#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;
static void set_Block_removable(ir_node *block, bool removable)
return get_Block_mark(block);
}
+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)
{
(void) ctx;
set_irn_link(node, NULL);
+ if (is_Block(node))
+ set_Block_removable(node, true);
}
/**
*/
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))
+ set_Block_removable(n, false);
return;
} else if (!is_Jmp(n)) { /* Check for non-empty block. */
ir_node *block = get_nodes_block(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);
}
}
}
**/
static unsigned test_whether_dispensable(ir_node *b, int pos)
{
- int i, j, n_preds = 1;
- ir_node *pred = get_Block_cfgpred_block(b, pos);
-
- /* Bad blocks will be optimized away, so we don't need space for them */
- if (is_Bad(pred))
- return 0;
-
- if (is_Block_removable(pred)) {
- /* Seems to be empty. At least we detected this in collect_nodes. */
- if (get_irn_link(b) == NULL) {
- /* There are no Phi nodes ==> all predecessors are dispensable. */
- n_preds = get_Block_n_cfgpreds(pred);
- } else {
- /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
- Handle all pred blocks with preds < pos as if they were already removed. */
- for (i = 0; i < pos; i++) {
- ir_node *b_pred = get_Block_cfgpred_block(b, i);
- if (! is_Bad(b_pred) && is_Block_removable(b_pred)) {
- for (j = get_Block_n_cfgpreds(b_pred) - 1; j >= 0; --j) {
- ir_node *b_pred_pred = get_Block_cfgpred_block(b_pred, j);
- if (is_pred_of(b_pred_pred, pred))
- goto non_dispensable;
- }
- } else {
- if (is_pred_of(b_pred, pred))
+ ir_node *pred = get_Block_cfgpred(b, pos);
+ ir_node *predb = get_nodes_block(pred);
+
+ if (is_Bad(pred) || !is_Block_removable(predb))
+ return 1;
+
+ /* can't remove self-loops */
+ if (predb == b)
+ goto non_dispensable;
+ if (is_unknown_jump(pred))
+ goto non_dispensable;
+
+ /* Seems to be empty. At least we detected this in collect_nodes. */
+ if (get_irn_link(b) != NULL) {
+ int n_cfgpreds = get_Block_n_cfgpreds(b);
+ int i;
+ /* there are Phi nodes */
+
+ /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
+ * Handle all pred blocks with preds < pos as if they were already
+ * removed. */
+ for (i = 0; i < pos; i++) {
+ ir_node *other_pred = get_Block_cfgpred(b, i);
+ ir_node *other_predb = get_nodes_block(other_pred);
+ if (is_Bad(other_pred))
+ continue;
+ if (is_Block_removable(other_predb)
+ && !Block_block_visited(other_predb)) {
+ int j;
+ for (j = get_Block_n_cfgpreds(other_predb) - 1; j >= 0; --j) {
+ ir_node *other_predpred
+ = get_Block_cfgpred_block(other_predb, j);
+ if (is_pred_of(other_predpred, predb))
goto non_dispensable;
}
+ } else if (is_pred_of(other_predb, predb)) {
+ goto non_dispensable;
}
- for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
- ir_node *b_pred = get_Block_cfgpred_block(b, i);
- if (is_pred_of(b_pred, pred))
- goto non_dispensable;
- }
- /* if we get here, the block is dispensable */
- n_preds = get_Block_n_cfgpreds(pred);
+ }
+ for (i = pos+1; i < n_cfgpreds; i++) {
+ ir_node *other_predb = get_Block_cfgpred_block(b, i);
+ if (is_pred_of(other_predb, predb))
+ goto non_dispensable;
}
}
-
- return n_preds;
+ /* we will not dispense already visited blocks */
+ if (Block_block_visited(predb))
+ return 1;
+ /* if we get here, the block is dispensable, count useful preds */
+ return get_irn_arity(predb);
non_dispensable:
- set_Block_removable(pred, false);
+ set_Block_removable(predb, false);
return 1;
}
/**
- * 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
/* Find the new predecessors for the Phi */
p_preds = 0;
for (i = 0, n = get_Block_n_cfgpreds(b); i < n; ++i) {
+ ir_graph *irg = get_irn_irg(b);
pred = get_Block_cfgpred_block(b, i);
if (is_Bad(pred)) {
- /* case Phi 1: Do nothing */
+ /* case Phi 1: maintain Bads, as somebody else is responsible to remove them */
+ in[p_preds++] = new_r_Bad(irg, get_irn_mode(phi));
} else if (is_Block_removable(pred) && !Block_block_visited(pred)) {
/* case Phi 2: It's an empty block and not yet visited. */
ir_node *phi_pred = get_Phi_pred(phi, i);
for (j = 0, k = get_Block_n_cfgpreds(pred); j < k; j++) {
- /* because of breaking loops, not all predecessors are
- * Bad-clean, so we must check this here again */
- if (! is_Bad(get_Block_cfgpred(pred, j))) {
- if (get_nodes_block(phi_pred) == pred) {
- /* case Phi 2a: */
- assert(is_Phi(phi_pred)); /* Block is empty!! */
-
- in[p_preds++] = get_Phi_pred(phi_pred, j);
- } else {
- /* case Phi 2b: */
- in[p_preds++] = phi_pred;
- }
+ ir_node *pred_pred = get_Block_cfgpred(pred, j);
+
+ if (is_Bad(pred_pred)) {
+ in[p_preds++] = new_r_Bad(irg, get_irn_mode(phi));
+ continue;
+ }
+
+ if (get_nodes_block(phi_pred) == pred) {
+ /* case Phi 2a: */
+ assert(is_Phi(phi_pred)); /* Block is empty!! */
+
+ in[p_preds++] = get_Phi_pred(phi_pred, j);
+ } else {
+ /* case Phi 2b: */
+ in[p_preds++] = phi_pred;
}
}
} else {
in[p_preds++] = get_Phi_pred(phi, i);
}
}
- assert(p_preds <= max_preds);
+ assert(p_preds == max_preds);
/* Fix the node */
if (p_preds == 1)
- /* By removal of Bad ins the Phi might be degenerated. */
exchange(phi, in[0]);
else
set_irn_in(phi, p_preds, in);
Moreover, it is only needed if predb is the direct dominator of b,
else there can be no uses of the Phi's in predb ... -*/
for (k = 0, n = get_Block_n_cfgpreds(b); k < n; ++k) {
- ir_node *predb = get_nodes_block(get_Block_cfgpred(b, k));
-
- if (is_Bad(predb))
+ ir_node *pred = get_Block_cfgpred(b, k);
+ ir_node *predb = get_nodes_block(pred);
+ if (is_Bad(pred))
continue;
if (is_Block_removable(predb) && !Block_block_visited(predb)) {
if (get_Block_idom(b) != predb) {
/* predb is not the dominator. There can't be uses of pred's Phi nodes, kill them .*/
- ir_graph *irg = get_irn_irg(b);
- exchange(phi, get_irg_bad(irg));
+ ir_graph *irg = get_irn_irg(b);
+ ir_mode *mode = get_irn_mode(phi);
+ exchange(phi, new_r_Bad(irg, mode));
} else {
/* predb is the direct dominator of b. There might be uses of the Phi nodes from
predb in further block, so move this phi from the predecessor into the block b */
pred = get_Block_cfgpred_block(b, i);
if (is_Bad(pred)) {
- /* Do nothing */
+ in[q_preds++] = pred;
} 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++) {
/* 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++) {
/*- Fix the block -*/
n_preds = 0;
for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
- pred = get_Block_cfgpred_block(b, i);
+ ir_node *pred = get_Block_cfgpred(b, i);
+ ir_node *predb = get_nodes_block(pred);
+ ir_graph *irg = get_irn_irg(pred);
+ /* case 1: Bad predecessor */
if (is_Bad(pred)) {
- /* case 1: Do nothing */
- } else if (is_Block_removable(pred) && !Block_block_visited(pred)) {
+ in[n_preds++] = new_r_Bad(irg, mode_X);
+ continue;
+ }
+ if (is_Block_removable(predb) && !Block_block_visited(predb)) {
/* case 2: It's an empty block and not yet visited. */
- assert(get_Block_n_cfgpreds(b) > 1 || has_Block_entity(b));
- /* Else it should be optimized by equivalent_node. */
- for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
- ir_node *pred_X = get_Block_cfgpred(pred, j);
-
- /* because of breaking loops, not all predecessors are Bad-clean,
- * so we must check this here again */
- if (! is_Bad(pred_X))
- in[n_preds++] = pred_X;
+ for (j = 0; j < get_Block_n_cfgpreds(predb); j++) {
+ ir_node *predpred = get_Block_cfgpred(predb, j);
+
+ if (is_Bad(predpred)) {
+ in[n_preds++] = new_r_Bad(irg, mode_X);
+ continue;
+ }
+
+ in[n_preds++] = predpred;
}
- /* Remove block as it might be kept alive. */
- exchange(pred, b/*get_irg_bad(irg)*/);
+ /* Remove block+jump as it might be kept alive. */
+ exchange(pred, new_r_Bad(get_irn_irg(b), mode_X));
+ exchange(predb, new_r_Bad(get_irn_irg(b), mode_BB));
} else {
/* case 3: */
- in[n_preds++] = get_Block_cfgpred(b, i);
+ in[n_preds++] = pred;
}
}
- assert(n_preds <= max_preds);
+ assert(n_preds == max_preds);
set_irn_in(b, n_preds, in);
env->changed = true;
- assert(get_irn_link(b) == NULL || p_preds == -1 || (n_preds == p_preds && "Wrong Phi Fix"));
+ /* see if phi-fix was correct */
+ assert(get_irn_link(b) == NULL || p_preds == -1 || (n_preds == p_preds));
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.
*
long num = get_tarval_long(tv);
long def_num = get_Cond_default_proj(cond);
ir_graph *irg = get_irn_irg(cond);
- ir_node *bad = get_irg_bad(irg);
+ ir_node *bad = new_r_Bad(irg, mode_X);
if (def_num == get_Proj_proj(proj1)) {
/* first one is the defProj */
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.
- *
- * 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.
+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
*
- * We use the mark flag to mark removable blocks in the first phase.
+ * 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;
+}
+
+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_Cmp(n) || is_Proj(n)) {
+ /* ignore */
+ } else {
+ ir_node *block = get_nodes_block(n);
+ set_has_operations(block_info, block);
+ }
+}
+
+typedef struct skip_env {
+ bool changed;
+ ir_phase *phase;
+} skip_env;
+
+static void optimize_conds(ir_node *b, void *x)
+{
+ skip_env *env = (skip_env*)x;
+ int i, j;
+ int n_preds = get_Block_n_cfgpreds(b);
+
+ if (has_phis(env->phase,b)) 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);
+ }
+ }
+}
+
+static void remove_empty_blocks(ir_node *b, void *x)
+{
+ skip_env *env = (skip_env*)x;
+ int i;
+ int n_preds = get_Block_n_cfgpreds(b);
+
+ for (i = 0; i < n_preds; ++i) {
+ ir_node *jmp, *jmp_block, *pred, *pred_block;
+
+ jmp = get_Block_cfgpred(b, i);
+ if (!is_Jmp(jmp)) continue;
+ if (is_unknown_jump(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 (get_Block_n_cfgpreds(jmp_block) != 1) continue;
+ pred = get_Block_cfgpred(jmp_block, 0);
+ exchange(jmp, pred);
+ env->changed = true;
+
+ /* cleanup: jmp_block might have a Keep edge! */
+ pred_block = get_nodes_block(pred);
+ exchange(jmp_block, pred_block);
+ }
+}
+
+/*
+ * Some cfg optimizations, which do not touch Phi nodes */
+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);
+
+ for(;;) {
+ env.changed = false;
+
+ /* Conds => Jmp optimization; might produce empty blocks */
+ irg_block_walk_graph(irg, optimize_conds, NULL, &env);
+
+ /* Remove empty blocks */
+ irg_block_walk_graph(irg, remove_empty_blocks, NULL, &env);
+ if (env.changed) {
+ set_irg_doms_inconsistent(irg);
+ /* Removing blocks might enable more Cond 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;
* here. Fix the edges! */
edges_deactivate(irg);
+ 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;
+ env.changed = false;
+ env.phis_moved = 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, 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);
+ }
+ DEL_ARR_F(switch_conds);
+
+ if (!env.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) {
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);
}
-
- /* the verifier doesn't work yet with floating nodes */
- if (get_irg_pinned(irg) == op_pin_state_pinned) {
- /* after optimize_cf(), only Bad data flow may remain. */
- if (irg_verify_bads(irg, BAD_DF | BAD_BLOCK | TUPLE)) {
- dump_ir_graph(irg, "-verify-cf");
- fprintf(stderr, "VERIFY_BAD in optimize_cf()\n");
- }
- }
}
/* Creates an ir_graph pass for optimize_cf. */