X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur4.c;h=4b903b20d34698af8d01c0da502592c0a05118b0;hb=a8d6e45f13a40f6c9269608bf937e3e641d0b715;hp=42bdee7a10e480f7a84a8f565b0b1948071705a8;hpb=0e5f781d021c0b1e20ba08f583d4971419fedb5d;p=libfirm diff --git a/ir/be/becopyheur4.c b/ir/be/becopyheur4.c index 42bdee7a1..4b903b20d 100644 --- a/ir/be/becopyheur4.c +++ b/ir/be/becopyheur4.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -29,9 +29,9 @@ * (also known as "heur3" :) * Performs simple copy minimization. */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif /* HAVE_CONFIG_H */ + +#define DISABLE_STATEV #include @@ -78,13 +78,16 @@ static firm_dbg_module_t *dbg = NULL; #endif -static int last_chunk_id = 0; -static int recolor_limit = 4; -static double dislike_influence = 0.1; +typedef float real_t; +#define REAL(C) (C ## f) + +static unsigned last_chunk_id = 0; +static int recolor_limit = 7; +static real_t dislike_influence = REAL(0.1); typedef struct _col_cost_t { - int col; - double cost; + int col; + real_t cost; } col_cost_t; /** @@ -92,14 +95,13 @@ typedef struct _col_cost_t { */ typedef struct _aff_chunk_t { const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */ - bitset_t *nodes; /**< A bitset containing all nodes inside this chunk. */ - bitset_t *interfere; /**< A bitset containing all interfering neighbours of the nodes in this chunk. */ + const ir_node **interfere; /**< An ARR_F containing all inference. */ int weight; /**< Weight of this chunk */ unsigned weight_consistent : 1; /**< Set if the weight is consistent. */ - unsigned deleted : 1; /**< Set if the was deleted. */ - int id; /**< For debugging: An id of this chunk. */ - int visited; - col_cost_t *color_affinity; + unsigned deleted : 1; /**< For debugging: Set if the was deleted. */ + unsigned id; /**< An id of this chunk. */ + unsigned visited; + col_cost_t color_affinity[1]; } aff_chunk_t; /** @@ -108,7 +110,7 @@ typedef struct _aff_chunk_t { typedef struct _aff_edge_t { const ir_node *src; /**< Source node. */ const ir_node *tgt; /**< Target node. */ - double weight; /**< The weight of this edge. */ + int weight; /**< The weight of this edge. */ } aff_edge_t; /* main coalescing environment */ @@ -117,12 +119,11 @@ typedef struct _co_mst_env_t { int k; /**< number of non-ignore registers in class */ bitset_t *ignore_regs; /**< set containing all global ignore registers */ ir_phase ph; /**< phase object holding data for nodes */ - pqueue *chunks; /**< priority queue for chunks */ + pqueue_t *chunks; /**< priority queue for chunks */ pset *chunkset; /**< set holding all chunks */ be_ifg_t *ifg; /**< the interference graph */ - const arch_env_t *aenv; /**< the arch environment */ copy_opt_t *co; /**< the copy opt object */ - int chunk_visited; + unsigned chunk_visited; col_cost_t **single_cols; } co_mst_env_t; @@ -139,7 +140,7 @@ typedef struct _co_mst_irn_t { int tmp_col; /**< a temporary assigned color */ unsigned fixed : 1; /**< the color is fixed */ struct list_head list; /**< Queue for coloring undo. */ - double constr_factor; + real_t constr_factor; } co_mst_irn_t; #define get_co_mst_irn(mst_env, irn) (phase_get_or_set_irn_data(&(mst_env)->ph, (irn))) @@ -152,12 +153,13 @@ typedef int decide_func_t(const co_mst_irn_t *node, int col); * Write a chunk to stderr for debugging. */ static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) { - bitset_pos_t idx; + int i, l; + (void) env; if (c->weight_consistent) ir_fprintf(stderr, " $%d ", c->weight); ir_fprintf(stderr, "{"); - bitset_foreach(c->nodes, idx) { - ir_node *n = get_idx_irn(env->co->irg, idx); + for (i = 0, l = ARR_LEN(c->n); i < l; ++i) { + const ir_node *n = c->n[i]; ir_fprintf(stderr, " %+F,", n); } ir_fprintf(stderr, "}"); @@ -173,8 +175,9 @@ static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *n if (bitset_popcnt(node->adm_colors) < 1) fprintf(stderr, "no admissible colors?!?"); else { - bitset_foreach(node->adm_colors, idx) + bitset_foreach(node->adm_colors, idx) { fprintf(stderr, " %d", idx); + } } } @@ -189,7 +192,7 @@ static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) { #endif /* DEBUG_libfirm */ -static INLINE int get_mst_irn_col(const co_mst_irn_t *node) { +static inline int get_mst_irn_col(const co_mst_irn_t *node) { return node->tmp_col >= 0 ? node->tmp_col : node->col; } @@ -235,29 +238,28 @@ static int cmp_aff_edge(const void *a, const void *b) { static __attribute__((unused)) int cmp_col_cost_lt(const void *a, const void *b) { const col_cost_t *c1 = a; const col_cost_t *c2 = b; - double diff = c1->cost - c2->cost; + real_t diff = c1->cost - c2->cost; return (diff > 0) - (diff < 0); } static int cmp_col_cost_gt(const void *a, const void *b) { const col_cost_t *c1 = a; const col_cost_t *c2 = b; - double diff = c2->cost - c1->cost; + real_t diff = c2->cost - c1->cost; return (diff > 0) - (diff < 0); } /** * Creates a new affinity chunk */ -static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) { - aff_chunk_t *c = xmalloc(sizeof(*c)); +static inline aff_chunk_t *new_aff_chunk(co_mst_env_t *env) { + aff_chunk_t *c = XMALLOCF(aff_chunk_t, color_affinity, env->n_regs); + c->n = NEW_ARR_F(const ir_node *, 0); + c->interfere = NEW_ARR_F(const ir_node *, 0); c->weight = -1; c->weight_consistent = 0; - c->n = NEW_ARR_F(const ir_node *, 0); - c->nodes = bitset_irg_malloc(env->co->irg); - c->interfere = bitset_irg_malloc(env->co->irg); - c->color_affinity = xmalloc(env->n_regs * sizeof(c->color_affinity[0])); - c->id = last_chunk_id++; + c->deleted = 0; + c->id = ++last_chunk_id; c->visited = 0; pset_insert(env->chunkset, c, c->id); return c; @@ -266,34 +268,84 @@ static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) { /** * Frees all memory allocated by an affinity chunk. */ -static INLINE void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { +static inline void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { pset_remove(env->chunkset, c, c->id); - bitset_free(c->nodes); - bitset_free(c->interfere); - xfree(c->color_affinity); + DEL_ARR_F(c->interfere); DEL_ARR_F(c->n); c->deleted = 1; free(c); } +/** + * binary search of sorted nodes. + * + * @return the position where n is found in the array arr or ~pos + * if the nodes is not here. + */ +static inline int nodes_bsearch(const ir_node **arr, const ir_node *n) { + int hi = ARR_LEN(arr); + int lo = 0; + + while (lo < hi) { + int md = lo + ((hi - lo) >> 1); + + if (arr[md] == n) + return md; + if (arr[md] < n) + lo = md + 1; + else + hi = md; + } + + return ~lo; +} + +/** Check if a node n can be found inside arr. */ +static int node_contains(const ir_node **arr, const ir_node *n) { + int i = nodes_bsearch(arr, n); + return i >= 0; +} + +/** + * Insert a node into the sorted nodes list. + * + * @return 1 if the node was inserted, 0 else + */ +static int nodes_insert(const ir_node ***arr, const ir_node *irn) { + int idx = nodes_bsearch(*arr, irn); + + if (idx < 0) { + int i, n = ARR_LEN(*arr); + const ir_node **l; + + ARR_APP1(const ir_node *, *arr, irn); + + /* move it */ + idx = ~idx; + l = *arr; + for (i = n - 1; i >= idx; --i) + l[i + 1] = l[i]; + l[idx] = irn; + return 1; + } + return 0; +} + /** * Adds a node to an affinity chunk */ -static INLINE void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) { +static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) { int i; - if (bitset_is_set(c->nodes, get_irn_idx(node->irn))) + if (! nodes_insert(&c->n, node->irn)) return; c->weight_consistent = 0; node->chunk = c; - bitset_set(c->nodes, get_irn_idx(node->irn)); - - ARR_APP1(ir_node *, c->n, node->irn); for (i = node->n_neighs - 1; i >= 0; --i) { ir_node *neigh = node->int_neighs[i]; - bitset_set(c->interfere, get_irn_idx(neigh)); + nodes_insert(&c->interfere, neigh); } } @@ -316,7 +368,7 @@ static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn, void *old) { res->tmp_col = -1; res->int_neighs = NULL; res->int_aff_neigh = 0; - res->col = arch_register_get_index(arch_get_irn_register(env->aenv, irn)); + res->col = arch_register_get_index(arch_get_irn_register(irn)); res->init_col = res->col; INIT_LIST_HEAD(&res->list); @@ -326,7 +378,7 @@ static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn, void *old) { res->adm_colors = bitset_obstack_alloc(phase_obst(ph), env->n_regs); /* Exclude colors not assignable to the irn */ - req = arch_get_register_req(env->aenv, irn, -1); + req = arch_get_register_req_out(irn); if (arch_register_req_is(req, limited)) rbitset_copy_to_bitset(req->limited, res->adm_colors); else @@ -336,7 +388,7 @@ static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn, void *old) { bitset_andnot(res->adm_colors, env->ignore_regs); /* compute the constraint factor */ - res->constr_factor = (double) (1 + env->n_regs - bitset_popcnt(res->adm_colors)) / env->n_regs; + res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcnt(res->adm_colors)) / env->n_regs; /* set the number of interfering affinity neighbours to -1, they are calculated later */ res->int_aff_neigh = -1; @@ -344,7 +396,7 @@ static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn, void *old) { /* build list of interfering neighbours */ len = 0; be_ifg_foreach_neighbour(env->ifg, nodes_it, irn, neigh) { - if (! arch_irn_is(env->aenv, neigh, ignore)) { + if (!arch_irn_is_ignore(neigh)) { obstack_ptr_grow(phase_obst(ph), neigh); ++len; } @@ -358,32 +410,37 @@ static void *co_mst_irn_init(ir_phase *ph, const ir_node *irn, void *old) { /** * Check if affinity chunk @p chunk interferes with node @p irn. */ -static INLINE int aff_chunk_interferes(co_mst_env_t *env, const aff_chunk_t *chunk, const ir_node *irn) { - (void) env; - return bitset_is_set(chunk->interfere, get_irn_idx(irn)); +static inline int aff_chunk_interferes(const aff_chunk_t *chunk, const ir_node *irn) { + return node_contains(chunk->interfere, irn); } /** * Check if there are interference edges from c1 to c2. - * @param env The global co_mst environment * @param c1 A chunk * @param c2 Another chunk * @return 1 if there are interferences between nodes of c1 and c2, 0 otherwise. */ -static INLINE int aff_chunks_interfere(co_mst_env_t *env, const aff_chunk_t *c1, const aff_chunk_t *c2) { - (void) env; +static inline int aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2) { + int i; + if (c1 == c2) return 0; /* check if there is a node in c2 having an interfering neighbor in c1 */ - return bitset_intersect(c1->interfere, c2->nodes); + for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) { + const ir_node *irn = c2->n[i]; + + if (node_contains(c1->interfere, irn)) + return 1; + } + return 0; } /** * Returns the affinity chunk of @p irn or creates a new * one with @p irn as element if there is none assigned. */ -static INLINE aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn) { +static inline aff_chunk_t *get_aff_chunk(co_mst_env_t *env, const ir_node *irn) { co_mst_irn_t *node = get_co_mst_irn(env, irn); return node->chunk; } @@ -398,13 +455,13 @@ static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node aff_chunk_t *c2 = get_aff_chunk(env, tgt); #ifdef DEBUG_libfirm - DB((dbg, LEVEL_4, "Attempt to let c1 (id %d): ", c1 ? c1->id : -1)); + DB((dbg, LEVEL_4, "Attempt to let c1 (id %u): ", c1 ? c1->id : 0)); if (c1) { DBG_AFF_CHUNK(env, LEVEL_4, c1); } else { DB((dbg, LEVEL_4, "{%+F}", src)); } - DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %d): ", c2 ? c2->id : -1)); + DB((dbg, LEVEL_4, "\n\tabsorb c2 (id %u): ", c2 ? c2->id : 0)); if (c2) { DBG_AFF_CHUNK(env, LEVEL_4, c2); } else { @@ -432,24 +489,28 @@ static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node } } else { /* c2 already exists */ - if (! aff_chunk_interferes(env, c2, src)) { + if (! aff_chunk_interferes(c2, src)) { aff_chunk_add_node(c2, get_co_mst_irn(env, src)); goto absorbed; } } } else if (c2 == NULL) { /* c1 already exists */ - if (! aff_chunk_interferes(env, c1, tgt)) { + if (! aff_chunk_interferes(c1, tgt)) { aff_chunk_add_node(c1, get_co_mst_irn(env, tgt)); goto absorbed; } - } else if (c1 != c2 && ! aff_chunks_interfere(env, c1, c2)) { + } else if (c1 != c2 && ! aff_chunks_interfere(c1, c2)) { int idx, len; for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx) aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx])); - bitset_or(c1->interfere, c2->interfere); + for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) { + const ir_node *irn = c2->interfere[idx]; + nodes_insert(&c1->interfere, irn); + } + c1->weight_consistent = 0; delete_aff_chunk(env, c2); @@ -473,7 +534,7 @@ static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) { for (i = 0; i < env->n_regs; ++i) { c->color_affinity[i].col = i; - c->color_affinity[i].cost = 0.0; + c->color_affinity[i].cost = REAL(0.0); } for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { @@ -482,7 +543,7 @@ static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) { co_mst_irn_t *node = get_co_mst_irn(env, n); node->chunk = c; - if (node->constr_factor > 0.0) { + if (node->constr_factor > REAL(0.0)) { bitset_pos_t col; bitset_foreach (node->adm_colors, col) c->color_affinity[col].cost += node->constr_factor; @@ -491,20 +552,18 @@ static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) { if (an != NULL) { neighb_t *neigh; co_gs_foreach_neighb(an, neigh) { - const ir_node *m = neigh->irn; - const int m_idx = get_irn_idx(m); + const ir_node *m = neigh->irn; - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, m, ignore)) + if (arch_irn_is_ignore(m)) continue; - w += bitset_is_set(c->nodes, m_idx) ? neigh->costs : 0; + w += node_contains(c->n, m) ? neigh->costs : 0; } } } for (i = 0; i < env->n_regs; ++i) - c->color_affinity[i].cost *= (1.0 / ARR_LEN(c->n)); + c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n)); c->weight = w; // c->weight = bitset_popcnt(c->nodes); @@ -525,8 +584,7 @@ static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t const ir_node *n = neigh->irn; int i; - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, n, ignore)) + if (arch_irn_is_ignore(n)) continue; /* check if the affinity neighbour interfere */ @@ -561,8 +619,7 @@ static void build_affinity_chunks(co_mst_env_t *env) { co_mst_irn_t *n1; affinity_node_t *an; - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, n, ignore)) + if (arch_irn_is_ignore(n)) continue; n1 = get_co_mst_irn(env, n); @@ -585,7 +642,7 @@ static void build_affinity_chunks(co_mst_env_t *env) { aff_edge_t edge; /* skip ignore nodes */ - if (arch_irn_is(env->aenv, m, ignore)) + if (arch_irn_is_ignore(m)) continue; edge.src = n; @@ -620,12 +677,13 @@ static void build_affinity_chunks(co_mst_env_t *env) { foreach_pset(env->chunkset, curr_chunk) { aff_chunk_assure_weight(env, curr_chunk); - DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id)); + DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id)); DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk); DBG((dbg, LEVEL_1, "\n")); pqueue_put(env->chunks, curr_chunk, curr_chunk->weight); } + foreach_phase_irn(&env->ph, n) { co_mst_irn_t *mirn = get_co_mst_irn(env, n); @@ -636,7 +694,7 @@ static void build_affinity_chunks(co_mst_env_t *env) { aff_chunk_assure_weight(env, curr_chunk); - DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id)); + DBG((dbg, LEVEL_1, "entry #%u", curr_chunk->id)); DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk); DBG((dbg, LEVEL_1, "\n")); @@ -649,7 +707,7 @@ static void build_affinity_chunks(co_mst_env_t *env) { static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk) { - pqueue *grow = new_pqueue(); + pqueue_t *grow = new_pqueue(); const ir_node *max_node = NULL; int max_weight = 0; int i; @@ -660,7 +718,7 @@ static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chu int w = 0; neighb_t *neigh; - if (arch_irn_is(env->aenv, irn, ignore)) + if (arch_irn_is_ignore(irn)) continue; if (an) { @@ -684,11 +742,11 @@ static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chu bitset_remv_irn(visited, max_node); i = 0; while (!pqueue_empty(grow)) { - ir_node *irn = pqueue_get(grow); + ir_node *irn = pqueue_pop_front(grow); affinity_node_t *an = get_affinity_info(env->co, irn); - neighb_t *neigh; + neighb_t *neigh; - if (arch_irn_is(env->aenv, irn, ignore)) + if (arch_irn_is_ignore(irn)) continue; assert(i <= ARR_LEN(chunk->n)); @@ -720,7 +778,7 @@ static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *v { waitq *nodes = new_waitq(); - DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%d) from %+F, color %d:", chunk->id, node->irn, col)); + DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col)); /* init queue and chunk */ waitq_put(nodes, node); @@ -741,17 +799,16 @@ static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *v int m_idx = get_irn_idx(m); co_mst_irn_t *n2; - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, m, ignore)) + if (arch_irn_is_ignore(m)) continue; n2 = get_co_mst_irn(env, m); - if (! bitset_is_set(visited, m_idx) && - decider(n2, col) && - ! n2->fixed && - ! aff_chunk_interferes(env, chunk, m) && - bitset_is_set(orig_chunk->nodes, m_idx)) + if (! bitset_is_set(visited, m_idx) && + decider(n2, col) && + ! n2->fixed && + ! aff_chunk_interferes(chunk, m) && + node_contains(orig_chunk->n, m)) { /* following conditions are met: @@ -800,19 +857,19 @@ static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, w if (get_mst_irn_col(node) == col) { decider = decider_has_color; check_for_best = 1; - DBG((dbg, LEVEL_4, "\tcolor %d wanted", col)); + DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col)); } else { decider = decider_hasnot_color; check_for_best = 0; - DBG((dbg, LEVEL_4, "\tcolor %d forbidden", col)); + DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col)); } /* create a new chunk starting at current node */ tmp_chunk = new_aff_chunk(env); waitq_put(tmp, tmp_chunk); expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col); - assert(bitset_popcnt(tmp_chunk->nodes) > 0 && "No nodes added to chunk"); + assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk"); /* remember the local best */ aff_chunk_assure_weight(env, tmp_chunk); @@ -829,7 +886,7 @@ static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, w * Resets the temporary fixed color of all nodes within wait queue @p nodes. * ATTENTION: the queue is empty after calling this function! */ -static INLINE void reject_coloring(struct list_head *nodes) { +static inline void reject_coloring(struct list_head *nodes) { co_mst_irn_t *n, *temp; DB((dbg, LEVEL_4, "\treject coloring for")); list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) { @@ -841,7 +898,7 @@ static INLINE void reject_coloring(struct list_head *nodes) { DB((dbg, LEVEL_4, "\n")); } -static INLINE void materialize_coloring(struct list_head *nodes) { +static inline void materialize_coloring(struct list_head *nodes) { co_mst_irn_t *n, *temp; list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) { assert(n->tmp_col >= 0); @@ -851,7 +908,7 @@ static INLINE void materialize_coloring(struct list_head *nodes) { } } -static INLINE void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed) +static inline void set_temp_color(co_mst_irn_t *node, int col, struct list_head *changed) { assert(col >= 0); assert(!node->fixed); @@ -863,7 +920,7 @@ static INLINE void set_temp_color(co_mst_irn_t *node, int col, struct list_head node->tmp_col = col; } -static INLINE int is_loose(co_mst_irn_t *node) +static inline int is_loose(co_mst_irn_t *node) { return !node->fixed && node->tmp_col < 0; } @@ -871,16 +928,17 @@ static INLINE int is_loose(co_mst_irn_t *node) /** * Determines the costs for each color if it would be assigned to node @p node. */ -static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) { - int *neigh_cols = alloca(env->n_regs * sizeof(*neigh_cols)); - int n_loose = 0; - double coeff; - int i; +static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) +{ + int *neigh_cols = ALLOCAN(int, env->n_regs); + int n_loose = 0; + real_t coeff; + int i; for (i = 0; i < env->n_regs; ++i) { neigh_cols[i] = 0; costs[i].col = i; - costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : 0.0; + costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0); } for (i = 0; i < node->n_neighs; ++i) { @@ -889,16 +947,14 @@ static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cos if (is_loose(n)) { ++n_loose; ++neigh_cols[col]; - } - - else - costs[col].cost = 0.0; + } else + costs[col].cost = REAL(0.0); } if (n_loose > 0) { - coeff = 1.0 / n_loose; + coeff = REAL(1.0) / n_loose; for (i = 0; i < env->n_regs; ++i) - costs[i].cost *= 1.0 - coeff * neigh_cols[i]; + costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i]; } } @@ -922,13 +978,13 @@ static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int /* The node has the color it should not have _and_ has not been visited yet. */ if (is_loose(node)) { - col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0])); + col_cost_t *costs = ALLOCAN(col_cost_t, env->n_regs); /* Get the costs for giving the node a specific color. */ determine_color_costs(env, node, costs); /* Since the node must not have the not_col, set the costs for that color to "infinity" */ - costs[exclude_col].cost = 0.0; + costs[exclude_col].cost = REAL(0.0); /* sort the colors according costs, cheapest first. */ qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt); @@ -953,21 +1009,25 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost if (depth > *max_depth) *max_depth = depth; - if (depth >= recolor_limit) - return 0; - DBG((dbg, LEVEL_4, "\tRecoloring %+F with color-costs", node->irn)); DBG_COL_COST(env, LEVEL_4, costs); DB((dbg, LEVEL_4, "\n")); + if (depth >= recolor_limit) { + DBG((dbg, LEVEL_4, "\tHit recolor limit\n")); + return 0; + } + for (i = 0; i < env->n_regs; ++i) { int tgt_col = costs[i].col; int neigh_ok = 1; int j; /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */ - if (costs[i].cost == 0.0) + if (costs[i].cost == REAL(0.0)) { + DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n")); return 0; + } /* Set the new color of the node and mark the node as temporarily fixed. */ assert(node->tmp_col < 0 && "Node must not have been temporary fixed."); @@ -982,8 +1042,7 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost neigh = node->int_neighs[j]; - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, neigh, ignore)) + if (arch_irn_is_ignore(neigh)) continue; nn = get_co_mst_irn(env, neigh); @@ -1020,6 +1079,7 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost } } + DBG((dbg, LEVEL_4, "\tAll colors failed\n")); return 0; } @@ -1085,29 +1145,27 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { int n_int_chunks = 0; waitq *tmp_chunks = new_waitq(); waitq *best_starts = NULL; - col_cost_t *order = alloca(env->n_regs * sizeof(order[0])); + col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs); bitset_t *visited; - int idx, len, i; + int idx, len, i, nidx, pos; struct list_head changed; - bitset_pos_t pos; - DB((dbg, LEVEL_2, "fragmentizing chunk #%d", c->id)); + DB((dbg, LEVEL_2, "fragmentizing chunk #%u", c->id)); DBG_AFF_CHUNK(env, LEVEL_2, c); DB((dbg, LEVEL_2, "\n")); - stat_ev_ctx_push_fmt("heur4_color_chunk", "%d", c->id); + stat_ev_ctx_push_fmt("heur4_color_chunk", "%u", c->id); ++env->chunk_visited; /* compute color preference */ - memset(order, 0, env->n_regs * sizeof(order[0])); - - bitset_foreach (c->interfere, pos) { - ir_node *n = get_idx_irn(env->co->irg, pos); + for (pos = 0, len = ARR_LEN(c->interfere); pos < len; ++pos) { + const ir_node *n = c->interfere[pos]; co_mst_irn_t *node = get_co_mst_irn(env, n); aff_chunk_t *chunk = node->chunk; - if (is_loose(node) && chunk && chunk->visited < env->chunk_visited && !chunk->deleted) { + if (is_loose(node) && chunk && chunk->visited < env->chunk_visited) { + assert(!chunk->deleted); chunk->visited = env->chunk_visited; ++n_int_chunks; @@ -1118,9 +1176,9 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { } for (i = 0; i < env->n_regs; ++i) { - double dislike = n_int_chunks > 0 ? 1.0 - order[i].cost / n_int_chunks : 0.0; + real_t dislike = n_int_chunks > 0 ? REAL(1.0) - order[i].cost / n_int_chunks : REAL(0.0); order[i].col = i; - order[i].cost = (1.0 - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike; + order[i].cost = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike; } qsort(order, env->n_regs, sizeof(order[0]), cmp_col_cost_gt); @@ -1198,7 +1256,7 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { if (local_best) { aff_chunk_assure_weight(env, local_best); - DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, col)); + DB((dbg, LEVEL_3, "\t\tlocal best chunk (id %u) for color %d: ", local_best->id, col)); DBG_AFF_CHUNK(env, LEVEL_3, local_best); if (! best_chunk || best_chunk->weight < local_best->weight) { @@ -1207,7 +1265,7 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { if (best_starts) del_waitq(best_starts); best_starts = good_starts; - DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %d), color %d\n", best_chunk->id, best_color)); + DB((dbg, LEVEL_3, "\n\t\t... setting global best chunk (id %u), color %d\n", best_chunk->id, best_color)); } else { DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n")); del_waitq(good_starts); @@ -1239,7 +1297,7 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { return; } - DB((dbg, LEVEL_2, "\tbest chunk #%d ", best_chunk->id)); + DB((dbg, LEVEL_2, "\tbest chunk #%u ", best_chunk->id)); DBG_AFF_CHUNK(env, LEVEL_2, best_chunk); DB((dbg, LEVEL_2, "using color %d\n", best_color)); @@ -1249,7 +1307,7 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { int res; /* bring the node to the color. */ - DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%d\n", best_color, node->irn, best_chunk->id)); + DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id)); INIT_LIST_HEAD(&changed); stat_ev_tim_push(); res = change_node_color(env, node, best_color, &changed); @@ -1262,18 +1320,24 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { } /* remove the nodes in best chunk from original chunk */ - bitset_andnot(c->nodes, best_chunk->nodes); - for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { - const ir_node *irn = c->n[idx]; + len = ARR_LEN(best_chunk->n); + for (idx = 0; idx < len; ++idx) { + const ir_node *irn = best_chunk->n[idx]; + int pos = nodes_bsearch(c->n, irn); - if (bitset_is_set(best_chunk->nodes, get_irn_idx(irn))) { - int last = ARR_LEN(c->n) - 1; + if (pos > 0) + c->n[pos] = NULL; + } + len = ARR_LEN(c->n); + for (idx = nidx = 0; idx < len; ++idx) { + const ir_node *irn = c->n[idx]; - c->n[idx] = c->n[last]; - ARR_SHRINKLEN(c->n, last); - len--; + if (irn != NULL) { + c->n[nidx++] = irn; } } + ARR_SHRINKLEN(c->n, nidx); + /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */ for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { @@ -1284,7 +1348,9 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { /* fragment the remaining chunk */ visited = bitset_irg_malloc(env->co->irg); - bitset_or(visited, best_chunk->nodes); + for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) + bitset_set(visited, get_irn_idx(best_chunk->n[idx])); + for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { const ir_node *irn = c->n[idx]; if (! bitset_is_set(visited, get_irn_idx(irn))) { @@ -1315,13 +1381,15 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { /** * Main driver for mst safe coalescing algorithm. */ -int co_solve_heuristic_mst(copy_opt_t *co) { +static int co_solve_heuristic_mst(copy_opt_t *co) { unsigned n_regs = co->cls->n_regs; bitset_t *ignore_regs = bitset_alloca(n_regs); unsigned i, j, k; ir_node *irn; co_mst_env_t mst_env; + last_chunk_id = 0; + stat_ev_tim_push(); /* init phase */ @@ -1336,7 +1404,6 @@ int co_solve_heuristic_mst(copy_opt_t *co) { mst_env.co = co; mst_env.ignore_regs = ignore_regs; mst_env.ifg = co->cenv->ifg; - mst_env.aenv = co->aenv; mst_env.chunkset = pset_new_ptr(512); mst_env.chunk_visited = 0; mst_env.single_cols = phase_alloc(&mst_env.ph, sizeof(*mst_env.single_cols) * n_regs); @@ -1347,11 +1414,11 @@ int co_solve_heuristic_mst(copy_opt_t *co) { mst_env.single_cols[i] = vec; for (j = 0; j < n_regs; ++j) { vec[j].col = j; - vec[j].cost = 0.0; + vec[j].cost = REAL(0.0); } vec[i].col = 0; vec[0].col = i; - vec[0].cost = 1.0; + vec[0].cost = REAL(1.0); } DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name)); @@ -1363,21 +1430,22 @@ int co_solve_heuristic_mst(copy_opt_t *co) { /* color chunks as long as there are some */ while (! pqueue_empty(mst_env.chunks)) { - aff_chunk_t *chunk = pqueue_get(mst_env.chunks); + aff_chunk_t *chunk = pqueue_pop_front(mst_env.chunks); color_aff_chunk(&mst_env, chunk); - DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%d) done\n", chunk->id)); + DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id)); delete_aff_chunk(&mst_env, chunk); } /* apply coloring */ foreach_phase_irn(&mst_env.ph, irn) { - co_mst_irn_t *mirn = get_co_mst_irn(&mst_env, irn); + co_mst_irn_t *mirn; const arch_register_t *reg; - if (arch_irn_is(mst_env.aenv, irn, ignore)) + if (arch_irn_is_ignore(irn)) continue; + mirn = get_co_mst_irn(&mst_env, irn); // assert(mirn->fixed && "Node should have fixed color"); /* skip nodes where color hasn't changed */ @@ -1385,7 +1453,7 @@ int co_solve_heuristic_mst(copy_opt_t *co) { continue; reg = arch_register_for_index(co->cls, mirn->col); - arch_set_irn_register(co->aenv, irn, reg); + arch_set_irn_register(irn, reg); DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col)); } @@ -1413,9 +1481,14 @@ void be_init_copyheur4(void) { lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co"); lc_opt_entry_t *heur4_grp = lc_opt_get_grp(co_grp, "heur4"); + static co_algo_info copyheur = { + co_solve_heuristic_mst, 0 + }; + lc_opt_add_table(heur4_grp, options); + be_register_copyopt("heur4", ©heur); + FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4"); } - BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4);