optimize polymorphic field accesses
[libfirm] / ir / opt / ifconv.c
index e989a30..daafa17 100644 (file)
@@ -5,10 +5,22 @@
  * @date 4.2.2005
  * $Id$
  */
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
 
+#ifdef HAVE_STDLIB_H
 #include <stdlib.h>
+#endif
+#ifdef HAVE_STRING_H
 #include <string.h>
+#endif
+#ifdef HAVE_ALLOCA_H
 #include <alloca.h>
+#endif
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
 
 #include "irgraph_t.h"
 #include "irnode_t.h"
@@ -27,6 +39,7 @@
 #include "set.h"
 #include "bitset.h"
 #include "bitfiddle.h"
+#include "irhooks.h"
 
 #define MAX_DEPTH                              20
 
@@ -234,43 +247,52 @@ static opt_if_conv_info_t default_info = {
 static firm_dbg_module_t *dbg;
 
 /**
- * A simple check for sde effects upton an opcode of a ir node.
+ * A simple check for side effects upto an opcode of a ir node.
  * @param irn The ir node to check,
  * @return 1 if the opcode itself may produce side effects, 0 if not.
  */
 static INLINE int has_side_effects(const ir_node *irn)
 {
-       opcode opc = get_irn_opcode(irn);
+       ir_op *op = get_irn_op(irn);
 
-       if(opc == iro_Cmp)
+       if (op == op_Cmp)
                return 0;
 
        return !mode_is_datab(get_irn_mode(irn));
 }
 
+enum failure_reason_t {
+  SUCCESS      = IF_RESULT_SUCCESS,
+  TO_DEEP      = IF_RESULT_TOO_DEEP,
+  SIDE_EFFECTS = IF_RESULT_SIDE_EFFECT,
+  PHI_FOUND    = IF_RESULT_SIDE_EFFECT_PHI
+};
+
 /**
- * Decdies, if a given expression and its subexpressions
+ * Decides, if a given expression and its subexpressions
  * (to certain, also given extent) can be moved to a block.
- * @param expr The expression to examine.
- * @param block The block where the expression should go.
- * @param depth The current depth, passed recursively. Use 0 for
- * non-recursive calls.
+ *
+ * @param expr      The expression to examine.
+ * @param block     The block where the expression should go.
+ * @param depth     The current depth, passed recursively. Use 0 for
+ *                  non-recursive calls.
  * @param max_depth The maximum depth to which the expression should be
  * examined.
+ *
+ * @return a failure reason
  */
 static int _can_move_to(ir_node *expr, ir_node *dest_block, int depth, int max_depth)
 {
        int i, n;
-       int res = 1;
+       int res = SUCCESS;
        ir_node *expr_block = get_nodes_block(expr);
 
-
        /*
         * If we are forced to look too deep into the expression,
         * treat it like it could not be moved.
         */
        if(depth >= max_depth) {
-               res = 0;
+               res = TO_DEEP;
                goto end;
        }
 
@@ -280,34 +302,34 @@ static int _can_move_to(ir_node *expr, ir_node *dest_block, int depth, int max_d
         * has side effects or anything else. It is executed on each
         * path the destination block is reached.
         */
-       if(block_dominates(expr_block, dest_block))
+       if (block_dominates(expr_block, dest_block))
                goto end;
 
        /*
         * We cannot move phis!
         */
-       if(is_Phi(expr)) {
-               res = 0;
+       if (is_Phi(expr)) {
+               res = PHI_FOUND;
                goto end;
        }
 
        /*
-        * This should be superflous and could be converted into a assertion.
+        * This should be superfluous and could be converted into a assertion.
         * The destination block _must_ dominate the block of the expression,
         * else the expression could be used without its definition.
         */
-       if(!block_dominates(dest_block, expr_block)) {
-               res = 0;
+       if (! block_dominates(dest_block, expr_block)) {
+               res = IF_RESULT_SIDE_EFFECT;
                goto end;
        }
 
        /*
         * Surely, if the expression does not have a data mode, it is not
-        * movable. Perhaps onw should also test the floating property of
+        * movable. Perhaps one should also test the floating property of
         * the opcode/node.
         */
-       if(has_side_effects(expr)) {
-               res = 0;
+       if (has_side_effects(expr)) {
+               res = IF_RESULT_SIDE_EFFECT;
                goto end;
        }
 
@@ -316,13 +338,14 @@ static int _can_move_to(ir_node *expr, ir_node *dest_block, int depth, int max_d
         * check them out. If one of them cannot be moved, this one
         * cannot be moved either.
         */
-       for(i = 0, n = get_irn_arity(expr); i < n; ++i) {
+       for (i = 0, n = get_irn_arity(expr); i < n; ++i) {
                ir_node *op = get_irn_n(expr, i);
                int new_depth = is_Proj(op) ? depth : depth + 1;
-               if(!_can_move_to(op, dest_block, new_depth, max_depth)) {
-                       res = 0;
+
+    res = _can_move_to(op, dest_block, new_depth, max_depth);
+
+    if (res != SUCCESS)
                        goto end;
-               }
        }
 
 end:
@@ -338,17 +361,24 @@ end:
  * deeper into an expression than a given threshold to examine if
  * it can be moved, the expression is rejected and the test returns
  * false.
- * @param expr The expression to check for.
+ *
+ * @param expr       The expression to check for.
  * @param dest_block The destination block you want @p expr to be.
- * @param max_depth The maximum depth @p expr should be investigated.
- * @return 1, if the expression can be moved to the destination block,
- * 0 if not.
+ * @param max_depth  The maximum depth @p expr should be investigated.
+ *
+ * @return return a failure reason
  */
 static INLINE int can_move_to(ir_node *expr, ir_node *dest_block, int max_depth)
 {
        return _can_move_to(expr, dest_block, 0, max_depth);
 }
 
+/**
+ * move a DAG given by a root node expr into a new block
+ *
+ * @param expr       the root of a dag
+ * @param dest_block the destination block
+ */
 static void move_to(ir_node *expr, ir_node *dest_block)
 {
        int i, n;
@@ -358,15 +388,18 @@ static void move_to(ir_node *expr, ir_node *dest_block)
         * If we reached the dominator, we are done.
         * We will never put code through the dominator
         */
-       if(block_dominates(expr_block, dest_block))
+       if (block_dominates(expr_block, dest_block))
                return;
 
-       for(i = 0, n = get_irn_arity(expr); i < n; ++i)
+       for (i = 0, n = get_irn_arity(expr); i < n; ++i)
                move_to(get_irn_n(expr, i), dest_block);
 
        set_nodes_block(expr, dest_block);
 }
 
+/**
+ * return the common dominator of two blocks
+ */
 static INLINE ir_node *common_idom(ir_node *b1, ir_node *b2)
 {
        if(block_dominates(b1, b2))
@@ -376,21 +409,20 @@ static INLINE ir_node *common_idom(ir_node *b1, ir_node *b2)
        else {
                ir_node *p;
 
-               for(p = b1; !block_dominates(p, b2); p = get_Block_idom(p));
+               for (p = get_Block_idom(b1); !block_dominates(p, b2); p = get_Block_idom(p));
                return p;
        }
 }
 
-
 /**
  * Information about a cond node.
  */
 typedef struct _cond_t {
        ir_node *cond;                                  /**< The cond node. */
-       struct list_head list;  /**< List head which is used for queueing this cond
+       struct list_head list;  /**< List head which is used for queuing this cond
                                                                                                                into the cond bunch it belongs to. */
-       unsigned in_list : 1;
        unsigned is_new : 1;
+       unsigned totally_covers : 1;
        struct _cond_t *link;
        long visited_nr;
 
@@ -410,6 +442,9 @@ typedef struct _cond_t {
        } cases[2];
 } cond_t;
 
+/**
+ * retrieve the conditional information from a Cond node
+ */
 static INLINE cond_t *get_cond(ir_node *irn, set *cond_set)
 {
        cond_t templ;
@@ -445,10 +480,11 @@ static void _walk_conds(cond_t *cond, cond_walker_t *pre, cond_walker_t *post,
                post(cond, env);
 }
 
+static long cond_visited_nr = 0;
+
 static void walk_conds(cond_t *cond, cond_walker_t *pre, cond_walker_t *post, void *env)
 {
-       static long visited_nr = 0;
-       _walk_conds(cond, pre, post, ++visited_nr, env);
+       _walk_conds(cond, pre, post, ++cond_visited_nr, env);
 }
 
 static void link_conds(cond_t *cond, void *env)
@@ -482,12 +518,13 @@ static int cond_cmp(const void *a, const void *b, size_t size)
 typedef struct _cond_info_t {
        struct list_head list;                  /**< Used to list all of these structs per class. */
 
-       struct list_head roots;                 /**< A list of non-depending conds. Two conds are
-                                                                                                                               independent, if yot can not reach the one from the
-                                                                                                                               other (all conds in this list have to dominate the
-                                                                                                                               block this struct is attached to. */
+       struct list_head roots;                 /**< A list of non-depending Conds. Two Conds are
+                                                                                                                               independent, if it's not possible not reach one from the
+                                                                                                                               other (all Conds in this list have to dominate the
+                                                                                                                               block this struct is attached to). */
 
-       set *cond_set;                                                  /**< A set of all dominating reachable conds. */
+       ir_node *first_phi;                                     /**< The first phi node this cond info was made for. */
+       set *cond_set;                                                  /**< A set of all dominating reachable Conds. */
 } cond_info_t;
 
 /**
@@ -497,7 +534,7 @@ static void _find_conds(ir_node *irn, long visited_nr,
                ir_node *dominator, cond_t *masked_by, int pos, int depth, cond_info_t *ci)
 {
        ir_node *block;
-       int is_modeb_cond = 0;
+       int saw_select_cond = 0;
 
        block = get_nodes_block(irn);
 
@@ -505,8 +542,8 @@ static void _find_conds(ir_node *irn, long visited_nr,
         * Only check this block if it is dominated by the specified
         * dominator or it has not been visited yet.
         */
-       if(block_dominates(dominator, block) && get_Block_block_visited(block) < visited_nr) {
-               cond_t *res = NULL;
+       if (block_dominates(dominator, block) && get_Block_block_visited(block) < visited_nr) {
+               cond_t *res = masked_by;
                int i, n;
 
                /* check, if we're on a ProjX
@@ -521,9 +558,11 @@ static void _find_conds(ir_node *irn, long visited_nr,
                        ir_node *cond = get_Proj_pred(irn);
 
                        /* true, if the mode is a mode_b cond _NO_ switch cond */
-                       is_modeb_cond = get_irn_opcode(cond) == iro_Cond
+                       int is_modeb_cond = get_irn_opcode(cond) == iro_Cond
                                && get_irn_mode(get_Cond_selector(cond)) == mode_b;
 
+      saw_select_cond = !is_modeb_cond;
+
                        /* Check, if the pred of the proj is a Cond
                         * with a Projb as selector.
                         */
@@ -539,13 +578,15 @@ static void _find_conds(ir_node *irn, long visited_nr,
                                /* get or insert the cond info into the set. */
                                res = set_insert(ci->cond_set, &c, sizeof(c), HASH_PTR(cond));
 
+                               /*
+                                * If this cond is already masked by the masked_by cond
+                                * return immediately, since we don't have anything to add.
+                                */
+                               if(masked_by && res->cases[proj].masked_by == masked_by)
+                                       return;
+
                                if(res->is_new) {
                                        res->is_new = 0;
-                                       INIT_LIST_HEAD(&res->list);
-                               }
-
-                               if(!res->in_list) {
-                                       res->in_list = 1;
                                        list_add(&res->list, &ci->roots);
                                }
 
@@ -566,18 +607,19 @@ static void _find_conds(ir_node *irn, long visited_nr,
                                 * Since the masked_by nodes masks a cond, remove it from the
                                 * root list of the conf trees.
                                 */
-                               else if(!list_empty(&masked_by->list)) {
+                               else {
+                                       assert(res->cases[proj].pos < 0);
                                        list_del_init(&masked_by->list);
                                }
 
-                               DBG((dbg, LEVEL_2, "%{firm:indent}found cond %n (%s branch) "
+                               DBG((dbg, LEVEL_2, "%D%n (%s branch) "
                                                        "for pos %d in block %n reached by %n\n",
-                                                       depth, cond, get_Proj_proj(irn) ? "true" : "false", pos,
+                                                       depth, cond, proj ? "true" : "false", pos,
                                                        block, masked_by ? masked_by->cond : NULL));
                        }
                }
 
-               if(get_Block_block_visited(block) < visited_nr) {
+               if(get_Block_block_visited(block) < visited_nr && !saw_select_cond) {
 
                        set_Block_block_visited(block, visited_nr);
 
@@ -591,7 +633,7 @@ static void _find_conds(ir_node *irn, long visited_nr,
                                 * as given by the caller. We also increase the depth for the
                                 * recursively called functions.
                                 */
-                               _find_conds(pred, visited_nr, dominator, res, pos, depth + 1, ci);
+                               _find_conds(pred, visited_nr, dominator, res, pos, depth + (res != masked_by), ci);
                        }
                }
        }
@@ -602,9 +644,10 @@ static void _find_conds(ir_node *irn, long visited_nr,
  * A convenience function for _find_conds.
  * It sets some parameters needed for recursion to appropriate start
  * values. Always use this function.
- * @param irn The node to start looking for conds from. This might
- *     be the phi node we are investigating.
- * @param conds The set to record the found conds in.
+ *
+ * @param irn   The node to start looking for Conds from. This might
+ *                 be the phi node we are investigating.
+ * @param conds The set to record the found Conds in.
  */
 static INLINE void find_conds(ir_node *irn, cond_info_t *ci)
 {
@@ -613,19 +656,18 @@ static INLINE void find_conds(ir_node *irn, cond_info_t *ci)
        ir_node *block = get_nodes_block(irn);
        ir_node *dom = get_Block_idom(block);
 
-
        for(i = 0, n = get_irn_arity(block); i < n; ++i) {
                ir_node *pred = get_irn_n(block, i);
 
                inc_irg_block_visited(current_ir_graph);
                visited_nr = get_irg_block_visited(current_ir_graph);
                set_Block_block_visited(block, visited_nr);
+
                DBG((dbg, LEVEL_2, "find conds at pred %d (%n) and idom %n\n", i, pred, dom));
                _find_conds(pred, visited_nr, dom, NULL, i, 0, ci);
        }
 }
 
-
 /**
  * Make the mux for a given cond.
  * @param phi The phi node which shall be replaced by a mux.
@@ -634,14 +676,15 @@ static INLINE void find_conds(ir_node *irn, cond_info_t *ci)
  * @return The mux node made for this cond.
  */
 static ir_node *make_mux_on_demand(ir_node *phi, ir_node *dom, cond_t *cond,
-               int max_depth, ir_node **mux, bitset_t *positions)
+               int max_depth, ir_node **mux, bitset_t *positions, int *muxes_made, long visited_nr)
 {
-       int i;
+       int i, can_move[2];
        ir_node *projb = get_Cond_selector(cond->cond);
        ir_node *bl = get_nodes_block(cond->cond);
        ir_node *operands[2];
        int set[2];
 
+       cond->visited_nr = visited_nr;
        DBG((dbg, LEVEL_2, "%n\n", cond->cond));
        for(i = 0; i < 2; ++i) {
                cond_t *masked_by = cond->cases[i].masked_by;
@@ -651,13 +694,15 @@ static ir_node *make_mux_on_demand(ir_node *phi, ir_node *dom, cond_t *cond,
                set[i] = -1;
 
                /*
-                * If this cond branch is masked by another cond, make the mux
-                * for that cond first, since the mux for this cond takes
+                * If this Cond branch is masked by another cond, make the mux
+                * for that Cond first, since the Mux for this cond takes
                 * it as an operand.
                 */
                if(masked_by) {
                        assert(pos < 0);
-                       operands[i] = make_mux_on_demand(phi, dom, masked_by, max_depth, mux, positions);
+                       DBG((dbg, LEVEL_2, "\tmasked by: %n\n", masked_by->cond));
+                       if(masked_by->visited_nr < visited_nr)
+                               operands[i] = make_mux_on_demand(phi, dom, masked_by, max_depth, mux, positions, muxes_made, visited_nr);
                }
 
                /*
@@ -672,28 +717,61 @@ static ir_node *make_mux_on_demand(ir_node *phi, ir_node *dom, cond_t *cond,
 
        /*
         * Move the operands to the dominator block if the cond
-        * made sense. Some conds found are not suitable for making a mux
+        * made sense. Some Conds found are not suitable for making a mux
         * out of them, since one of their branches cannot be reached from
         * the phi block. In that case we do not make a mux and return NULL.
         */
-       if(operands[0] && operands[1]
-                       && can_move_to(operands[0], bl, max_depth)
-                       && can_move_to(operands[1], bl, max_depth)) {
-
-               move_to(operands[0], bl);
-               move_to(operands[1], bl);
-
-               /* Make the mux. */
-               *mux = new_r_Mux(current_ir_graph, bl, projb,
-                               operands[0], operands[1], get_irn_mode(operands[0]));
-
-               DBG((dbg, LEVEL_2, "\t%n(%n, %n, %n)[%d, %d]\n",
-                                       *mux, projb, operands[0], operands[1], set[0], set[1]));
-
-               for(i = 0; i < 2; ++i)
-                       if(set[i] >= 0)
-                               bitset_set(positions, set[i]);
+  if(operands[0] && operands[1]) {
+    if (operands[0] == operands[1]) {
+      /* there is no gain in using mux in this case, as
+         it will be optimized away. We will NOT move the
+         content of the blocks either
+        */
+      for (i = 0; i < 2; ++i)
+        if(set[i] >= 0)
+          bitset_set(positions, set[i]);
+
+      *mux = operands[0];
+      return *mux;
+    }
+
+               can_move[0] = can_move_to(operands[0], bl, max_depth);
+               can_move[1] = can_move_to(operands[1], bl, max_depth);
+
+    if (can_move[0] == SUCCESS && can_move[1] == SUCCESS) {
+                 move_to(operands[0], bl);
+                 move_to(operands[1], bl);
+
+                 /* Make the mux. */
+                 *mux = new_r_Mux(current_ir_graph, bl, projb,
+                                 operands[0], operands[1], get_irn_mode(operands[0]));
+
+                 *muxes_made += 1;
+
+                 DBG((dbg, LEVEL_2, "\t%n(%n, %n, %n)[%d, %d]\n",
+                                         *mux, projb, operands[0], operands[1], set[0], set[1]));
+
+                 for(i = 0; i < 2; ++i)
+        if(set[i] >= 0) {
+                                 bitset_set(positions, set[i]);
+
+          /* we have done one */
+          hook_if_conversion(current_ir_graph, phi, set[i], *mux, IF_RESULT_SUCCESS);
+        }
+    }
+    else {
+      if(can_move[0] != SUCCESS)
+        hook_if_conversion(current_ir_graph, phi, set[0], NULL, can_move[0]);
+      if(can_move[1] != SUCCESS)
+        hook_if_conversion(current_ir_graph, phi, set[1], NULL, can_move[1]);
+    }
        }
+  else {
+    if(operands[0] != SUCCESS)
+      hook_if_conversion(current_ir_graph, phi, set[0], NULL, IF_RESULT_BAD_CF);
+    if(operands[1] != SUCCESS)
+      hook_if_conversion(current_ir_graph, phi, set[1], NULL, IF_RESULT_BAD_CF);
+  }
 
        return *mux;
 }
@@ -710,7 +788,7 @@ typedef struct _phi_info_t {
  * @param irn A phi node.
  * @param info Parameters for the if conversion algorithm.
  */
-static void check_out_phi(phi_info_t *phi_info, opt_if_conv_info_t *info)
+static int check_out_phi(phi_info_t *phi_info, opt_if_conv_info_t *info)
 {
        int max_depth = info->max_depth;
        ir_node *irn = phi_info->irn;
@@ -718,6 +796,7 @@ static void check_out_phi(phi_info_t *phi_info, opt_if_conv_info_t *info)
        cond_info_t *cond_info = phi_info->cond_info;
        cond_t *cond;
        int i, arity;
+       int muxes_made = 0;
        bitset_t *positions;
 
        block = get_nodes_block(irn);
@@ -755,12 +834,11 @@ static void check_out_phi(phi_info_t *phi_info, opt_if_conv_info_t *info)
                }
 
                DBG((dbg, LEVEL_2, "\tcommon idom: %n\n", cidom));
-               make_mux_on_demand(irn, cidom, cond, max_depth, &mux, positions);
+               make_mux_on_demand(irn, cidom, cond, max_depth, &mux, positions, &muxes_made, ++cond_visited_nr);
 
                if(mux) {
-                       bitset_foreach(positions, pos) {
+                       bitset_foreach(positions, pos)
                                set_irn_n(irn, (int) pos, mux);
-                       }
                }
        }
 
@@ -771,6 +849,8 @@ static void check_out_phi(phi_info_t *phi_info, opt_if_conv_info_t *info)
        nw = optimize_in_place_2(irn);
        if(nw != irn)
                exchange(irn, nw);
+
+       return muxes_made;
 }
 
 typedef struct _cond_walk_info_t {
@@ -804,6 +884,8 @@ static void annotate_cond_info_post(ir_node *irn, void *data)
                if(!ci) {
                        ci = obstack_alloc(cwi->obst, sizeof(*ci));
                        ci->cond_set = new_set(cond_cmp, log2_ceil(get_irn_arity(block)));
+                       ci->first_phi = irn;
+
                        INIT_LIST_HEAD(&ci->roots);
                        INIT_LIST_HEAD(&ci->list);
 
@@ -835,7 +917,7 @@ static void annotate_cond_info_post(ir_node *irn, void *data)
                }
 
                else
-                       DBG((dbg, LEVEL_2, "conds already computed for %n\n", irn));
+                       DBG((dbg, LEVEL_2, "conds already computed for %n (look at %n)\n", irn, ci->first_phi));
 
                set_irn_link(block, ci);
 
@@ -895,21 +977,9 @@ static void vcg_dump_conds(ir_graph *irg, cond_walk_info_t *cwi)
        }
 }
 
-#if 0
-/**
- * Free the sets which are put at some blocks.
- */
-static void free_sets(ir_node *irn, void *data)
-{
-       if(is_Block(irn) && get_irn_link(irn)) {
-               set *conds = get_irn_link(irn);
-               del_set(conds);
-       }
-}
-#endif
-
 void opt_if_conv(ir_graph *irg, opt_if_conv_info_t *params)
 {
+       int muxes_made = 0;
        struct obstack obst;
        phi_info_t *phi_info;
        cond_info_t *cond_info;
@@ -929,13 +999,13 @@ void opt_if_conv(ir_graph *irg, opt_if_conv_info_t *params)
        /* Init the debug stuff. */
        dbg = firm_dbg_register("firm.opt.ifconv");
 #if 0
-       firm_dbg_set_mask(dbg, LEVEL_1 | LEVEL_2 | LEVEL_3);
+       firm_dbg_set_mask(dbg, LEVEL_1);
 #endif
 
        /* Ensure, that the dominators are computed. */
        compute_doms(irg);
 
-       DBG((dbg, LEVEL_2, "if conversion for irg %s(%p)\n",
+       DBG((dbg, LEVEL_1, "if conversion for irg %s(%p)\n",
                                get_entity_name(get_irg_entity(irg)), irg));
 
        /*
@@ -953,13 +1023,15 @@ void opt_if_conv(ir_graph *irg, opt_if_conv_info_t *params)
 
        /* Process each suitable phi found. */
        list_for_each_entry(phi_info_t, phi_info, &cwi.phi_head, list) {
-               DBG((dbg, LEVEL_4, "phi node %n\n", phi_info->irn));
-               check_out_phi(phi_info, p);
+               DBG((dbg, LEVEL_2, "phi node %n\n", phi_info->irn));
+               muxes_made += check_out_phi(phi_info, p);
        }
 
        list_for_each_entry(cond_info_t, cond_info, &cwi.cond_info_head, list) {
                del_set(cond_info->cond_set);
        }
 
+       DBG((dbg, LEVEL_1, "muxes made: %d\n", muxes_made));
+
        obstack_free(&obst, NULL);
 }