X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopystat.c;h=6a9c602d0e31413452d0c516ad8dd85315092f9e;hb=d14c6378674f36728eacaf5dc7e4bb045ff9fbab;hp=d48e09bd688c8fa4980e4232ebf1df577df14b68;hpb=b9776d1f90b9376855ea7f3f875ce22e65b39272;p=libfirm

diff --git a/ir/be/becopystat.c b/ir/be/becopystat.c
index d48e09bd6..6a9c602d0 100644
--- a/ir/be/becopystat.c
+++ b/ir/be/becopystat.c
@@ -9,27 +9,34 @@
 #endif
 
 #include <string.h>
+#include <libcore/lc_timing.h>
+
+#include "xmalloc.h"
 #include "irgraph.h"
+#include "irgwalk.h"
 #include "irprog.h"
+#include "iredges_t.h"
 #include "phiclass_t.h"
-#include "becopyopt.h"
+#include "bechordal_t.h"
+#include "beutil.h"
+#include "becopyopt_t.h"
 #include "becopystat.h"
-#include "xmalloc.h"
+#include "beirg_t.h"
+#include "bemodule.h"
 
-#ifdef DO_STAT
+#define DEBUG_LVL SET_LEVEL_1
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
-#define DEBUG_LVL 0 //SET_LEVEL_1
-static firm_dbg_module_t *dbg = NULL;
-
-#define MAX_ARITY 10
-#define MAX_CLS_SIZE 10
-#define MAX_CLS_PHIS 10
-#define MAX_PHASE 2
+#define MAX_ARITY 20
+#define MAX_CLS_SIZE 20
+#define MAX_CLS_PHIS 20
 
 /**
  * For an explanation of these values see the code of copystat_dump_pretty
  */
 enum vals_t {
+	/* FROM HERE: PROBLEM CHARACTERIZATION */
+
 	I_ALL_NODES = 0,
 	I_BLOCKS,
 
@@ -56,15 +63,23 @@ enum vals_t {
 	I_CLS_PHIS_S,
 	I_CLS_PHIS_E = I_CLS_PHIS_S+MAX_CLS_PHIS,
 
+	/* FROM HERE: RESULT VLAUES */
+	/* all of them are external set */
+
 	/* ilp values */
-	I_ILP_TIME,			/* !external set! solving time in seconds */
-	I_ILP_ITER,			/* !external set! number of simplex iterations */
+	I_HEUR_TIME,		/* solving time in milli seconds */
+	I_ILP_TIME,			/* solving time in milli seconds */
+	I_ILP_VARS,
+	I_ILP_CSTR,
+	I_ILP_ITER,			/* number of simplex iterations */
 
 	/* copy instructions */
-	I_COPIES_MAX,		/* !external set! max possible costs of copies*/
-	I_COPIES_INIT,		/* !external set! number of copies in initial allocation */
-	I_COPIES_HEUR,		/* !external set! number of copies after heuristic */
-	I_COPIES_OPT,		/* !external set! number of copies after ilp */
+	I_COPIES_MAX,		/* max possible costs of copies*/
+	I_COPIES_INIT,		/* number of copies in initial allocation */
+	I_COPIES_HEUR,		/* number of copies after heuristic */
+	I_COPIES_5SEC,		/* number of copies after ilp with max n sec */
+	I_COPIES_30SEC,		/* number of copies after ilp with max n sec */
+	I_COPIES_OPT,		/* number of copies after ilp */
 	I_COPIES_IF,		/* number of copies inevitable due to root-arg-interf */
 
 	ASIZE
@@ -78,27 +93,28 @@ int curr_vals[ASIZE];
 static pset *all_phi_nodes;
 static pset *all_phi_classes;
 static pset *all_copy_nodes;
+static ir_graph *last_irg;
 
-void copystat_init(void) {
-	dbg = firm_dbg_register("ir.be.copystat");
-	firm_dbg_set_mask(dbg, DEBUG_LVL);
+void be_init_copystat(void) {
+	FIRM_DBG_REGISTER(dbg, "firm.be.copystat");
 
-	all_phi_nodes = pset_new_ptr_default();
+	all_phi_nodes   = pset_new_ptr_default();
 	all_phi_classes = pset_new_ptr_default();
-	all_copy_nodes = pset_new_ptr_default();
-	phi_class_init();
+	all_copy_nodes  = pset_new_ptr_default();
+	memset(curr_vals, 0, sizeof(curr_vals));
 }
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copystat);
 
-void copystat_reset(void) {
-	int i;
-	for (i = 0; i < ASIZE; ++i)
-		curr_vals[i] = 0;
+void be_quit_copystat(void) {
 	del_pset(all_phi_nodes);
 	del_pset(all_phi_classes);
 	del_pset(all_copy_nodes);
-	all_phi_nodes = pset_new_ptr_default();
-	all_phi_classes = pset_new_ptr_default();
-	all_copy_nodes = pset_new_ptr_default();
+}
+BE_REGISTER_MODULE_DESTRUCTOR(be_quit_copystat);
+
+void copystat_reset(void) {
+	be_quit_copystat();
+	be_init_copystat();
 }
 
 /**
@@ -108,20 +124,46 @@ static void irg_stat_walker(ir_node *node, void *env) {
 	arch_env_t *arch_env = env;
 	curr_vals[I_ALL_NODES]++; /* count all nodes */
 
- 	if (is_Block(node)) /* count all blocks */
- 		curr_vals[I_BLOCKS]++;
+	if (is_Block(node)) /* count all blocks */
+		curr_vals[I_BLOCKS]++;
+
+	if (is_Reg_Phi(node)) /* collect phis */
+		pset_insert_ptr(all_phi_nodes, node);
 
- 	if (is_Phi(node)) /* collect phis */
- 		pset_insert_ptr(all_phi_nodes, node);
+	if (is_Perm_Proj(arch_env, node))
+		pset_insert_ptr(all_copy_nodes, node);
 
- 	if (is_Copy(arch_env, node))
- 		pset_insert_ptr(all_copy_nodes, node);
+	/* TODO: Add 2-Addr-Code nodes */
 }
 
-void copystat_collect_irg(ir_graph *irg, arch_env_t *arch_env) {
+static void copystat_collect_irg(ir_graph *irg, arch_env_t *arch_env) {
 	irg_walk_graph(irg, irg_stat_walker, NULL, arch_env);
-	curr_vals[I_BLOCKS] -= 2; /* substract 2 for start and end block */
 	all_phi_classes = phi_class_compute_by_phis(all_phi_nodes);
+	last_irg = irg;
+}
+
+/**
+ * @return 1 if the block at pos @p pos removed a critical edge
+ * 		   0 else
+ */
+static INLINE int was_edge_critical(const ir_node *bl, int pos) {
+	const ir_edge_t *edge;
+	const ir_node *bl_at_pos, *bl_before;
+	assert(is_Block(bl));
+
+	/* Does bl have several predecessors ?*/
+	if (get_irn_arity(bl) <= 1)
+		return 0;
+
+	/* Does the pred have exactly one predecessor */
+	bl_at_pos = get_irn_n(bl, pos);
+	if (get_irn_arity(bl_at_pos) != 1)
+		return 0;
+
+	/* Does the pred of the pred have several successors */
+	bl_before = get_irn_n(bl_at_pos, 0);
+	edge = get_block_succ_first(bl_before);
+	return get_block_succ_next(bl_before, edge) ? 1 : 0;
 }
 
 /**
@@ -129,6 +171,7 @@ void copystat_collect_irg(ir_graph *irg, arch_env_t *arch_env) {
  */
 static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi) {
  	int arity, i;
+ 	ir_node *phi_bl;
 	assert(is_Phi(phi));
 
 	/* count all phi phis */
@@ -142,10 +185,11 @@ static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi) {
 	else
 		curr_vals[I_PHI_ARITY_S + arity]++;
 
+	phi_bl = get_nodes_block(phi);
 	/* type of argument {self, const, pred, glob} */
 	for (i = 0; i < arity; i++) {
         ir_node *block_of_arg, *block_ith_pred;
-		ir_node *cfg_node, *arg = get_irn_n(phi, i);
+		ir_node *arg = get_irn_n(phi, i);
 
 		if (arg == phi) {
 			curr_vals[I_PHI_ARG_SELF]++;
@@ -157,17 +201,12 @@ static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi) {
 			continue;
 		}
 
-		block_of_arg = get_nodes_block(arg);
-
 		/* get the pred block skipping blocks on critical edges */
-		cfg_node = get_irn_n(get_nodes_block(phi), i);
-		block_ith_pred = get_nodes_block(cfg_node);
-		if (get_irn_opcode(cfg_node) == iro_Jmp && get_irn_arity(block_ith_pred) == 1) {
-			/* Then cfg_node_block has exactly 1 pred and 1 succ block,
-			 * thus it must have been inserted during remove_critical_edges */
+		block_ith_pred = get_Block_cfgpred_block(phi_bl, i);
+		if (was_edge_critical(phi_bl, i))
 			block_ith_pred = get_Block_cfgpred_block(block_ith_pred, 0);
-		}
 
+		block_of_arg = get_nodes_block(arg);
 		if (block_of_arg == block_ith_pred) {
 			curr_vals[I_PHI_ARG_PRED]++;
 			continue;
@@ -181,9 +220,10 @@ static void stat_phi_node(be_chordal_env_t *chordal_env, ir_node *phi) {
  * Collect register-constrained node data
  */
 static void stat_copy_node(be_chordal_env_t *chordal_env, ir_node *root) {
+	be_lv_t *lv = be_get_birg_liveness(chordal_env->birg);
 	curr_vals[I_CPY_CNT]++;
 	curr_vals[I_COPIES_MAX]++;
-	if (nodes_interfere(chordal_env, root, get_Copy_src(root))) {
+	if (values_interfere(lv, root, get_Perm_src(root))) {
 		curr_vals[I_COPIES_IF]++;
 		assert(0 && "A Perm pair (in/out) should never interfere!");
 	}
@@ -195,6 +235,7 @@ static void stat_copy_node(be_chordal_env_t *chordal_env, ir_node *root) {
 static void stat_phi_class(be_chordal_env_t *chordal_env, pset *pc) {
 	int i, o, size, if_free, phis;
 	ir_node **members, *p;
+	be_lv_t *lv = be_get_birg_liveness(chordal_env->birg);
 
 	/* phi class count */
 	curr_vals[I_CLS_CNT]++;
@@ -208,13 +249,16 @@ static void stat_phi_class(be_chordal_env_t *chordal_env, pset *pc) {
 
 	/* get an array of all members for double iterating */
 	members = xmalloc(size * sizeof(*members));
-	for (i = 0, p = pset_first(pc); p; p = pset_next(pc))
+	DBG((dbg, LEVEL_2, "Phi-class:\n"));
+	for (i = 0, p = pset_first(pc); p; p = pset_next(pc)) {
+		DBG((dbg, LEVEL_2, "  %+F\n", p));
 		members[i++] = p;
+	}
 	assert(i == size);
 
 	/* determine number of phis on this class */
 	phis = 0;
-	for (i = 0; i < size-1; ++i)
+	for (i = 0; i < size; ++i)
 		if (is_Phi(members[i]))
 			phis++;
 	if (phis > MAX_CLS_PHIS)
@@ -227,7 +271,7 @@ static void stat_phi_class(be_chordal_env_t *chordal_env, pset *pc) {
 	if_free = 1;
 	for (i = 0; i < size-1; ++i)
 		for (o = i+1; o < size; ++o)
-			if (nodes_interfere(chordal_env, members[i], members[o])) {
+			if (values_interfere(lv, members[i], members[o])) {
 				if_free = 0;
 				curr_vals[I_CLS_IF_CNT]++;
 			}
@@ -238,27 +282,28 @@ static void stat_phi_class(be_chordal_env_t *chordal_env, pset *pc) {
 	xfree(members);
 }
 
-#define is_curr_reg_class(irn) \
-  (arch_get_irn_reg_class(chordal_env->session_env->main_env->arch_env, irn, \
-                          arch_pos_make_out(0)) == chordal_env->cls)
-
-void copystat_collect_cls(be_chordal_env_t *chordal_env) {
+void copystat_collect_cls(be_chordal_env_t *cenv) {
 	ir_node *n;
 	pset *pc;
+	ir_graph *irg = cenv->irg;
+	arch_env_t *aenv = cenv->birg->main_env->arch_env;
+
+	copystat_reset();
+	copystat_collect_irg(irg, aenv);
 
 	for (n = pset_first(all_phi_nodes); n; n = pset_next(all_phi_nodes))
-		if (is_curr_reg_class(n))
-			stat_phi_node(chordal_env, n);
+		if (arch_get_irn_reg_class(aenv, n, -1) == cenv->cls)
+			stat_phi_node(cenv, n);
 
 	for (n = pset_first(all_copy_nodes); n; n = pset_next(all_copy_nodes))
-		if (is_curr_reg_class(n))
-			stat_copy_node(chordal_env, n);
+		if (arch_get_irn_reg_class(aenv, n, -1) == cenv->cls)
+			stat_copy_node(cenv, n);
 
 	for (pc = pset_first(all_phi_classes); pc; pc = pset_next(all_phi_classes)) {
 		ir_node *member = pset_first(pc);
 		pset_break(pc);
-		if (is_curr_reg_class(member))
-			stat_phi_class(chordal_env, pc);
+		if (arch_get_irn_reg_class(aenv, member, -1) == cenv->cls)
+			stat_phi_class(cenv, pc);
 	}
 }
 
@@ -274,35 +319,58 @@ void copystat_add_init_costs(int costs) {
 void copystat_add_heur_costs(int costs) {
 	curr_vals[I_COPIES_HEUR] += costs;
 }
+void copystat_add_ilp_5_sec_costs(int costs) {
+	curr_vals[I_COPIES_5SEC] += costs;
+}
+void copystat_add_ilp_30_sec_costs(int costs) {
+	curr_vals[I_COPIES_30SEC] += costs;
+}
 void copystat_add_opt_costs(int costs) {
 	curr_vals[I_COPIES_OPT] += costs;
 }
+void copystat_add_heur_time(int time) {
+	curr_vals[I_HEUR_TIME] += time;
+}
+
+#ifdef WITH_ILP
+
 void copystat_add_ilp_time(int time) {
 	curr_vals[I_ILP_TIME] += time;
 }
+void copystat_add_ilp_vars(int vars) {
+	curr_vals[I_ILP_VARS] += vars;
+}
+void copystat_add_ilp_csts(int csts) {
+	curr_vals[I_ILP_CSTR] += csts;
+}
 void copystat_add_ilp_iter(int iters) {
 	curr_vals[I_ILP_ITER] += iters;
 }
 
+#endif /* WITH_ILP */
+
 void copystat_dump(ir_graph *irg) {
 	int i;
 	char buf[1024];
 	FILE *out;
 
 	snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
+	buf[sizeof(buf) - 1] = '\0';
 	out = ffopen(buf, "stat", "wt");
 
-	fprintf(out, "%s\n", get_irp_prog_name());
+	fprintf(out, "%d\n", ASIZE);
 	for (i = 0; i < ASIZE; i++) {
+#if 0
 		if (i >= I_PHI_ARITY_S && i <= I_PHI_ARITY_E)
 			fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_PHI_CNT]);
 		else if (i >= I_CLS_SIZE_S && i <= I_CLS_SIZE_E)
 			fprintf(out, "%i %i\n", curr_vals[i], curr_vals[I_CLS_CNT]);
 		else
+#endif
 			fprintf(out, "%i\n", curr_vals[i]);
 	}
 
-    fclose(out);
+	fclose(out);
 }
 
 void copystat_dump_pretty(ir_graph *irg) {
@@ -311,6 +379,7 @@ void copystat_dump_pretty(ir_graph *irg) {
 	FILE *out;
 
 	snprintf(buf, sizeof(buf), "%s__%s", get_irp_prog_name(), get_entity_name(get_irg_entity(irg)));
+	buf[sizeof(buf) - 1] = '\0';
 	out = ffopen(buf, "pstat", "wt");
 
 	fprintf(out, "Nodes     %4d\n", curr_vals[I_ALL_NODES]);
@@ -352,4 +421,150 @@ void copystat_dump_pretty(ir_graph *irg) {
 	fclose(out);
 }
 
+/**
+ * Helpers for saving and restoring colors of nodes.
+ * Used to get dependable and comparable benchmark results.
+ */
+typedef struct color_saver {
+	arch_env_t *arch_env;
+	be_chordal_env_t *chordal_env;
+	pmap *saved_colors;
+	int flag; /* 0 save, 1 load */
+} color_save_t;
+
+static void save_load(ir_node *irn, void *env) {
+	color_save_t *saver = env;
+	if (saver->chordal_env->cls == arch_get_irn_reg_class(saver->arch_env, irn, -1)) {
+		if (saver->flag == 0) { /* save */
+			const arch_register_t *reg = arch_get_irn_register(saver->arch_env, irn);
+			pmap_insert(saver->saved_colors, irn, (void *) reg);
+		} else { /*load */
+			arch_register_t *reg = pmap_get(saver->saved_colors, irn);
+			arch_set_irn_register(saver->arch_env, irn, reg);
+		}
+	}
+}
+
+static void save_colors(color_save_t *color_saver) {
+	color_saver->flag = 0;
+	irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
+}
+
+#ifdef WITH_ILP
+static void load_colors(color_save_t *color_saver) {
+	color_saver->flag = 1;
+	irg_walk_graph(color_saver->chordal_env->irg, save_load, NULL, color_saver);
+}
 #endif
+
+/**
+ * Main compare routine
+ */
+void co_compare_solvers(be_chordal_env_t *chordal_env) {
+	copy_opt_t *co;
+	lc_timer_t *timer;
+	color_save_t saver;
+	int costs_inevit, costs_init, costs_solved, lower_bound;
+
+	phi_class_compute(chordal_env->irg);
+	copystat_collect_cls(chordal_env);
+
+	co = new_copy_opt(chordal_env, co_get_costs_loop_depth);
+	co_build_ou_structure(co);
+	co_build_graph_structure(co);
+	DBG((dbg, LEVEL_1, "----> CO: %s\n", co->name));
+
+	/* save colors */
+	saver.arch_env     = chordal_env->birg->main_env->arch_env;
+	saver.chordal_env  = chordal_env;
+	saver.saved_colors = pmap_create();
+	save_colors(&saver);
+
+	/* initial values */
+	costs_inevit = co_get_inevit_copy_costs(co);
+	lower_bound  = co_get_lower_bound(co);
+	costs_init   = co_get_copy_costs(co);
+
+	DBG((dbg, LEVEL_1, "Inevit Costs: %3d\n", costs_inevit));
+	DBG((dbg, LEVEL_1, "Lower Bound: %3d\n", lower_bound));
+	DBG((dbg, LEVEL_1, "Init costs: %3d\n", costs_init));
+
+	copystat_add_inevit_costs(costs_inevit);
+	copystat_add_init_costs(costs_init);
+	copystat_add_max_costs(co_get_max_copy_costs(co));
+
+	/* heuristic 1 (Daniel Grund) */
+	timer = lc_timer_register("heur1", NULL);
+	lc_timer_reset_and_start(timer);
+
+	co_solve_heuristic(co);
+
+	lc_timer_stop(timer);
+
+	costs_solved = co_get_copy_costs(co);
+	DBG((dbg, LEVEL_1, "HEUR1 costs: %3d\n", costs_solved));
+	copystat_add_heur_time(lc_timer_elapsed_msec(timer));
+	copystat_add_heur_costs(costs_solved);
+	assert(lower_bound <= costs_solved);
+
+	/* heuristic 2 (Sebastian Hack) */
+	timer = lc_timer_register("heur2", NULL);
+	lc_timer_reset_and_start(timer);
+
+	co_solve_heuristic_new(co);
+
+	lc_timer_stop(timer);
+
+	costs_solved = co_get_copy_costs(co);
+	DBG((dbg, LEVEL_1, "HEUR2 costs: %3d\n", costs_solved));
+	copystat_add_heur_time(lc_timer_elapsed_msec(timer));
+	copystat_add_heur_costs(costs_solved);
+	assert(lower_bound <= costs_solved);
+
+	/* Park & Moon register coalescing (Kimon Hoffmann) */
+	timer = lc_timer_register("park", NULL);
+	lc_timer_reset_and_start(timer);
+
+	co_solve_park_moon(co);
+
+	lc_timer_stop(timer);
+
+	costs_solved = co_get_copy_costs(co);
+	DBG((dbg, LEVEL_1, "Park/Moon costs: %3d\n", costs_solved));
+	copystat_add_heur_time(lc_timer_elapsed_msec(timer));
+	copystat_add_heur_costs(costs_solved);
+	assert(lower_bound <= costs_solved);
+
+
+#ifdef WITH_ILP
+
+	/* ILP 1 is not yet implemented, so it makes no sense to compare */
+#if 0
+	load_colors(&saver);
+
+	co_solve_ilp1(co, 60.0);
+
+	costs_solved = co_get_copy_costs(co);
+	DBG((dbg, LEVEL_1, "ILP1 costs: %3d\n", costs_solved));
+	copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
+	assert(lower_bound <= costs_solved);
+#endif /* 0 */
+
+	/* ILP 2 */
+	load_colors(&saver);
+
+	co_solve_ilp2(co);
+
+	costs_solved = co_get_copy_costs(co);
+	DBG((dbg, LEVEL_1, "ILP2 costs: %3d\n", costs_solved));
+	copystat_add_opt_costs(costs_solved); /* TODO: ADAPT */
+	assert(lower_bound <= costs_solved);
+
+#endif /* WITH_ILP */
+
+	/* free memory for statistic structures */
+	pmap_destroy(saver.saved_colors);
+	co_free_graph_structure(co);
+	co_free_ou_structure(co);
+	free_copy_opt(co);
+}