#include "vector.h"
#include "vector_t.h"
#include "heuristical_co.h"
+#include "heuristical_co_ld.h"
#include "pbqp_t.h"
#include "html_dumper.h"
#include "pbqp_node_t.h"
#include "pbqp_node.h"
-#define TIMER 0
+#define TIMER 1
+#define PRINT_RPEO 1
-static bool use_exec_freq = true;
+static int use_exec_freq = true;
+static int use_late_decision = true;
typedef struct _be_pbqp_alloc_env_t {
pbqp *pbqp_inst; /**< PBQP instance for register allocation */
static const lc_opt_table_entry_t options[] = {
LC_OPT_ENT_BOOL ("exec_freq", "use exec_freq", &use_exec_freq),
+ LC_OPT_ENT_BOOL ("late_decision", "use late decision for register allocation", &use_late_decision),
LC_OPT_LAST
};
ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
xfree(tu_name);
result = fopen(buf, "wt");
- if(result == NULL) {
+ if (result == NULL) {
panic("Couldn't open '%s' for writing.", buf);
}
/* set costs depending on register constrains */
unsigned idx;
- for(idx = 0; idx < colors_n; idx++) {
- if(bitset_is_set(ignored_regs, idx) || !arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, idx))) {
+ for (idx = 0; idx < colors_n; idx++) {
+ if (bitset_is_set(ignored_regs, idx) || !arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, idx))) {
vector_set(costs_vector, idx, INF_COSTS);
cntConstrains++;
}
pbqp_matrix *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
unsigned *restr_nodes = pbqp_alloc_env->restr_nodes;
- if(get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
+ if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
/* increase ife edge counter */
pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
/* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
- if(get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
+ if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
unsigned src_idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
unsigned trg_idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
assert(src_idx != trg_idx && "Interfering nodes could not have the same register!");
return;
}
- if(get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
- if(get_free_regs(restr_nodes, cls, src_node) == 1) {
+ if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
+ if (get_free_regs(restr_nodes, cls, src_node) == 1) {
unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
}
pbqp_matrix *afe_matrix = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
unsigned colors_n = arch_register_class_n_regs(cls);
- if(get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
- if(use_exec_freq) {
+ if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
+ if (use_exec_freq) {
/* get exec_freq for copy_block */
ir_node *root_bl = get_nodes_block(src_node);
ir_node *copy_bl = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
/* create afe-matrix */
unsigned row, col;
- for(row = 0; row < colors_n; row++) {
- for(col = 0; col < colors_n; col++) {
- if(row != col)
+ for (row = 0; row < colors_n; row++) {
+ for (col = 0; col < colors_n; col++) {
+ if (row != col)
pbqp_matrix_set(afe_matrix, row, col, (num)res);
}
}
}
/* do useful optimization to speed up pbqp solving */
- if(get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
+ if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
return;
}
- if(get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
- if(get_free_regs(restr_nodes, cls, src_node) == 1) {
+ if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
+ if (get_free_regs(restr_nodes, cls, src_node) == 1) {
unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
}
continue;
/* no edges to itself */
- if(irn == arg) {
+ if (irn == arg) {
continue;
}
continue;
/* no edges to itself */
- if(irn == other) {
+ if (irn == other) {
continue;
}
pqueue_t *queue = new_pqueue();
pqueue_t *restr_nodes_queue = new_pqueue();
plist_t *temp_list = plist_new();
+ plist_t *sorted_list = plist_new();
ir_node *irn;
ir_nodeset_t live_nodes;
+ plist_element_t *el;
+ ir_node *last_element = NULL;
/* first, determine the pressure */
/* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
continue;
/* create pbqp source node if it dosn't exist */
- if(get_node(pbqp_inst, get_irn_idx(proj)) == NULL) {
+ if (get_node(pbqp_inst, get_irn_idx(proj)) == NULL) {
create_pbqp_node(pbqp_alloc_env, proj);
}
/* create nodes and interference edges */
foreach_ir_nodeset(&live_nodes, live, iter) {
/* create pbqp source node if it dosn't exist */
- if(get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
+ if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
create_pbqp_node(pbqp_alloc_env, live);
}
/* no edges to itself */
- if(proj == live) {
+ if (proj == live) {
continue;
}
else {
if (arch_irn_consider_in_reg_alloc(cls, irn)) {
/* create pbqp source node if it dosn't exist */
- if(get_node(pbqp_inst, get_irn_idx(irn)) == NULL) {
+ if (get_node(pbqp_inst, get_irn_idx(irn)) == NULL) {
create_pbqp_node(pbqp_alloc_env, irn);
}
/* create nodes and interference edges */
foreach_ir_nodeset(&live_nodes, live, iter) {
/* create pbqp source node if it dosn't exist */
- if(get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
+ if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
create_pbqp_node(pbqp_alloc_env, live);
}
/* no edges to itself */
- if(irn == live) {
+ if (irn == live) {
continue;
}
/* order nodes for perfect elimination order */
if (get_irn_mode(irn) == mode_T) {
- plist_element_t *first = plist_first(temp_list);
- const ir_edge_t *edge;
+ bool allHaveIFEdges = true;
+ const ir_edge_t *edge;
foreach_out_edge(irn, edge) {
ir_node *proj = get_edge_src_irn(edge);
if (!arch_irn_consider_in_reg_alloc(cls, proj))
continue;
- // insert proj node into priority queue (descending by the number of interference edges)
- if(get_free_regs(restr_nodes, cls, proj) <= 4/*bitset_is_set(restr_nodes, get_irn_idx(proj))*/) {
+ /* insert proj node into priority queue (descending by the number of interference edges) */
+ if (get_free_regs(restr_nodes, cls, proj) <= 4) {
pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
}
else {
- pqueue_put(queue,proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
+ pqueue_put(queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
+ }
+
+ /* skip last step if there is no last_element */
+ if(last_element == NULL)
+ continue;
+
+ /* check if proj has an if edge to last_element (at this time pbqp contains only if edges) */
+ if(get_edge(pbqp_inst, proj->node_idx, last_element->node_idx) == NULL && get_edge(pbqp_inst, last_element->node_idx, proj->node_idx) == NULL) {
+ allHaveIFEdges = false; /* there is no if edge between proj and last_element */
}
}
- /* first insert all restricted nodes */
- while(!pqueue_empty(restr_nodes_queue)) {
- if(first == NULL) {
- plist_insert_back(temp_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(restr_nodes_queue))));
- first = plist_first(temp_list);
- } else {
- plist_insert_before(temp_list, first, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(restr_nodes_queue))));
+ if(last_element != NULL && allHaveIFEdges) {
+ if (get_free_regs(restr_nodes, cls, last_element) <= 4) {
+ pqueue_put(restr_nodes_queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
}
+ else {
+ pqueue_put(queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
+ }
+ plist_erase(temp_list, plist_find_value(temp_list, get_node(pbqp_inst, last_element->node_idx)));
+ last_element = NULL;
+ }
+
+ /* first insert all restricted proj nodes */
+ while (!pqueue_empty(restr_nodes_queue)) {
+ plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(restr_nodes_queue))));
}
/* insert proj nodes descending by their number of interference edges */
- while(!pqueue_empty(queue)) {
- if(first == NULL) {
- plist_insert_back(temp_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(queue))));
- first = plist_first(temp_list);
- } else {
- plist_insert_before(temp_list, first, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(queue))));
- }
+ while (!pqueue_empty(queue)) {
+ plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(queue))));
}
+
+ /* invert sorted list */
+ foreach_plist(sorted_list, el) {
+ plist_insert_front(temp_list, el->data);
+ }
+
+ plist_clear(sorted_list);
+
}
else {
if (arch_irn_consider_in_reg_alloc(cls, irn)) {
+ // remember last colorable node
+ last_element = irn;
plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
}
+ else {
+ // node not colorable, so ignore it
+ last_element = NULL;
+ }
}
}
/* insert nodes into reverse perfect elimination order */
- plist_element_t *el;
foreach_plist(temp_list, el) {
plist_insert_back(rpeo, el->data);
}
/* free reserved memory */
ir_nodeset_destroy(&live_nodes);
plist_free(temp_list);
+ plist_free(sorted_list);
del_pqueue(queue);
del_pqueue(restr_nodes_queue);
}
}
}
-void be_pbqp_coloring(be_chordal_env_t *env)
+static void be_pbqp_coloring(be_chordal_env_t *env)
{
- ir_graph *irg = env->irg;
- be_irg_t *birg = env->birg;
- const arch_register_class_t *cls = env->cls;
- unsigned colors_n = arch_register_class_n_regs(cls);
- be_pbqp_alloc_env_t pbqp_alloc_env;
- unsigned idx, row, col;
- be_lv_t *lv;
+ ir_graph *irg = env->irg;
+ be_irg_t *birg = env->birg;
+ const arch_register_class_t *cls = env->cls;
+ be_lv_t *lv = NULL;
+ plist_element_t *element = NULL;
+ unsigned colors_n = arch_register_class_n_regs(cls);
+ be_pbqp_alloc_env_t pbqp_alloc_env;
+ unsigned row, col;
+
#if TIMER
ir_timer_t *t_ra_pbqp_alloc_create = ir_timer_new();
/* create costs matrix template for interference edges */
struct pbqp_matrix *ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
/* set costs */
- for(row = 0, col=0; row < colors_n; row++, col++)
+ for (row = 0, col=0; row < colors_n; row++, col++)
pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
pbqp_alloc_env.ife_matrix_template = ife_matrix;
- if(!use_exec_freq) {
+ if (!use_exec_freq) {
/* create costs matrix template for affinity edges */
struct pbqp_matrix *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
/* set costs */
- for(row = 0; row < colors_n; row++) {
- for(col = 0; col < colors_n; col++) {
- if(row != col)
+ for (row = 0; row < colors_n; row++) {
+ for (col = 0; col < colors_n; col++) {
+ if (row != col)
pbqp_matrix_set(afe_matrix, row, col, 2);
}
}
#if TIMER
ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
#endif
- plist_element_t *el;
- foreach_plist(pbqp_alloc_env.rpeo, el) {
- pbqp_node *node = el->data;
- idx = node->index;
- ir_node *irn = get_idx_irn(irg, idx);
+ foreach_plist(pbqp_alloc_env.rpeo, element) {
+ pbqp_node *node = element->data;
+ ir_node *irn = get_idx_irn(irg, node->index);
+
create_affinity_edges(irn, &pbqp_alloc_env);
}
#if TIMER
set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
#endif
+ /* print out reverse perfect eleminiation order */
+#if PRINT_RPEO
+ plist_element_t *elements;
+ foreach_plist(pbqp_alloc_env.rpeo, elements) {
+ pbqp_node *node = elements->data;
+ printf(" %d(%lu);", node->index, get_idx_irn(irg, node->index)->node_nr);
+ }
+ printf("\n");
+#endif
+
/* solve pbqp instance */
#if TIMER
ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
#endif
- solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
+ if(use_late_decision) {
+ solve_pbqp_heuristical_co_ld(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
+ }
+ else {
+ solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
+ }
#if TIMER
ir_timer_stop(t_ra_pbqp_alloc_solve);
#endif
/* assign colors */
- plist_element_t *element;
foreach_plist(pbqp_alloc_env.rpeo, element) {
- pbqp_node *node = element->data;
- idx = node->index;
- ir_node *irn = get_idx_irn(irg, idx);
- num color = get_node_solution(pbqp_alloc_env.pbqp_inst, idx);
- const arch_register_t *reg = arch_register_for_index(cls, color);
+ pbqp_node *node = element->data;
+ ir_node *irn = get_idx_irn(irg, node->index);
+ num color = get_node_solution(pbqp_alloc_env.pbqp_inst, node->index);
+ const arch_register_t *reg = arch_register_for_index(cls, color);
arch_set_irn_register(irn, reg);
}
/**
* Initializes this module.
*/
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pbqp_coloring);
void be_init_pbqp_coloring(void)
{
lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
be_register_chordal_coloring("pbqp", &coloring);
}
-BE_REGISTER_MODULE_CONSTRUCTOR(be_pbqp_alloc);
-
#endif
projects / libfirm / blobdiff