* @brief Control flow optimizations.
* @author Goetz Lindenmaier, Michael Beck, Sebastian Hack
* @version $Id$
+ *
+ * Removes Bad control flow predecessors and empty blocks. A block is empty
+ * if it contains only a Jmp node. Blocks can only be removed if they are not
+ * needed for the semantics of Phi nodes. Further, we NEVER remove labeled
+ * blocks (even if we could move the label).
*/
#include "config.h"
#include "iroptimize.h"
#include <assert.h>
+#include <stdbool.h>
-#include "plist.h"
#include "xmalloc.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irflag_t.h"
#include "firmstat.h"
#include "irpass.h"
+#include "irphase_t.h"
#include "iropt_dbg.h"
-/*------------------------------------------------------------------*/
-/* Control flow optimization. */
-/* */
-/* Removes Bad control flow predecessors and empty blocks. A block */
-/* is empty if it contains only a Jmp node. */
-/* Blocks can only be removed if they are not needed for the */
-/* semantics of Phi nodes. */
-/* Further, we NEVER remove labeled blocks (even if we could move */
-/* the label. */
-/*------------------------------------------------------------------*/
-
-#define set_Block_removable(block) set_Block_mark(block, 1)
-#define set_Block_non_removable(block) set_Block_mark(block, 0)
-#define is_Block_removable(block) (get_Block_mark(block) != 0)
-
-/**
- * Replace binary Conds that jumps twice into the same block
- * by a simple Jmp.
- * E.g.
- * @verbatim
- * Cond Jmp Bad
- * / \ | /
- * ProjX True ProjX False ==> | /
- * \ / | /
- * Block Block
- * @endverbatim
- *
- * Such pattern are the result of if-conversion.
- *
- * Note that the simple case that Block has only these two
- * predecessors are already handled in equivalent_node_Block().
- */
-static int remove_senseless_conds(ir_node *bl)
-{
- int i, j;
- int n = get_Block_n_cfgpreds(bl);
- int changed = 0;
-
- for (i = 0; i < n; ++i) {
- ir_node *pred_i = get_Block_cfgpred(bl, i);
- ir_node *cond_i = skip_Proj(pred_i);
-
- /* binary Cond */
- if (is_Cond(cond_i) && get_irn_mode(get_Cond_selector(cond_i)) == mode_b) {
-
- for (j = i + 1; j < n; ++j) {
- ir_node *pred_j = get_Block_cfgpred(bl, j);
- ir_node *cond_j = skip_Proj(pred_j);
-
- if (cond_j == cond_i) {
- ir_graph *irg = get_irn_irg(bl);
- ir_node *jmp = new_r_Jmp(get_nodes_block(cond_i));
- set_irn_n(bl, i, jmp);
- set_irn_n(bl, j, new_r_Bad(irg));
-
- DBG_OPT_IFSIM2(cond_i, jmp);
- changed = 1;
- break;
- }
- }
- }
- }
- return changed;
-}
-
/** An environment for merge_blocks and collect nodes. */
typedef struct merge_env {
- int changed; /**< Set if the graph was changed. */
- int phis_moved; /**< Set if Phi nodes were moved. */
- plist_t *list; /**< Helper list for all found Switch Conds. */
+ bool changed; /**< Set if the graph was changed. */
+ bool phis_moved; /**< Set if Phi nodes were moved. */
} merge_env;
-/**
- * Removes Tuples from Block control flow predecessors.
- * Optimizes blocks with equivalent_node(). This is tricky,
- * as we want to avoid nodes that have as block predecessor Bads.
- * Therefore we also optimize at control flow operations, depending
- * how we first reach the Block.
- */
-static void merge_blocks(ir_node *node, void *ctx)
+/** set or reset the removable property of a block. */
+static void set_Block_removable(ir_node *block, bool removable)
{
- int i;
- ir_node *new_block;
- merge_env *env = ctx;
-
- /* clear the link field for ALL nodes first */
- set_irn_link(node, NULL);
-
- if (is_Block(node)) {
- /* Remove Tuples */
- for (i = get_Block_n_cfgpreds(node) - 1; i >= 0; --i) {
- ir_node *pred = get_Block_cfgpred(node, i);
- ir_node *skipped = skip_Tuple(pred);
- if (pred != skipped) {
- set_Block_cfgpred(node, i, skipped);
- env->changed = 1;
- }
- }
-
- /* see below */
- new_block = equivalent_node(node);
- if (new_block != node && ! is_Block_dead(new_block)) {
- exchange(node, new_block);
- env->changed = 1;
- }
-
- } else if (get_opt_optimize() && (get_irn_mode(node) == mode_X)) {
- /* We will soon visit a block. Optimize it before visiting! */
- ir_node *b = get_nodes_block(skip_Proj(node));
-
- if (!is_Block_dead(b)) {
- new_block = equivalent_node(b);
-
- while (!irn_visited(b) && !is_Block_dead(new_block) && new_block != b) {
- /* We would have to run gigo() if new is bad, so we
- promote it directly below. Nevertheless, we sometimes reach a block
- the first time through a dataflow node. In this case we optimized the
- block as such and have to promote the Bad here. */
- exchange(b, new_block);
- env->changed = 1;
- b = new_block;
- new_block = equivalent_node(b);
- }
-
- /* normally, we would create a Bad block here, but this must be
- * prevented, so just set it's cf to Bad.
- */
- if (is_Block_dead(new_block)) {
- ir_graph *irg = get_irn_irg(node);
- exchange(node, new_r_Bad(irg));
- env->changed = 1;
- }
- }
- }
+ set_Block_mark(block, removable);
}
-/**
- * Block walker removing control flow from dead block by
- * inspecting dominance info.
- * Do not replace blocks by Bad. This optimization shall
- * ensure, that all Bad control flow predecessors are
- * removed, and no new other Bads are introduced.
- * Further removed useless Conds and clear the mark of all blocks.
- *
- * Must be run in the post walker.
- */
-static void remove_unreachable_blocks_and_conds(ir_node *block, void *env)
+/** check if a block has the removable property set. */
+static bool is_Block_removable(ir_node *block)
{
- int i;
- int *changed = env;
-
- /* Check block predecessors and turn control flow into bad.
- Beware of Tuple, kill them. */
- for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) {
- ir_node *pred_X = get_Block_cfgpred(block, i);
- ir_node *skipped = skip_Tuple(pred_X);
-
- if (! is_Bad(skipped)) {
- ir_node *pred_bl = get_nodes_block(skip_Proj(skipped));
-
- if (is_Block_dead(pred_bl) || (get_Block_dom_depth(pred_bl) < 0)) {
- ir_graph *irg = get_irn_irg(block);
- set_Block_dead(pred_bl);
- exchange(pred_X, new_r_Bad(irg));
- *changed = 1;
- } else if (skipped != pred_X) {
- set_Block_cfgpred(block, i, skipped);
- *changed = 1;
- }
- }
- }
+ return get_Block_mark(block);
+}
- *changed |= remove_senseless_conds(block);
+/** 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;
+}
- /* clear the block mark of all non labeled blocks */
- if (has_Block_entity(block))
- set_Block_non_removable(block);
- else
- set_Block_removable(block);
+/** 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);
+ if (is_Block(node))
+ set_Block_removable(node, true);
}
/**
*/
static void collect_nodes(ir_node *n, void *ctx)
{
- ir_opcode code = get_irn_opcode(n);
- merge_env *env = ctx;
-
- if (code == iro_Block) {
- /* mark the block as non-removable if it is labeled */
- if (has_Block_entity(n))
- set_Block_non_removable(n);
+ ir_node ***switch_conds = (ir_node***)ctx;
+
+ if (is_Phi(n)) {
+ /* Collect Phi nodes to compact ins along with block's ins. */
+ ir_node *block = get_nodes_block(n);
+ set_irn_link(n, get_irn_link(block));
+ set_irn_link(block, n);
+ } else if (is_Block(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 {
- ir_node *b = get_nodes_block(n);
-
- if (code == iro_Phi && get_irn_arity(n) > 0) {
- /* Collect Phi nodes to compact ins along with block's ins. */
- set_irn_link(n, get_irn_link(b));
- set_irn_link(b, n);
- } else if (code != iro_Jmp && !is_Bad(b)) { /* Check for non-empty block. */
- set_Block_non_removable(b);
-
- if (code == iro_Proj) { /* link Proj nodes */
- ir_node *pred = get_Proj_pred(n);
-
- set_irn_link(n, get_irn_link(pred));
- set_irn_link(pred, n);
- } else if (code == iro_Cond) {
- ir_node *sel = get_Cond_selector(n);
- if (mode_is_int(get_irn_mode(sel))) {
- /* found a switch-Cond, collect */
- plist_insert_back(env->list, n);
- }
- }
+ /* 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)) {
+ /* link Proj nodes */
+ 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) && is_switch_Cond(n)) {
+ /* found a switch-Cond, collect */
+ ARR_APP1(ir_node*, *switch_conds, n);
}
}
}
-/** Returns true if pred is predecessor of block. */
-static int is_pred_of(ir_node *pred, ir_node *b)
+/** Returns true if pred is predecessor of block b. */
+static bool is_pred_of(ir_node *pred, ir_node *b)
{
int i;
for (i = get_Block_n_cfgpreds(b) - 1; i >= 0; --i) {
ir_node *b_pred = get_Block_cfgpred_block(b, i);
if (b_pred == pred)
- return 1;
+ return true;
}
- return 0;
+ return false;
}
/** Test whether we can optimize away pred block pos of b.
* To perform the test for pos, we must regard predecessors before pos
* as already removed.
**/
-static int test_whether_dispensable(ir_node *b, int pos)
+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_Block_dead(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_Block_dead(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_non_removable(pred);
+ 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
* for all nodes, not regarding whether there is a possibility for optimization.
*
* For each predecessor p of a Block b there are three cases:
- * -1. The predecessor p is a Bad node: just skip it. The in array of b shrinks by one.
- * -2. The predecessor p is empty. Remove p. All predecessors of p are now
- * predecessors of b.
- * -3. The predecessor p is a block containing useful code. Just keep p as is.
+ * - The predecessor p is a Bad node: just skip it. The in array of b shrinks
+ * by one.
+ * - The predecessor p is empty. Remove p. All predecessors of p are now
+ * predecessors of b.
+ * - The predecessor p is a block containing useful code. Just keep p as is.
*
* For Phi nodes f we have to check the conditions at the Block of f.
* For cases 1 and 3 we proceed as for Blocks. For case 2 we can have two
* cases:
- * -2a: The old predecessor of the Phi f is a Phi pred_f IN THE BLOCK REMOVED. In this
- * case we proceed as for blocks. We remove pred_f. All
- * predecessors of pred_f now are predecessors of f.
- * -2b: The old predecessor of f is NOT in the block removed. It might be a Phi, too.
- * We have to replicate f for each predecessor of the removed block. Or, with
- * other words, the removed predecessor block has exactly one predecessor.
+ * -2a: The old predecessor of the Phi f is a Phi pred_f IN THE BLOCK REMOVED.
+ * In this case we proceed as for blocks. We remove pred_f. All
+ * predecessors of pred_f now are predecessors of f.
+ * -2b: The old predecessor of f is NOT in the block removed. It might be a Phi
+ * too. We have to replicate f for each predecessor of the removed block.
+ * Or, with other words, the removed predecessor block has exactly one
+ * predecessor.
*
* Further there is a special case for self referencing blocks:
* @verbatim
* If there is a Phi in pred_b, but we remove pred_b, we have to generate a
* Phi in loop_b, that has the ins of the Phi in pred_b and a self referencing
* backedge.
- * @@@ It is negotiable whether we should do this ... there might end up a copy
- * from the Phi in the loop when removing the Phis.
*/
static void optimize_blocks(ir_node *b, void *ctx)
{
int i, j, k, n, max_preds, n_preds, p_preds = -1;
ir_node *pred, *phi, *next;
ir_node **in;
- merge_env *env = ctx;
+ 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. */
in = XMALLOCN(ir_node*, max_preds);
/*- Fix the Phi nodes of the current block -*/
- for (phi = get_irn_link(b); phi != NULL; phi = next) {
+ for (phi = (ir_node*)get_irn_link(b); phi != NULL; phi = (ir_node*)next) {
assert(is_Phi(phi));
- next = get_irn_link(phi);
+ next = (ir_node*)get_irn_link(phi);
/* 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_Block_dead(pred)) {
- /* case Phi 1: Do nothing */
+ if (is_Bad(pred)) {
+ /* 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);
- env->changed = 1;
+ env->changed = true;
}
/*- This happens only if merge between loop backedge and single loop entry.
- 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 ... -*/
+ 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));
+ 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)) {
ir_node *next_phi;
/* we found a predecessor block at position k that will be removed */
- for (phi = get_irn_link(predb); phi; phi = next_phi) {
+ for (phi = (ir_node*)get_irn_link(predb); phi; phi = next_phi) {
int q_preds = 0;
- next_phi = get_irn_link(phi);
+ next_phi = (ir_node*)get_irn_link(phi);
assert(is_Phi(phi));
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, new_r_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 */
set_nodes_block(phi, b);
set_irn_link(phi, get_irn_link(b));
set_irn_link(b, phi);
- env->phis_moved = 1;
+ env->phis_moved = true;
/* first, copy all 0..k-1 predecessors */
for (i = 0; i < k; i++) {
pred = get_Block_cfgpred_block(b, i);
- if (is_Block_dead(pred)) {
- /* Do nothing */
+ if (is_Bad(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;
/* 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 */
for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
pred = get_Block_cfgpred_block(b, i);
- if (is_Block_dead(pred)) {
- /* Do nothing */
+ if (is_Bad(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;
exchange(phi, in[0]);
else
set_irn_in(phi, q_preds, in);
- env->changed = 1;
+ env->changed = true;
assert(q_preds <= max_preds);
// assert(p_preds == q_preds && "Wrong Phi Fix");
/*- Fix the block -*/
n_preds = 0;
for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
- pred = get_Block_cfgpred_block(b, i);
-
- if (is_Block_dead(pred)) {
- /* case 1: Do nothing */
- } else if (is_Block_removable(pred) && !Block_block_visited(pred)) {
+ 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)) {
+ 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/*new_r_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 = 1;
+ 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 that with only a Jmp predecessor.
- */
-static void remove_simple_blocks(ir_node *block, void *ctx)
-{
- ir_node *new_blk = equivalent_node(block);
- merge_env *env = ctx;
-
- if (new_blk != block) {
- exchange(block, new_blk);
- env->changed = 1;
- }
-}
-
-/**
- * Handle pre-optimized table switch Cond's.
- * During iropt, all Projs from a switch-Cond are already removed except
- * the defProj and maybe the taken one.
- * The defProj cannot be removed WITHOUT looking backwards, so we do this here.
+ * Optimize table-switch Conds.
*
* @param cond the switch-Cond
- *
- * @return non-zero if a switch-Cond was optimized
- *
- * Expects all Proj's linked to the cond node
+ * @return true if the switch-Cond was optimized
*/
-static int handle_switch_cond(ir_node *cond)
+static bool handle_switch_cond(ir_node *cond)
{
- ir_node *sel = get_Cond_selector(cond);
-
- ir_node *proj1 = get_irn_link(cond);
- ir_node *proj2 = get_irn_link(proj1);
- ir_node *jmp, *blk;
-
- blk = get_nodes_block(cond);
+ ir_node *sel = get_Cond_selector(cond);
+ ir_node *proj1 = (ir_node*)get_irn_link(cond);
+ ir_node *proj2 = (ir_node*)get_irn_link(proj1);
+ ir_node *blk = get_nodes_block(cond);
+ /* exactly 1 Proj on the Cond node: must be the defaultProj */
if (proj2 == NULL) {
- /* this Cond has only one Proj: must be the defProj */
+ ir_node *jmp = new_r_Jmp(blk);
assert(get_Cond_default_proj(cond) == get_Proj_proj(proj1));
/* convert it into a Jmp */
- jmp = new_r_Jmp(blk);
exchange(proj1, jmp);
- return 1;
- } else if (get_irn_link(proj2) == NULL) {
- /* We have two Proj's here. Check if the Cond has
- a constant argument */
+ return true;
+ }
+
+ /* 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 + 1 case
+ * (this one case should be the one taken) */
+ if (get_irn_link(proj2) == NULL) {
ir_tarval *tv = value_of(sel);
if (tv != tarval_bad) {
long num = get_tarval_long(tv);
long def_num = get_Cond_default_proj(cond);
ir_graph *irg = get_irn_irg(cond);
+ ir_node *bad = new_r_Bad(irg, mode_X);
if (def_num == get_Proj_proj(proj1)) {
/* first one is the defProj */
if (num == get_Proj_proj(proj2)) {
- jmp = new_r_Jmp(blk);
+ ir_node *jmp = new_r_Jmp(blk);
exchange(proj2, jmp);
- exchange(proj1, new_r_Bad(irg));
- return 1;
+ exchange(proj1, bad);
+ return true;
}
} else if (def_num == get_Proj_proj(proj2)) {
/* second one is the defProj */
if (num == get_Proj_proj(proj1)) {
- jmp = new_r_Jmp(blk);
+ ir_node *jmp = new_r_Jmp(blk);
exchange(proj1, jmp);
- exchange(proj2, new_r_Bad(irg));
- return 1;
+ exchange(proj2, bad);
+ return true;
}
} else {
/* neither: strange, Cond was not optimized so far */
if (num == get_Proj_proj(proj1)) {
- jmp = new_r_Jmp(blk);
+ ir_node *jmp = new_r_Jmp(blk);
exchange(proj1, jmp);
- exchange(proj2, new_r_Bad(irg));
- return 1;
+ exchange(proj2, bad);
+ return true;
} else if (num == get_Proj_proj(proj2)) {
- jmp = new_r_Jmp(blk);
+ ir_node *jmp = new_r_Jmp(blk);
exchange(proj2, jmp);
- exchange(proj1, new_r_Bad(irg));
- return 1;
+ exchange(proj1, bad);
+ return true;
}
}
}
}
- return 0;
+ 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, changed;
+ int i, j, n;
ir_node **in = NULL;
- ir_node *cond, *end = get_irg_end(irg);
- plist_element_t *el;
+ ir_node *end = get_irg_end(irg);
+ 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);
- /* we use the mark flag to mark removable blocks */
- ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK);
-restart:
- env.changed = 0;
- env.phis_moved = 0;
+ /* First the "simple" optimizations, which do not touch Phis */
+ cfgopt_ignoring_phis(irg);
- /* ALWAYS kill unreachable control flow. Backend cannot handle it anyway.
- Use dominator info to kill blocks. Also optimize useless Conds. */
- assure_doms(irg);
- irg_block_walk_graph(irg, NULL, remove_unreachable_blocks_and_conds, &env.changed);
-
- /* fix the keep-alives */
- changed = 0;
- for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
- ir_node *ka = get_End_keepalive(end, i);
-
- if (is_Block(ka)) {
- /* do NOT keep dead blocks */
- if (is_Block_dead(ka) || get_Block_dom_depth(ka) < 0) {
- set_End_keepalive(end, i, new_r_Bad(irg));
- changed = 1;
- }
- } else {
- ir_node *block = get_nodes_block(ka);
+ /* we use the mark flag to mark removable blocks */
+ ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK);
- if (is_Bad(block) || is_Block_dead(block) || get_Block_dom_depth(block) < 0) {
- /* do NOT keep nodes in dead blocks */
- set_End_keepalive(end, i, new_r_Bad(irg));
- changed = 1;
- }
- }
- }
- env.changed |= changed;
+ /* The switch Cond optimization might expose unreachable code, so we loop */
+ for (;;) {
+ int length;
+ ir_node **switch_conds = NULL;
+ bool changed = false;
- ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
+ assure_doms(irg);
- env.list = plist_new();
- irg_walk(end, merge_blocks, collect_nodes, &env);
+ /*
+ * 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);
- ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+ 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);
- env.changed = 0;
}
- /* handle all collected switch-Conds */
- foreach_plist(env.list, el) {
- cond = plist_element_get_value(el);
- env.changed |= handle_switch_cond(cond);
- }
- plist_free(env.list);
-
- if (env.changed) {
- /* The Cond optimization might generate unreachable code, so restart if
- it happens. */
- goto restart;
- }
+ /* assert due to collect_nodes:
+ * 1. removable blocks are now marked as such
+ * 2. phi lists are up to date
+ */
- /* Optimize the standard code. */
- env.changed = 0;
+ /* 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);
-
- /* in rare cases a node may be kept alive more than once, use the visited flag to detect this */
- ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
- inc_irg_visited(irg);
-
- /* fix the keep-alives again */
- changed = 0;
- for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
- ir_node *ka = get_End_keepalive(end, i);
-
- if (is_Block(ka)) {
- /* do NOT keep dead blocks */
- if (is_Block_dead(ka) || get_Block_dom_depth(ka) < 0) {
- set_End_keepalive(end, i, new_r_Bad(irg));
- changed = 1;
- }
- } else {
- ir_node *block = get_nodes_block(ka);
+ irg_block_walk_graph(irg, optimize_blocks, NULL, &env);
- if (is_Bad(block) || is_Block_dead(block) || get_Block_dom_depth(block) < 0) {
- /* do NOT keep nodes in dead blocks */
- set_End_keepalive(end, i, new_r_Bad(irg));
- changed = 1;
- }
- }
+ new_end = optimize_in_place(end);
+ if (new_end != end) {
+ set_irg_end(irg, new_end);
+ end = new_end;
}
- env.changed |= changed;
-
remove_End_Bads_and_doublets(end);
-
- ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_VISITED);
+ ir_free_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK);
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 (j != n) {
set_End_keepalives(end, j, in);
- env.changed = 1;
+ env.changed = true;
}
}
}
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. */
ir_graph_pass_t *optimize_cf_pass(const char *name)
{
return def_graph_pass(name ? name : "optimize_cf", optimize_cf);
-} /* optimize_cf_pass */
+}