Added is_Const
[libfirm] / ir / opt / ifconv.c
index 5a59e0e..0c907d3 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"
 #include "set.h"
 #include "bitset.h"
 #include "bitfiddle.h"
+#include "irhooks.h"
+#include "return.h"
 
-#define MAX_DEPTH                              20
-
-/*
- * Mux optimization routines.
- */
-
-#if 0
-static ir_node *local_optimize_mux(ir_node *mux)
-{
-       int i, n;
-       ir_node *res = mux;
-  ir_node *sel = get_Mux_sel(mux);
-       ir_node *cmp = skip_Proj(sel);
-
-       /* Optimize the children  */
-       for(i = 1, n = get_irn_arity(mux); i < n; ++i) {
-               ir_node *operand = get_irn_n(mux, i);
-               if(get_irn_op(operand) == op_Mux)
-                       optimize_mux(operand);
-       }
-
-       /* If we have no cmp above the mux, get out. */
-       if(is_Proj(sel) && get_irn_mode(sel) == mode_b && get_irn_opcode(cmp) == iro_Cmp) {
-
-               pn_Cmp cc = get_Proj_proj(sel);
-               ir_mode *mode = get_irn_mode(mux);
-               ir_node *block = get_nodes_block(n);
-               ir_node *cmp_left = get_Cmp_left(cmp);
-               ir_node *cmp_right = get_Cmp_right(cmp);
-               ir_node *mux_true = get_Mux_true(mux);
-               ir_node *mux_false = get_Mux_false(mux);
-
-               /*
-                * Check for comparisons with signed integers.
-                */
-               if(mode_is_int(mode)                                    /* We need an integral mode */
-                               && mode_is_signed(mode)   /* which is signed */
-                               && cc == Lt) {                                          /* and have to compare for < */
-
-                       /*
-                        * Mux(x:T < 0, -1, 0) -> Shrs(x, sizeof_bits(T) - 1)
-                        * Conditions:
-                        * T must be signed.
-                        */
-                       if(classify_Const(cmp_right) == CNST_NULL
-                               && classify_Const(mux_true) == CNST_ALL_ONE
-                               && classify_Const(mux_false) == CNST_NULL) {
-
-                               ir_mode *u_mode = find_unsigned_mode(mode);
-
-                               res = new_r_Shrs(current_ir_graph, block, cmp_left,
-                                               new_r_Const_long(current_ir_graph, block, u_mode,
-                                                       get_mode_size_bits(mode) - 1),
-                                               mode);
-                       }
-
-                       /*
-                        * Mux(0 < x:T, 1, 0) -> Shr(-x, sizeof_bits(T) - 1)
-                        * Conditions:
-                        * T must be signed.
-                        */
-                       else if(classify_Const(cmp_left) == CNST_NULL
-                               && classify_Const(mux_true) == CNST_ONE
-                               && classify_Const(mux_false) == CNST_NULL) {
-
-                               ir_mode *u_mode = find_unsigned_mode(mode);
-
-                               res = new_r_Shr(current_ir_graph, block,
-
-                                               /* -x goes to 0 - x in Firm (cmp_left is 0, see the if) */
-                                               new_r_Sub(current_ir_graph, block, cmp_left, cmp_right, mode),
-
-                                               /* This is sizeof_bits(T) - 1 */
-                                               new_r_Const_long(current_ir_graph, block, u_mode,
-                                                       get_mode_size_bits(mode) - 1),
-                                               mode);
-                       }
-               }
-       }
-
-       return res;
-}
-#endif
+#define MAX_DEPTH                              4
 
 /**
  * check, if a node is const and return its tarval or
@@ -234,43 +166,52 @@ 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 +221,34 @@ 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 +257,14 @@ 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 +280,24 @@ 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 +307,18 @@ 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))
@@ -376,18 +328,17 @@ 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 is_new : 1;
        unsigned totally_covers : 1;
@@ -410,6 +361,9 @@ 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;
@@ -483,23 +437,23 @@ 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). */
 
        ir_node *first_phi;                                     /**< The first phi node this cond info was made for. */
-       set *cond_set;                                                  /**< A set of all dominating reachable conds. */
+       set *cond_set;                                                  /**< A set of all dominating reachable Conds. */
 } cond_info_t;
 
 /**
  * @see find_conds.
  */
-static void _find_conds(ir_node *irn, long visited_nr,
+static void _find_conds(ir_node *irn, unsigned 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);
 
@@ -507,7 +461,7 @@ 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) {
+       if (block_dominates(dominator, block) && get_Block_block_visited(block) < visited_nr) {
                cond_t *res = masked_by;
                int i, n;
 
@@ -523,9 +477,11 @@ 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.
                         */
@@ -582,7 +538,7 @@ static void _find_conds(ir_node *irn, long visited_nr,
                        }
                }
 
-               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);
 
@@ -607,18 +563,18 @@ 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)
 {
        int i, n;
-       long visited_nr;
+       unsigned long visited_nr;
        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);
 
@@ -641,7 +597,7 @@ static INLINE void find_conds(ir_node *irn, cond_info_t *ci)
 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 *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];
@@ -657,8 +613,8 @@ 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) {
@@ -680,30 +636,61 @@ 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]));
-
-               *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]);
+  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])
+      hook_if_conversion(current_ir_graph, phi, set[0], NULL, IF_RESULT_BAD_CF);
+    if(operands[1])
+      hook_if_conversion(current_ir_graph, phi, set[1], NULL, IF_RESULT_BAD_CF);
+  }
 
        return *mux;
 }
@@ -934,6 +921,9 @@ void opt_if_conv(ir_graph *irg, opt_if_conv_info_t *params)
        firm_dbg_set_mask(dbg, LEVEL_1);
 #endif
 
+       /* if-conversion works better with normalized returns */
+       normalize_one_return(irg);
+
        /* Ensure, that the dominators are computed. */
        compute_doms(irg);