X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur.c;h=3b1c1d3231893751b17b4b9c6c818e60b32fbb13;hb=5bb8cd35a8074112fa2da8b8f68d7c77918119e5;hp=a859c5845dcdb33c45916d3a100a82e91250d087;hpb=4794cb5f95f18ed6afee5c1cc33aaf5f39d3e4a2;p=libfirm diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c index a859c5845..3b1c1d323 100644 --- a/ir/be/becopyheur.c +++ b/ir/be/becopyheur.c @@ -1,35 +1,51 @@ -/** - * 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_t.h" #include "becopystat.h" -#include "bitset.h" +#include "beintlive_t.h" +#include "beirg.h" +#include "bemodule.h" -static firm_dbg_module_t *dbg = NULL; +DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) #define SEARCH_FREE_COLORS @@ -38,7 +54,7 @@ static firm_dbg_module_t *dbg = NULL; #define SLOTS_CHANGED_NODES 32 #define list_entry_queue(lh) list_entry(lh, qnode_t, queue) -#define HASH_CONFLICT(c) (HASH_PTR(c.n1) ^ HASH_PTR(c.n2)) +#define HASH_CONFLICT(c) (hash_irn(c.n1) ^ hash_irn(c.n2)) /** * Modeling additional conflicts between nodes. NOT live range interference @@ -53,41 +69,53 @@ typedef struct _conflict_t { */ typedef struct _node_stat_t { ir_node *irn; - int new_color; - int pinned_local :1; + int new_color; + int pinned_local :1; } node_stat_t; /** * Represents a node in the optimization queue. */ typedef struct _qnode_t { - struct list_head queue; /**< chaining of unit_t->queue */ - 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 int set_cmp_conflict_t(const void *x, const void *y, size_t size) { +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 be_values_interfere(env->birg->lv, 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 { @@ -100,62 +128,69 @@ 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->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; +static int set_cmp_node_stat_t(const void *x, const void *y, size_t size) +{ + (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)); @@ -166,8 +201,9 @@ 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 @@ -178,11 +214,12 @@ 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); } @@ -191,7 +228,6 @@ static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *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 @@ -211,11 +247,11 @@ static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) { * Else the first conflicting ir_node encountered is returned. * */ -static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) { +static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) +{ 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; @@ -242,18 +278,19 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const */ if (irn != trigger) { bitset_t *free_cols = bitset_alloca(cls->n_regs); - arch_register_req_t req; + const 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); + bitset_copy(free_cols, co->cenv->ignore_colors); + bitset_flip_all(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)) { + req = arch_get_register_req_out(irn); + if (arch_register_req_is(req, limited)) { bitset_t *limited = bitset_alloca(cls->n_regs); - req.limited(req.limited_env, limited); + rbitset_copy_to_bitset(req->limited, limited); bitset_and(free_cols, limited); } @@ -274,7 +311,7 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const #endif /* SEARCH_FREE_COLORS */ /* 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))) { + 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; } @@ -309,7 +346,8 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const * @returns 1 iff all members colors could be set * 0 else */ -static int qnode_try_color(const qnode_t *qn) { +static int qnode_try_color(const qnode_t *qn) +{ int i; for (i=0; imis_size; ++i) { ir_node *test_node, *confl_node; @@ -324,6 +362,7 @@ 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 */ @@ -343,10 +382,8 @@ 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; } @@ -355,25 +392,27 @@ 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; + unsigned pos; + int 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) { + for (i=1; inode_count; ++i) { int is_safe = 1; - for(o=1; onode_count; ++o) { + for (o=1; onode_count; ++o) { if (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) { if (i!=o) { unsafe_costs[unsafe_count] = ou->costs[i]; @@ -413,7 +452,7 @@ static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) { /* Exact Algorithm: Brute force */ curr = bitset_alloca(unsafe_count); bitset_set_all(curr); - while ((max = bitset_popcnt(curr)) != 0) { + while (!bitset_is_empty(curr)) { /* 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) */ @@ -438,7 +477,7 @@ no_stable_set: } /* transfer the best set into the qn */ - qn->mis_size = 1+safe_count+bitset_popcnt(best); + qn->mis_size = 1+safe_count+bitset_popcount(best); qn->mis_costs = safe_costs+best_weight; qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */ next = 1; @@ -451,12 +490,13 @@ 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 = xmalloc(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; } @@ -464,7 +504,8 @@ 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); @@ -475,7 +516,8 @@ 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])) { @@ -504,13 +546,15 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) { * case for approximately 80% of all phi classes and 100% of register constrained * nodes. (All other phi classes are reduced to this case.) */ -static void ou_optimize(unit_t *ou) { - int i; - 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); +static void ou_optimize(unit_t *ou) +{ + qnode_t *curr = NULL; + qnode_t *tmp; + const arch_register_req_t *req; + bitset_t const* ignore; + unsigned n_regs; + unsigned idx; + int i; DBG((dbg, LEVEL_1, "\tOptimizing unit:\n")); for (i=0; inode_count; ++i) @@ -519,20 +563,32 @@ static void ou_optimize(unit_t *ou) { /* init queue */ INIT_LIST_HEAD(&ou->queue); - arch_get_allocatable_regs(aenv, ou->nodes[0], -1, pos_regs); + 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)) { + unsigned const* limited = req->limited; - /* exclude ingore colors */ - arch_put_non_ignore_regs(aenv, cls, ign_regs); - bitset_and(pos_regs, ign_regs); + for (idx = 0; idx != n_regs; ++idx) { + if (bitset_is_set(ignore, idx)) + continue; + if (!rbitset_is_set(limited, idx)) + continue; - assert(bitset_popcnt(pos_regs) != 0 && "No register is allowed for this node !!?"); + ou_insert_qnode(ou, new_qnode(ou, idx)); + } + } else { + for (idx = 0; idx != n_regs; ++idx) { + if (bitset_is_set(ignore, idx)) + continue; - /* create new qnode */ - bitset_foreach(pos_regs, i) - ou_insert_qnode(ou, new_qnode(ou, i)); + 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); @@ -578,9 +634,13 @@ static void ou_optimize(unit_t *ou) { free_qnode(curr); } -int co_solve_heuristic(copy_opt_t *co) { +/** + * Solves the problem using a heuristic approach + * Uses the OU data structure + */ +int co_solve_heuristic(copy_opt_t *co) +{ unit_t *curr; - dbg = firm_dbg_register("ir.be.copyoptheur"); ASSERT_OU_AVAIL(co); @@ -590,6 +650,16 @@ int co_solve_heuristic(copy_opt_t *co) { ou_optimize(curr); del_pset(pinned_global); - return 0; } + +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur); +void be_init_copyheur(void) +{ + static co_algo_info copyheur = { + co_solve_heuristic, 0 + }; + + be_register_copyopt("heur1", ©heur); + FIRM_DBG_REGISTER(dbg, "ir.be.copyoptheur"); +}