*/
#ifdef HAVE_CONFIG_H
-# include <config.h>
+# include "config.h"
#endif
#include <assert.h>
+#include "xmalloc.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "irprog_t.h"
#include "firmstat.h"
#include "cfopt.h"
+#include "iropt_dbg.h"
/*------------------------------------------------------------------*/
/* Control flow optimization. */
/* semantics of Phi nodes. */
/*------------------------------------------------------------------*/
+/**
+ * Replace binary Conds that jumps twice into the same block
+ * by a simple Jmp.
+ * E.g.
+ * @code
+ * Cond Jmp Bad
+ * / \ | /
+ * ProjX True ProjX False ==> | /
+ * \ / | /
+ * Block Block
+ * @endcode
+ *
+ * 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 void remove_senseless_conds(ir_node *bl, void *data)
+{
+ int i, j;
+ int n = get_Block_n_cfgpreds(bl);
+
+ assert(is_Block(bl));
+
+ for (i = 0; i < n; ++i) {
+ ir_node *pred_i = get_Block_cfgpred(bl, i);
+ ir_node *cond_i = skip_Proj(pred_i);
+
+ 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
+ && get_irn_op(cond_i) == op_Cond
+ && get_irn_mode(get_Cond_selector(cond_i)) == mode_b) {
+
+ ir_node *jmp = new_r_Jmp(current_ir_graph, get_nodes_block(cond_i));
+ set_irn_n(bl, i, jmp);
+ set_irn_n(bl, j, new_Bad());
+
+ DBG_OPT_IFSIM2(cond_i, jmp);
+ break;
+ }
+ }
+ }
+}
+
+
/**
* Removes Tuples from Block control flow predecessors.
* Optimizes blocks with equivalent_node(). This is tricky,
if (get_opt_normalize())
set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
}
+
+ /* see below */
new_block = equivalent_node(n);
- if (new_block != n)
+ if (new_block != n && ! is_Bad(new_block))
exchange (n, new_block);
} else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
/* We will soon visit a block. Optimize it before visiting! */
- ir_node *b = get_nodes_block(n);
+ ir_node *b = get_nodes_block(skip_Proj(n));
if (!is_Bad(b)) {
new_block = equivalent_node(b);
while (irn_not_visited(b) && (!is_Bad(new_block)) && (new_block != b)) {
- /* We would have to run gigo if new is bad, so we
+ /* 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. */
- assert(((b == new_block) ||
- get_opt_control_flow_straightening() ||
- get_opt_control_flow_weak_simplification()) &&
- ("strange flag setting"));
+ assert((get_opt_control_flow_straightening() ||
+ get_opt_control_flow_weak_simplification()) &&
+ ("strange flag setting"));
exchange (b, new_block);
b = new_block;
new_block = equivalent_node(b);
}
- b = new_block;
- }
- /*
- * BEWARE: do not kill floating notes here as they might be needed in
- * valid blocks because of global CSE.
- */
- if (is_Bad(b) && get_opt_normalize() &&
- get_op_pinned(get_irn_op(n)) == op_pin_state_pinned)
- exchange(n, new_Bad());
+ /* normally, we would create a Bad block here, but this must be
+ * prevented, so just set it's cf to Bad.
+ */
+ if (is_Bad(new_block))
+ exchange(n, new_Bad());
+ }
}
}
/**
- * Remove dead block by inspecting dominance info
+ * Remove cf 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.
+ *
+ * Must be run in the post walker.
*/
-static void remove_dead_blocks(ir_node *block, void *env) {
- /* delete dead blocks: if we have dominator information, this can easily be detected.
- * Here, new Bad blocks my be introduced.
- *
- * BEWARE: don't kill the end block */
- if (block != get_irg_end_block(current_ir_graph) &&
- get_Block_dom_depth(block) == -1 &&
- get_opt_unreachable_code()) {
- exchange (block, new_Bad());
+static void remove_dead_block_cf(ir_node *block, void *env)
+{
+ int i, n;
+
+ /* check block predecessors and turn control flow into bad */
+ for (i = 0, n = get_Block_n_cfgpreds(block); i < n; ++i) {
+ ir_node *pred_X = get_Block_cfgpred(block, i);
+
+ if (! is_Bad(pred_X)) {
+ ir_node *pred_bl = get_nodes_block(skip_Proj(pred_X));
+
+ if (is_Bad(pred_bl) || (get_Block_dom_depth(pred_bl) == -1))
+ exchange (pred_X, new_Bad());
+ }
}
}
if (is_no_Block(n)) {
ir_node *b = get_nodes_block(n);
- /*
- * BEWARE: do not kill floating notes here as they might be needed in
- * valid blocks because of global CSE.
- */
- if (is_Bad(b) &&
- get_op_pinned(get_irn_op(n)) == op_pin_state_pinned) {
- /* previous merge_blocks() may have killed dead blocks */
- exchange(n, new_Bad());
- }
- else if ((get_irn_op(n) == op_Phi)) {
+ if ((get_irn_op(n) == op_Phi)) {
/* Collect Phi nodes to compact ins along with block's ins. */
set_irn_link(n, get_irn_link(b));
set_irn_link(b, n);
}
-/** Test wether we can optimize away pred block pos of b.
+/** Test whether we can optimize away pred block pos of b.
*
* @param b A block node.
* @param pos The position of the predecessor block to judge about.
* requires that a copy is added before the merge. We have to
* keep one of the case blocks to place the copies in.
*
- * To perform the test for pos, we must regard preds before pos
+ * 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) {
ir_node *cfop = get_Block_cfgpred(b, pos);
ir_node *pred = get_nodes_block(cfop);
+ /* Bad blocks will be optimized away, so we don't need space for them */
+ if (is_Bad(pred))
+ return 0;
+
if (get_Block_block_visited(pred) + 1
< get_irg_block_visited(current_ir_graph)) {
/**
- * This method removed Bad cf preds from Blocks and Phis, and removes
+ * 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.
*
* We first adapt Phi nodes, then Block nodes, as we need the old ins
* 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 precessor of the Phi f is a Phi pred_f IN THE BLOCK REMOVED. In this
+ * 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.
for (i = 0, k = get_Block_n_cfgpreds(b); i < k; ++i) {
max_preds += test_whether_dispensable(b, i);
}
- in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
+ in = xmalloc(max_preds * sizeof(*in));
/*-
printf(" working on "); DDMN(b);
Therefore we replace the old phi by the new one.
Further we have to remove the old Phi node by replacing it
- by Bad. Else it will remain in the keepalive array of End
+ by Bad. Else it will remain in the keep alive array of End
and cause illegal situations. So if there is no loop, we should
replace it by Bad.
*/
in[p_preds++] = get_Phi_pred(phi, i);
}
}
+ assert(p_preds <= max_preds);
/* Fix the node */
if (p_preds == 1)
else
set_irn_in(phi, q_preds, in);
-// assert(p_preds == q_preds && "Wrong Phi Fix");
+ assert(q_preds <= max_preds);
+// assert(p_preds == q_preds && "Wrong Phi Fix");
}
phi = get_irn_link(phi);
}
in[n_preds++] = get_Block_cfgpred(b, i);
}
}
+ assert(n_preds <= max_preds);
+
set_irn_in(b, n_preds, in);
assert(get_irn_link(b) == NULL || (n_preds == p_preds && "Wrong Phi Fix"));
- free(in);
+ xfree(in);
}
* phase.
* @@@ It would be better to add a struct in the link field
* that keeps the Phi list and the mark. Place it on an obstack, as
- * we will lose blocks and thereby generate mem leaks.
+ * we will lose blocks and thereby generate memory leaks.
*/
void optimize_cf(ir_graph *irg) {
- int i;
+ int i, n;
ir_node **in;
ir_node *end = get_irg_end(irg);
ir_graph *rem = current_ir_graph;
if (get_irg_dom_state(current_ir_graph) == dom_consistent)
set_irg_dom_inconsistent(current_ir_graph);
- if (dom_state == dom_consistent) {
- /* we have dominace info, we can kill dead block */
- irg_block_walk(get_irg_end_block(irg), NULL, remove_dead_blocks, NULL);
+ if (dom_state == dom_consistent && get_opt_optimize() && get_opt_unreachable_code()) {
+ ir_node *end = get_irg_end(irg);
+
+ /* we have dominance info, we can kill dead block */
+ irg_block_walk_graph(irg, NULL, remove_dead_block_cf, NULL);
+
+ /* fix the keep-alives */
+ for (i = 0, n = get_End_n_keepalives(end); i < n; ++i) {
+ ir_node *ka = get_End_keepalive(end, i);
+
+ if (is_Block(ka) && (get_Block_dom_depth(ka) == -1))
+ set_End_keepalive(end, i, new_Bad());
+ if (is_Phi(ka) && (get_Block_dom_depth(get_nodes_block(ka)) == -1))
+ set_End_keepalive(end, i, new_Bad());
+ }
}
+ irg_block_walk_graph(current_ir_graph, NULL, remove_senseless_conds, NULL);
+
/* Use block visited flag to mark non-empty blocks. */
inc_irg_block_visited(irg);
irg_walk(end, merge_blocks, collect_nodes, NULL);
in[0] = get_nodes_block(end);
inc_irg_visited(current_ir_graph);
- for(i = 0; i < get_End_n_keepalives(end); i++) {
+ for (i = 0; i < get_End_n_keepalives(end); i++) {
ir_node *ka = get_End_keepalive(end, i);
if (irn_not_visited(ka)) {
/* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
end->in = in;
- /* after optimize_cf(), only Bad data flow may remain. */
- if (irg_vrfy_bads(irg, BAD_DF | BAD_BLOCK | TUPLE)) {
- dump_ir_block_graph(irg, "-vrfy-cf");
- dump_ir_graph(irg, "-vrfy-cf");
- fprintf(stderr, "VRFY_BAD in optimize_cf()\n");
+
+ /* 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_vrfy_bads(irg, BAD_DF | BAD_BLOCK | TUPLE)) {
+ dump_ir_block_graph(irg, "-vrfy-cf");
+ dump_ir_graph(irg, "-vrfy-cf");
+ fprintf(stderr, "VRFY_BAD in optimize_cf()\n");
+ }
}
current_ir_graph = rem;
projects / libfirm / blobdiff