/* * This file is part of libFirm. * Copyright (C) 2012 University of Karlsruhe. */ /** * @file * @brief Simple copy minimization heuristics. * @author Christian Wuerdig * @date 27.04.2007 * * This is the C implementation of the mst algorithm * originally written in Java by Sebastian Hack. * (also known as "heur3" :) * Performs simple copy minimization. */ #define DISABLE_STATEV #include #include "array.h" #include "debug.h" #include "irnode_t.h" #include "bitset.h" #include "raw_bitset.h" #include "irnodemap.h" #include "pqueue.h" #include "xmalloc.h" #include "pdeq.h" #include "irprintf.h" #include "util.h" #include "irtools.h" #include "list.h" #include "statev_t.h" #include "bearch.h" #include "beifg.h" #include "be_t.h" #include "becopyopt_t.h" #include "bemodule.h" #ifdef DEBUG_libfirm #define DBG_AFF_CHUNK(env, level, chunk) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_aff_chunk((env), (chunk)); } while (0) #define DBG_COL_COST(env, level, cost) do { if (firm_dbg_get_mask(dbg) & (level)) dbg_col_cost((env), (cost)); } while (0) static firm_dbg_module_t *dbg = NULL; #else #define DBG_AFF_CHUNK(env, level, chunk) ((void)(env), (void)(chunk), (void)0) #define DBG_COL_COST(env, level, cost) ((void)(env), (void)(cost), (void)0) #endif typedef float real_t; #define REAL(C) (C ## f) static unsigned last_chunk_id; static int recolor_limit = 7; static double dislike_influence = REAL(0.1); typedef struct col_cost_t { unsigned col; real_t cost; } col_cost_t; /** * An affinity chunk. */ typedef struct aff_chunk_t { const ir_node **n; /**< An ARR_F containing all nodes of the chunk. */ const ir_node **interfere; /**< An ARR_F containing all inference. */ int weight; /**< Weight of this chunk */ unsigned id; /**< An id of this chunk. */ unsigned visited; list_head list; bool weight_consistent : 1; /**< Set if the weight is consistent. */ #ifndef NDEBUG bool deleted : 1; /**< For debugging: Set if the was deleted. */ #endif col_cost_t color_affinity[]; } aff_chunk_t; /** * An affinity edge. */ typedef struct aff_edge_t { const ir_node *src; /**< Source node. */ const ir_node *tgt; /**< Target node. */ int weight; /**< The weight of this edge. */ } aff_edge_t; /* main coalescing environment */ typedef struct co_mst_env_t { bitset_t const *allocatable_regs; /**< set containing all global ignore registers */ ir_nodemap map; /**< phase object holding data for nodes */ struct obstack obst; pqueue_t *chunks; /**< priority queue for chunks */ list_head chunklist; /**< list holding all chunks */ be_ifg_t *ifg; /**< the interference graph */ copy_opt_t *co; /**< the copy opt object */ col_cost_t **single_cols; unsigned n_regs; /**< number of regs in class */ unsigned chunk_visited; } co_mst_env_t; /* stores coalescing related information for a node */ typedef struct co_mst_irn_t { const ir_node *irn; /**< the irn this information belongs to */ aff_chunk_t *chunk; /**< the chunk this irn belongs to */ bitset_t *adm_colors; /**< set of admissible colors for this irn */ ir_node **int_neighs; /**< array of all interfering neighbours (cached for speed reasons) */ unsigned n_neighs; /**< length of the interfering neighbours array. */ int int_aff_neigh; /**< number of interfering affinity neighbours */ unsigned col; /**< color currently assigned */ unsigned init_col; /**< the initial color */ int tmp_col; /**< a temporary assigned color */ struct list_head list; /**< Queue for coloring undo. */ real_t constr_factor; bool fixed:1; /**< the color is fixed */ } co_mst_irn_t; /** * In case there is no phase information for irn, initialize it. */ static co_mst_irn_t *co_mst_irn_init(co_mst_env_t *env, const ir_node *irn) { co_mst_irn_t *const res = OALLOC(&env->obst, co_mst_irn_t); res->irn = irn; res->chunk = NULL; res->fixed = 0; res->tmp_col = -1; res->int_neighs = NULL; /* set the number of interfering affinity neighbours to -1, they are calculated later */ res->int_aff_neigh = -1; res->col = arch_get_irn_register(irn)->index; res->init_col = res->col; INIT_LIST_HEAD(&res->list); DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn)); /* set admissible registers */ unsigned const n_regs = env->n_regs; bitset_t *const adm = bitset_obstack_alloc(&env->obst, n_regs); res->adm_colors = adm; bitset_copy(adm, env->allocatable_regs); /* Exclude colors not assignable to the irn */ arch_register_req_t const *const req = arch_get_irn_register_req(irn); if (req->limited != NULL) rbitset_and(adm->data, req->limited, n_regs); /* compute the constraint factor */ res->constr_factor = (real_t)(1 + n_regs - bitset_popcount(adm)) / n_regs; /* build list of interfering neighbours */ unsigned len = 0; neighbours_iter_t nodes_it; be_ifg_foreach_neighbour(env->ifg, &nodes_it, irn, neigh) { if (!arch_irn_is_ignore(neigh)) { obstack_ptr_grow(&env->obst, neigh); ++len; } } res->int_neighs = (ir_node**)obstack_finish(&env->obst); res->n_neighs = len; return res; } static co_mst_irn_t *get_co_mst_irn(co_mst_env_t *env, const ir_node *node) { co_mst_irn_t *res = ir_nodemap_get(co_mst_irn_t, &env->map, node); if (res == NULL) { res = co_mst_irn_init(env, node); ir_nodemap_insert(&env->map, node, res); } return res; } typedef bool decide_func_t(const co_mst_irn_t *node, unsigned col); #ifdef DEBUG_libfirm /** * Write a chunk to stderr for debugging. */ static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) { (void)env; if (c->weight_consistent) ir_fprintf(stderr, " $%d ", c->weight); ir_fprintf(stderr, "{"); for (size_t 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, "}"); } /** * Dump all admissible colors to stderr. */ static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) { (void)env; if (bitset_popcount(node->adm_colors) < 1) { fprintf(stderr, "no admissible colors?!?"); } else { bitset_foreach(node->adm_colors, idx) { ir_fprintf(stderr, " %zu", idx); } } } /** * Dump color-cost pairs to stderr. */ static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) { for (unsigned i = 0, n = env->n_regs; i < n; ++i) fprintf(stderr, " (%u, %.4f)", cost[i].col, cost[i].cost); } #endif /* DEBUG_libfirm */ static inline unsigned get_mst_irn_col(const co_mst_irn_t *node) { return node->tmp_col >= 0 ? (unsigned)node->tmp_col : node->col; } /** * @return true if node @p node has color @p col, false otherwise. */ static bool decider_has_color(const co_mst_irn_t *node, unsigned col) { return get_mst_irn_col(node) == col; } /** * @return true if node @p node has not color @p col, false otherwise. */ static bool decider_hasnot_color(const co_mst_irn_t *node, unsigned col) { return get_mst_irn_col(node) != col; } /** * Always returns true. */ static bool decider_always_yes(const co_mst_irn_t *node, unsigned col) { (void)node; (void)col; return true; } /** compares two affinity edges by its weight */ static int cmp_aff_edge(const void *a, const void *b) { const aff_edge_t *e1 = (const aff_edge_t*)a; const aff_edge_t *e2 = (const aff_edge_t*)b; if (e2->weight == e1->weight) { if (e2->src->node_idx == e1->src->node_idx) return QSORT_CMP(e2->tgt->node_idx, e1->tgt->node_idx); else return QSORT_CMP(e2->src->node_idx, e1->src->node_idx); } /* sort in descending order */ return QSORT_CMP(e2->weight, e1->weight); } static int cmp_col_cost_gt(const void *a, const void *b) { const col_cost_t *c1 = (const col_cost_t*)a; const col_cost_t *c2 = (const col_cost_t*)b; real_t diff = c2->cost - c1->cost; if (diff > 0) return 1; if (diff < 0) return -1; return QSORT_CMP(c1->col, c2->col); } /** * Creates a new affinity chunk */ 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 = false; #ifndef NDEBUG c->deleted = false; #endif c->id = ++last_chunk_id; c->visited = 0; list_add(&c->list, &env->chunklist); return c; } /** * Frees all memory allocated by an affinity chunk. */ static inline void delete_aff_chunk(aff_chunk_t *c) { list_del(&c->list); DEL_ARR_F(c->interfere); DEL_ARR_F(c->n); #ifndef NDEBUG c->deleted = true; #endif 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) { unsigned hi = ARR_LEN(arr); unsigned lo = 0; while (lo < hi) { unsigned md = lo + ((hi - lo) / 2); 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 bool 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 true if the node was inserted */ static bool nodes_insert(const ir_node ***arr, const ir_node *irn) { int idx = nodes_bsearch(*arr, irn); if (idx >= 0) return false; ARR_APP1(const ir_node *, *arr, irn); /* move it */ idx = ~idx; const ir_node **l = *arr; for (int i = ARR_LEN(*arr) - 2; i >= idx; --i) l[i + 1] = l[i]; l[idx] = irn; return true; } /** * Adds a node to an affinity chunk */ static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) { if (!nodes_insert(&c->n, node->irn)) return; c->weight_consistent = false; node->chunk = c; for (unsigned i = node->n_neighs; i-- > 0; ) { ir_node *neigh = node->int_neighs[i]; nodes_insert(&c->interfere, neigh); } } /** * Check if affinity chunk @p chunk interferes with node @p irn. */ static inline bool 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 c1 A chunk * @param c2 Another chunk * @return true if there are interferences between nodes of c1 and c2 */ static inline bool aff_chunks_interfere(const aff_chunk_t *c1, const aff_chunk_t *c2) { if (c1 == c2) return false; /* check if there is a node in c2 having an interfering neighbor in c1 */ for (size_t i = ARR_LEN(c2->n); i-- > 0; ) { const ir_node *irn = c2->n[i]; if (node_contains(c1->interfere, irn)) return true; } return false; } /** * 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) { co_mst_irn_t *node = get_co_mst_irn(env, irn); return node->chunk; } /** * Let chunk(src) absorb the nodes of chunk(tgt) (only possible when there * are no interference edges from chunk(src) to chunk(tgt)). */ static void aff_chunk_absorb(co_mst_env_t *const env, ir_node const *const src, ir_node const *const tgt) { aff_chunk_t *c1 = get_aff_chunk(env, src); aff_chunk_t *c2 = get_aff_chunk(env, tgt); #ifdef DEBUG_libfirm 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 %u): ", c2 ? c2->id : 0)); if (c2) { DBG_AFF_CHUNK(env, LEVEL_4, c2); } else { DB((dbg, LEVEL_4, "{%+F}", tgt)); } DB((dbg, LEVEL_4, "\n")); #endif if (c1 == NULL) { if (c2 == NULL) { /* no chunk exists */ co_mst_irn_t *mirn = get_co_mst_irn(env, src); int i; for (i = mirn->n_neighs; i-- > 0; ) { if (mirn->int_neighs[i] == tgt) break; } if (i < 0) { /* create one containing both nodes */ c1 = new_aff_chunk(env); aff_chunk_add_node(c1, get_co_mst_irn(env, src)); aff_chunk_add_node(c1, get_co_mst_irn(env, tgt)); goto absorbed; } } else { /* c2 already exists */ 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(c1, tgt)) { aff_chunk_add_node(c1, get_co_mst_irn(env, tgt)); goto absorbed; } } else if (c1 != c2 && !aff_chunks_interfere(c1, c2)) { for (size_t idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx) aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx])); for (size_t 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 = false; delete_aff_chunk(c2); absorbed: DB((dbg, LEVEL_4, " ... absorbed\n")); return; } DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n")); } /** * Assures that the weight of the given chunk is consistent. */ static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) { if (!c->weight_consistent) { for (unsigned i = 0, n = env->n_regs; i < n; ++i) { c->color_affinity[i].col = i; c->color_affinity[i].cost = REAL(0.0); } int w = 0; for (unsigned idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { const ir_node *n = c->n[idx]; const affinity_node_t *an = get_affinity_info(env->co, n); co_mst_irn_t *node = get_co_mst_irn(env, n); node->chunk = c; if (node->constr_factor > REAL(0.0)) { bitset_foreach (node->adm_colors, col) c->color_affinity[col].cost += node->constr_factor; } if (an != NULL) { co_gs_foreach_neighb(an, neigh) { const ir_node *m = neigh->irn; if (arch_irn_is_ignore(m)) continue; w += node_contains(c->n, m) ? neigh->costs : 0; } } } for (unsigned i = 0, n = env->n_regs; i < n; ++i) c->color_affinity[i].cost *= (REAL(1.0) / ARR_LEN(c->n)); c->weight = w; // c->weight = bitset_popcount(c->nodes); c->weight_consistent = true; } } /** * Count the number of interfering affinity neighbours */ static unsigned count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an) { const ir_node *irn = an->irn; const co_mst_irn_t *node = get_co_mst_irn(env, irn); unsigned res = 0; co_gs_foreach_neighb(an, neigh) { const ir_node *n = neigh->irn; if (arch_irn_is_ignore(n)) continue; /* check if the affinity neighbour interfere */ for (unsigned i = 0, n_neighs = node->n_neighs; i < n_neighs; ++i) { if (node->int_neighs[i] == n) { ++res; break; } } } return res; } /** * Build chunks of nodes connected by affinity edges. * We start at the heaviest affinity edge. * The chunks of the two edge-defining nodes will be * merged if there are no interference edges from one * chunk to the other. */ static void build_affinity_chunks(co_mst_env_t *env) { aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0); /* at first we create the affinity edge objects */ be_ifg_foreach_node(env->ifg, n) { co_mst_irn_t *n1 = get_co_mst_irn(env, n); affinity_node_t *an = get_affinity_info(env->co, n); unsigned n_idx = get_irn_idx(n); if (an != NULL) { if (n1->int_aff_neigh < 0) n1->int_aff_neigh = count_interfering_aff_neighs(env, an); /* build the affinity edges */ co_gs_foreach_neighb(an, neigh) { const ir_node *m = neigh->irn; unsigned m_idx = get_irn_idx(m); /* record the edge in only one direction */ if (n_idx < m_idx) { aff_edge_t edge; edge.src = n; edge.tgt = m; co_mst_irn_t *n2 = get_co_mst_irn(env, m); if (n2->int_aff_neigh < 0) { affinity_node_t *am = get_affinity_info(env->co, m); n2->int_aff_neigh = count_interfering_aff_neighs(env, am); } /* * these weights are pure hackery ;-). * It's not chriswue's fault but mine. */ edge.weight = neigh->costs; ARR_APP1(aff_edge_t, edges, edge); } } } } /* now: sort edges and build the affinity chunks */ QSORT_ARR(edges, cmp_aff_edge); for (size_t i = 0, len = ARR_LEN(edges); i < len; ++i) { DBG((dbg, LEVEL_1, "edge (%u,%u) %f\n", edges[i].src->node_idx, edges[i].tgt->node_idx, edges[i].weight)); aff_chunk_absorb(env, edges[i].src, edges[i].tgt); } /* now insert all chunks into a priority queue */ list_for_each_entry(aff_chunk_t, curr_chunk, &env->chunklist, list) { aff_chunk_assure_weight(env, curr_chunk); 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); } for (size_t pn = 0, n = ARR_LEN(env->map.data); pn < n; ++pn) { co_mst_irn_t *mirn = (co_mst_irn_t*)env->map.data[pn]; if (mirn == NULL) continue; if (mirn->chunk != NULL) continue; /* no chunk is allocated so far, do it now */ aff_chunk_t *curr_chunk = new_aff_chunk(env); aff_chunk_add_node(curr_chunk, mirn); aff_chunk_assure_weight(env, curr_chunk); 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); } DEL_ARR_F(edges); } /** * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node. */ static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *visited, aff_chunk_t *chunk, aff_chunk_t *orig_chunk, decide_func_t *decider, unsigned col) { deq_t nodes; deq_init(&nodes); DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%u) from %+F, color %d:", chunk->id, node->irn, col)); /* init queue and chunk */ deq_push_pointer_right(&nodes, node); bitset_set(visited, get_irn_idx(node->irn)); aff_chunk_add_node(chunk, node); DB((dbg, LEVEL_1, " %+F", node->irn)); /* as long as there are nodes in the queue */ while (!deq_empty(&nodes)) { co_mst_irn_t *n = deq_pop_pointer_left(co_mst_irn_t, &nodes); affinity_node_t *an = get_affinity_info(env->co, n->irn); /* check all affinity neighbors */ if (an != NULL) { co_gs_foreach_neighb(an, neigh) { const ir_node *m = neigh->irn; co_mst_irn_t *n2 = get_co_mst_irn(env, m); unsigned m_idx = get_irn_idx(m); 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: - neighbour is not visited - neighbour likes the color - neighbour has not yet a fixed color - the new chunk doesn't interfere with the neighbour - neighbour belongs or belonged once to the original chunk */ bitset_set(visited, m_idx); aff_chunk_add_node(chunk, n2); DB((dbg, LEVEL_1, " %+F", n2->irn)); /* enqueue for further search */ deq_push_pointer_right(&nodes, n2); } } } } DB((dbg, LEVEL_1, "\n")); deq_free(&nodes); } /** * Fragment the given chunk into chunks having given color and not having given * color. */ static aff_chunk_t *fragment_chunk(co_mst_env_t *env, unsigned col, aff_chunk_t *c, deq_t *tmp) { bitset_t *visited = bitset_malloc(get_irg_last_idx(env->co->irg)); aff_chunk_t *best = NULL; for (unsigned 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))) continue; co_mst_irn_t *node = get_co_mst_irn(env, irn); decide_func_t *decider; bool check_for_best; if (get_mst_irn_col(node) == col) { decider = decider_has_color; check_for_best = true; DBG((dbg, LEVEL_4, "\tcolor %d wanted\n", col)); } else { decider = decider_hasnot_color; check_for_best = false; DBG((dbg, LEVEL_4, "\tcolor %d forbidden\n", col)); } /* create a new chunk starting at current node */ aff_chunk_t *tmp_chunk = new_aff_chunk(env); deq_push_pointer_right(tmp, tmp_chunk); expand_chunk_from(env, node, visited, tmp_chunk, c, decider, col); assert(ARR_LEN(tmp_chunk->n) > 0 && "No nodes added to chunk"); /* remember the local best */ aff_chunk_assure_weight(env, tmp_chunk); if (check_for_best && (!best || best->weight < tmp_chunk->weight)) best = tmp_chunk; } assert(best && "No chunk found?"); free(visited); return best; } /** * 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) { DB((dbg, LEVEL_4, "\treject coloring for")); list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) { DB((dbg, LEVEL_4, " %+F", n->irn)); assert(n->tmp_col >= 0); n->tmp_col = -1; list_del_init(&n->list); } DB((dbg, LEVEL_4, "\n")); } static inline void materialize_coloring(struct list_head *nodes) { list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) { assert(n->tmp_col >= 0); n->col = n->tmp_col; n->tmp_col = -1; list_del_init(&n->list); } } static inline void set_temp_color(co_mst_irn_t *node, unsigned col, struct list_head *changed) { assert(!node->fixed); assert(node->tmp_col < 0); assert(node->list.next == &node->list && node->list.prev == &node->list); assert(bitset_is_set(node->adm_colors, col)); list_add_tail(&node->list, changed); node->tmp_col = col; } static inline bool is_loose(const co_mst_irn_t *node) { return !node->fixed && node->tmp_col < 0; } /** * 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) { const unsigned n_regs = env->n_regs; unsigned *neigh_cols = ALLOCAN(unsigned, n_regs); unsigned n_loose = 0; for (unsigned i = 0; i < n_regs; ++i) { neigh_cols[i] = 0; costs[i].col = i; costs[i].cost = bitset_is_set(node->adm_colors, i) ? node->constr_factor : REAL(0.0); } for (unsigned i = node->n_neighs; i-- != 0;) { co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]); unsigned col = get_mst_irn_col(n); assert(col < n_regs); if (is_loose(n)) { ++n_loose; ++neigh_cols[col]; } else costs[col].cost = REAL(0.0); } if (n_loose > 0) { real_t coeff = REAL(1.0) / n_loose; for (unsigned i = 0; i < n_regs; ++i) costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i]; } } /* need forward declaration due to recursive call */ static bool recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones, unsigned depth, unsigned *max_depth, unsigned *trip); /** * Tries to change node to a color but @p explude_col. * @return true if succeeded, false otherwise. */ static bool change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, unsigned exclude_col, struct list_head *changed, unsigned depth, unsigned *max_depth, unsigned *trip) { unsigned col = get_mst_irn_col(node); /* neighbours has already a different color -> good, temporary fix it */ if (col != exclude_col) { if (is_loose(node)) set_temp_color(node, col, changed); return true; } /* The node has the color it should not have _and_ has not been visited yet. */ if (is_loose(node)) { 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 = REAL(0.0); /* sort the colors according costs, cheapest first. */ QSORT(costs, env->n_regs, cmp_col_cost_gt); /* Try recoloring the node using the color list. */ return recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip); } return false; } /** * Tries to bring node @p node to cheapest color and color all interfering * neighbours with other colors. * ATTENTION: Expect @p costs already sorted by increasing costs. * @return true if coloring could be applied, false otherwise. */ static bool recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed, unsigned depth, unsigned *max_depth, unsigned *trip) { ++*trip; if (depth > *max_depth) *max_depth = depth; 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 >= (unsigned)recolor_limit) { DBG((dbg, LEVEL_4, "\tHit recolor limit\n")); return false; } struct list_head local_changed; for (unsigned i = 0, n = env->n_regs; i < n; ++i) { unsigned tgt_col = costs[i].col; /* If the costs for that color (and all successive) are infinite, bail out we won't make it anyway. */ if (costs[i].cost == REAL(0.0)) { DBG((dbg, LEVEL_4, "\tAll further colors forbidden\n")); return false; } /* 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."); INIT_LIST_HEAD(&local_changed); set_temp_color(node, tgt_col, &local_changed); DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col)); /* try to color all interfering neighbours with current color forbidden */ bool neigh_ok = true; for (unsigned j = node->n_neighs; j-- != 0;) { ir_node *neigh = node->int_neighs[j]; co_mst_irn_t *nn = get_co_mst_irn(env, neigh); DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_col: %d, col: %d)\n", neigh, j, nn->fixed, nn->tmp_col, nn->col)); /* Try to change the color of the neighbor and record all nodes which get changed in the tmp list. Add this list to the "changed" list for that color. If we did not succeed to change the color of the neighbor, we bail out and try the next color. */ if (get_mst_irn_col(nn) == tgt_col) { /* try to color neighbour with tgt_col forbidden */ neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip); if (!neigh_ok) break; } } /* We managed to assign the target color to all neighbors, so from the perspective of the current node, every thing was ok and we can return safely. */ if (neigh_ok) { /* append the local_changed ones to global ones */ list_splice(&local_changed, changed); return true; } else { /* coloring of neighbours failed, so we try next color */ reject_coloring(&local_changed); } } DBG((dbg, LEVEL_4, "\tAll colors failed\n")); return false; } /** * Tries to bring node @p node and all its neighbours to color @p tgt_col. * @return true if color @p col could be applied, false otherwise */ static bool change_node_color(co_mst_env_t *env, co_mst_irn_t *node, unsigned tgt_col, struct list_head *changed) { /* if node already has the target color -> good, temporary fix it */ unsigned col = get_mst_irn_col(node); if (col == tgt_col) { DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col)); if (is_loose(node)) set_temp_color(node, tgt_col, changed); return true; } /* Node has not yet a fixed color and target color is admissible -> try to recolor node and its affinity neighbours */ if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) { col_cost_t *costs = env->single_cols[tgt_col]; unsigned max_depth = 0; unsigned trip = 0; DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn)); bool res = recolor_nodes(env, node, costs, changed, 0, &max_depth, &trip); DBG((dbg, LEVEL_4, "\t\tCNC: <<=== Recoloring of %+F %s\n", node->irn, res ? "succeeded" : "failed")); stat_ev_int("heur4_recolor_depth_max", max_depth); stat_ev_int("heur4_recolor_trip", trip); return res; } #ifdef DEBUG_libfirm if (firm_dbg_get_mask(dbg) & LEVEL_4) { if (is_loose(node)) { DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn)); dbg_admissible_colors(env, node); DB((dbg, LEVEL_4, ")\n")); } else { DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col)); } } #endif return false; } /** * Tries to color an affinity chunk (or at least a part of it). * Inserts uncolored parts of the chunk as a new chunk into the priority queue. */ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { 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", "%u", c->id); ++env->chunk_visited; /* compute color preference */ col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs); unsigned n_int_chunks = 0; for (size_t 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) { assert(!chunk->deleted); chunk->visited = env->chunk_visited; ++n_int_chunks; aff_chunk_assure_weight(env, chunk); for (unsigned i = 0, n = env->n_regs; i < n; ++i) order[i].cost += chunk->color_affinity[i].cost; } } for (unsigned i = 0, n = env->n_regs; i < n; ++i) { 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 = (REAL(1.0) - dislike_influence) * c->color_affinity[i].cost + dislike_influence * dislike; } QSORT(order, env->n_regs, cmp_col_cost_gt); DBG_COL_COST(env, LEVEL_2, order); DB((dbg, LEVEL_2, "\n")); /* check which color is the "best" for the given chunk. * If we found a color which was ok for all nodes, we take it * and do not look further. (see n_succeeded usage below.) * If we have many colors which fit all nodes it is hard to decide * which one to take anyway. * TODO Sebastian: Perhaps we should at all nodes and figure out * a suitable color using costs as done above (determine_color_costs). */ deq_t tmp_chunks; deq_init(&tmp_chunks); unsigned n_nodes = ARR_LEN(c->n); aff_chunk_t *best_chunk = NULL; int best_color = -1; for (unsigned i = 0, n = env->n_regs; i < n; ++i) { /* skip ignore colors */ unsigned col = order[i].col; if (!bitset_is_set(env->allocatable_regs, col)) continue; DB((dbg, LEVEL_2, "\ttrying color %d\n", col)); /* try to bring all nodes of given chunk to the current color. */ unsigned n_succeeded = 0; for (size_t idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { const ir_node *irn = c->n[idx]; co_mst_irn_t *node = get_co_mst_irn(env, irn); assert(!node->fixed && "Node must not have a fixed color."); DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, col)); /* The order of the colored nodes is important, so we record the * successfully colored ones in the order they appeared. */ struct list_head changed; INIT_LIST_HEAD(&changed); stat_ev_tim_push(); bool good = change_node_color(env, node, col, &changed); stat_ev_tim_pop("heur4_recolor"); if (good) { materialize_coloring(&changed); node->fixed = 1; } else { reject_coloring(&changed); } n_succeeded += good; DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, good ? "succeeded" : "failed")); } /* unfix all nodes */ for (size_t idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { co_mst_irn_t *node = get_co_mst_irn(env, c->n[idx]); node->fixed = 0; } /* try next color when failed */ if (n_succeeded == 0) continue; /* fragment the chunk according to the coloring */ aff_chunk_t *local_best = fragment_chunk(env, col, c, &tmp_chunks); /* search the best of the good list and make it the new best if it is better than the current */ if (local_best) { aff_chunk_assure_weight(env, local_best); 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) { best_chunk = local_best; best_color = col; 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")); } } /* if all nodes were recolored, bail out */ if (n_succeeded == n_nodes) break; } stat_ev_int("heur4_colors_tried", i); /* free all intermediate created chunks except best one */ while (!deq_empty(&tmp_chunks)) { aff_chunk_t *tmp = deq_pop_pointer_left(aff_chunk_t, &tmp_chunks); if (tmp != best_chunk) delete_aff_chunk(tmp); } deq_free(&tmp_chunks); /* return if coloring failed */ if (!best_chunk) return; 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)); for (size_t idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) { const ir_node *irn = best_chunk->n[idx]; co_mst_irn_t *node = get_co_mst_irn(env, irn); /* bring the node to the color. */ DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%u\n", best_color, node->irn, best_chunk->id)); struct list_head changed; INIT_LIST_HEAD(&changed); stat_ev_tim_push(); bool res = change_node_color(env, node, best_color, &changed); stat_ev_tim_pop("heur4_recolor"); if (res) { materialize_coloring(&changed); node->fixed = 1; } assert(list_empty(&changed)); } /* remove the nodes in best chunk from original chunk */ for (size_t idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) { const ir_node *irn = best_chunk->n[idx]; int pos = nodes_bsearch(c->n, irn); if (pos > 0) c->n[pos] = NULL; } size_t nidx = 0; for (size_t idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { const ir_node *irn = c->n[idx]; 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 (size_t idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { const ir_node *n = c->n[idx]; co_mst_irn_t *nn = get_co_mst_irn(env, n); nn->chunk = c; } /* fragment the remaining chunk */ bitset_t *visited = bitset_malloc(get_irg_last_idx(env->co->irg)); for (size_t idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) bitset_set(visited, get_irn_idx(best_chunk->n[idx])); for (size_t 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))) { aff_chunk_t *new_chunk = new_aff_chunk(env); co_mst_irn_t *node = get_co_mst_irn(env, irn); expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0); aff_chunk_assure_weight(env, new_chunk); pqueue_put(env->chunks, new_chunk, new_chunk->weight); } } for (size_t idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) { const ir_node *n = best_chunk->n[idx]; co_mst_irn_t *nn = get_co_mst_irn(env, n); nn->chunk = NULL; } /* clear obsolete chunks and free some memory */ delete_aff_chunk(best_chunk); free(visited); stat_ev_ctx_pop("heur4_color_chunk"); } /** * Main driver for mst safe coalescing algorithm. */ static int co_solve_heuristic_mst(copy_opt_t *co) { last_chunk_id = 0; stat_ev_tim_push(); /* init phase */ co_mst_env_t mst_env; ir_nodemap_init(&mst_env.map, co->irg); obstack_init(&mst_env.obst); unsigned const n_regs = co->cls->n_regs; mst_env.n_regs = n_regs; mst_env.chunks = new_pqueue(); mst_env.co = co; mst_env.allocatable_regs = co->cenv->allocatable_regs; mst_env.ifg = co->cenv->ifg; INIT_LIST_HEAD(&mst_env.chunklist); mst_env.chunk_visited = 0; mst_env.single_cols = OALLOCN(&mst_env.obst, col_cost_t*, n_regs); for (unsigned i = 0; i < n_regs; ++i) { col_cost_t *vec = OALLOCN(&mst_env.obst, col_cost_t, n_regs); mst_env.single_cols[i] = vec; for (unsigned j = 0; j < n_regs; ++j) { vec[j].col = j; vec[j].cost = REAL(0.0); } vec[i].col = 0; vec[0].col = i; vec[0].cost = REAL(1.0); } DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name)); /* build affinity chunks */ stat_ev_tim_push(); build_affinity_chunks(&mst_env); stat_ev_tim_pop("heur4_initial_chunk"); /* color chunks as long as there are some */ while (!pqueue_empty(mst_env.chunks)) { aff_chunk_t *chunk = (aff_chunk_t*)pqueue_pop_front(mst_env.chunks); color_aff_chunk(&mst_env, chunk); DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id)); delete_aff_chunk(chunk); } /* apply coloring */ for (size_t pn = 0; pn < ARR_LEN(mst_env.map.data); ++pn) { co_mst_irn_t *mirn = (co_mst_irn_t*)mst_env.map.data[pn]; if (mirn == NULL) continue; /* skip nodes where color hasn't changed */ if (mirn->init_col == mirn->col) continue; ir_node *const irn = get_idx_irn(co->irg, pn); arch_set_irn_register_idx(irn, mirn->col); DB((dbg, LEVEL_1, "%+F set color from %d to %d\n", irn, mirn->init_col, mirn->col)); } /* free allocated memory */ del_pqueue(mst_env.chunks); obstack_free(&mst_env.obst, NULL); ir_nodemap_destroy(&mst_env.map); stat_ev_tim_pop("heur4_total"); return 0; } static const lc_opt_table_entry_t options[] = { LC_OPT_ENT_INT("limit", "limit recoloring", &recolor_limit), LC_OPT_ENT_DBL("di", "dislike influence", &dislike_influence), LC_OPT_LAST }; BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyheur4) void be_init_copyheur4(void) { lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra"); lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal"); 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 }; lc_opt_add_table(heur4_grp, options); be_register_copyopt("heur4", ©heur); FIRM_DBG_REGISTER(dbg, "firm.be.co.heur4"); }