bool phis_moved; /**< Set if Phi nodes were moved. */
} merge_env;
+/** set or reset the removable property of a block. */
static void set_Block_removable(ir_node *block, bool removable)
{
set_Block_mark(block, removable);
}
+/** check if a block has the removable property set. */
static bool is_Block_removable(ir_node *block)
{
return get_Block_mark(block);
}
-static bool is_switch_Cond(ir_node *cond) {
+/** checks if a given Cond node is a switch Cond. */
+static bool is_switch_Cond(ir_node *cond)
+{
ir_node *sel = get_Cond_selector(cond);
return get_irn_mode(sel) != mode_b;
}
-static void clear_link(ir_node *node, void *ctx)
+/** Walker: clear link fields and mark all blocks as removable. */
+static void clear_link_and_mark_blocks_removable(ir_node *node, void *ctx)
{
(void) ctx;
set_irn_link(node, NULL);
set_irn_link(n, get_irn_link(block));
set_irn_link(block, n);
} else if (is_Block(n)) {
- if (has_Block_entity(n))
+ if (has_Block_entity(n)) {
+ /* block with a jump label attached cannot be removed. */
set_Block_removable(n, false);
+ }
+ } else if (is_Bad(n) || is_Jmp(n)) {
+ /* ignore these */
return;
- } else if (!is_Jmp(n)) { /* Check for non-empty block. */
+ } else {
+ /* Check for non-empty block. */
ir_node *block = get_nodes_block(n);
+ if (is_Bad(block))
+ return;
+
set_Block_removable(block, false);
if (is_Proj(n)) {
}
}
-/** Returns true if pred is predecessor of block. */
+/** Returns true if pred is predecessor of block b. */
static bool is_pred_of(ir_node *pred, ir_node *b)
{
int i;
ir_node **in;
merge_env *env = (merge_env*)ctx;
+ if (get_Block_dom_depth(b) < 0) {
+ /* ignore unreachable blocks */
+ return;
+ }
+
/* Count the number of predecessor if this block is merged with pred blocks
that are empty. */
max_preds = 0;
pred = get_Block_cfgpred_block(b, i);
if (is_Bad(pred)) {
- in[q_preds++] = pred;
+ ir_graph *irg = get_irn_irg(b);
+ ir_mode *mode = get_irn_mode(phi);
+ in[q_preds++] = new_r_Bad(irg, mode);
} else if (is_Block_removable(pred) && !Block_block_visited(pred)) {
/* It's an empty block and not yet visited. */
for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
- if (! is_Bad(get_Block_cfgpred(pred, j)))
+ if (! is_Bad(get_Block_cfgpred(pred, j))) {
in[q_preds++] = phi;
+ } else {
+ ir_graph *irg = get_irn_irg(b);
+ ir_mode *mode = get_irn_mode(phi);
+ in[q_preds++] = new_r_Bad(irg, mode);
+ }
}
} else {
in[q_preds++] = phi;
} else if (is_Block_removable(pred) && !Block_block_visited(pred)) {
/* It's an empty block and not yet visited. */
for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
- if (! is_Bad(get_Block_cfgpred(pred, j)))
+ if (! is_Bad(get_Block_cfgpred(pred, j))) {
in[q_preds++] = phi;
+ } else {
+ ir_graph *irg = get_irn_irg(b);
+ ir_mode *mode = get_irn_mode(phi);
+ in[q_preds++] = new_r_Bad(irg, mode);
+ }
}
} else {
in[q_preds++] = phi;
/* handle Cond nodes with constant argument. In this case the localopt rules
* should have killed all obviously impossible cases.
- * So the only case left to handle here is 1 defaultProj + 1case
+ * So the only case left to handle here is 1 defaultProj + 1 case
* (this one case should be the one taken) */
if (get_irn_link(proj2) == NULL) {
ir_tarval *tv = value_of(sel);
return false;
}
-static bool get_phase_flag(ir_phase *block_info, ir_node *block, int offset) {
- return ((int)phase_get_irn_data(block_info, block)) & (1<<offset);
-}
-static void set_phase_flag(ir_phase *block_info, ir_node *block, int offset) {
- int data = (int)phase_get_irn_data(block_info, block);
- data |= (1<<offset);
- phase_set_irn_data(block_info, block, (void*)data);
-}
-
-static bool has_operations(ir_phase *block_info, ir_node *block) {
- return get_phase_flag(block_info, block, 1);
-}
-static void set_has_operations(ir_phase *block_info, ir_node *block) {
- set_phase_flag(block_info, block, 1);
-}
-
-static bool has_phis(ir_phase *block_info, ir_node *block) {
- return get_phase_flag(block_info, block, 2);
-}
-static void set_has_phis(ir_phase *block_info, ir_node *block) {
- set_phase_flag(block_info, block, 2);
-}
-
-static bool is_unknown_jump_target(ir_phase *block_info, ir_node *block) {
- return get_phase_flag(block_info, block, 3);
-}
-static void set_is_unknown_jump_target(ir_phase *block_info, ir_node *block) {
- set_phase_flag(block_info, block, 3);
-}
-
/**
- * Optimize Conds, where true and false jump to the same block into a Jmp
+ * Optimize boolean Conds, where true and false jump to the same block into a Jmp
+ * Block must contain no Phi nodes.
*
* Cond
* / \
return true;
}
+typedef enum block_flags_t {
+ BF_HAS_OPERATIONS = 1 << 0,
+ BF_HAS_PHIS = 1 << 1,
+ BF_IS_UNKNOWN_JUMP_TARGET = 1 << 2,
+} block_flags_t;
+
+static bool get_phase_flag(ir_phase *block_info, ir_node *block, int flag)
+{
+ return PTR_TO_INT(phase_get_irn_data(block_info, block)) & flag;
+}
+
+static void set_phase_flag(ir_phase *block_info, ir_node *block,
+ block_flags_t flag)
+{
+ int data = PTR_TO_INT(phase_get_irn_data(block_info, block));
+ data |= flag;
+ phase_set_irn_data(block_info, block, INT_TO_PTR(data));
+}
+
+static void clear_phase_flag(ir_phase *block_info, ir_node *block)
+{
+ phase_set_irn_data(block_info, block, NULL);
+}
+
+static bool has_operations(ir_phase *block_info, ir_node *block)
+{
+ return get_phase_flag(block_info, block, BF_HAS_OPERATIONS);
+}
+
+static void set_has_operations(ir_phase *block_info, ir_node *block)
+{
+ set_phase_flag(block_info, block, BF_HAS_OPERATIONS);
+}
+
+static bool has_phis(ir_phase *block_info, ir_node *block)
+{
+ return get_phase_flag(block_info, block, BF_HAS_PHIS);
+}
+
+static void set_has_phis(ir_phase *block_info, ir_node *block)
+{
+ set_phase_flag(block_info, block, BF_HAS_PHIS);
+}
+
+static bool is_unknown_jump_target(ir_phase *block_info, ir_node *block)
+{
+ return get_phase_flag(block_info, block, BF_IS_UNKNOWN_JUMP_TARGET);
+}
+
+static void set_is_unknown_jump_target(ir_phase *block_info, ir_node *block)
+{
+ set_phase_flag(block_info, block, BF_IS_UNKNOWN_JUMP_TARGET);
+}
+
+/**
+ * Pre-Walker: fill block info information.
+ */
static void compute_block_info(ir_node *n, void *x)
{
ir_phase *block_info = (ir_phase *)x;
} else if (is_Phi(n)) {
ir_node *block = get_nodes_block(n);
set_has_phis(block_info, block);
- } else if (is_Jmp(n) || is_Cond(n) || is_Cmp(n) || is_Proj(n)) {
+ } else if (is_Jmp(n) || is_Cond(n) || is_Proj(n)) {
/* ignore */
} else {
ir_node *block = get_nodes_block(n);
}
}
+static void clear_block_info(ir_node *block, void *x)
+{
+ ir_phase *block_info = (ir_phase *)x;
+ clear_phase_flag(block_info, block);
+}
+
typedef struct skip_env {
bool changed;
ir_phase *phase;
} skip_env;
-static void optimize_conds(ir_node *b, void *x)
+/**
+ * Post-Block-walker: Optimize useless if's (boolean Cond nodes
+ * with same true/false target)
+ * away.
+ */
+static void optimize_ifs(ir_node *block, void *x)
{
skip_env *env = (skip_env*)x;
int i, j;
- int n_preds = get_Block_n_cfgpreds(b);
+ int n_preds = get_Block_n_cfgpreds(block);
- if (has_phis(env->phase,b)) return;
+ if (has_phis(env->phase, block))
+ return;
/* optimize Cond predecessors (might produce Bad predecessors) */
- for (i = 0; i < n_preds; i++) {
- for (j = i+1; j < n_preds; j++) {
- optimize_pred_cond(b, i, j);
+ for (i = 0; i < n_preds; ++i) {
+ for (j = i+1; j < n_preds; ++j) {
+ optimize_pred_cond(block, i, j);
}
}
}
-static void remove_empty_blocks(ir_node *b, void *x)
+/**
+ * Pre-Block walker: remove empty blocks (only contain a Jmp)
+ * that are control flow predecessors of the current block.
+ */
+static void remove_empty_blocks(ir_node *block, void *x)
{
skip_env *env = (skip_env*)x;
int i;
- int n_preds = get_Block_n_cfgpreds(b);
+ int n_preds = get_Block_n_cfgpreds(block);
for (i = 0; i < n_preds; ++i) {
ir_node *jmp, *jmp_block, *pred, *pred_block;
+ int n_jpreds = 0;
- jmp = get_Block_cfgpred(b, i);
- if (!is_Jmp(jmp)) continue;
- if (is_unknown_jump(jmp)) continue;
+ jmp = get_Block_cfgpred(block, i);
+ if (!is_Jmp(jmp))
+ continue;
jmp_block = get_nodes_block(jmp);
- if (is_unknown_jump_target(env->phase, jmp_block)) continue;
- if (has_operations(env->phase,jmp_block)) continue;
- /* jmp_block is an empty block! */
+ if (jmp_block == block)
+ continue; /* this infinite loop cannot be optimized any further */
+ if (is_unknown_jump_target(env->phase, jmp_block))
+ continue; /* unknown jump target must not be optimized */
+ if (has_operations(env->phase,jmp_block))
+ continue; /* this block contains operations and cannot be skipped */
+ if (has_phis(env->phase,jmp_block))
+ continue; /* this block contains Phis and is not skipped */
+ if (Block_block_visited(jmp_block)) {
+ continue;
+ /* otherwise we could break the walker,
+ * if block was reached via KeepAlive edge -> jmp_block -> A ---> block,
+ * because the walker cannot handle Id nodes.
+ *
+ * A B
+ * \ /
+ * jmp_block
+ * / \
+ * block End
+ */
+ }
- if (get_Block_n_cfgpreds(jmp_block) != 1) continue;
- pred = get_Block_cfgpred(jmp_block, 0);
- exchange(jmp, pred);
- env->changed = true;
+ /* jmp_block is an empty block and can be optimized! */
- /* cleanup: jmp_block might have a Keep edge! */
- pred_block = get_nodes_block(pred);
- exchange(jmp_block, pred_block);
+ n_jpreds = get_Block_n_cfgpreds(jmp_block);
+ /**
+ * If the jmp block has only one predecessor this is straightforward.
+ * However, if there are more predecessors, we only handle this,
+ * if block has no Phis.
+ */
+ if (n_jpreds == 1) {
+ /* skip jmp block by rerouting its predecessor to block
+ *
+ * A A
+ * | |
+ * jmp_block => |
+ * | |
+ * block block
+ */
+ pred = get_Block_cfgpred(jmp_block, 0);
+ exchange(jmp, pred);
+
+ /* cleanup: jmp_block might have a Keep edge! */
+ pred_block = get_nodes_block(pred);
+ exchange(jmp_block, pred_block);
+ env->changed = true;
+ } else if (! has_phis(env->phase, block)) {
+ /* all predecessors can skip the jmp block, so block gets some new predecessors
+ *
+ * A B A B
+ * \ / | |
+ * jmp_block C => Bad C | |
+ * \ / \ | | /
+ * block block
+ */
+ ir_node **ins = NULL;
+ int j;
+ NEW_ARR_A(ir_node *, ins, n_preds+n_jpreds);
+ /* first copy the old predecessors, because the outer loop (i) still walks over them */
+ for (j = 0; j < n_preds; ++j) {
+ ins[j] = get_Block_cfgpred(block, j);
+ }
+ /* now append the new predecessors */
+ for (j = 0; j < n_jpreds; ++j) {
+ pred = get_Block_cfgpred(jmp_block, j);
+ ins[n_preds+j] = pred;
+ }
+ set_irn_in(block, n_preds+n_jpreds, ins);
+ /* convert the jmp_block to Bad */
+ ir_graph *irg = get_irn_irg(block);
+ exchange(jmp_block, new_r_Bad(irg, mode_BB));
+ exchange(jmp, new_r_Bad(irg, mode_X));
+ /* let the outer loop walk over the new predecessors as well */
+ n_preds += n_jpreds;
+ env->changed = true;
+ // TODO What if jmp_block had a KeepAlive edge?
+ } else {
+ /* This would involve Phis ... */
+ }
}
}
/*
- * Some cfg optimizations, which do not touch Phi nodes */
-static void cfgopt_ignoring_phis(ir_graph *irg) {
+ * All cfg optimizations, which do not touch Phi nodes.
+ *
+ * Note that this might create critical edges.
+ */
+static void cfgopt_ignoring_phis(ir_graph *irg)
+{
ir_phase *block_info = new_phase(irg, NULL);
- skip_env env = { false, block_info };
-
- irg_walk_graph(irg, compute_block_info, NULL, block_info);
+ skip_env env = { true, block_info };
- for(;;) {
+ while (env.changed) {
+ irg_walk_graph(irg, compute_block_info, NULL, block_info);
env.changed = false;
- /* Conds => Jmp optimization; might produce empty blocks */
- irg_block_walk_graph(irg, optimize_conds, NULL, &env);
-
- /* Remove empty blocks */
+ /* Remove blocks, which only consist of a Jmp */
irg_block_walk_graph(irg, remove_empty_blocks, NULL, &env);
+
+ /* Optimize Cond->Jmp, where then- and else-block are the same. */
+ irg_block_walk_graph(irg, NULL, optimize_ifs, &env);
+
if (env.changed) {
set_irg_doms_inconsistent(irg);
- /* Removing blocks might enable more Cond optimizations */
+ /* clear block info, because it must be recomputed */
+ irg_block_walk_graph(irg, clear_block_info, NULL, block_info);
+ /* Removing blocks and Conds might enable more optimizations */
continue;
} else {
break;
ir_node *new_end;
merge_env env;
+ env.changed = false;
+ env.phis_moved = false;
+
assert(get_irg_phase_state(irg) != phase_building);
/* if the graph is not pinned, we cannot determine empty blocks */
assert(get_irg_pinned(irg) != op_pin_state_floats &&
"Control flow optimization need a pinned graph");
- /* FIXME: control flow opt destroys block edges. So edges are deactivated
- * here. Fix the edges! */
edges_deactivate(irg);
+ /* First the "simple" optimizations, which do not touch Phis */
cfgopt_ignoring_phis(irg);
/* we use the mark flag to mark removable blocks */
for (;;) {
int length;
ir_node **switch_conds = NULL;
- env.changed = false;
- env.phis_moved = false;
+ bool changed = false;
assure_doms(irg);
* computations, i.e., these blocks might be removed.
*/
switch_conds = NEW_ARR_F(ir_node*, 0);
- irg_walk(end, clear_link, collect_nodes, &switch_conds);
+ irg_walk(end, clear_link_and_mark_blocks_removable, collect_nodes, &switch_conds);
/* handle all collected switch-Conds */
length = ARR_LEN(switch_conds);
for (i = 0; i < length; ++i) {
ir_node *cond = switch_conds[i];
- env.changed |= handle_switch_cond(cond);
+ changed |= handle_switch_cond(cond);
}
DEL_ARR_F(switch_conds);
- if (!env.changed) break;
+ if (!changed)
+ break;
set_irg_doms_inconsistent(irg);
set_irg_extblk_inconsistent(irg);
if (env.phis_moved) {
/* Bad: when we moved Phi's, we might produce dead Phi nodes
that are kept-alive.
- Some other phases cannot copy with this, so will them.
+ Some other phases cannot copy with this, so kill them.
*/
n = get_End_n_keepalives(end);
if (n > 0) {