]->num_nodes);
err = 1;
}
}
free (sizes);
/* Check get_loop_body. */
for (i = 1; i < loops->num; i++)
{
loop = loops->parray[i];
if (!loop)
continue;
bbs = get_loop_body (loop);
for (j = 0; j < loop->num_nodes; j++)
if (!flow_bb_inside_loop_p (loop, bbs[j]))
{
error ("Bb %d do not belong to loop %d.",
bbs[j]->index, i);
err = 1;
}
free (bbs);
}
/* Check headers and latches. */
for (i = 1; i < loops->num; i++)
{
loop = loops->parray[i];
if (!loop)
continue;
if ((loops->state & LOOPS_HAVE_PREHEADERS)
&& (!loop->header->pred->pred_next
|| loop->header->pred->pred_next->pred_next))
{
error ("Loop %d's header does not have exactly 2 entries.", i);
err = 1;
}
if (loops->state & LOOPS_HAVE_SIMPLE_LATCHES)
{
if (!loop->latch->succ
|| loop->latch->succ->succ_next)
{
error ("Loop %d's latch does not have exactly 1 successor.", i);
err = 1;
}
if (loop->latch->succ->dest != loop->header)
{
error ("Loop %d's latch does not have header as successor.", i);
err = 1;
}
if (loop->latch->loop_father != loop)
{
error ("Loop %d's latch does not belong directly to it.", i);
err = 1;
}
}
if (loop->header->loop_father != loop)
{
error ("Loop %d's header does not belong directly to it.", i);
err = 1;
}
if ((loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
&& (loop_latch_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP))
{
error ("Loop %d's latch is marked as part of irreducible region.", i);
err = 1;
}
}
/* Check irreducible loops. */
if (loops->state & LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
{
/* Record old info. */
irreds = sbitmap_alloc (last_basic_block);
FOR_EACH_BB (bb)
{
if (bb->flags & BB_IRREDUCIBLE_LOOP)
SET_BIT (irreds, bb->index);
else
RESET_BIT (irreds, bb->index);
for (e = bb->succ; e; e = e->succ_next)
if (e->flags & EDGE_IRREDUCIBLE_LOOP)
e->flags |= EDGE_ALL_FLAGS + 1;
}
/* Recount it. */
mark_irreducible_loops (loops);
/* Compare. */
FOR_EACH_BB (bb)
{
if ((bb->flags & BB_IRREDUCIBLE_LOOP)
&& !TEST_BIT (irreds, bb->index))
{
error ("Basic block %d should be marked irreducible.", bb->index);
err = 1;
}
else if (!(bb->flags & BB_IRREDUCIBLE_LOOP)
&& TEST_BIT (irreds, bb->index))
{
error ("Basic block %d should not be marked irreducible.", bb->index);
err = 1;
}
for (e = bb->succ; e; e = e->succ_next)
{
if ((e->flags & EDGE_IRREDUCIBLE_LOOP)
&& !(e->flags & (EDGE_ALL_FLAGS + 1)))
{
error ("Edge from %d to %d should be marked irreducible.",
e->src->index, e->dest->index);
err = 1;
}
else if (!(e->flags & EDGE_IRREDUCIBLE_LOOP)
&& (e->flags & (EDGE_ALL_FLAGS + 1)))
{
error ("Edge from %d to %d should not be marked irreducible.",
e->src->index, e->dest->index);
err = 1;
}
e->flags &= ~(EDGE_ALL_FLAGS + 1);
}
}
free (irreds);
}
static void
copy_loop_headers (void)
{
struct loops *loops;
unsigned i;
struct loop *loop;
basic_block header;
edge exit;
basic_block *bbs;
unsigned n_bbs;
loops = loop_optimizer_init (dump_file);
if (!loops)
return;
rewrite_into_loop_closed_ssa ();
/* We do not try to keep the information about irreducible regions
up-to-date. */
loops->state &= ~LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS;
#ifdef ENABLE_CHECKING
verify_loop_structure (loops);
#endif
bbs = xmalloc (sizeof (basic_block) * n_basic_blocks);
for (i = 1; i < loops->num; i++)
{
/* Copy at most 20 insns. */
int limit = 20;
loop = loops->parray[i];
if (!loop)
continue;
header = loop->header;
/* If the loop is already a do-while style one (either because it was
written as such, or because jump threading transformed it into one),
we might be in fact peeling the first iteration of the loop. This
in general is not a good idea. */
if (do_while_loop_p (loop))
continue;
/* Iterate the header copying up to limit; this takes care of the cases
like while (a && b) {...}, where we want to have both of the conditions
copied. TODO -- handle while (a || b) - like cases, by not requiring
the header to have just a single successor and copying up to
postdominator. */
exit = NULL;
n_bbs = 0;
while (should_duplicate_loop_header_p (header, loop, &limit))
{
/* Find a successor of header that is inside a loop; i.e. the new
header after the condition is copied. */
if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest))
exit = EDGE_SUCC (header, 0);
else
exit = EDGE_SUCC (header, 1);
bbs[n_bbs++] = header;
header = exit->dest;
}
if (!exit)
continue;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"Duplicating header of the loop %d up to edge %d->%d.\n",
loop->num, exit->src->index, exit->dest->index);
/* Ensure that the header will have just the latch as a predecessor
inside the loop. */
if (EDGE_COUNT (exit->dest->preds) > 1)
exit = EDGE_SUCC (loop_split_edge_with (exit, NULL), 0);
if (!tree_duplicate_sese_region (loop_preheader_edge (loop), exit,
bbs, n_bbs, NULL))
{
fprintf (dump_file, "Duplication failed.\n");
continue;
}
/* Ensure that the latch and the preheader is simple (we know that they
are not now, since there was the loop exit condition. */
loop_split_edge_with (loop_preheader_edge (loop), NULL);
loop_split_edge_with (loop_latch_edge (loop), NULL);
}
free (bbs);
#ifdef ENABLE_CHECKING
verify_loop_closed_ssa ();
#endif
loop_optimizer_finalize (loops, NULL);
}
static unsigned int
copy_loop_headers (void)
{
struct loops *loops;
unsigned i;
struct loop *loop;
basic_block header;
edge exit, entry;
basic_block *bbs, *copied_bbs;
unsigned n_bbs;
unsigned bbs_size;
loops = loop_optimizer_init (LOOPS_HAVE_PREHEADERS
| LOOPS_HAVE_SIMPLE_LATCHES);
if (!loops)
return 0;
#ifdef ENABLE_CHECKING
verify_loop_structure (loops);
#endif
bbs = XNEWVEC (basic_block, n_basic_blocks);
copied_bbs = XNEWVEC (basic_block, n_basic_blocks);
bbs_size = n_basic_blocks;
for (i = 1; i < loops->num; i++)
{
/* Copy at most 20 insns. */
int limit = 20;
loop = loops->parray[i];
if (!loop)
continue;
header = loop->header;
/* If the loop is already a do-while style one (either because it was
written as such, or because jump threading transformed it into one),
we might be in fact peeling the first iteration of the loop. This
in general is not a good idea. */
if (do_while_loop_p (loop))
continue;
/* Iterate the header copying up to limit; this takes care of the cases
like while (a && b) {...}, where we want to have both of the conditions
copied. TODO -- handle while (a || b) - like cases, by not requiring
the header to have just a single successor and copying up to
postdominator. */
exit = NULL;
n_bbs = 0;
while (should_duplicate_loop_header_p (header, loop, &limit))
{
/* Find a successor of header that is inside a loop; i.e. the new
header after the condition is copied. */
if (flow_bb_inside_loop_p (loop, EDGE_SUCC (header, 0)->dest))
exit = EDGE_SUCC (header, 0);
else
exit = EDGE_SUCC (header, 1);
bbs[n_bbs++] = header;
gcc_assert (bbs_size > n_bbs);
header = exit->dest;
}
if (!exit)
continue;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"Duplicating header of the loop %d up to edge %d->%d.\n",
loop->num, exit->src->index, exit->dest->index);
/* Ensure that the header will have just the latch as a predecessor
inside the loop. */
if (!single_pred_p (exit->dest))
exit = single_pred_edge (loop_split_edge_with (exit, NULL));
entry = loop_preheader_edge (loop);
if (!tree_duplicate_sese_region (entry, exit, bbs, n_bbs, copied_bbs))
{
fprintf (dump_file, "Duplication failed.\n");
continue;
}
/* If the loop has the form "for (i = j; i < j + 10; i++)" then
this copying can introduce a case where we rely on undefined
signed overflow to eliminate the preheader condition, because
we assume that "j < j + 10" is true. We don't want to warn
about that case for -Wstrict-overflow, because in general we
don't warn about overflow involving loops. Prevent the
warning by setting TREE_NO_WARNING. */
if (warn_strict_overflow > 0)
{
unsigned int i;
for (i = 0; i < n_bbs; ++i)
{
tree last;
last = last_stmt (copied_bbs[i]);
if (TREE_CODE (last) == COND_EXPR)
TREE_NO_WARNING (last) = 1;
}
}
/* Ensure that the latch and the preheader is simple (we know that they
are not now, since there was the loop exit condition. */
loop_split_edge_with (loop_preheader_edge (loop), NULL);
loop_split_edge_with (loop_latch_edge (loop), NULL);
}
free (bbs);
free (copied_bbs);
loop_optimizer_finalize (loops);
return 0;
}
