Simplify collect_and_rewire_users(): Use get_edge_src_pos() instead of searching...
[libfirm] / ir / opt / cfopt.c
index 30910cb..7ca15da 100644 (file)
@@ -21,7 +21,6 @@
  * @file
  * @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
@@ -47,7 +46,6 @@
 #include "irdump.h"
 #include "irverify.h"
 #include "iredges.h"
-#include "opt_manage.h"
 
 #include "array_t.h"
 
@@ -109,14 +107,13 @@ static void clear_link_and_mark_blocks_removable(ir_node *node, void *ctx)
  */
 static void collect_nodes(ir_node *n, void *ctx)
 {
-       ir_node ***switch_conds = (ir_node***)ctx;
-
+       (void) ctx;
        if (is_Phi(n)) {
                /* Collect Phi nodes to compact ins along with block's ins. */
                ir_node *block = get_nodes_block(n);
                add_Block_phi(block, n);
        } else if (is_Block(n)) {
-               if (has_Block_entity(n)) {
+               if (get_Block_entity(n) != NULL) {
                        /* block with a jump label attached cannot be removed. */
                        set_Block_removable(n, false);
                }
@@ -134,9 +131,6 @@ static void collect_nodes(ir_node *n, void *ctx)
                        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);
                }
        }
 }
@@ -239,6 +233,26 @@ non_dispensable:
        return 1;
 }
 
+/**
+ * This method merges blocks. A block is applicable to be merged, if it
+ * has only one predecessor with an unconditional jump to this block;
+ * and if this block does not contain any phis.
+ */
+static void merge_blocks(ir_node *b, void *env)
+{
+       (void) env;
+
+       if (get_Block_n_cfgpreds(b) == 1) {
+               ir_node* pred = get_Block_cfgpred(b, 0);
+               if (is_Jmp(pred)) {
+                       ir_node* pred_block = get_nodes_block(pred);
+                       if (get_Block_phis(b) == NULL) {
+                               exchange(b, pred_block);
+                       }
+               }
+       }
+}
+
 /**
  * This method removes empty blocks.  A block is empty if it only contains Phi
  * and Jmp nodes.
@@ -513,77 +527,6 @@ static void optimize_blocks(ir_node *b, void *ctx)
        xfree(in);
 }
 
-/**
- * Optimize table-switch Conds.
- *
- * @param cond the switch-Cond
- * @return true if the switch-Cond was optimized
- */
-static bool handle_switch_cond(ir_node *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) {
-               ir_node *jmp = new_r_Jmp(blk);
-               assert(get_Cond_default_proj(cond) == get_Proj_proj(proj1));
-               /* convert it into a Jmp */
-               exchange(proj1, jmp);
-               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) {
-                       /* we have a constant switch */
-                       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)) {
-                                       ir_node *jmp = new_r_Jmp(blk);
-                                       exchange(proj2, jmp);
-                                       exchange(proj1, bad);
-                                       return true;
-                               }
-                       } else if (def_num == get_Proj_proj(proj2)) {
-                               /* second one is the defProj */
-                               if (num == get_Proj_proj(proj1)) {
-                                       ir_node *jmp = new_r_Jmp(blk);
-                                       exchange(proj1, jmp);
-                                       exchange(proj2, bad);
-                                       return true;
-                               }
-                       } else {
-                               /* neither: strange, Cond was not optimized so far */
-                               if (num == get_Proj_proj(proj1)) {
-                                       ir_node *jmp = new_r_Jmp(blk);
-                                       exchange(proj1, jmp);
-                                       exchange(proj2, bad);
-                                       return true;
-                               } else if (num == get_Proj_proj(proj2)) {
-                                       ir_node *jmp = new_r_Jmp(blk);
-                                       exchange(proj2, jmp);
-                                       exchange(proj1, bad);
-                                       return true;
-                               }
-                       }
-               }
-       }
-       return false;
-}
-
 /**
  * Optimize boolean Conds, where true and false jump to the same block into a Jmp
  * Block must contain no Phi nodes.
@@ -629,13 +572,13 @@ typedef enum block_flags_t {
 static bool get_block_flag(const ir_nodehashmap_t *infos, const ir_node *block,
                            int flag)
 {
-       return PTR_TO_INT(ir_nodehashmap_get(infos, block)) & flag;
+       return PTR_TO_INT(ir_nodehashmap_get(void, infos, block)) & flag;
 }
 
 static void set_block_flag(ir_nodehashmap_t *infos, ir_node *block,
                            block_flags_t flag)
 {
-       int data = PTR_TO_INT(ir_nodehashmap_get(infos, block));
+       int data = PTR_TO_INT(ir_nodehashmap_get(void, infos, block));
        data |= flag;
        ir_nodehashmap_insert(infos, block, INT_TO_PTR(data));
 }
@@ -744,7 +687,7 @@ static void remove_empty_blocks(ir_node *block, void *x)
        int n_preds = get_Block_n_cfgpreds(block);
 
        for (i = 0; i < n_preds; ++i) {
-               ir_node *jmp, *jmp_block, *pred, *pred_block;
+               ir_node *jmp, *jmp_block;
                int n_jpreds = 0;
 
                jmp = get_Block_cfgpred(block, i);
@@ -755,8 +698,6 @@ static void remove_empty_blocks(ir_node *block, void *x)
                        continue; /* this infinite loop cannot be optimized any further */
                if (is_unknown_jump_target(&env->block_infos, jmp_block))
                        continue; /* unknown jump target must not be optimized */
-               if (has_operations(&env->block_infos,jmp_block))
-                       continue; /* this block contains operations and cannot be skipped */
                if (has_phis(&env->block_infos,jmp_block))
                        continue; /* this block contains Phis and is not skipped */
                if (Block_block_visited(jmp_block)) {
@@ -783,6 +724,15 @@ static void remove_empty_blocks(ir_node *block, void *x)
                 * if block has no Phis.
                 */
                if (n_jpreds == 1) {
+                       ir_node *pred        = get_Block_cfgpred(jmp_block, 0);
+                       ir_node *pred_block  = get_nodes_block(pred);
+                       if (has_operations(&env->block_infos,jmp_block)) {
+                               if (get_irg_start_block(get_irn_irg(pred_block)) == pred_block)
+                                       continue; /* must not merge operations into start block */
+                               if (!is_Jmp(pred))
+                                       continue; /* must not create partially dead code, especially when it is mode_M */
+                       }
+
                        /* skip jmp block by rerouting its predecessor to block
                         *
                         *     A              A
@@ -791,14 +741,14 @@ static void remove_empty_blocks(ir_node *block, void *x)
                         *     |              |
                         *   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->block_infos, block)) {
+               } else if ( !has_phis(&env->block_infos, block) &&
+                           !has_operations(&env->block_infos,jmp_block))
+               {
                        /* all predecessors can skip the jmp block, so block gets some new
                         * predecessors
                         *
@@ -817,7 +767,7 @@ static void remove_empty_blocks(ir_node *block, void *x)
                        }
                        /* now append the new predecessors */
                        for (j = 0; j < n_jpreds; ++j) {
-                               pred = get_Block_cfgpred(jmp_block, j);
+                               ir_node *pred = get_Block_cfgpred(jmp_block, j);
                                ins[n_preds+j] = pred;
                        }
                        set_irn_in(block, n_preds+n_jpreds, ins);
@@ -858,12 +808,13 @@ static void cfgopt_ignoring_phis(ir_graph *irg)
                irg_block_walk_graph(irg, NULL, optimize_ifs, &env);
 
                if (env.changed) {
-                       clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE);
+                       confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
                        /* clear block info, because it must be recomputed */
                        irg_block_walk_graph(irg, clear_block_info, NULL, &env.block_infos);
                        /* Removing blocks and Conds might enable more optimizations */
                        continue;
                } else {
+                       confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
                        break;
                }
        }
@@ -872,7 +823,7 @@ static void cfgopt_ignoring_phis(ir_graph *irg)
 }
 
 /* Optimizations of the control flow that also require changes of Phi nodes.  */
-static ir_graph_state_t do_cfopt(ir_graph *irg)
+void optimize_cf(ir_graph *irg)
 {
        int i, j, n;
        ir_node **in = NULL;
@@ -883,13 +834,11 @@ static ir_graph_state_t do_cfopt(ir_graph *irg)
        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");
 
-       edges_deactivate(irg);
+       assure_irg_properties(irg, IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE);
 
        /* First the "simple" optimizations, which do not touch Phis */
        cfgopt_ignoring_phis(irg);
@@ -898,38 +847,13 @@ static ir_graph_state_t do_cfopt(ir_graph *irg)
        ir_reserve_resources(irg, IR_RESOURCE_BLOCK_MARK | IR_RESOURCE_IRN_LINK
                             | IR_RESOURCE_PHI_LIST);
 
-       /* The switch Cond optimization might expose unreachable code, so we loop */
-       for (;;) {
-               int length;
-               ir_node **switch_conds = NULL;
-               bool changed = false;
-
-               assure_doms(irg);
-
-               /*
-                * 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;
-
-               clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE
-                                  | IR_GRAPH_STATE_VALID_EXTENDED_BLOCKS
-                                  | IR_GRAPH_STATE_CONSISTENT_ENTITY_USAGE);
-       }
+       /*
+        * 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.
+        */
+       irg_walk(end, clear_link_and_mark_blocks_removable, collect_nodes, NULL);
 
        /* assert due to collect_nodes:
         * 1. removable blocks are now marked as such
@@ -937,11 +861,11 @@ static ir_graph_state_t do_cfopt(ir_graph *irg)
         */
 
        /* 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, NULL, &env);
+        * 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_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
+       irg_block_walk_graph(irg, optimize_blocks, merge_blocks, &env);
 
        new_end = optimize_in_place(end);
        if (new_end != end) {
@@ -989,18 +913,8 @@ static ir_graph_state_t do_cfopt(ir_graph *irg)
                }
        }
 
-       return 0;
-}
-
-static optdesc_t opt_cf = {
-       "control-flow",
-       IR_GRAPH_STATE_NO_UNREACHABLE_CODE,
-       do_cfopt,
-};
-
-void optimize_cf(ir_graph *irg)
-{
-       perform_irg_optimization(irg, &opt_cf);
+       confirm_irg_properties(irg,
+               env.changed ? IR_GRAPH_PROPERTIES_NONE : IR_GRAPH_PROPERTIES_ALL);
 }
 
 /* Creates an ir_graph pass for optimize_cf. */