#include "irflag_t.h"
#include "firmstat.h"
#include "irpass.h"
+#include "irphase_t.h"
#include "iropt_dbg.h"
typedef struct merge_env {
bool changed; /**< Set if the graph was changed. */
bool phis_moved; /**< Set if Phi nodes were moved. */
- ir_node **switch_conds; /**< Helper list for all found Switch Conds. */
} 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 void clear_link(ir_node *node, void *ctx)
+/** 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;
+}
+
+/** 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);
*/
static void collect_nodes(ir_node *n, void *ctx)
{
- merge_env *env = (merge_env*)ctx;
+ ir_node ***switch_conds = (ir_node***)ctx;
if (is_Phi(n)) {
/* Collect Phi nodes to compact ins along with block's ins. */
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)) {
ir_node *pred = get_Proj_pred(n);
set_irn_link(n, get_irn_link(pred));
set_irn_link(pred, n);
- } else if (is_Cond(n)) {
- ir_node *sel = get_Cond_selector(n);
- if (get_irn_mode(sel) != mode_b) {
- /* found a switch-Cond, collect */
- ARR_APP1(ir_node*, env->switch_conds, n);
- }
+ } else if (is_Cond(n) && is_switch_Cond(n)) {
+ /* found a switch-Cond, collect */
+ ARR_APP1(ir_node*, *switch_conds, 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;
}
/**
- * This method removed Bad cf predecessors from Blocks and Phis, and removes
- * empty blocks. A block is empty if it only contains Phi and Jmp nodes.
+ * This method removes empty blocks. A block is empty if it only contains Phi
+ * and Jmp nodes.
*
* We first adapt Phi nodes, then Block nodes, as we need the old ins
* of the Block to adapt the Phi nodes. We do this by computing new
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)) {
- /* Do nothing */
+ 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;
/* now we are at k, copy the phi predecessors */
pred = get_nodes_block(get_Block_cfgpred(b, k));
for (i = 0; i < get_Phi_n_preds(phi); i++) {
- if (! is_Bad(get_Block_cfgpred(pred, i)))
- in[q_preds++] = get_Phi_pred(phi, i);
+ in[q_preds++] = get_Phi_pred(phi, i);
}
/* and now all the rest */
pred = get_Block_cfgpred_block(b, i);
if (is_Bad(pred)) {
- /* Do nothing */
+ 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;
ir_node *predb = get_nodes_block(pred);
ir_graph *irg = get_irn_irg(pred);
- /* case 1: Do nothing */
+ /* case 1: Bad predecessor */
if (is_Bad(pred)) {
in[n_preds++] = new_r_Bad(irg, mode_X);
continue;
xfree(in);
}
-/**
- * Block walker: optimize all blocks using the default optimizations.
- * This removes Blocks with only a Jmp predecessor.
- */
-static void remove_simple_blocks(ir_node *block, void *ctx)
-{
- merge_env *env = (merge_env*)ctx;
- ir_node *new_blk = equivalent_node(block);
-
- if (new_blk != block) {
- exchange(block, new_blk);
- env->changed = true;
- }
-}
-
/**
* Optimize table-switch Conds.
*
/* 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;
}
-/* Optimizations of the control flow that also require changes of Phi nodes.
- *
- * This optimization performs two passes over the graph.
- *
- * The first pass collects all Phi nodes in a link list in the block
- * nodes. Further it performs simple control flow optimizations.
- * Finally it marks all blocks that do not contain useful
- * computations, i.e., these blocks might be removed.
+/**
+ * Optimize boolean Conds, where true and false jump to the same block into a Jmp
+ * Block must contain no Phi nodes.
*
- * The second pass performs the optimizations intended by this algorithm.
- * It walks only over block nodes and adapts these and the Phi nodes in these
- * blocks, which it finds in a linked list computed by the first pass.
+ * Cond
+ * / \
+ * projA projB => Jmp Bad
+ * \ / \ /
+ * block block
+ */
+static bool optimize_pred_cond(ir_node *block, int i, int j)
+{
+ ir_node *projA, *projB, *cond, *pred_block, *jmp, *bad;
+ assert(i != j);
+
+ projA = get_Block_cfgpred(block, i);
+ if (!is_Proj(projA)) return false;
+ projB = get_Block_cfgpred(block, j);
+ if (!is_Proj(projB)) return false;
+ cond = get_Proj_pred(projA);
+ if (!is_Cond(cond)) return false;
+
+ if (cond != get_Proj_pred(projB)) return false;
+ if (is_switch_Cond(cond)) return false;
+
+ /* cond should actually be a Jmp */
+ pred_block = get_nodes_block(cond);
+ jmp = new_r_Jmp(pred_block);
+ bad = new_r_Bad(get_irn_irg(block), mode_X);
+
+ assert(projA != projB);
+ exchange(projA, jmp);
+ exchange(projB, bad);
+ 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;
+
+ if (is_Block(n)) {
+ int i, max = get_Block_n_cfgpreds(n);
+ for (i=0; i<max; i++) {
+ ir_node *pred = get_Block_cfgpred(n,i);
+ if (is_unknown_jump(pred)) {
+ set_is_unknown_jump_target(block_info, n);
+ }
+ }
+ } 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_Proj(n)) {
+ /* ignore */
+ } else {
+ ir_node *block = get_nodes_block(n);
+ set_has_operations(block_info, block);
+ }
+}
+
+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;
+
+/**
+ * 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(block);
+
+ 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(block, i, j);
+ }
+ }
+}
+
+/**
+ * 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(block);
+
+ for (i = 0; i < n_preds; ++i) {
+ ir_node *jmp, *jmp_block, *pred, *pred_block;
+ int n_jpreds = 0;
+
+ jmp = get_Block_cfgpred(block, i);
+ if (!is_Jmp(jmp))
+ continue;
+ jmp_block = get_nodes_block(jmp);
+ 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
+ */
+ }
+
+ /* jmp_block is an empty block and can be optimized! */
+
+ 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 ... */
+ }
+ }
+}
+
+/*
+ * All cfg optimizations, which do not touch Phi nodes.
*
- * We use the mark flag to mark removable blocks in the first phase.
+ * 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 = { true, block_info };
+
+ while (env.changed) {
+ irg_walk_graph(irg, compute_block_info, NULL, block_info);
+ env.changed = false;
+
+ /* 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);
+ /* 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;
+ }
+ }
+
+ phase_free(block_info);
+}
+
+/* Optimizations of the control flow that also require changes of Phi nodes. */
void optimize_cf(ir_graph *irg)
{
int i, j, n;
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 */
ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK);
-restart:
- env.changed = false;
- env.phis_moved = false;
- assure_doms(irg);
+ /* The switch Cond optimization might expose unreachable code, so we loop */
+ for (;;) {
+ int length;
+ ir_node **switch_conds = NULL;
+ bool changed = false;
- env.switch_conds = NEW_ARR_F(ir_node*, 0);
- irg_walk(end, clear_link, collect_nodes, &env);
+ assure_doms(irg);
- /* handle all collected switch-Conds */
- n = ARR_LEN(env.switch_conds);
- for (i = 0; i < n; ++i) {
- ir_node *cond = env.switch_conds[i];
- env.changed |= handle_switch_cond(cond);
- }
- DEL_ARR_F(env.switch_conds);
+ /*
+ * This pass collects all Phi nodes in a link list in the block
+ * nodes. Further it performs simple control flow optimizations.
+ * Finally it marks all blocks that do not contain useful
+ * computations, i.e., these blocks might be removed.
+ */
+ switch_conds = NEW_ARR_F(ir_node*, 0);
+ 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];
+ changed |= handle_switch_cond(cond);
+ }
+ DEL_ARR_F(switch_conds);
+
+ if (!changed)
+ break;
- if (env.changed) {
- /* Handle graph state if was changed. */
- set_irg_outs_inconsistent(irg);
set_irg_doms_inconsistent(irg);
set_irg_extblk_inconsistent(irg);
- set_irg_loopinfo_inconsistent(irg);
set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
-
- /* The Cond optimization might generate unreachable code, so restart if
- it happens. */
- goto restart;
}
- /* Optimize the standard code. */
+ /* assert due to collect_nodes:
+ * 1. removable blocks are now marked as such
+ * 2. phi lists are up to date
+ */
+
+ /* Optimize the standard code.
+ * It walks only over block nodes and adapts these and the Phi nodes in these
+ * blocks, which it finds in a linked list computed before.
+ * */
assure_doms(irg);
- irg_block_walk_graph(irg, optimize_blocks, remove_simple_blocks, &env);
+ irg_block_walk_graph(irg, optimize_blocks, NULL, &env);
new_end = optimize_in_place(end);
if (new_end != end) {
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) {
NEW_ARR_A(ir_node *, in, n);
- if (env.changed) {
- /* Handle graph state if was changed. */
- set_irg_outs_inconsistent(irg);
- }
assure_irg_outs(irg);
for (i = j = 0; i < n; ++i) {
if (env.changed) {
/* Handle graph state if was changed. */
- set_irg_outs_inconsistent(irg);
set_irg_doms_inconsistent(irg);
set_irg_extblk_inconsistent(irg);
- set_irg_loopinfo_inconsistent(irg);
set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
}
}