- stack alignment is now power of two

[libfirm] / ir / be / beblocksched.c

diff --git a/ir/be/beblocksched.c b/ir/be/beblocksched.c
index ee4669a..38fc4de 100644 (file)
--- a/ir/be/beblocksched.c
+++ b/ir/be/beblocksched.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -23,6 +23,16 @@
  * @author      Matthias Braun, Christoph Mallon
  * @date        27.09.2006
  * @version     $Id$
+ *
+ * The goals of the greedy (and ILP) algorithm here works by assuming that
+ * we want to change as many jumps to fallthroughs as possible (executed jumps
+ * actually, we have to look at the execution frequencies). The algorithms
+ * do this by collecting execution frequencies of all branches (which is easily
+ * possible when all critical edges are split) then removes critical edges where
+ * possible as we don't need and want them anymore now. The algorithms then try
+ * to change as many edges to fallthroughs as possible, this is done by setting
+ * a next and prev pointers on blocks. The greedy algorithm sorts the edges by
+ * execution frequencies and tries to transform them to fallthroughs in this order
  */
 #ifdef HAVE_CONFIG_H
 #include "config.h"
@@ -49,9 +59,8 @@
 #include "bemodule.h"
 #include "be.h"
 
-#include <libcore/lc_opts.h>
-#include <libcore/lc_opts_enum.h>
-#include <libcore/lc_timing.h>
+#include "lc_opts.h"
+#include "lc_opts_enum.h"
 
 #ifdef WITH_ILP
 #include <lpp/lpp.h>
@@ -134,12 +143,15 @@ static void collect_egde_frequency(ir_node *block, void *data)
        entry->prev  = NULL;
        set_irn_link(block, entry);
 
-       if (block == get_irg_start_block(env->irg))
-               return;
-
        arity = get_Block_n_cfgpreds(block);
 
-       if (arity == 1) {
+       if (arity == 0) {
+               assert(block == get_irg_start_block(env->irg)
+                               || block == get_irg_end_block(env->irg));
+               /* must be the start block (or end-block for endless loops), nothing to
+                * do here */
+               return;
+       } else if (arity == 1) {
                edge.block            = block;
                edge.pos              = 0;
                edge.execfreq         = get_block_execfreq(env->execfreqs, block);
@@ -190,18 +202,19 @@ static void coalesce_blocks(blocksched_env_t *env)
        for (i = 0; i < edge_count; ++i) {
                const edge_t *edge  = &env->edges[i];
                ir_node      *block = edge->block;
+               int           pos   = edge->pos;
                ir_node      *pred_block;
                blocksched_entry_t *entry, *pred_entry;
 
-               /* the block might have been removed already... */
-               if (is_Bad(get_Block_cfgpred(block, 0)))
-                       continue;
-
                /* only check edge with highest frequency */
                if (! edge->highest_execfreq)
                        continue;
 
-               pred_block = get_Block_cfgpred_block(block, edge->pos);
+               /* the block might have been removed already... */
+               if (is_Bad(get_Block_cfgpred(block, 0)))
+                       continue;
+
+               pred_block = get_Block_cfgpred_block(block, pos);
                entry      = get_irn_link(block);
                pred_entry = get_irn_link(pred_block);
 
@@ -223,6 +236,7 @@ static void coalesce_blocks(blocksched_env_t *env)
        for (i = 0; i < edge_count; ++i) {
                const edge_t *edge  = &env->edges[i];
                ir_node      *block = edge->block;
+               int           pos   = edge->pos;
                ir_node      *pred_block;
                blocksched_entry_t *entry, *pred_entry;
 
@@ -230,7 +244,11 @@ static void coalesce_blocks(blocksched_env_t *env)
                if (is_Bad(get_Block_cfgpred(block, 0)))
                        continue;
 
-               pred_block = get_Block_cfgpred_block(block, edge->pos);
+               /* we can't do fallthroughs in backedges */
+               if (is_backedge(block, pos))
+                       continue;
+
+               pred_block = get_Block_cfgpred_block(block, pos);
                entry      = get_irn_link(block);
                pred_entry = get_irn_link(pred_block);
 
@@ -343,7 +361,7 @@ static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
        ir_node            *startblock = get_irg_start_block(irg);
        blocksched_entry_t *entry      = get_irn_link(startblock);
 
-       set_using_visited(irg);
+       set_using_irn_visited(irg);
        inc_irg_visited(irg);
 
        env->worklist = new_pdeq();
@@ -351,7 +369,7 @@ static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env)
        assert(pdeq_empty(env->worklist));
        del_pdeq(env->worklist);
 
-       clear_using_visited(irg);
+       clear_using_irn_visited(irg);
 
        return entry;
 }
@@ -695,7 +713,7 @@ static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfr
        list.n_blks = 0;
 
        set_using_irn_link(irg);
-       set_using_visited(irg);
+       set_using_irn_visited(irg);
        inc_irg_block_visited(irg);
 
        create_block_list(get_irg_start_block(irg), &list);
@@ -709,7 +727,7 @@ static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfr
        }
 
        clear_using_irn_link(irg);
-       clear_using_visited(irg);
+       clear_using_irn_visited(irg);
 
        return blk_list;
 }