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;
}
static void
doloop_modify (struct loop *loop, struct niter_desc *desc,
rtx doloop_seq, rtx condition, rtx count)
{
rtx counter_reg;
rtx tmp, noloop = NULL_RTX;
rtx sequence;
rtx jump_insn;
rtx jump_label;
int nonneg = 0, irr;
bool increment_count;
basic_block loop_end = desc->out_edge->src;
enum machine_mode mode;
jump_insn = BB_END (loop_end);
if (dump_file)
{
fprintf (dump_file, "Doloop: Inserting doloop pattern (");
if (desc->const_iter)
fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, desc->niter);
else
fputs ("runtime", dump_file);
fputs (" iterations).\n", dump_file);
}
/* Discard original jump to continue loop. The original compare
result may still be live, so it cannot be discarded explicitly. */
delete_insn (jump_insn);
counter_reg = XEXP (condition, 0);
if (GET_CODE (counter_reg) == PLUS)
counter_reg = XEXP (counter_reg, 0);
mode = GET_MODE (counter_reg);
increment_count = false;
switch (GET_CODE (condition))
{
case NE:
/* Currently only NE tests against zero and one are supported. */
if (XEXP (condition, 1) == const1_rtx)
{
increment_count = true;
noloop = const1_rtx;
}
else if (XEXP (condition, 1) == const0_rtx)
noloop = const0_rtx;
else
abort ();
break;
case GE:
/* Currently only GE tests against zero are supported. */
if (XEXP (condition, 1) != const0_rtx)
abort ();
noloop = constm1_rtx;
/* The iteration count does not need incrementing for a GE test. */
increment_count = false;
/* Determine if the iteration counter will be non-negative.
Note that the maximum value loaded is iterations_max - 1. */
if (desc->niter_max
<= ((unsigned HOST_WIDEST_INT) 1
<< (GET_MODE_BITSIZE (mode) - 1)))
nonneg = 1;
break;
/* Abort if an invalid doloop pattern has been generated. */
default:
abort ();
}
if (increment_count)
count = simplify_gen_binary (PLUS, mode, count, const1_rtx);
/* Insert initialization of the count register into the loop header. */
start_sequence ();
tmp = force_operand (count, counter_reg);
convert_move (counter_reg, tmp, 1);
sequence = get_insns ();
end_sequence ();
emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src));
if (desc->noloop_assumptions)
{
rtx ass = copy_rtx (desc->noloop_assumptions);
basic_block preheader = loop_preheader_edge (loop)->src;
basic_block set_zero
= loop_split_edge_with (loop_preheader_edge (loop), NULL_RTX);
basic_block new_preheader
= loop_split_edge_with (loop_preheader_edge (loop), NULL_RTX);
basic_block bb;
edge te;
gcov_type cnt;
/* Expand the condition testing the assumptions and if it does not pass,
reset the count register to 0. */
add_test (XEXP (ass, 0), preheader, set_zero);
EDGE_SUCC (preheader, 0)->flags &= ~EDGE_FALLTHRU;
cnt = EDGE_SUCC (preheader, 0)->count;
EDGE_SUCC (preheader, 0)->probability = 0;
EDGE_SUCC (preheader, 0)->count = 0;
irr = EDGE_SUCC (preheader, 0)->flags & EDGE_IRREDUCIBLE_LOOP;
te = make_edge (preheader, new_preheader, EDGE_FALLTHRU | irr);
te->probability = REG_BR_PROB_BASE;
te->count = cnt;
set_immediate_dominator (CDI_DOMINATORS, new_preheader, preheader);
set_zero->count = 0;
set_zero->frequency = 0;
for (ass = XEXP (ass, 1); ass; ass = XEXP (ass, 1))
{
bb = loop_split_edge_with (te, NULL_RTX);
te = EDGE_SUCC (bb, 0);
add_test (XEXP (ass, 0), bb, set_zero);
make_edge (bb, set_zero, irr);
}
start_sequence ();
convert_move (counter_reg, noloop, 0);
sequence = get_insns ();
end_sequence ();
emit_insn_after (sequence, BB_END (set_zero));
}
/* Some targets (eg, C4x) need to initialize special looping
registers. */
#ifdef HAVE_doloop_begin
{
rtx init;
unsigned level = get_loop_level (loop) + 1;
init = gen_doloop_begin (counter_reg,
desc->const_iter ? desc->niter_expr : const0_rtx,
desc->niter_max,
GEN_INT (level));
if (init)
{
start_sequence ();
emit_insn (init);
sequence = get_insns ();
end_sequence ();
emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src));
}
}
#endif
/* Insert the new low-overhead looping insn. */
emit_jump_insn_after (doloop_seq, BB_END (loop_end));
jump_insn = BB_END (loop_end);
jump_label = block_label (desc->in_edge->dest);
JUMP_LABEL (jump_insn) = jump_label;
LABEL_NUSES (jump_label)++;
/* Ensure the right fallthru edge is marked, for case we have reversed
the condition. */
desc->in_edge->flags &= ~EDGE_FALLTHRU;
desc->out_edge->flags |= EDGE_FALLTHRU;
/* Add a REG_NONNEG note if the actual or estimated maximum number
of iterations is non-negative. */
if (nonneg)
{
REG_NOTES (jump_insn)
= gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX, REG_NOTES (jump_insn));
}
}
static struct loop *
unswitch_loop (struct loops *loops, struct loop *loop, basic_block unswitch_on,
rtx cond, rtx cinsn)
{
edge entry, latch_edge, true_edge, false_edge, e;
basic_block switch_bb, unswitch_on_alt;
struct loop *nloop;
sbitmap zero_bitmap;
int irred_flag, prob;
rtx seq;
/* Some sanity checking. */
gcc_assert (flow_bb_inside_loop_p (loop, unswitch_on));
gcc_assert (EDGE_COUNT (unswitch_on->succs) == 2);
gcc_assert (just_once_each_iteration_p (loop, unswitch_on));
gcc_assert (!loop->inner);
gcc_assert (flow_bb_inside_loop_p (loop, EDGE_SUCC (unswitch_on, 0)->dest));
gcc_assert (flow_bb_inside_loop_p (loop, EDGE_SUCC (unswitch_on, 1)->dest));
entry = loop_preheader_edge (loop);
/* Make a copy. */
irred_flag = entry->flags & EDGE_IRREDUCIBLE_LOOP;
entry->flags &= ~EDGE_IRREDUCIBLE_LOOP;
zero_bitmap = sbitmap_alloc (2);
sbitmap_zero (zero_bitmap);
if (!duplicate_loop_to_header_edge (loop, entry, loops, 1,
zero_bitmap, NULL, NULL, NULL, 0))
{
sbitmap_free (zero_bitmap);
return NULL;
}
sbitmap_free (zero_bitmap);
entry->flags |= irred_flag;
/* Record the block with condition we unswitch on. */
unswitch_on_alt = get_bb_copy (unswitch_on);
true_edge = BRANCH_EDGE (unswitch_on_alt);
false_edge = FALLTHRU_EDGE (unswitch_on);
latch_edge = single_succ_edge (get_bb_copy (loop->latch));
/* Create a block with the condition. */
prob = true_edge->probability;
switch_bb = create_empty_bb (EXIT_BLOCK_PTR->prev_bb);
seq = compare_and_jump_seq (XEXP (cond, 0), XEXP (cond, 1), GET_CODE (cond),
block_label (true_edge->dest),
prob, cinsn);
emit_insn_after (seq, BB_END (switch_bb));
e = make_edge (switch_bb, true_edge->dest, 0);
e->probability = prob;
e->count = latch_edge->count * prob / REG_BR_PROB_BASE;
e = make_edge (switch_bb, FALLTHRU_EDGE (unswitch_on)->dest, EDGE_FALLTHRU);
e->probability = false_edge->probability;
e->count = latch_edge->count * (false_edge->probability) / REG_BR_PROB_BASE;
if (irred_flag)
{
switch_bb->flags |= BB_IRREDUCIBLE_LOOP;
EDGE_SUCC (switch_bb, 0)->flags |= EDGE_IRREDUCIBLE_LOOP;
EDGE_SUCC (switch_bb, 1)->flags |= EDGE_IRREDUCIBLE_LOOP;
}
else
{
switch_bb->flags &= ~BB_IRREDUCIBLE_LOOP;
EDGE_SUCC (switch_bb, 0)->flags &= ~EDGE_IRREDUCIBLE_LOOP;
EDGE_SUCC (switch_bb, 1)->flags &= ~EDGE_IRREDUCIBLE_LOOP;
}
/* Loopify from the copy of LOOP body, constructing the new loop. */
nloop = loopify (loops, latch_edge,
single_pred_edge (get_bb_copy (loop->header)), switch_bb,
BRANCH_EDGE (switch_bb), FALLTHRU_EDGE (switch_bb), true);
/* Remove branches that are now unreachable in new loops. */
remove_path (loops, true_edge);
remove_path (loops, false_edge);
/* One of created loops do not have to be subloop of the outer loop now,
so fix its placement in loop data structure. */
fix_loop_placement (loop);
fix_loop_placement (nloop);
/* Preserve the simple loop preheaders. */
loop_split_edge_with (loop_preheader_edge (loop), NULL_RTX);
loop_split_edge_with (loop_preheader_edge (nloop), NULL_RTX);
return nloop;
}
