#ifdef HAVE_ALLOCA_H
#include <alloca.h>
#endif
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
#include "irprog.h"
+#include "irloop_t.h"
#include "xmalloc.h"
+#include "beutil.h"
#include "bechordal_t.h"
#include "becopyopt.h"
#include "becopystat.h"
-#define DEBUG_LVL 0 //SET_LEVEL_1
static firm_dbg_module_t *dbg = NULL;
-#define is_curr_reg_class(irn) (arch_get_irn_reg_class(co->chordal_env->arch_env, irn, arch_pos_make_out(0)) == co->chordal_env->cls)
+#define is_curr_reg_class(irn) \
+ (arch_get_irn_reg_class(get_arch_env(co), \
+ irn, -1) == co->chordal_env->cls)
#define MIN(a,b) ((a<b)?(a):(b))
+#define MAX(a,b) ((a<b)?(b):(a))
+
/**
- * Computes a 'max independent set' wrt. ifg-edges only (no coloring conflicts, no register constraints)
- * @return The size of such a mis
- * NOTE: Code adapted from becopyheur
- * BETTER: Here we can be sure having a chordal graph to work on, so for 'larger' opt-units we
- * could use a special algorithm.
+ * Determines a maximum weighted independent set with respect to
+ * the interference and conflict edges of all nodes in a qnode.
*/
-static int get_ifg_mis_size(unit_t *ou) {
- int all_size, curr_size, i, o;
- int *which;
- ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
-
- /* all contains all nodes */
- all_size = 0;
- for (i=0; i<ou->node_count; ++i)
- all[all_size++] = ou->nodes[i];
-
- /* which[i] says which element to take out of all[] and put into curr[i] */
- which = alloca(all_size*sizeof(*which));
- for (curr_size=0; curr_size<all_size; ++curr_size)
- which[curr_size] = curr_size;
-
- /* stores the currently examined set */
- curr = alloca(all_size*sizeof(*curr));
-
- while (1) { /* this loop will terminate because at least a single node will be a max indep. set */
- /* build current set */
- for (i=0; i<curr_size; ++i)
- curr[i] = all[which[all_size-curr_size+i]];
-
- /* check current set */
- for (i=0; i<curr_size; ++i)
- for (o=i+1; o<curr_size; ++o)
- if (nodes_interfere(ou->co->chordal_env, curr[i], curr[o]))
- goto conflict_found;
-
- /* We had no conflict. This is the (one) max indep. set */
- return curr_size;
-
-conflict_found:
- /* We had a conflict. Generate next set */
- if (which[all_size-curr_size+1] == all_size-curr_size+1) {
- curr_size--;
- for (i=0; i<curr_size; ++i)
- which[all_size-curr_size+i] = i;
- } else {
- int redo = 1;
- while (redo) {
- int pos = all_size;
- do {
- pos--;
- } while (!(which[pos] = (which[pos]+1) % all_size));
-
- for (i=pos+1; i<all_size; ++i)
- which[i] = MIN(which[i-1]+1, all_size-1);
-
- redo = 0;
- for (i=all_size-curr_size; i<all_size-1; ++i)
- if (which[i]>=which[i+1]) {
- redo = 1;
- break;
- }
+static int ou_max_ind_set_costs(unit_t *ou) {
+ be_chordal_env_t *chordal_env = ou->co->chordal_env;
+ ir_node **safe, **unsafe;
+ int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
+ bitset_t *curr;
+ int max, pos, curr_weight, best_weight = 0;
+
+ /* assign the nodes into two groups.
+ * safe: node has no interference, hence it is in every max stable set.
+ * unsafe: node has an interference
+ */
+ safe = alloca((ou->node_count-1) * sizeof(*safe));
+ safe_costs = 0;
+ safe_count = 0;
+ unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
+ unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
+ unsafe_count = 0;
+ for(i=1; i<ou->node_count; ++i) {
+ int is_safe = 1;
+ for(o=1; o<ou->node_count; ++o) {
+ if (i==o)
+ continue;
+ if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
+ unsafe_costs[unsafe_count] = ou->costs[i];
+ unsafe[unsafe_count] = ou->nodes[i];
+ ++unsafe_count;
+ is_safe = 0;
+ break;
}
}
+ if (is_safe) {
+ safe_costs += ou->costs[i];
+ safe[safe_count++] = ou->nodes[i];
+ }
}
- assert(0 && "How did you get here?");
+
+
+ /* now compute the best set out of the unsafe nodes*/
+ if (unsafe_count > MIS_HEUR_TRIGGER) {
+ bitset_t *best = bitset_alloca(unsafe_count);
+ /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
+ for (i=0; i<unsafe_count; ++i) {
+ bitset_set(best, i);
+ /* check if it is a stable set */
+ for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
+ if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
+ bitset_clear(best, i); /* clear the bit and try next one */
+ break;
+ }
+ }
+ /* compute the weight */
+ bitset_foreach(best, pos)
+ best_weight += unsafe_costs[pos];
+ } else {
+ /* Exact Algorithm: Brute force */
+ curr = bitset_alloca(unsafe_count);
+ bitset_set_all(curr);
+ while ((max = bitset_popcnt(curr)) != 0) {
+ /* check if curr is a stable set */
+ for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
+ for (o=bitset_next_set(curr, i+1); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to qnode_max_ind_set(): NOT (curr, i) */
+ if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
+ goto no_stable_set;
+
+ /* if we arrive here, we have a stable set */
+ /* compute the weigth of the stable set*/
+ curr_weight = 0;
+ bitset_foreach(curr, pos)
+ curr_weight += unsafe_costs[pos];
+
+ /* any better ? */
+ if (curr_weight > best_weight) {
+ best_weight = curr_weight;
+ }
+
+ no_stable_set:
+ bitset_minus1(curr);
+ }
+ }
+
+ return safe_costs+best_weight;
}
-/**
- * Builds an optimization unit for a given optimizable irn (root).
- * This opt-unit is inserted in the main structure co.
- * If an arg of root itself is optimizable process this arg before with a
- * recursive call. For handling this situation and loops co->root is used
- * to remember all roots.
- */
-static void co_append_unit(copy_opt_t *co, ir_node *root) {
- int i, arity;
+static void co_collect_units(ir_node *irn, void *env) {
+ copy_opt_t *co = env;
unit_t *unit;
- DBG((dbg, LEVEL_1, "\t Root: %n\n", root));
- /* check if we encountered this root earlier */
- if (pset_find_ptr(co->roots, root))
- return;
- pset_insert_ptr(co->roots, root);
+ arch_register_req_t req;
- assert(is_curr_reg_class(root) && "node is in wrong register class!");
+ if (!is_curr_reg_class(irn))
+ return;
+ if (!is_optimizable(get_arch_env(co), irn, &req))
+ return;
- /* init unit */
- arity = get_irn_arity(root);
+ /* Init a new unit */
unit = xcalloc(1, sizeof(*unit));
unit->co = co;
- unit->interf = 0;
unit->node_count = 1;
- unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
- unit->nodes[0] = root;
INIT_LIST_HEAD(&unit->queue);
- /* check all args */
- for (i=0; i<arity; ++i) {
- ir_node *arg = get_irn_n(root, i);
- assert(is_curr_reg_class(arg) && "Argument not in same register class.");
- if (arg != root) {
- if (!nodes_interfere(co->chordal_env, root, arg)) {
- DBG((dbg, LEVEL_1, "\t Member: %n\n", arg));
- if (is_optimizable(arg))
- co_append_unit(co, arg);
- unit->nodes[unit->node_count++] = arg;
- } else
- unit->interf++;
+ /* Phi with some/all of its arguments */
+ if (is_Reg_Phi(irn)) {
+ int i, arity;
+
+ /* init */
+ arity = get_irn_arity(irn);
+ unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
+ unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
+ unit->nodes[0] = irn;
+
+ /* fill */
+ for (i=0; i<arity; ++i) {
+ int o, arg_pos;
+ ir_node *arg = get_irn_n(irn, i);
+
+ assert(is_curr_reg_class(arg) && "Argument not in same register class.");
+ if (arg == irn)
+ continue;
+ if (nodes_interfere(co->chordal_env, irn, arg)) {
+ unit->inevitable_costs += co->get_costs(irn, arg, i);
+ continue;
+ }
+
+ /* Else insert the argument of the phi to the members of this ou */
+ DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
+
+ /* Check if arg has occurred at a prior position in the arg/list */
+ arg_pos = 0;
+ for (o=0; o<unit->node_count; ++o)
+ if (unit->nodes[o] == arg) {
+ arg_pos = o;
+ break;
+ }
+
+ if (!arg_pos) { /* a new argument */
+ /* insert node, set costs */
+ unit->nodes[unit->node_count] = arg;
+ unit->costs[unit->node_count] = co->get_costs(irn, arg, i);
+ unit->node_count++;
+ } else { /* arg has occured before in same phi */
+ /* increase costs for existing arg */
+ unit->costs[arg_pos] += co->get_costs(irn, arg, i);
+ }
+ }
+ unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
+ unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
+ } else
+
+ /* Proj of a perm with corresponding arg */
+ if (is_Perm_Proj(get_arch_env(co), irn)) {
+ assert(!nodes_interfere(co->chordal_env, irn, get_Copy_src(irn)));
+ unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
+ unit->costs = xmalloc(2 * sizeof(*unit->costs));
+ unit->node_count = 2;
+ unit->nodes[0] = irn;
+ unit->nodes[1] = get_Copy_src(irn);
+ unit->costs[1] = co->get_costs(irn, unit->nodes[1], -1);
+ } else
+
+ /* Src == Tgt of a 2-addr-code instruction */
+ if (is_2addr_code(get_arch_env(co), irn, &req)) {
+ ir_node *other = req.other;
+ if (!nodes_interfere(co->chordal_env, irn, other)) {
+ unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
+ unit->costs = xmalloc(2 * sizeof(*unit->costs));
+ unit->node_count = 2;
+ unit->nodes[0] = irn;
+ unit->nodes[1] = other;
+ unit->costs[1] = co->get_costs(irn, other, -120480);
+ }
+ } else
+ assert(0 && "This is not an optimizable node!");
+
+ /* Insert the new unit at a position according to its costs */
+ if (unit->node_count > 1) {
+ int i;
+ struct list_head *tmp;
+
+ /* Determine the maximum costs this unit can cause: all_nodes_cost */
+ for(i=1; i<unit->node_count; ++i) {
+ unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
+ unit->all_nodes_costs += unit->costs[i];
}
- }
- unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
- list_add_tail(&unit->units, &co->units);
- /* Init ifg_mis_size to node_count. So get_lower_bound returns correct results. */
- unit->ifg_mis_size = get_ifg_mis_size(unit);
-}
-static void co_collect_in_block(ir_node *block, void *env) {
- copy_opt_t *co = env;
- struct list_head *head = get_block_border_head(co->chordal_env, block);
- border_t *curr;
+ /* Determine the minimal costs this unit will cause: min_nodes_costs */
+ unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
- list_for_each_entry_reverse(border_t, curr, head, list)
- if (curr->is_def && curr->is_real && is_optimizable(curr->irn))
- co_append_unit(co, curr->irn);
-}
-
-static void co_collect_units(copy_opt_t *co) {
- DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
- co->roots = pset_new_ptr(64);
- dom_tree_walk_irg(co->chordal_env->irg, co_collect_in_block, NULL, co);
- del_pset(co->roots);
+ /* Insert the new ou according to its sort_key */
+ tmp = &co->units;
+ while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
+ tmp = tmp->next;
+ list_add(&unit->units, tmp);
+ } else {
+ free(unit);
+ }
}
-copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env) {
+copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)) {
const char *s1, *s2, *s3;
int len;
- copy_opt_t *co;
+ copy_opt_t *co;
dbg = firm_dbg_register("ir.be.copyopt");
- firm_dbg_set_mask(dbg, DEBUG_LVL);
co = xcalloc(1, sizeof(*co));
co->chordal_env = chordal_env;
+ co->get_costs = get_costs;
s1 = get_irp_prog_name();
- s2 = get_entity_name(get_irg_entity(co->chordal_env->irg));
+ s2 = get_entity_name(get_irg_entity(get_irg(co)));
s3 = chordal_env->cls->name;
len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
co->name = xmalloc(len);
- if (!strcmp(co->name, DEBUG_IRG))
- firm_dbg_set_mask(dbg, -1);
snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
+ DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
INIT_LIST_HEAD(&co->units);
- co_collect_units(co);
+ irg_walk_graph(get_irg(co), co_collect_units, NULL, co);
return co;
}
xfree(co->name);
list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
xfree(curr->nodes);
+ xfree(curr->costs);
xfree(curr);
}
}
int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
- int i, max;
- for(i=0, max=get_irn_n_outs(irn); i<max; ++i) {
- ir_node *n = get_irn_out(irn, i);
- if (is_optimizable(n) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
- return 1;
+ arch_env_t *aenv = co->chordal_env->main_env->arch_env;
+ const ir_edge_t *edge;
+
+ foreach_out_edge(irn, edge) {
+ ir_node *n = edge->src;
+ arch_register_req_t req;
+
+ arch_get_register_req(aenv, &req, n, -1);
+
+ if( ( (req.type == arch_register_req_type_should_be_same && req.other == irn) ||
+ is_Reg_Phi(n) ||
+ is_Perm(get_arch_env(co), n)
+ ) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
+ return 1;
}
+
return 0;
}
+int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos) {
+ int cost = 0;
+ ir_loop *loop;
+ ir_node *root_block = get_nodes_block(root);
+
+ if (is_Phi(root)) {
+ /* for phis the copies are placed in the corresponding pred-block */
+ loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
+ } else {
+ /* a perm places the copy in the same block as it resides */
+ loop = get_irn_loop(root_block);
+ }
+ if (loop) {
+ int d = get_loop_depth(loop);
+ cost = d*d;
+ }
+ return cost+1;
+}
+
+int get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
+ return 1;
+}
-int co_get_copy_count(const copy_opt_t *co) {
+int co_get_max_copy_costs(const copy_opt_t *co) {
int i, res = 0;
unit_t *curr;
+
list_for_each_entry(unit_t, curr, &co->units, units) {
- int root_col = get_irn_col(co, curr->nodes[0]);
- res += curr->interf;
+ res += curr->inevitable_costs;
for (i=1; i<curr->node_count; ++i)
- if (root_col != get_irn_col(co, curr->nodes[i]))
- res++;
+ res += curr->costs[i];
}
return res;
}
-int co_get_lower_bound(const copy_opt_t *co) {
+int co_get_inevit_copy_costs(const copy_opt_t *co) {
int res = 0;
unit_t *curr;
- list_for_each_entry(unit_t, curr, &co->units, units)
- res += curr->interf + curr->node_count - curr->ifg_mis_size;
- return res;
-}
-int co_get_interferer_count(const copy_opt_t *co) {
- int res = 0;
- unit_t *curr;
list_for_each_entry(unit_t, curr, &co->units, units)
- res += curr->interf;
+ res += curr->inevitable_costs;
return res;
}
-/**
- * Needed for result checking
- */
-static void co_collect_for_checker(ir_node *block, void *env) {
- copy_opt_t *co = env;
- struct list_head *head = get_block_border_head(co->chordal_env, block);
- border_t *curr;
-
- list_for_each_entry_reverse(border_t, curr, head, list)
- if (curr->is_def && curr->is_real && is_curr_reg_class(curr->irn))
- obstack_ptr_grow(&co->ob, curr->irn);
-}
+int co_get_copy_costs(const copy_opt_t *co) {
+ int i, res = 0;
+ unit_t *curr;
-/**
- * This O(n^2) checker checks if
- * two ifg-connected nodes have the same color
- * register constraint are satisfied
- */
-void co_check_allocation(copy_opt_t *co) {
- ir_node **nodes, *n1, *n2;
- int i, o;
-
- obstack_init(&co->ob);
- dom_tree_walk_irg(co->chordal_env->irg, co_collect_for_checker, NULL, co);
- obstack_ptr_grow(&co->ob, NULL);
-
- nodes = (ir_node **) obstack_finish(&co->ob);
- for (i = 0, n1 = nodes[i]; n1; n1 = nodes[++i]) {
- assert(arch_reg_is_allocatable(co->chordal_env->arch_env, n1, arch_pos_make_out(0),
- arch_get_irn_register(co->chordal_env->arch_env, n1, 0)) && "Constraint does not hold");
- for (o = i+1, n2 = nodes[o]; n2; n2 = nodes[++o])
- if (nodes_interfere(co->chordal_env, n1, n2)
- && get_irn_col(co, n1) == get_irn_col(co, n2)) {
- DBG((dbg, 0, "Error: %n %d and %n %d have the same color %d.\n", n1, get_irn_graph_nr(n1), n2, get_irn_graph_nr(n2), get_irn_col(co, n1)));
- assert(0 && "Interfering values have the same color!");
+ list_for_each_entry(unit_t, curr, &co->units, units) {
+ int root_col = get_irn_col(co, curr->nodes[0]);
+ DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
+ res += curr->inevitable_costs;
+ for (i=1; i<curr->node_count; ++i) {
+ int arg_col = get_irn_col(co, curr->nodes[i]);
+ if (root_col != arg_col) {
+ DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
+ res += curr->costs[i];
}
+ }
}
- obstack_free(&co->ob, NULL);
- DBG((dbg, 2, "The checker seems to be happy!\n"));
+ return res;
+}
+
+int co_get_lower_bound(const copy_opt_t *co) {
+ int res = 0;
+ unit_t *curr;
+ list_for_each_entry(unit_t, curr, &co->units, units)
+ res += curr->inevitable_costs + curr->min_nodes_costs;
+ return res;
}
projects / libfirm / blobdiff