X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur.c;h=58158c17a1099a7cb4a37c9f619b5200f20f74f0;hb=1ef3d57f913e2f533aba0ab6b22f4d66223b86ed;hp=a76b904384a13e42f648c2c7fed499f14008a3f4;hpb=c7c9e568001eed38d54ef24036784ef0fb90df03;p=libfirm diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c index a76b90438..58158c17a 100644 --- a/ir/be/becopyheur.c +++ b/ir/be/becopyheur.c @@ -1,42 +1,59 @@ -/** - * 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. */ -#ifdef HAVE_CONFIG_H #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; } /** * If a local pinned conflict occurs, a new edge in the conflict graph is added. * The next maximum independent set build, will regard it. */ -static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) { +static inline void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) { 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 { @@ -98,62 +125,63 @@ static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, cons /** * Checks if two nodes are in a conflict. */ -static INLINE int qnode_are_conflicting(const qnode_t *qn, const ir_node *n1, const ir_node *n2) { +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)); } /** * Finds a node status entry of a node if existent. Otherwise it will return * an initialized new entry for this node. */ -static INLINE node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) { +static inline node_stat_t *qnode_find_or_insert_node(const qnode_t *qn, ir_node *irn) { node_stat_t find; 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); +static inline int qnode_get_new_color(const qnode_t *qn, ir_node *irn) { + const node_stat_t *found = qnode_find_node(qn, irn); if (found) return found->new_color; else - return get_irn_col(qn->ou->co, irn); + return get_irn_col(irn); } /** * Sets the virtual color of a node. */ -static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) { +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)); @@ -164,8 +192,8 @@ static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int colo * to the same optimization unit and has been optimized before the current * 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); +static inline int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) { + const node_stat_t *found = qnode_find_node(qn, irn); if (found) return found->pinned_local; else @@ -176,88 +204,130 @@ static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) { * Local-pins a node, so optimizations of further nodes of the same opt unit * can handle situations in which a color change would undo prior optimizations. */ -static INLINE void qnode_pin_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); + found->new_color = get_irn_col(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) { - const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env; - const arch_register_class_t *cls = chordal_env->cls; - const arch_env_t *arch_env = chordal_env->main_env->arch_env; + copy_opt_t *co = qn->ou->co; + const be_chordal_env_t *chordal_env = co->cenv; + const arch_register_class_t *cls = co->cls; 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_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)) { 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))) { + +#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_out(irn); + 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); + } + + /* Exclude the color of the irn, because it must _change_ its color */ + bitset_clear(free_cols, irn_col); + + /* 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; + } + } +#endif /* SEARCH_FREE_COLORS */ + + /* If target color is not allocatable changing color is impossible */ + if (!arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, col))) { DBG((dbg, LEVEL_3, "\t %+F impossible\n", irn)); return CHANGE_IMPOSSIBLE; } /* - * Process all nodes which would conflict with this change + * 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_t *ifg = chordal_env->ifg; - void *iter = be_ifg_neighbours_iter_alloca(ifg); - ir_node *sub_res, *curr; - - /* - * 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; - } + 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 have been applied. - * So it's save to change this irn + * 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; } + /** * Tries to set the colors for all members of this queue node; * to the target color qn->color @@ -279,6 +349,7 @@ 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 */ @@ -298,10 +369,8 @@ 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; } @@ -310,21 +379,22 @@ static int qnode_try_color(const qnode_t *qn) { * Determines a maximum weighted independent set with respect to * the interference and conflict edges of all nodes in a qnode. */ -static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) { +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. * 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)); + safe = ALLOCAN(ir_node*, ou->node_count - 1); + safe_costs = 0; + safe_count = 0; + unsafe = ALLOCAN(ir_node*, ou->node_count - 1); + unsafe_costs = ALLOCAN(int, ou->node_count - 1); unsafe_count = 0; for(i=1; inode_count; ++i) { int is_safe = 1; @@ -406,12 +476,12 @@ no_stable_set: /** * Creates a new qnode */ -static INLINE qnode_t *new_qnode(const unit_t *ou, int color) { - qnode_t *qn = xmalloc(sizeof(*qn)); - qn->ou = ou; - qn->color = color; - qn->mis = malloc(ou->node_count * sizeof(*qn->mis)); - qn->conflicts = new_set(set_cmp_conflict_t, SLOTS_CONFLICTS); +static inline qnode_t *new_qnode(const unit_t *ou, int color) { + qnode_t *qn = XMALLOC(qnode_t); + qn->ou = ou; + qn->color = color; + qn->mis = XMALLOCN(ir_node*, ou->node_count); + 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; } @@ -419,7 +489,7 @@ static INLINE qnode_t *new_qnode(const unit_t *ou, int color) { /** * Frees space used by a queue node */ -static INLINE void free_qnode(qnode_t *qn) { +static inline void free_qnode(qnode_t *qn) { del_set(qn->conflicts); del_set(qn->changed_nodes); xfree(qn->mis); @@ -430,7 +500,7 @@ static INLINE void free_qnode(qnode_t *qn) { * Inserts a qnode in the sorted queue of the optimization unit. Queue is * ordered by field 'size' (the size of the mis) in decreasing order. */ -static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { +static inline void ou_insert_qnode(unit_t *ou, qnode_t *qn) { struct list_head *lh; if (qnode_are_conflicting(qn, ou->nodes[0], ou->nodes[0])) { @@ -460,9 +530,13 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { * nodes. (All other phi classes are reduced to this case.) */ 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); + qnode_t *curr = NULL; + qnode_t *tmp; + const arch_register_req_t *req; + bitset_t const* ignore; + bitset_pos_t n_regs; + bitset_pos_t idx; + int i; DBG((dbg, LEVEL_1, "\tOptimizing unit:\n")); for (i=0; inode_count; ++i) @@ -470,12 +544,33 @@ 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, pos_regs); - bitset_foreach(pos_regs, i) - ou_insert_qnode(ou, new_qnode(ou, i)); + + req = arch_get_register_req_out(ou->nodes[0]); + ignore = ou->co->cenv->ignore_colors; + n_regs = req->cls->n_regs; + if (arch_register_req_is(req, limited)) { + rawbs_base_t const* limited = req->limited; + + for (idx = 0; idx != n_regs; ++idx) { + if (bitset_is_set(ignore, idx)) + continue; + if (!rbitset_is_set(limited, idx)) + continue; + + ou_insert_qnode(ou, new_qnode(ou, idx)); + } + } else { + for (idx = 0; idx != n_regs; ++idx) { + if (bitset_is_set(ignore, idx)) + continue; + + 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); @@ -521,9 +616,11 @@ 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"); + 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) @@ -531,4 +628,5 @@ void co_heur_opt(copy_opt_t *co) { ou_optimize(curr); del_pset(pinned_global); + return 0; }