static void
build_constructors (gimple swtch)
{
unsigned i, branch_num = gimple_switch_num_labels (swtch);
tree pos = info.range_min;
for (i = 1; i < branch_num; i++)
{
tree cs = gimple_switch_label (swtch, i);
basic_block bb = label_to_block (CASE_LABEL (cs));
edge e;
tree high;
gimple_stmt_iterator gsi;
int j;
if (bb == info.final_bb)
e = find_edge (info.switch_bb, bb);
else
e = single_succ_edge (bb);
gcc_assert (e);
while (tree_int_cst_lt (pos, CASE_LOW (cs)))
{
int k;
for (k = 0; k < info.phi_count; k++)
{
constructor_elt *elt;
elt = VEC_quick_push (constructor_elt,
info.constructors[k], NULL);
elt->index = int_const_binop (MINUS_EXPR, pos,
info.range_min, 0);
elt->value = info.default_values[k];
}
pos = int_const_binop (PLUS_EXPR, pos, integer_one_node, 0);
}
gcc_assert (tree_int_cst_equal (pos, CASE_LOW (cs)));
j = 0;
if (CASE_HIGH (cs))
high = CASE_HIGH (cs);
else
high = CASE_LOW (cs);
for (gsi = gsi_start_phis (info.final_bb);
!gsi_end_p (gsi); gsi_next (&gsi))
{
gimple phi = gsi_stmt (gsi);
tree val = PHI_ARG_DEF_FROM_EDGE (phi, e);
tree low = CASE_LOW (cs);
pos = CASE_LOW (cs);
do
{
constructor_elt *elt;
elt = VEC_quick_push (constructor_elt,
info.constructors[j], NULL);
elt->index = int_const_binop (MINUS_EXPR, pos, info.range_min, 0);
elt->value = val;
pos = int_const_binop (PLUS_EXPR, pos, integer_one_node, 0);
} while (!tree_int_cst_lt (high, pos) && tree_int_cst_lt (low, pos));
j++;
}
}
}
static bool
value_replacement (basic_block cond_bb, basic_block middle_bb,
edge e0, edge e1, gimple phi,
tree arg0, tree arg1)
{
gimple cond;
edge true_edge, false_edge;
enum tree_code code;
/* If the type says honor signed zeros we cannot do this
optimization. */
if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
return false;
if (!empty_block_p (middle_bb))
return false;
cond = last_stmt (cond_bb);
code = gimple_cond_code (cond);
/* This transformation is only valid for equality comparisons. */
if (code != NE_EXPR && code != EQ_EXPR)
return false;
/* We need to know which is the true edge and which is the false
edge so that we know if have abs or negative abs. */
extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
/* At this point we know we have a COND_EXPR with two successors.
One successor is BB, the other successor is an empty block which
falls through into BB.
The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
There is a single PHI node at the join point (BB) with two arguments.
We now need to verify that the two arguments in the PHI node match
the two arguments to the equality comparison. */
if ((operand_equal_for_phi_arg_p (arg0, gimple_cond_lhs (cond))
&& operand_equal_for_phi_arg_p (arg1, gimple_cond_rhs (cond)))
|| (operand_equal_for_phi_arg_p (arg1, gimple_cond_lhs (cond))
&& operand_equal_for_phi_arg_p (arg0, gimple_cond_rhs (cond))))
{
edge e;
tree arg;
/* For NE_EXPR, we want to build an assignment result = arg where
arg is the PHI argument associated with the true edge. For
EQ_EXPR we want the PHI argument associated with the false edge. */
e = (code == NE_EXPR ? true_edge : false_edge);
/* Unfortunately, E may not reach BB (it may instead have gone to
OTHER_BLOCK). If that is the case, then we want the single outgoing
edge from OTHER_BLOCK which reaches BB and represents the desired
path from COND_BLOCK. */
if (e->dest == middle_bb)
e = single_succ_edge (e->dest);
/* Now we know the incoming edge to BB that has the argument for the
RHS of our new assignment statement. */
if (e0 == e)
arg = arg0;
else
arg = arg1;
replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
/* Note that we optimized this PHI. */
return true;
}
return false;
}
static int
optimize_mode_switching (void)
{
int e;
basic_block bb;
bool need_commit = false;
static const int num_modes[] = NUM_MODES_FOR_MODE_SWITCHING;
#define N_ENTITIES ARRAY_SIZE (num_modes)
int entity_map[N_ENTITIES];
struct bb_info *bb_info[N_ENTITIES];
int i, j;
int n_entities = 0;
int max_num_modes = 0;
bool emitted ATTRIBUTE_UNUSED = false;
basic_block post_entry = 0;
basic_block pre_exit = 0;
struct edge_list *edge_list = 0;
/* These bitmaps are used for the LCM algorithm. */
sbitmap *kill, *del, *insert, *antic, *transp, *comp;
sbitmap *avin, *avout;
for (e = N_ENTITIES - 1; e >= 0; e--)
if (OPTIMIZE_MODE_SWITCHING (e))
{
int entry_exit_extra = 0;
/* Create the list of segments within each basic block.
If NORMAL_MODE is defined, allow for two extra
blocks split from the entry and exit block. */
if (targetm.mode_switching.entry && targetm.mode_switching.exit)
entry_exit_extra = 3;
bb_info[n_entities]
= XCNEWVEC (struct bb_info,
last_basic_block_for_fn (cfun) + entry_exit_extra);
entity_map[n_entities++] = e;
if (num_modes[e] > max_num_modes)
max_num_modes = num_modes[e];
}
if (! n_entities)
return 0;
/* Make sure if MODE_ENTRY is defined MODE_EXIT is defined. */
gcc_assert ((targetm.mode_switching.entry && targetm.mode_switching.exit)
|| (!targetm.mode_switching.entry
&& !targetm.mode_switching.exit));
if (targetm.mode_switching.entry && targetm.mode_switching.exit)
{
/* Split the edge from the entry block, so that we can note that
there NORMAL_MODE is supplied. */
post_entry = split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
pre_exit = create_pre_exit (n_entities, entity_map, num_modes);
}
df_analyze ();
/* Create the bitmap vectors. */
antic = sbitmap_vector_alloc (last_basic_block_for_fn (cfun),
n_entities * max_num_modes);
transp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun),
n_entities * max_num_modes);
comp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun),
n_entities * max_num_modes);
avin = sbitmap_vector_alloc (last_basic_block_for_fn (cfun),
n_entities * max_num_modes);
avout = sbitmap_vector_alloc (last_basic_block_for_fn (cfun),
n_entities * max_num_modes);
kill = sbitmap_vector_alloc (last_basic_block_for_fn (cfun),
n_entities * max_num_modes);
bitmap_vector_ones (transp, last_basic_block_for_fn (cfun));
bitmap_vector_clear (antic, last_basic_block_for_fn (cfun));
bitmap_vector_clear (comp, last_basic_block_for_fn (cfun));
for (j = n_entities - 1; j >= 0; j--)
{
int e = entity_map[j];
int no_mode = num_modes[e];
struct bb_info *info = bb_info[j];
rtx insn;
/* Determine what the first use (if any) need for a mode of entity E is.
This will be the mode that is anticipatable for this block.
Also compute the initial transparency settings. */
FOR_EACH_BB_FN (bb, cfun)
{
struct seginfo *ptr;
int last_mode = no_mode;
bool any_set_required = false;
HARD_REG_SET live_now;
info[bb->index].mode_out = info[bb->index].mode_in = no_mode;
REG_SET_TO_HARD_REG_SET (live_now, df_get_live_in (bb));
/* Pretend the mode is clobbered across abnormal edges. */
{
edge_iterator ei;
edge eg;
FOR_EACH_EDGE (eg, ei, bb->preds)
if (eg->flags & EDGE_COMPLEX)
break;
if (eg)
{
rtx ins_pos = BB_HEAD (bb);
if (LABEL_P (ins_pos))
ins_pos = NEXT_INSN (ins_pos);
gcc_assert (NOTE_INSN_BASIC_BLOCK_P (ins_pos));
if (ins_pos != BB_END (bb))
ins_pos = NEXT_INSN (ins_pos);
ptr = new_seginfo (no_mode, ins_pos, bb->index, live_now);
add_seginfo (info + bb->index, ptr);
for (i = 0; i < no_mode; i++)
clear_mode_bit (transp[bb->index], j, i);
}
}
FOR_BB_INSNS (bb, insn)
{
if (INSN_P (insn))
{
int mode = targetm.mode_switching.needed (e, insn);
rtx link;
if (mode != no_mode && mode != last_mode)
{
any_set_required = true;
last_mode = mode;
ptr = new_seginfo (mode, insn, bb->index, live_now);
add_seginfo (info + bb->index, ptr);
for (i = 0; i < no_mode; i++)
clear_mode_bit (transp[bb->index], j, i);
}
if (targetm.mode_switching.after)
last_mode = targetm.mode_switching.after (e, last_mode,
insn);
/* Update LIVE_NOW. */
for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
if (REG_NOTE_KIND (link) == REG_DEAD)
reg_dies (XEXP (link, 0), &live_now);
note_stores (PATTERN (insn), reg_becomes_live, &live_now);
for (link = REG_NOTES (insn); link; link = XEXP (link, 1))
if (REG_NOTE_KIND (link) == REG_UNUSED)
reg_dies (XEXP (link, 0), &live_now);
}
}
info[bb->index].computing = last_mode;
/* Check for blocks without ANY mode requirements.
N.B. because of MODE_AFTER, last_mode might still
be different from no_mode, in which case we need to
mark the block as nontransparent. */
if (!any_set_required)
{
ptr = new_seginfo (no_mode, BB_END (bb), bb->index, live_now);
add_seginfo (info + bb->index, ptr);
if (last_mode != no_mode)
for (i = 0; i < no_mode; i++)
clear_mode_bit (transp[bb->index], j, i);
}
}
if (targetm.mode_switching.entry && targetm.mode_switching.exit)
{
int mode = targetm.mode_switching.entry (e);
info[post_entry->index].mode_out =
info[post_entry->index].mode_in = no_mode;
if (pre_exit)
{
info[pre_exit->index].mode_out =
info[pre_exit->index].mode_in = no_mode;
}
if (mode != no_mode)
{
bb = post_entry;
/* By always making this nontransparent, we save
an extra check in make_preds_opaque. We also
need this to avoid confusing pre_edge_lcm when
antic is cleared but transp and comp are set. */
for (i = 0; i < no_mode; i++)
clear_mode_bit (transp[bb->index], j, i);
/* Insert a fake computing definition of MODE into entry
blocks which compute no mode. This represents the mode on
entry. */
info[bb->index].computing = mode;
if (pre_exit)
info[pre_exit->index].seginfo->mode =
targetm.mode_switching.exit (e);
}
}
/* Set the anticipatable and computing arrays. */
for (i = 0; i < no_mode; i++)
{
int m = targetm.mode_switching.priority (entity_map[j], i);
FOR_EACH_BB_FN (bb, cfun)
{
if (info[bb->index].seginfo->mode == m)
set_mode_bit (antic[bb->index], j, m);
if (info[bb->index].computing == m)
set_mode_bit (comp[bb->index], j, m);
}
}
}
/* Calculate the optimal locations for the
placement mode switches to modes with priority I. */
FOR_EACH_BB_FN (bb, cfun)
bitmap_not (kill[bb->index], transp[bb->index]);
edge_list = pre_edge_lcm_avs (n_entities * max_num_modes, transp, comp, antic,
kill, avin, avout, &insert, &del);
for (j = n_entities - 1; j >= 0; j--)
{
int no_mode = num_modes[entity_map[j]];
/* Insert all mode sets that have been inserted by lcm. */
for (int ed = NUM_EDGES (edge_list) - 1; ed >= 0; ed--)
{
edge eg = INDEX_EDGE (edge_list, ed);
eg->aux = (void *)(intptr_t)-1;
for (i = 0; i < no_mode; i++)
{
int m = targetm.mode_switching.priority (entity_map[j], i);
if (mode_bit_p (insert[ed], j, m))
{
eg->aux = (void *)(intptr_t)m;
break;
}
}
}
FOR_EACH_BB_FN (bb, cfun)
{
struct bb_info *info = bb_info[j];
int last_mode = no_mode;
/* intialize mode in availability for bb. */
for (i = 0; i < no_mode; i++)
if (mode_bit_p (avout[bb->index], j, i))
{
if (last_mode == no_mode)
last_mode = i;
if (last_mode != i)
{
last_mode = no_mode;
break;
}
}
info[bb->index].mode_out = last_mode;
/* intialize mode out availability for bb. */
last_mode = no_mode;
for (i = 0; i < no_mode; i++)
if (mode_bit_p (avin[bb->index], j, i))
{
if (last_mode == no_mode)
last_mode = i;
if (last_mode != i)
{
last_mode = no_mode;
break;
}
}
info[bb->index].mode_in = last_mode;
for (i = 0; i < no_mode; i++)
if (mode_bit_p (del[bb->index], j, i))
info[bb->index].seginfo->mode = no_mode;
}
/* Now output the remaining mode sets in all the segments. */
/* In case there was no mode inserted. the mode information on the edge
might not be complete.
Update mode info on edges and commit pending mode sets. */
need_commit |= commit_mode_sets (edge_list, entity_map[j], bb_info[j]);
/* Reset modes for next entity. */
clear_aux_for_edges ();
FOR_EACH_BB_FN (bb, cfun)
{
struct seginfo *ptr, *next;
int cur_mode = bb_info[j][bb->index].mode_in;
for (ptr = bb_info[j][bb->index].seginfo; ptr; ptr = next)
{
next = ptr->next;
if (ptr->mode != no_mode)
{
rtx mode_set;
rtl_profile_for_bb (bb);
start_sequence ();
targetm.mode_switching.emit (entity_map[j], ptr->mode,
cur_mode, ptr->regs_live);
mode_set = get_insns ();
end_sequence ();
/* modes kill each other inside a basic block. */
cur_mode = ptr->mode;
/* Insert MODE_SET only if it is nonempty. */
if (mode_set != NULL_RTX)
{
emitted = true;
if (NOTE_INSN_BASIC_BLOCK_P (ptr->insn_ptr))
/* We need to emit the insns in a FIFO-like manner,
i.e. the first to be emitted at our insertion
point ends up first in the instruction steam.
Because we made sure that NOTE_INSN_BASIC_BLOCK is
only used for initially empty basic blocks, we
can achieve this by appending at the end of
the block. */
emit_insn_after
(mode_set, BB_END (NOTE_BASIC_BLOCK (ptr->insn_ptr)));
else
emit_insn_before (mode_set, ptr->insn_ptr);
}
default_rtl_profile ();
}
free (ptr);
}
}
free (bb_info[j]);
}
free_edge_list (edge_list);
/* Finished. Free up all the things we've allocated. */
sbitmap_vector_free (del);
sbitmap_vector_free (insert);
sbitmap_vector_free (kill);
sbitmap_vector_free (antic);
sbitmap_vector_free (transp);
sbitmap_vector_free (comp);
sbitmap_vector_free (avin);
sbitmap_vector_free (avout);
if (need_commit)
commit_edge_insertions ();
if (targetm.mode_switching.entry && targetm.mode_switching.exit)
cleanup_cfg (CLEANUP_NO_INSN_DEL);
else if (!need_commit && !emitted)
return 0;
return 1;
}
static bool
check_process_case (tree cs, struct switch_conv_info *info)
{
tree ldecl;
basic_block label_bb, following_bb;
edge e;
ldecl = CASE_LABEL (cs);
label_bb = label_to_block (ldecl);
e = find_edge (info->switch_bb, label_bb);
gcc_assert (e);
if (CASE_LOW (cs) == NULL_TREE)
{
/* Default branch. */
info->default_prob = e->probability;
info->default_count = e->count;
}
else
{
int i;
info->other_count += e->count;
for (i = 0; i < 2; i++)
if (info->bit_test_bb[i] == label_bb)
break;
else if (info->bit_test_bb[i] == NULL)
{
info->bit_test_bb[i] = label_bb;
info->bit_test_uniq++;
break;
}
if (i == 2)
info->bit_test_uniq = 3;
if (CASE_HIGH (cs) != NULL_TREE
&& ! tree_int_cst_equal (CASE_LOW (cs), CASE_HIGH (cs)))
info->bit_test_count += 2;
else
info->bit_test_count++;
}
if (!label_bb)
{
info->reason = "bad case - cs BB label is NULL";
return false;
}
if (!single_pred_p (label_bb))
{
if (info->final_bb && info->final_bb != label_bb)
{
info->reason = "bad case - a non-final BB has two predecessors";
return false; /* sth complex going on in this branch */
}
following_bb = label_bb;
}
else
{
if (!empty_block_p (label_bb))
{
info->reason = "bad case - a non-final BB not empty";
return false;
}
e = single_succ_edge (label_bb);
following_bb = single_succ (label_bb);
}
if (!info->final_bb)
info->final_bb = following_bb;
else if (info->final_bb != following_bb)
{
info->reason = "bad case - different final BB";
return false; /* the only successor is not common for all the branches */
}
return true;
}
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;
}
static struct loop *
unswitch_loop (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;
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;
if (!duplicate_loop_to_header_edge (loop, entry, 1,
NULL, NULL, NULL, 0))
return NULL;
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 = apply_probability (latch_edge->count, prob);
e = make_edge (switch_bb, FALLTHRU_EDGE (unswitch_on)->dest, EDGE_FALLTHRU);
e->probability = false_edge->probability;
e->count = apply_probability (latch_edge->count, false_edge->probability);
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 (latch_edge,
single_pred_edge (get_bb_copy (loop->header)), switch_bb,
BRANCH_EDGE (switch_bb), FALLTHRU_EDGE (switch_bb), true,
prob, REG_BR_PROB_BASE - prob);
copy_loop_info (loop, nloop);
/* Remove branches that are now unreachable in new loops. */
remove_path (true_edge);
remove_path (false_edge);
/* Preserve the simple loop preheaders. */
split_edge (loop_preheader_edge (loop));
split_edge (loop_preheader_edge (nloop));
return nloop;
}
static void
eliminate_tail_call (struct tailcall *t)
{
tree param, rslt;
gimple *stmt, *call;
tree arg;
size_t idx;
basic_block bb, first;
edge e;
gphi *phi;
gphi_iterator gpi;
gimple_stmt_iterator gsi;
gimple *orig_stmt;
stmt = orig_stmt = gsi_stmt (t->call_gsi);
bb = gsi_bb (t->call_gsi);
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Eliminated tail recursion in bb %d : ",
bb->index);
print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
fprintf (dump_file, "\n");
}
gcc_assert (is_gimple_call (stmt));
first = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
/* Remove the code after call_gsi that will become unreachable. The
possibly unreachable code in other blocks is removed later in
cfg cleanup. */
gsi = t->call_gsi;
gimple_stmt_iterator gsi2 = gsi_last_bb (gimple_bb (gsi_stmt (gsi)));
while (gsi_stmt (gsi2) != gsi_stmt (gsi))
{
gimple *t = gsi_stmt (gsi2);
/* Do not remove the return statement, so that redirect_edge_and_branch
sees how the block ends. */
if (gimple_code (t) != GIMPLE_RETURN)
{
gimple_stmt_iterator gsi3 = gsi2;
gsi_prev (&gsi2);
gsi_remove (&gsi3, true);
release_defs (t);
}
else
gsi_prev (&gsi2);
}
/* Number of executions of function has reduced by the tailcall. */
e = single_succ_edge (gsi_bb (t->call_gsi));
decrease_profile (EXIT_BLOCK_PTR_FOR_FN (cfun), e->count, EDGE_FREQUENCY (e));
decrease_profile (ENTRY_BLOCK_PTR_FOR_FN (cfun), e->count,
EDGE_FREQUENCY (e));
if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e));
/* Replace the call by a jump to the start of function. */
e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)),
first);
gcc_assert (e);
PENDING_STMT (e) = NULL;
/* Add phi node entries for arguments. The ordering of the phi nodes should
be the same as the ordering of the arguments. */
for (param = DECL_ARGUMENTS (current_function_decl),
idx = 0, gpi = gsi_start_phis (first);
param;
param = DECL_CHAIN (param), idx++)
{
if (!arg_needs_copy_p (param))
continue;
arg = gimple_call_arg (stmt, idx);
phi = gpi.phi ();
gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi)));
add_phi_arg (phi, arg, e, gimple_location (stmt));
gsi_next (&gpi);
}
/* Update the values of accumulators. */
adjust_accumulator_values (t->call_gsi, t->mult, t->add, e);
call = gsi_stmt (t->call_gsi);
rslt = gimple_call_lhs (call);
if (rslt != NULL_TREE)
{
/* Result of the call will no longer be defined. So adjust the
SSA_NAME_DEF_STMT accordingly. */
SSA_NAME_DEF_STMT (rslt) = gimple_build_nop ();
}
gsi_remove (&t->call_gsi, true);
release_defs (call);
}
static void
find_tail_calls (basic_block bb, struct tailcall **ret)
{
tree ass_var = NULL_TREE, ret_var, func, param;
gimple *stmt;
gcall *call = NULL;
gimple_stmt_iterator gsi, agsi;
bool tail_recursion;
struct tailcall *nw;
edge e;
tree m, a;
basic_block abb;
size_t idx;
tree var;
if (!single_succ_p (bb))
return;
for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
{
stmt = gsi_stmt (gsi);
/* Ignore labels, returns, nops, clobbers and debug stmts. */
if (gimple_code (stmt) == GIMPLE_LABEL
|| gimple_code (stmt) == GIMPLE_RETURN
|| gimple_code (stmt) == GIMPLE_NOP
|| gimple_clobber_p (stmt)
|| is_gimple_debug (stmt))
continue;
/* Check for a call. */
if (is_gimple_call (stmt))
{
call = as_a <gcall *> (stmt);
ass_var = gimple_call_lhs (call);
break;
}
/* If the statement references memory or volatile operands, fail. */
if (gimple_references_memory_p (stmt)
|| gimple_has_volatile_ops (stmt))
return;
}
if (gsi_end_p (gsi))
{
edge_iterator ei;
/* Recurse to the predecessors. */
FOR_EACH_EDGE (e, ei, bb->preds)
find_tail_calls (e->src, ret);
return;
}
/* If the LHS of our call is not just a simple register, we can't
transform this into a tail or sibling call. This situation happens,
in (e.g.) "*p = foo()" where foo returns a struct. In this case
we won't have a temporary here, but we need to carry out the side
effect anyway, so tailcall is impossible.
??? In some situations (when the struct is returned in memory via
invisible argument) we could deal with this, e.g. by passing 'p'
itself as that argument to foo, but it's too early to do this here,
and expand_call() will not handle it anyway. If it ever can, then
we need to revisit this here, to allow that situation. */
if (ass_var && !is_gimple_reg (ass_var))
return;
/* We found the call, check whether it is suitable. */
tail_recursion = false;
func = gimple_call_fndecl (call);
if (func
&& !DECL_BUILT_IN (func)
&& recursive_call_p (current_function_decl, func))
{
tree arg;
for (param = DECL_ARGUMENTS (func), idx = 0;
param && idx < gimple_call_num_args (call);
param = DECL_CHAIN (param), idx ++)
{
arg = gimple_call_arg (call, idx);
if (param != arg)
{
/* Make sure there are no problems with copying. The parameter
have a copyable type and the two arguments must have reasonably
equivalent types. The latter requirement could be relaxed if
we emitted a suitable type conversion statement. */
if (!is_gimple_reg_type (TREE_TYPE (param))
|| !useless_type_conversion_p (TREE_TYPE (param),
TREE_TYPE (arg)))
break;
/* The parameter should be a real operand, so that phi node
created for it at the start of the function has the meaning
of copying the value. This test implies is_gimple_reg_type
from the previous condition, however this one could be
relaxed by being more careful with copying the new value
of the parameter (emitting appropriate GIMPLE_ASSIGN and
updating the virtual operands). */
if (!is_gimple_reg (param))
break;
}
}
if (idx == gimple_call_num_args (call) && !param)
tail_recursion = true;
}
/* Make sure the tail invocation of this function does not refer
to local variables. */
FOR_EACH_LOCAL_DECL (cfun, idx, var)
{
if (TREE_CODE (var) != PARM_DECL
&& auto_var_in_fn_p (var, cfun->decl)
&& (ref_maybe_used_by_stmt_p (call, var)
|| call_may_clobber_ref_p (call, var)))
return;
}
/* Now check the statements after the call. None of them has virtual
operands, so they may only depend on the call through its return
value. The return value should also be dependent on each of them,
since we are running after dce. */
m = NULL_TREE;
a = NULL_TREE;
abb = bb;
agsi = gsi;
while (1)
{
tree tmp_a = NULL_TREE;
tree tmp_m = NULL_TREE;
gsi_next (&agsi);
while (gsi_end_p (agsi))
{
ass_var = propagate_through_phis (ass_var, single_succ_edge (abb));
abb = single_succ (abb);
agsi = gsi_start_bb (abb);
}
stmt = gsi_stmt (agsi);
if (gimple_code (stmt) == GIMPLE_LABEL
|| gimple_code (stmt) == GIMPLE_NOP)
continue;
if (gimple_code (stmt) == GIMPLE_RETURN)
break;
if (gimple_clobber_p (stmt))
continue;
if (is_gimple_debug (stmt))
continue;
if (gimple_code (stmt) != GIMPLE_ASSIGN)
return;
/* This is a gimple assign. */
if (! process_assignment (as_a <gassign *> (stmt), gsi, &tmp_m,
&tmp_a, &ass_var))
return;
if (tmp_a)
{
tree type = TREE_TYPE (tmp_a);
if (a)
a = fold_build2 (PLUS_EXPR, type, fold_convert (type, a), tmp_a);
else
a = tmp_a;
}
if (tmp_m)
{
tree type = TREE_TYPE (tmp_m);
if (m)
m = fold_build2 (MULT_EXPR, type, fold_convert (type, m), tmp_m);
else
m = tmp_m;
if (a)
a = fold_build2 (MULT_EXPR, type, fold_convert (type, a), tmp_m);
}
}
/* See if this is a tail call we can handle. */
ret_var = gimple_return_retval (as_a <greturn *> (stmt));
/* We may proceed if there either is no return value, or the return value
is identical to the call's return. */
if (ret_var
&& (ret_var != ass_var))
return;
/* If this is not a tail recursive call, we cannot handle addends or
multiplicands. */
if (!tail_recursion && (m || a))
return;
/* For pointers only allow additions. */
if (m && POINTER_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl))))
return;
nw = XNEW (struct tailcall);
nw->call_gsi = gsi;
nw->tail_recursion = tail_recursion;
nw->mult = m;
nw->add = a;
nw->next = *ret;
*ret = nw;
}
