Handle multiple control dependencies in a more sophisticated (and correct) way

[libfirm] / ir / opt / return.c

diff --git a/ir/opt/return.c b/ir/opt/return.c
index 59530d9..3b5338a 100644 (file)
--- a/ir/opt/return.c
+++ b/ir/opt/return.c
@@ -33,7 +33,7 @@
 /*
  * Normalize the Returns of a graph by creating a new End block
  * with One Return(Phi).
- * This is the prefered input for the if-conversion.
+ * This is the preferred input for the if-conversion.
  *
  * In pseudocode, it means:
  *
@@ -55,19 +55,25 @@ void normalize_one_return(ir_graph *irg)
   ir_node *endbl = get_irg_end_block(irg);
   int i, j, k, n, last_idx, n_rets, n_ret_vals = -1;
   unsigned char *returns;
-  ir_node **in, **retvals;
+  ir_node **in, **retvals, **endbl_in;
 
-  ir_node *block, *new_ret;
+  ir_node *block;
 
   /* look, if we have more than one return */
-  n       = get_Block_n_cfgpreds(endbl);
+  n = get_Block_n_cfgpreds(endbl);
+  if (n <= 0) {
+    /* The end block has no predecessors, we have an endless
+       loop. In that case, no returns exists. */
+    return;
+  }
+
   returns = alloca((n + 7) >> 3);
   memset(returns, 0, (n + 7) >> 3);
 
   for (n_rets = i = 0; i < n; ++i) {
     ir_node *node = get_Block_cfgpred(endbl, i);
 
-    if (get_irn_op(node) == op_Return) {
+    if (is_Return(node)) {
       ++n_rets;
 
       set_bit(i);
@@ -81,12 +87,15 @@ void normalize_one_return(ir_graph *irg)
   if (n_rets <= 1)
     return;
 
-  in      = alloca(sizeof(*in) * IMAX(n_rets, n_ret_vals));
-  retvals = alloca(sizeof(*in) * n_rets * n_ret_vals);
+  in       = alloca(sizeof(*in)       * IMAX(n_rets, n_ret_vals));
+  retvals  = alloca(sizeof(*retvals)  * n_rets * n_ret_vals);
+  endbl_in = alloca(sizeof(*endbl_in) * n);
 
+  last_idx = 0;
   for (j = i = 0; i < n; ++i) {
+    ir_node *ret = get_Block_cfgpred(endbl, i);
+
     if (get_bit(i)) {
-      ir_node *ret  = get_Block_cfgpred(endbl, i);
       ir_node *block = get_nodes_block(ret);
 
       /* create a new Jmp for every Ret and place the in in */
@@ -96,11 +105,10 @@ void normalize_one_return(ir_graph *irg)
       for (k = 0; k < n_ret_vals; ++k)
         retvals[j + k*n_rets] = get_irn_n(ret, k);
 
-      set_Block_cfgpred(endbl, i, new_r_Bad(irg));
-      last_idx = i;
-
       ++j;
     }
+    else
+      endbl_in[last_idx++] = ret;
   }
 
   /* ok, create a new block with all created in's */
@@ -108,21 +116,38 @@ void normalize_one_return(ir_graph *irg)
 
   /* now create the Phi nodes */
   for (j = i = 0; i < n_ret_vals; ++i, j += n_rets) {
+    int k;
+    ir_node *first;
     /* the return values are already shuffled */
-    in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
+
+    /* Beware: normally the Phi constructor automatically replaces a Phi(a,...a) into a
+       but NOT, if a is Unknown. Here, we known that this case can be optimize also,
+       so do it here */
+    first = retvals[j + 0];
+    for (k = 1; k < n_rets; ++k) {
+      if (retvals[j + k] != first) {
+        first = NULL;
+        break;
+      }
+    }
+    if (first)
+      in[i] = first;
+    else
+      in[i] = new_r_Phi(irg, block, n_rets, &retvals[j], get_irn_mode(retvals[j]));
   }
 
-  new_ret = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);
+  endbl_in[last_idx++] = new_r_Return(irg, block, in[0], n_ret_vals-1, &in[1]);
 
-  set_Block_cfgpred(endbl, last_idx, new_ret);
+  set_irn_in(endbl, last_idx, endbl_in);
 
   /* invalidate analysis information:
    * a new Block was added, so dominator, outs and loop are inconsistent,
    * trouts and callee-state should be still valid
    */
-  set_irg_dom_inconsistent(irg);
+  set_irg_doms_inconsistent(irg);
   set_irg_outs_inconsistent(irg);
-  set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
+  set_irg_extblk_inconsistent(irg);
+  set_irg_loopinfo_state(irg, loopinfo_cf_inconsistent);
 }
 
 /**
@@ -169,7 +194,7 @@ static int can_move_ret(ir_node *ret)
 /*
  * Normalize the Returns of a graph by moving
  * the Returns upwards as much as possible.
- * This might be prefered for code generation.
+ * This might be preferred for code generation.
  *
  * In pseudocode, it means:
  *
@@ -205,7 +230,7 @@ void normalize_n_returns(ir_graph *irg)
   for (n_finals = n_rets = i = 0; i < n; ++i) {
     ir_node *ret = get_Block_cfgpred(endbl, i);
 
-    if (get_irn_op(ret) == op_Return && can_move_ret(ret)) {
+    if (is_Return(ret) && can_move_ret(ret)) {
       /*
        * Ok, all conditions met, we can move this Return, put it
        * on our work list.
@@ -312,14 +337,14 @@ void normalize_n_returns(ir_graph *irg)
 
   exchange(endbl, new_r_Block(irg, n_finals, in));
 
-  /* the end block is not automatically skiped, so do it here */
+  /* the end block is not automatically skipped, so do it here */
   set_irg_end_block(irg, skip_Id(get_irg_end_block(irg)));
 
   /* Invalidate analysis information:
-   * Blocks become dead and new Eeturns were deleted, so dominator, outs and loop are inconsistent,
+   * Blocks become dead and new Returns were deleted, so dominator, outs and loop are inconsistent,
    * trouts and callee-state should be still valid
    */
-  set_irg_dom_inconsistent(irg);
+  set_irg_doms_inconsistent(irg);
   set_irg_outs_inconsistent(irg);
   set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
 }