#include <malloc.h>
#endif
+#include "debug.h"
#include "xmalloc.h"
-#include "becopyopt.h"
+#include "becopyopt_t.h"
#include "becopystat.h"
#include "bitset.h"
static firm_dbg_module_t *dbg = NULL;
-#define SLOTS_PINNED_GLOBAL 256
+#define SEARCH_FREE_COLORS
+
+#define SLOTS_PINNED_GLOBAL 64
#define SLOTS_CONFLICTS 8
#define SLOTS_CHANGED_NODES 32
-#define MIN(a,b) ((a<b)?(a):(b))
#define list_entry_queue(lh) list_entry(lh, qnode_t, queue)
#define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2))
set *changed_nodes; /**< contains node_stat_t's. */
} qnode_t;
-pset *pinned_global; /**< optimized nodes should not be altered any more */
+static pset *pinned_global; /**< optimized nodes should not be altered any more */
static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
const conflict_t *xx = x;
static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
conflict_t c;
/* search for live range interference */
- if (n1!=n2 && nodes_interfere(qn->ou->co->chordal_env, n1, n2))
+ if (n1!=n2 && nodes_interfere(qn->ou->co->cenv, n1, n2))
return 1;
/* search for recoloring conflicts */
if ((int)n1 < (int)n2) {
static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
node_stat_t *found = qnode_find_or_insert_node(qn, irn);
found->new_color = color;
+ DBG((dbg, LEVEL_3, "\t col(%+F) := %d\n", irn, color));
}
/**
static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
node_stat_t *found = qnode_find_or_insert_node(qn, irn);
found->pinned_local = 1;
+ if (found->new_color == NO_COLOR)
+ found->new_color = get_irn_col(qn->ou->co, irn);
}
+
/**
* Possible return values of qnode_color_irn()
*/
* Performs virtual re-coloring of node @p n to color @p col. Virtual colors of
* other nodes are changed too, as required to preserve correctness. Function is
* aware of local and global pinning. Recursive.
- * @param irn The node to set the color for
- * @param col The color to set
+ *
+ * If irn == trigger the color @p col must be used. (the first recoloring)
+ * If irn != trigger an arbitrary free color may be used. If no color is free, @p col is used.
+ *
+ * @param irn The node to set the color for
+ * @param col The color to set
* @param trigger The irn that caused the wish to change the color of the irn
+ * External callers must call with trigger = irn
+ *
* @return CHANGE_SAVE iff setting the color is possible, with all transitive effects.
* CHANGE_IMPOSSIBLE iff conflicts with reg-constraintsis occured.
* Else the first conflicting ir_node encountered is returned.
*
- * ASSUMPTION: Assumes that a life range of a single value can't be split into
- * several smaller intervals where other values can live in between.
- * This should be true in SSA.
*/
static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
- ir_node *res;
- struct obstack confl_ob;
- ir_node **confl, *cn;
- int i, irn_col;
- const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env;
- const arch_env_t *arch_env = get_arch_env(qn->ou->co);
- const arch_register_class_t *cls = chordal_env->cls;
-
- DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
- obstack_init(&confl_ob);
- irn_col = qnode_get_new_color(qn, irn);
+ copy_opt_t *co = qn->ou->co;
+ const be_chordal_env_t *chordal_env = co->cenv;
+ const arch_register_class_t *cls = co->cls;
+ const arch_env_t *arch_env = co->aenv;
+ int irn_col = qnode_get_new_color(qn, irn);
+ ir_node *sub_res, *curr;
+ be_ifg_t *ifg = chordal_env->ifg;
+ void *iter = be_ifg_neighbours_iter_alloca(ifg);
+
+ DBG((dbg, LEVEL_3, "\t %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
+
+ /* If the target color is already set do nothing */
if (irn_col == col) {
- DBG((dbg, LEVEL_4, "\t Already same color.\n"));
- goto ret_save;
+ DBG((dbg, LEVEL_3, "\t %+F same color\n", irn));
+ return CHANGE_SAVE;
}
+
+ /* If the irn is pinned, changing color is impossible */
if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
- res = irn;
- goto ret_confl;
+ DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
+ return irn;
}
- if (!arch_reg_is_allocatable(arch_env,
- irn,
- arch_pos_make_out(0),
- arch_register_for_index(cls, col)))
- goto ret_imposs;
-
- /* get all nodes which would conflict with this change */
- {
- struct obstack q;
- int in, out;
- ir_node *irn_bl;
-
- irn_bl = get_nodes_block(irn);
-
- /* first check for a conflicting node which is 'living in' the irns block */
- {
- ir_node *n;
- pset *live_ins = put_live_in(irn_bl, pset_new_ptr_default());
- for (n = pset_first(live_ins); n; n = pset_next(live_ins)) {
- DBG((dbg, LEVEL_4, "Checking %+F which is live-in at the block\n", n));
- if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
- && n != trigger && qnode_get_new_color(qn, n) == col
- && nodes_interfere(chordal_env, irn, n)) {
-
- DBG((dbg, LEVEL_4, "\t %+F\ttroubles\n", n));
- obstack_ptr_grow(&confl_ob, n);
- pset_break(live_ins);
- break;
- }
- }
- del_pset(live_ins);
+
+#ifdef SEARCH_FREE_COLORS
+ /* If we resolve conflicts (recursive calls) we can use any unused color.
+ * In case of the first call @p col must be used.
+ */
+ if (irn != trigger) {
+ bitset_t *free_cols = bitset_alloca(cls->n_regs);
+ arch_register_req_t req;
+ ir_node *curr;
+ int free_col;
+
+ /* Get all possible colors */
+ arch_put_non_ignore_regs(arch_env, cls, free_cols);
+
+ /* Exclude colors not assignable to the irn */
+ arch_get_register_req(arch_env, &req, irn, -1);
+ if (arch_register_req_is(&req, limited)) {
+ bitset_t *limited = bitset_alloca(cls->n_regs);
+ req.limited(req.limited_env, limited);
+ bitset_and(free_cols, limited);
}
- /* setup the queue of blocks. */
- obstack_init(&q);
- obstack_ptr_grow(&q, irn_bl);
- in = 1;
- out = 0;
-
- /* process the queue. The code below checks for every block dominated
- * by the irns one, and in which the irn is live, if there are
- * conflicting nodes */
- while (out < in) {
- ir_node *curr_bl, *sub_bl;
- int i, max;
-
- curr_bl = ((ir_node **)obstack_base(&q))[out++];
-
- /* Add to the result all nodes in the block, which have
- * the target color and interfere with the irn */
- for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
- ir_node *n = get_irn_out(curr_bl, i);
- DBG((dbg, LEVEL_4, "Checking %+F defined in same block\n", n));
- if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
- && n != trigger && qnode_get_new_color(qn, n) == col
- && nodes_interfere(chordal_env, irn, n)) {
- DBG((dbg, LEVEL_4, "\t %+F\ttroubles\n", n));
- obstack_ptr_grow(&confl_ob, n);
- }
- }
+ /* Exclude the color of the irn, because it must _change_ its color */
+ bitset_clear(free_cols, irn_col);
- /* If irn lives out check i-dominated blocks where the irn lives in */
- /* Fill the queue */
- if (is_live_out(curr_bl, irn)) {
- dominates_for_each(curr_bl, sub_bl)
- if (is_live_in(sub_bl, irn)) {
- obstack_ptr_grow(&q, sub_bl);
- in++;
- }
- }
+ /* Exclude all colors used by adjacent nodes */
+ be_ifg_foreach_neighbour(ifg, iter, irn, curr)
+ bitset_clear(free_cols, qnode_get_new_color(qn, curr));
+
+ free_col = bitset_next_set(free_cols, 0);
+
+ if (free_col != -1) {
+ qnode_set_new_color(qn, irn, free_col);
+ return CHANGE_SAVE;
}
- obstack_free(&q, NULL);
- obstack_ptr_grow(&confl_ob, NULL);
- confl = (ir_node **) obstack_finish(&confl_ob);
}
+#endif /* SEARCH_FREE_COLORS */
- /* process all nodes which would conflict with this change */
- for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
- ir_node *sub_res;
+ /* If target color is not allocatable changing color is impossible */
+ if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
+ DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
+ return CHANGE_IMPOSSIBLE;
+ }
- /* try to color the conflicting node cn with the color of the irn itself */
- sub_res = qnode_color_irn(qn, cn, irn_col, irn);
- if (sub_res != CHANGE_SAVE) {
- res = sub_res;
- goto ret_confl;
+ /*
+ * If we arrive here changing color may be possible, but there may be conflicts.
+ * Try to color all conflicting nodes 'curr' with the color of the irn itself.
+ */
+ be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
+ DBG((dbg, LEVEL_3, "\t Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
+ if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
+ sub_res = qnode_color_irn(qn, curr, irn_col, irn);
+ if (sub_res != CHANGE_SAVE) {
+ be_ifg_neighbours_break(ifg, iter);
+ return sub_res;
+ }
}
}
- /* if we arrive here all sub changes can be applied, so it's save to change this irn */
-ret_save:
- DBG((dbg, LEVEL_3, "\t %+F save\n", irn));
- obstack_free(&confl_ob, NULL);
+ /*
+ * If we arrive here, all conflicts were resolved.
+ * So it is save to change this irn
+ */
qnode_set_new_color(qn, irn, col);
return CHANGE_SAVE;
-
-ret_imposs:
- DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn));
- obstack_free(&confl_ob, NULL);
- return CHANGE_IMPOSSIBLE;
-
-ret_confl:
- DBG((dbg, LEVEL_3, "\t %+F conflicting\n", irn));
- obstack_free(&confl_ob, NULL);
- return res;
}
+
/**
* Tries to set the colors for all members of this queue node;
* to the target color qn->color
/**
* Determines a maximum weighted independent set with respect to
* the interference and conflict edges of all nodes in a qnode.
- * TODO: This runs in 2^n in worst case. Use a heuristic iff n>???
*/
static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
ir_node **safe, **unsafe;
- /* now brute force the best set out of the unsafe nodes*/
+ /* now compute the best set out of the unsafe nodes*/
best = bitset_alloca(unsafe_count);
- 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); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to ou_max_ind_set_costs(): NOT (curr, i+1) */
- if (qnode_are_conflicting(qn, 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;
- bitset_copy(best, curr);
+
+ if (unsafe_count > MIS_HEUR_TRIGGER) {
+ /* Heuristic: 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 (qnode_are_conflicting(qn, 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); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to ou_max_ind_set_costs(): NOT (curr, i+1) */
+ if (qnode_are_conflicting(qn, 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;
+ bitset_copy(best, curr);
+ }
no_stable_set:
- bitset_minus1(curr);
+ bitset_minus1(curr);
+ }
}
/* transfer the best set into the qn */
qn->mis_size = 1+safe_count+bitset_popcnt(best);
qn->mis_costs = safe_costs+best_weight;
- qn->mis[0] = ou->nodes[0]; /* the root is alwazs in a max stable set */
+ qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
next = 1;
for (i=0; i<safe_count; ++i)
qn->mis[next++] = safe[i];
*/
static void ou_optimize(unit_t *ou) {
int i;
- qnode_t *curr, *tmp;
- bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs);
+ qnode_t *curr = NULL, *tmp;
+ const arch_env_t *aenv = ou->co->aenv;
+ const arch_register_class_t *cls = ou->co->cls;
+ bitset_t *pos_regs = bitset_alloca(cls->n_regs);
+ bitset_t *ign_regs = bitset_alloca(cls->n_regs);
DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
for (i=0; i<ou->node_count; ++i)
/* init queue */
INIT_LIST_HEAD(&ou->queue);
- arch_get_allocatable_regs(get_arch_env(ou->co), ou->nodes[0], arch_pos_make_out(0), ou->co->chordal_env->cls, pos_regs);
+
+ arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs);
+
+ /* exclude ingore colors */
+ arch_put_non_ignore_regs(aenv, cls, ign_regs);
+ bitset_and(pos_regs, ign_regs);
+
+ /* create new qnode */
bitset_foreach(pos_regs, i)
ou_insert_qnode(ou, new_qnode(ou, i));
/* try */
if (qnode_try_color(curr))
break;
+
/* no success, so re-insert */
del_set(curr->changed_nodes);
curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
/* apply the best found qnode */
if (curr->mis_size >= 2) {
node_stat_t *ns;
+ int root_col = qnode_get_new_color(curr, ou->nodes[0]);
DBG((dbg, LEVEL_1, "\t Best color: %d Costs: %d << %d << %d\n", curr->color, ou->min_nodes_costs, ou->all_nodes_costs - curr->mis_costs, ou->all_nodes_costs));
/* globally pin root and all args which have the same color */
pset_insert_ptr(pinned_global, ou->nodes[0]);
for (i=1; i<ou->node_count; ++i) {
ir_node *irn = ou->nodes[i];
int nc = qnode_get_new_color(curr, irn);
- if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
+ if (nc != NO_COLOR && nc == root_col)
pset_insert_ptr(pinned_global, irn);
}
set_irn_col(ou->co, ns->irn, ns->new_color);
}
}
- /*
- * Enable for checking register allocation after each ou
- * be_ra_chordal_check(ou->co->chordal_env);
- */
}
/* free best qnode (curr) and queue */
free_qnode(curr);
}
-void co_heur_opt(copy_opt_t *co) {
+int co_solve_heuristic(copy_opt_t *co) {
unit_t *curr;
dbg = firm_dbg_register("ir.be.copyoptheur");
- if (!strcmp(co->name, DEBUG_IRG))
- firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_HEUR);
- else
- firm_dbg_set_mask(dbg, DEBUG_LVL_HEUR);
pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
list_for_each_entry(unit_t, curr, &co->units, units)
ou_optimize(curr);
del_pset(pinned_global);
+
+ return 0;
}