X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur.c;h=7c669b5e3fdca20b01075eea622774ce7a6ef62d;hb=ccdf802c598f7adc5e35fff42fed029c59f48a57;hp=bf75e77f5f8a28c361ce19e3a6d9b36e415349d7;hpb=73567efae10b02832d0d1817eba56ecc96178424;p=libfirm diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c index bf75e77f5..7c669b5e3 100644 --- a/ir/be/becopyheur.c +++ b/ir/be/becopyheur.c @@ -1,13 +1,33 @@ -/** - * Author: Daniel Grund - * Date: 12.04.2005 - * Copyright: (c) Universitaet Karlsruhe - * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. +/* + * Copyright (C) 1995-2008 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 First simple copy minimization heuristics. + * @author Daniel Grund + * @date 12.04.2005 + * @version $Id$ + * * Heuristic for minimizing copies using a queue which holds 'qnodes' not yet * examined. A qnode has a 'target color', nodes out of the opt unit and * a 'conflict graph'. 'Conflict graph' = "Interference graph' + 'conflict edges' - * A 'max indep set' is determined form these. We try to color this mis using a + * A 'max indep set' is determined from these. We try to color this mis using a * color-exchanging mechanism. Occuring conflicts are modeled with 'conflict edges' * and the qnode is reinserted in the queue. The first qnode colored without * conflicts is the best one. @@ -16,27 +36,26 @@ #include "config.h" #endif -#ifdef HAVE_ALLOCA_H -#include -#endif -#ifdef HAVE_MALLOC_H -#include -#endif - +#include "debug.h" +#include "bitset.h" +#include "raw_bitset.h" #include "xmalloc.h" -#include "becopyopt.h" + +#include "becopyopt_t.h" #include "becopystat.h" -#include "bitset.h" +#include "beintlive_t.h" +#include "beirg_t.h" + +DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) -static firm_dbg_module_t *dbg = NULL; +#define SEARCH_FREE_COLORS -#define SLOTS_PINNED_GLOBAL 256 +#define SLOTS_PINNED_GLOBAL 64 #define SLOTS_CONFLICTS 8 #define SLOTS_CHANGED_NODES 32 -#define MIN(a,b) ((aqueue */ - const unit_t *ou; /**< the opt unit this qnode belongs to */ - int color; /**< target color */ - set *conflicts; /**< contains conflict_t's. All internal conflicts */ - int mis_costs; /**< costs of nodes/copies in the mis. */ - int mis_size; /**< size of the array below */ - ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */ - set *changed_nodes; /**< contains node_stat_t's. */ + struct list_head queue; /**< chaining of unit_t->queue */ + const unit_t *ou; /**< the opt unit this node belongs to */ + int color; /**< target color */ + set *conflicts; /**< contains conflict_t's. All internal conflicts */ + int mis_costs; /**< costs of nodes/copies in the mis. */ + int mis_size; /**< size of the array below */ + ir_node **mis; /**< the nodes of unit_t->nodes[] being part of the max independent set */ + set *changed_nodes; /**< contains node_stat_t's. */ } qnode_t; static pset *pinned_global; /**< optimized nodes should not be altered any more */ +static INLINE int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b) +{ + if (env->ifg) + return be_ifg_connected(env->ifg, a, b); + else + return values_interfere(env->birg, a, b); +} + static int set_cmp_conflict_t(const void *x, const void *y, size_t size) { const conflict_t *xx = x; const conflict_t *yy = y; - return ! (xx->n1 == yy->n1 && xx->n2 == yy->n2); + (void) size; + + return xx->n1 != yy->n1 || xx->n2 != yy->n2; } /** @@ -85,7 +114,7 @@ static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, cons conflict_t c; DBG((dbg, LEVEL_4, "\t %+F -- %+F\n", n1, n2)); - if ((int)n1 < (int)n2) { + if (get_irn_idx(n1) < get_irn_idx(n2)) { c.n1 = n1; c.n2 = n2; } else { @@ -101,30 +130,31 @@ static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, cons 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) { + if (get_irn_idx(n1) < get_irn_idx(n2)) { c.n1 = n1; c.n2 = n2; } else { c.n1 = n2; c.n2 = n1; } - return (int) set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)); + return set_find(qn->conflicts, &c, sizeof(c), HASH_CONFLICT(c)) != 0; } static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) { - return ((node_stat_t *)x)->irn != ((node_stat_t *)y)->irn; + (void) size; + return ((const node_stat_t*)x)->irn != ((const node_stat_t*)y)->irn; } /** * Finds a node status entry of a node if existent. Otherwise return NULL */ -static INLINE node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) { +static INLINE const node_stat_t *qnode_find_node(const qnode_t *qn, ir_node *irn) { node_stat_t find; find.irn = irn; - return set_find(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn)); + return set_find(qn->changed_nodes, &find, sizeof(find), hash_irn(irn)); } /** @@ -136,14 +166,14 @@ static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node find.irn = irn; find.new_color = NO_COLOR; find.pinned_local = 0; - return set_insert(qn->changed_nodes, &find, sizeof(find), HASH_PTR(irn)); + return set_insert(qn->changed_nodes, &find, sizeof(find), hash_irn(irn)); } /** * Returns the virtual color of a node if set before, else returns the real color. */ static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) { - node_stat_t *found = qnode_find_node(qn, irn); + const node_stat_t *found = qnode_find_node(qn, irn); if (found) return found->new_color; else @@ -156,6 +186,7 @@ static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) { 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)); } /** @@ -164,7 +195,7 @@ static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int colo * processed node. */ static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) { - node_stat_t *found = qnode_find_node(qn, irn); + const node_stat_t *found = qnode_find_node(qn, irn); if (found) return found->pinned_local; else @@ -178,157 +209,128 @@ static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) { 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() */ #define CHANGE_SAVE NULL #define CHANGE_IMPOSSIBLE (ir_node *)1 -#define is_conflicting_node(n) (((int)n) > 1) /** * 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, - -1, - 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, -1, 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); + const arch_register_req_t *req; + ir_node *curr; + int free_col; + + /* Get all possible colors */ + bitset_copy(free_cols, co->cenv->ignore_colors); + bitset_flip_all(free_cols); + + /* Exclude colors not assignable to the irn */ + req = arch_get_register_req(arch_env, irn, -1); + if (arch_register_req_is(req, limited)) { + bitset_t *limited = bitset_alloca(cls->n_regs); + rbitset_copy_to_bitset(req->limited, 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, -1, 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 @@ -350,6 +352,7 @@ static int qnode_try_color(const qnode_t *qn) { } else if (confl_node == CHANGE_IMPOSSIBLE) { DBG((dbg, LEVEL_3, "\t Impossible --> remove from qnode\n")); qnode_add_conflict(qn, test_node, test_node); + return 0; } else { if (qnode_is_pinned_local(qn, confl_node)) { /* changing test_node would change back a node of current ou */ @@ -369,29 +372,12 @@ static int qnode_try_color(const qnode_t *qn) { DBG((dbg, LEVEL_3, "\t Conflicting global --> remove from qnode\n")); qnode_add_conflict(qn, test_node, test_node); } - } - - if (confl_node != CHANGE_SAVE) return 0; + } } return 1; } -typedef int(*confl_f)(const ir_node *a, const ir_node *b, void *data); - -/** - * @param result Gets filled with the computed maximum independent set. - * @param count The size of input arrays / the number of nodes - * @param nodes A set of nodes to copmute the max. ind. set for - * @param weights Weights associated to the nodes in @p nodes - * @param confl Callback function to decide if two values interfere - * @param data Passed into all callbacks - * @return The size of the computed set - */ -//int max_ind_set(ir_node **result, int count, ir_node **nodes, int *weights, confl_f confl, void *data) { -// -//} - /** * Determines a maximum weighted independent set with respect to * the interference and conflict edges of all nodes in a qnode. @@ -400,7 +386,8 @@ static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) { ir_node **safe, **unsafe; int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs; bitset_t *curr, *best; - int max, next, pos, curr_weight, best_weight = 0; + bitset_pos_t pos; + int max, next, curr_weight, best_weight = 0; /* assign the nodes into two groups. * safe: node has no interference, hence it is in every max stable set. @@ -437,7 +424,7 @@ static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) { best = bitset_alloca(unsafe_count); if (unsafe_count > MIS_HEUR_TRIGGER) { - /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */ + /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */ for (i=0; iou = ou; qn->color = color; - qn->mis = malloc(ou->node_count * sizeof(*qn->mis)); + qn->mis = xmalloc(ou->node_count * sizeof(*qn->mis)); qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS); qn->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES); return qn; @@ -548,7 +535,10 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { static void ou_optimize(unit_t *ou) { int i; qnode_t *curr = NULL, *tmp; - bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs); + const arch_env_t *aenv = ou->co->aenv; + const arch_register_class_t *cls = ou->co->cls; + bitset_pos_t idx; + bitset_t *pos_regs = bitset_alloca(cls->n_regs); DBG((dbg, LEVEL_1, "\tOptimizing unit:\n")); for (i=0; inode_count; ++i) @@ -556,12 +546,21 @@ static void ou_optimize(unit_t *ou) { /* init queue */ INIT_LIST_HEAD(&ou->queue); - arch_get_allocatable_regs(get_arch_env(ou->co), ou->nodes[0], -1, ou->co->chordal_env->cls, pos_regs); - bitset_foreach(pos_regs, i) - ou_insert_qnode(ou, new_qnode(ou, i)); + + arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs); + + /* exclude ingore colors */ + bitset_andnot(pos_regs, ou->co->cenv->ignore_colors); + + assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?"); + + /* create new qnode */ + bitset_foreach(pos_regs, idx) + ou_insert_qnode(ou, new_qnode(ou, idx)); /* search best */ - while (!list_empty(&ou->queue)) { + for (;;) { + assert(!list_empty(&ou->queue)); /* get head of queue */ curr = list_entry_queue(ou->queue.next); list_del(&curr->queue); @@ -570,6 +569,7 @@ static void ou_optimize(unit_t *ou) { /* 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); @@ -579,13 +579,14 @@ static void ou_optimize(unit_t *ou) { /* 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; inode_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); } @@ -597,10 +598,6 @@ static void ou_optimize(unit_t *ou) { 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 */ @@ -609,13 +606,11 @@ static void ou_optimize(unit_t *ou) { 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); + FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur"); + + ASSERT_OU_AVAIL(co); pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL); list_for_each_entry(unit_t, curr, &co->units, units) @@ -623,4 +618,5 @@ void co_heur_opt(copy_opt_t *co) { ou_optimize(curr); del_pset(pinned_global); + return 0; }