moved ia32 architecture configuration to a separate file and clean it up a bit
[libfirm] / ir / be / bespillslots.c
index 4459e80..52d20fd 100644 (file)
@@ -1,23 +1,42 @@
 /*
- * Author:      Matthias Braun
- * Date:               26.7.06
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+ * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
+/**
+ * @file
+ * @brief       Spillslot coalescer.
+ * @author      Matthias Braun
+ * @date        26.07.2006
+ * @version     $Id$
  */
 #ifdef HAVE_CONFIG_H
-#include <config.h>
+#include "config.h"
 #endif
 
 #include <stdlib.h>
 
 #include "set.h"
-
+#include "array.h"
 #include "irgwalk.h"
 #include "ircons.h"
 #include "irprintf.h"
 #include "execfreq.h"
 #include "unionfind.h"
-#include "type.h"
 #include "irdump_t.h"
 
 #include "benode_t.h"
 #include "benodesets.h"
 #include "bestatevent.h"
 #include "bespilloptions.h"
+#include "bemodule.h"
+#include "beintlive_t.h"
+#include "beirg_t.h"
+#include "bearch_t.h"
 
 #define DBG_COALESCING         1
 #define DBG_INTERFERENCES      2
@@ -75,6 +98,8 @@ static int cmp_spill(const void* d1, const void* d2, size_t size)
 {
        const spill_t* s1 = d1;
        const spill_t* s2 = d2;
+       (void) size;
+
        return s1->spill != s2->spill;
 }
 
@@ -120,13 +145,13 @@ static spill_t *collect_spill(be_fec_env_t *env, ir_node *node,
 
        /* insert into set of spills if not already there */
        spill.spill = node;
-       res = set_find(env->spills, &spill, sizeof(spill), hash);
+       res         = set_find(env->spills, &spill, sizeof(spill), hash);
 
        if(res == NULL) {
                spill.spillslot = set_count(env->spills);
-               spill.mode = mode;
+               spill.mode      = mode;
                spill.alignment = align;
-               res = set_insert(env->spills, &spill, sizeof(spill), hash);
+               res             = set_insert(env->spills, &spill, sizeof(spill), hash);
        } else {
                assert(res->mode == mode);
                assert(res->alignment == align);
@@ -154,9 +179,9 @@ static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
        }
 
        spill.spillslot = set_count(env->spills);
-       spill.mode = mode;
+       spill.mode      = mode;
        spill.alignment = align;
-       res = set_insert(env->spills, &spill, sizeof(spill), hash);
+       res             = set_insert(env->spills, &spill, sizeof(spill), hash);
 
        // collect attached spills and mem-phis
        arity = get_irn_arity(node);
@@ -172,10 +197,10 @@ static spill_t *collect_memphi(be_fec_env_t *env, ir_node *node,
                }
 
                // add an affinity edge
-               affinty_edge = obstack_alloc(&env->obst, sizeof(affinty_edge[0]));
+               affinty_edge           = obstack_alloc(&env->obst, sizeof(affinty_edge[0]));
                affinty_edge->affinity = get_block_execfreq(exec_freq, get_nodes_block(arg));
-               affinty_edge->slot1 = res->spillslot;
-               affinty_edge->slot2 = arg_spill->spillslot;
+               affinty_edge->slot1    = res->spillslot;
+               affinty_edge->slot2    = arg_spill->spillslot;
                ARR_APP1(affinity_edge_t*, env->affinity_edges, affinty_edge);
        }
 
@@ -236,46 +261,97 @@ static int merge_interferences(be_fec_env_t *env, bitset_t** interferences,
        return res;
 }
 
-#if 0
-static void dump_interference_graph(be_fec_env_t *env, bitset_t **interferences, const char* suffix) {
-       char name[256];
-       int i;
-       int spillcount;
-       spill_t *spill;
-       FILE *f;
-       static int cnt = 0;
-
-       snprintf(name, sizeof(name), "%d-%s-spillslots-%s.vcg", cnt++, get_irg_dump_name(env->chordal_env->birg->irg), suffix);
-
-       f = fopen(name, "w");
-       assert(f != NULL);
-
-       fprintf(f, "graph: {\n");
-
-       spillcount = set_count(env->spills);
-       for(spill = set_first(env->spills), i = 0; spill != NULL; spill = set_next(env->spills), ++i) {
-               int slotid = spill->spillslot;
-               fprintf(f, "\tnode: { title: \"n%d\" label: \"%d\" }\n", i, slotid);
-       }
+static int my_values_interfere2(be_irg_t *birg, const ir_node *a,
+                                const ir_node *b)
+{
+       be_lv_t *lv = be_get_birg_liveness(birg);
+
+    int a2b = _value_dominates(a, b);
+    int b2a = _value_dominates(b, a);
+
+    /* If there is no dominance relation, they do not interfere. */
+    if((a2b | b2a) > 0) {
+        const ir_edge_t *edge;
+        ir_node *bb;
+
+        /*
+         * Adjust a and b so, that a dominates b if
+         * a dominates b or vice versa.
+         */
+        if(b2a) {
+            const ir_node *t = a;
+            a = b;
+            b = t;
+        }
+
+        bb = get_nodes_block(b);
+
+        /*
+         * If a is live end in b's block it is
+         * live at b's definition (a dominates b)
+         */
+        if(be_is_live_end(lv, bb, a))
+            return 1;
+
+        /*
+         * Look at all usages of a.
+         * If there's one usage of a in the block of b, then
+         * we check, if this use is dominated by b, if that's true
+         * a and b interfere. Note that b must strictly dominate the user,
+         * since if b is the last user of in the block, b and a do not
+         * interfere.
+         * Uses of a not in b's block can be disobeyed, because the
+         * check for a being live at the end of b's block is already
+         * performed.
+         */
+        foreach_out_edge(a, edge) {
+            const ir_node *user = get_edge_src_irn(edge);
+                       if(is_Sync(user)) {
+                               const ir_edge_t *edge2;
+                               foreach_out_edge(user, edge2) {
+                                       const ir_node *user2 = get_edge_src_irn(edge2);
+                                       assert(!is_Sync(user2));
+                                       if(get_nodes_block(user2) == bb && !is_Phi(user2) &&
+                                          _value_strictly_dominates(b, user2))
+                                               return 1;
+                               }
+                       } else {
+                               if(get_nodes_block(user) == bb && !is_Phi(user) &&
+                                               _value_strictly_dominates(b, user))
+                return 1;
+                       }
+        }
+    }
 
-       for(i = 0; i < ARR_LEN(env->affinity_edges); ++i) {
-               affinity_edge_t *edge = env->affinity_edges[i];
-               fprintf(f, "\tedge: { sourcename: \"n%d\" targetname: \"n%d\" color: green }\n", edge->slot1, edge->slot2);
-       }
+       return 0;
+}
 
-       for(i = 0; i < spillcount; ++i) {
-               int i2;
-               for(i2 = 0; i2 < spillcount; ++i2) {
-                       if(bitset_is_set(interferences[i], i2)) {
-                               fprintf(f, "\tedge: { sourcename: \"n%d\" targetname: \"n%d\" color: red }\n", i, i2);
-                       }
+/**
+ * same as values_interfere but with special handling for Syncs
+ */
+static int my_values_interfere(be_irg_t *birg, ir_node *a, ir_node *b)
+{
+       if(is_Sync(a)) {
+               int i, arity = get_irn_arity(a);
+               for(i = 0; i < arity; ++i) {
+                       ir_node *in = get_irn_n(a, i);
+                       if(my_values_interfere(birg, in, b))
+                               return 1;
+               }
+               return 0;
+       } else if(is_Sync(b)) {
+               int i, arity = get_irn_arity(b);
+               for(i = 0; i < arity; ++i) {
+                       ir_node *in = get_irn_n(b, i);
+                       /* a is not a sync, so no need for my_values_interfere */
+                       if(my_values_interfere2(birg, a, in))
+                               return 1;
                }
+               return 0;
        }
 
-       fprintf(f, "}\n");
-       fclose(f);
+       return my_values_interfere2(birg, a, b);
 }
-#endif
 
 /**
  * A greedy coalescing algorithm for spillslots:
@@ -292,7 +368,6 @@ static void do_greedy_coalescing(be_fec_env_t *env)
        int affinity_edge_count;
        bitset_t **interferences;
        int* spillslot_unionfind;
-       const be_lv_t *lv = be_get_birg_liveness(env->birg);
 
        spillcount = set_count(env->spills);
        if(spillcount == 0)
@@ -319,10 +394,20 @@ static void do_greedy_coalescing(be_fec_env_t *env)
                interferences[i] = bitset_alloca(spillcount);
        }
 
-       // construct interferences
-       for(i = 0; i < spillcount; ++i) {
+       /* construct interferences */
+       for (i = 0; i < spillcount; ++i) {
+               ir_node *spill1 = spilllist[i]->spill;
+
+               if (is_NoMem(spill1))
+                       continue;
+
                for(i2 = i+1; i2 < spillcount; ++i2) {
-                       if(values_interfere(lv, spilllist[i]->spill, spilllist[i2]->spill)) {
+                       ir_node *spill2 = spilllist[i2]->spill;
+
+                       if (is_NoMem(spill2))
+                               continue;
+
+                       if (my_values_interfere(env->birg, spill1, spill2)) {
                                DBG((dbg, DBG_INTERFERENCES, "Slot %d and %d interfere\n", i, i2));
                                bitset_set(interferences[i], i2);
                                bitset_set(interferences[i2], i);
@@ -330,20 +415,20 @@ static void do_greedy_coalescing(be_fec_env_t *env)
                }
        }
 
-       // sort affinity edges
+       /* sort affinity edges */
        affinity_edge_count = ARR_LEN(env->affinity_edges);
        qsort(env->affinity_edges, affinity_edge_count, sizeof(env->affinity_edges[0]), cmp_affinity);
 
        //dump_interference_graph(env, interferences, "before");
 
-       // try to merge affine nodes
+       /* try to merge affine nodes */
        for(i = 0; i < affinity_edge_count; ++i) {
                const affinity_edge_t *edge = env->affinity_edges[i];
                int s1 = uf_find(spillslot_unionfind, edge->slot1);
                int s2 = uf_find(spillslot_unionfind, edge->slot2);
 
                /* test if values interfere */
-               if(bitset_is_set(interferences[s1], s2)) {
+               if (bitset_is_set(interferences[s1], s2)) {
                        assert(bitset_is_set(interferences[s2], s1));
                        continue;
                }
@@ -426,6 +511,8 @@ static int cmp_memperm(const void* d1, const void* d2, size_t size)
 {
        const memperm_t* e1 = d1;
        const memperm_t* e2 = d2;
+       (void) size;
+
        return e1->block != e2->block;
 }
 
@@ -454,7 +541,7 @@ static ir_entity* create_stack_entity(be_fec_env_t *env, spill_slot_t *slot)
        ir_type *frame = get_irg_frame_type(irg);
        ir_entity *res = frame_alloc_area(frame, slot->size, slot->align, 0);
 
-       // adjust size of the entity type...
+       /* adjust size of the entity type... */
        ir_type *enttype = get_entity_type(res);
        set_type_size_bytes(enttype, slot->size);
 
@@ -482,6 +569,28 @@ static void enlarge_spillslot(spill_slot_t *slot, int otheralign, int othersize)
        }
 }
 
+
+static void assign_spill_entity(const arch_env_t *arch_env, ir_node *node, ir_entity *entity)
+{
+       if(is_NoMem(node))
+               return;
+       if(is_Sync(node)) {
+               int i, arity;
+
+               arity = get_irn_arity(node);
+               for(i = 0; i < arity; ++i) {
+                       ir_node *in = get_irn_n(node, i);
+                       assert(!is_Phi(in));
+
+                       assign_spill_entity(arch_env, in, entity);
+               }
+               return;
+       }
+
+       assert(arch_get_frame_entity(arch_env, node) == NULL);
+       arch_set_frame_entity(arch_env, node, entity);
+}
+
 /**
  * Create stack entities for the spillslots and assign them to the spill and
  * reload nodes.
@@ -563,7 +672,7 @@ static void assign_spillslots(be_fec_env_t *env)
                                }
                        }
                } else {
-                       arch_set_frame_entity(arch_env, node, slot->entity);
+                       assign_spill_entity(arch_env, node, slot->entity);
                }
        }
 
@@ -628,7 +737,7 @@ static void create_memperms(be_fec_env_t *env)
                // insert node into schedule
                blockend = get_end_of_block_insertion_point(memperm->block);
                sched_add_before(blockend, mempermnode);
-               be_stat_ev("mem_perm", memperm->entrycount);
+               stat_ev_dbl("mem_perm", memperm->entrycount);
 
                i = 0;
                for(entry = memperm->entries; entry != NULL; entry = entry->next, ++i) {
@@ -639,7 +748,6 @@ static void create_memperms(be_fec_env_t *env)
                        be_set_MemPerm_out_entity(mempermnode, i, entry->out);
                        set_irg_current_block(irg, memperm->block);
                        proj = new_Proj(mempermnode, get_irn_mode(arg), i);
-                       sched_add_before(blockend, proj);
 
                        set_irn_n(entry->node, entry->pos, proj);
                }
@@ -669,19 +777,17 @@ static int count_spillslots(const be_fec_env_t *env)
 be_fec_env_t *be_new_frame_entity_coalescer(be_irg_t *birg)
 {
        const arch_env_t *arch_env = birg->main_env->arch_env;
+       be_fec_env_t     *env      = xmalloc(sizeof(env[0]));
 
-       be_fec_env_t *env = xmalloc(sizeof(env[0]));
-
-       be_assure_liveness(birg);
+       be_liveness_assure_chk(be_assure_liveness(birg));
 
        obstack_init(&env->obst);
-       env->arch_env = arch_env;
-       env->birg = birg;
-       env->spills = new_set(cmp_spill, 10);
-       env->reloads = NEW_ARR_F(ir_node*, 0);
+       env->arch_env       = arch_env;
+       env->birg           = birg;
+       env->spills         = new_set(cmp_spill, 10);
+       env->reloads        = NEW_ARR_F(ir_node*, 0);
        env->affinity_edges = NEW_ARR_F(affinity_edge_t*, 0);
-       env->memperms = new_set(cmp_memperm, 10);
-       FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");
+       env->memperms       = new_set(cmp_memperm, 10);
 
        return env;
 }
@@ -699,19 +805,13 @@ void be_free_frame_entity_coalescer(be_fec_env_t *env)
 
 void be_assign_entities(be_fec_env_t *env)
 {
-       if(be_stat_ev_is_active()) {
-               int count = set_count(env->spills);
-               be_stat_ev("spillslots", count);
-       }
+       stat_ev_dbl("spillslots", set_count(env->spills));
 
        if(be_coalesce_spill_slots) {
                do_greedy_coalescing(env);
        }
 
-       if(be_stat_ev_is_active()) {
-               int count = count_spillslots(env);
-               be_stat_ev("spillslots_after_coalescing", count);
-       }
+       stat_ev_dbl("spillslots_after_coalescing", count_spillslots(env));
 
        assign_spillslots(env);
 
@@ -737,21 +837,28 @@ static void collect_spills_walker(ir_node *node, void *data)
        if (!arch_irn_class_is(arch_env, node, reload))
                return;
 
-       mode = get_irn_mode(node);
-       cls = arch_get_irn_reg_class(arch_env, node, -1);
+       mode  = get_irn_mode(node);
+       cls   = arch_get_irn_reg_class(arch_env, node, -1);
        align = arch_isa_get_reg_class_alignment(arch_env_get_isa(arch_env), cls);
 
        be_node_needs_frame_entity(env, node, mode, align);
 }
 
-void be_coalesce_spillslots(const be_chordal_env_t *chordal_env)
+void be_coalesce_spillslots(be_irg_t *birg)
 {
-       be_fec_env_t *env = be_new_frame_entity_coalescer(chordal_env->birg);
+       be_fec_env_t *env = be_new_frame_entity_coalescer(birg);
 
        /* collect reloads */
-       irg_walk_graph(chordal_env->irg, NULL, collect_spills_walker, env);
+       irg_walk_graph(birg->irg, NULL, collect_spills_walker, env);
 
        be_assign_entities(env);
 
        be_free_frame_entity_coalescer(env);
 }
+
+void be_init_spillslots(void)
+{
+       FIRM_DBG_REGISTER(dbg, "firm.be.spillslots");
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillslots);