condeval is called jump threading now
[libfirm] / ir / opt / code_placement.c
index 5227145..577ad98 100644 (file)
@@ -25,9 +25,7 @@
  *           Michael Beck
  * @version  $Id$
  */
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
 
 #include "adt/pdeq.h"
 #include "irnode_t.h"
@@ -68,7 +66,7 @@ place_floats_early(ir_node *n, waitq *worklist) {
        int i, irn_arity;
 
        /* we must not run into an infinite loop */
-       assert(irn_not_visited(n));
+       assert(!irn_visited(n));
        mark_irn_visited(n);
 
        /* Place floating nodes. */
@@ -92,7 +90,7 @@ place_floats_early(ir_node *n, waitq *worklist) {
                        ir_node *pred = get_irn_n(n, i);
                        ir_node *pred_block;
 
-                       if ((irn_not_visited(pred))
+                       if (!irn_visited(pred)
                            && (get_irn_pinned(pred) == op_pin_state_floats)) {
 
                                /*
@@ -127,7 +125,8 @@ place_floats_early(ir_node *n, waitq *worklist) {
                                b = pred_block;
                                depth = get_Block_dom_depth(pred_block);
                        }
-                       /* Avoid that the node is placed in the Start block */
+                       /* Avoid that the node is placed in the Start block if we are not in the
+                          backend phase. */
                        if (depth == 1 &&
                                        get_Block_dom_depth(get_nodes_block(n)) > 1 &&
                                        get_irg_phase_state(current_ir_graph) != phase_backend) {
@@ -153,7 +152,7 @@ place_floats_early(ir_node *n, waitq *worklist) {
                 */
                for (i = -1; i < irn_arity; ++i) {
                        ir_node *pred = get_irn_n(n, i);
-                       if (irn_not_visited(pred))
+                       if (!irn_visited(pred))
                                waitq_put(worklist, pred);
                }
        } else if (is_Block(n)) {
@@ -163,7 +162,7 @@ place_floats_early(ir_node *n, waitq *worklist) {
                 */
                for (i = irn_arity - 1; i >= 0; --i) {
                        ir_node *pred = get_irn_n(n, i);
-                       if (irn_not_visited(pred))
+                       if (!irn_visited(pred))
                                waitq_put(worklist, pred);
                }
        } else if (is_Phi(n)) {
@@ -176,13 +175,13 @@ place_floats_early(ir_node *n, waitq *worklist) {
                 * of the Phi-input if the Phi is not in a bad block.
                 */
                pred = get_nodes_block(n);
-               if (irn_not_visited(pred))
+               if (!irn_visited(pred))
                        waitq_put(worklist, pred);
 
                for (i = irn_arity - 1; i >= 0; --i) {
                        ir_node *pred = get_irn_n(n, i);
 
-                       if (irn_not_visited(pred)) {
+                       if (!irn_visited(pred)) {
                                if (! in_dead_block &&
                                        get_irn_pinned(pred) == op_pin_state_floats &&
                                        is_Block_unreachable(get_nodes_block(pred))) {
@@ -200,13 +199,13 @@ place_floats_early(ir_node *n, waitq *worklist) {
                 * All other nodes: move nodes from dead blocks into the same block.
                 */
                pred = get_nodes_block(n);
-               if (irn_not_visited(pred))
+               if (!irn_visited(pred))
                        waitq_put(worklist, pred);
 
                for (i = irn_arity - 1; i >= 0; --i) {
                        ir_node *pred = get_irn_n(n, i);
 
-                       if (irn_not_visited(pred)) {
+                       if (!irn_visited(pred)) {
                                if (! in_dead_block &&
                                        get_irn_pinned(pred) == op_pin_state_floats &&
                                        is_Block_unreachable(get_nodes_block(pred))) {
@@ -236,18 +235,22 @@ static void place_early(waitq *worklist) {
        /* Work the content of the worklist. */
        while (!waitq_empty(worklist)) {
                ir_node *n = waitq_get(worklist);
-               if (irn_not_visited(n))
+               if (!irn_visited(n))
                        place_floats_early(n, worklist);
        }
-
-       set_irg_outs_inconsistent(current_ir_graph);
        set_irg_pinned(current_ir_graph, op_pin_state_pinned);
 }
 
 /**
- * Compute the deepest common ancestor of block and dca.
+ * Compute the deepest common dominator tree ancestor of block and dca.
+ *
+ * @param dca    the deepest common dominator tree ancestor so far,
+ *               might be NULL
+ * @param block  a block
+ *
+ * @return  the deepest common dominator tree ancestor of block and dca
  */
-static ir_node *calc_dca(ir_node *dca, ir_node *block) {
+static ir_node *calc_dom_dca(ir_node *dca, ir_node *block) {
        assert(block);
 
        /* we do not want to place nodes in dead blocks */
@@ -268,11 +271,11 @@ static ir_node *calc_dca(ir_node *dca, ir_node *block) {
        while (block != dca) {
                block = get_Block_idom(block); dca = get_Block_idom(dca);
        }
-
        return dca;
 }
 
-/** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
+/**
+ * Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
  * I.e., DCA is the block where we might place PRODUCER.
  * A data flow edge points from producer to consumer.
  */
@@ -292,20 +295,17 @@ static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *produ
                                ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
 
                                if (!is_Block_unreachable(new_block))
-                                       dca = calc_dca(dca, new_block);
+                                       dca = calc_dom_dca(dca, new_block);
                        }
                }
        } else {
-               dca = calc_dca(dca, get_nodes_block(consumer));
+               dca = calc_dom_dca(dca, get_nodes_block(consumer));
        }
-
        return dca;
 }
 
-/* FIXME: the name clashes here with the function from ana/field_temperature.c
- * please rename. */
-static INLINE int get_irn_loop_depth(ir_node *n) {
-       return get_loop_depth(get_irn_loop(n));
+static inline int get_block_loop_depth(ir_node *block) {
+       return get_loop_depth(get_irn_loop(block));
 }
 
 /**
@@ -329,25 +329,25 @@ static void move_out_of_loops(ir_node *n, ir_node *early) {
        while (dca != early) {
                dca = get_Block_idom(dca);
                if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
-               if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
+               if (get_block_loop_depth(dca) < get_block_loop_depth(best)) {
                        best = dca;
                }
        }
-       if (best != get_nodes_block(n)) {
-               /* debug output
-               printf("Moving out of loop: "); DDMN(n);
-               printf(" Outermost block: "); DDMN(early);
-               printf(" Best block: "); DDMN(best);
-               printf(" Innermost block: "); DDMN(get_nodes_block(n));
-               */
+       if (best != get_nodes_block(n))
                set_nodes_block(n, best);
-       }
 }
 
-/* deepest common ancestor in the dominator tree of all nodes'
-   blocks depending on us; our final placement has to dominate DCA. */
-static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
-{
+/**
+ * Calculate the deepest common ancestor in the dominator tree of all nodes'
+ * blocks depending on node; our final placement has to dominate DCA.
+ *
+ * @param node  the definition node
+ * @param dca   the deepest common ancestor block so far, initially
+ *              NULL
+ *
+ * @return the deepest common dominator ancestor of all blocks of node's users
+ */
+static ir_node *get_deepest_common_dom_ancestor(ir_node *node, ir_node *dca) {
        int i;
 
        for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
@@ -362,7 +362,9 @@ static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
                }
 
                if (is_Proj(succ)) {
-                       dca = get_deepest_common_ancestor(succ, dca);
+                       /* Proj nodes are in the same block as node, so
+                        * the users of Proj are our users. */
+                       dca = get_deepest_common_dom_ancestor(succ, dca);
                } else {
                        /* ignore if succ is in dead code */
                        ir_node *succ_blk = get_nodes_block(succ);
@@ -371,12 +373,16 @@ static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
                        dca = consumer_dom_dca(dca, succ, node);
                }
        }
-
        return dca;
 }
 
-static void set_projs_block(ir_node *node, ir_node *block)
-{
+/**
+ * Put all the Proj nodes of a node into a given block.
+ *
+ * @param node   the mode_T node
+ * @param block  the block to put the Proj nodes to
+ */
+static void set_projs_block(ir_node *node, ir_node *block) {
        int i;
 
        for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
@@ -384,7 +390,7 @@ static void set_projs_block(ir_node *node, ir_node *block)
 
                assert(is_Proj(succ));
 
-               if(get_irn_mode(succ) == mode_T) {
+               if (get_irn_mode(succ) == mode_T) {
                        set_projs_block(succ, block);
                }
                set_nodes_block(succ, block);
@@ -407,7 +413,7 @@ static void place_floats_late(ir_node *n, pdeq *worklist) {
        int i, n_outs;
        ir_node *early_blk;
 
-       assert(irn_not_visited(n)); /* no multiple placement */
+       assert(!irn_visited(n)); /* no multiple placement */
 
        mark_irn_visited(n);
 
@@ -435,7 +441,7 @@ static void place_floats_late(ir_node *n, pdeq *worklist) {
                    producer of one of their inputs in the same block anyway. */
                for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
                        ir_node *succ = get_irn_out(n, i);
-                       if (irn_not_visited(succ) && !is_Phi(succ))
+                       if (!irn_visited(succ) && !is_Phi(succ))
                                place_floats_late(succ, worklist);
                }
 
@@ -453,7 +459,7 @@ static void place_floats_late(ir_node *n, pdeq *worklist) {
                                /* deepest common ancestor in the dominator tree of all nodes'
                                   blocks depending on us; our final placement has to dominate
                                   DCA. */
-                               ir_node *dca = get_deepest_common_ancestor(n, NULL);
+                               ir_node *dca = get_deepest_common_dom_ancestor(n, NULL);
                                if (dca != NULL) {
                                        set_nodes_block(n, dca);
                                        move_out_of_loops(n, early_blk);
@@ -470,7 +476,7 @@ static void place_floats_late(ir_node *n, pdeq *worklist) {
        n_outs = get_irn_n_outs(n);
        for (i = 0; i < n_outs; i++) {
                ir_node *succ = get_irn_out(n, i);
-               if (irn_not_visited(get_irn_out(n, i))) {
+               if (!irn_visited(succ)) {
                        pdeq_putr(worklist, succ);
                }
        }
@@ -492,7 +498,7 @@ static void place_late(waitq *worklist) {
        /* And now empty the worklist again... */
        while (!waitq_empty(worklist)) {
                ir_node *n = waitq_get(worklist);
-               if (irn_not_visited(n))
+               if (!irn_visited(n))
                        place_floats_late(n, worklist);
        }
 }
@@ -503,6 +509,7 @@ void place_code(ir_graph *irg) {
        ir_graph *rem = current_ir_graph;
 
        current_ir_graph = irg;
+       remove_critical_cf_edges(irg);
 
        /* Handle graph state */
        assert(get_irg_phase_state(irg) != phase_building);
@@ -518,15 +525,15 @@ void place_code(ir_graph *irg) {
        worklist = new_waitq();
        place_early(worklist);
 
-       /* place_early() invalidates the outs, place_late needs them. */
-       compute_irg_outs(irg);
+       /* Note: place_early changes only blocks, no data edges. So, the
+        * data out edges are still valid, no need to recalculate them here. */
 
        /* Now move the nodes down in the dominator tree. This reduces the
           unnecessary executions of the node. */
        place_late(worklist);
 
-       set_irg_outs_inconsistent(current_ir_graph);
-       set_irg_loopinfo_inconsistent(current_ir_graph);
+       set_irg_outs_inconsistent(irg);
+       set_irg_loopinfo_inconsistent(irg);
        del_waitq(worklist);
        current_ir_graph = rem;
 }