X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur4.c;h=ee97153b601fc6dab49a33b8f2e8b1efd3a93b98;hb=b0e09fe0aada2e5d9dc0299e969b3c93c1092d97;hp=21d54436c05283b834ccbe9d3a1225d71dd5b0e7;hpb=11108f4ca68e22a5f9cebee6cbd0a36b4efbe9dc;p=libfirm diff --git a/ir/be/becopyheur4.c b/ir/be/becopyheur4.c index 21d54436c..ee97153b6 100644 --- a/ir/be/becopyheur4.c +++ b/ir/be/becopyheur4.c @@ -41,11 +41,12 @@ #include "raw_bitset.h" #include "irphase_t.h" #include "pqueue.h" -#include "pset_new.h" #include "xmalloc.h" #include "pdeq.h" +#include "pset.h" #include "irprintf.h" #include "irbitset.h" +#include "error.h" #include "bearch.h" #include "beifg.h" @@ -73,9 +74,12 @@ typedef struct _col_cost_t { * An affinity chunk. */ typedef struct _aff_chunk_t { + 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. */ 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. */ } aff_chunk_t; @@ -93,10 +97,9 @@ typedef struct _co_mst_env_t { int n_regs; /**< number of regs in class */ int k; /**< number of non-ignore registers in class */ bitset_t *ignore_regs; /**< set containing all global ignore registers */ - int *map_regs; /**< map the available colors to the available registers */ ir_phase ph; /**< phase object holding data for nodes */ pqueue *chunks; /**< priority queue for chunks */ - pset_new_t chunkset; /**< set holding all 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 */ @@ -127,7 +130,7 @@ 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) { - int idx; + bitset_pos_t idx; if (c->weight_consistent) ir_fprintf(stderr, " $%d ", c->weight); ir_fprintf(stderr, "{"); @@ -142,7 +145,9 @@ static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) { * Dump all admissible colors to stderr. */ static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) { - int idx; + bitset_pos_t idx; + (void) env; + if (bitset_popcnt(node->adm_colors) < 1) fprintf(stderr, "no admissible colors?!?"); else { @@ -188,6 +193,8 @@ static int decider_hasnot_color(const co_mst_irn_t *node, int col) { * Always returns true. */ static int decider_always_yes(const co_mst_irn_t *node, int col) { + (void) node; + (void) col; return 1; } @@ -221,9 +228,11 @@ static INLINE aff_chunk_t *new_aff_chunk(co_mst_env_t *env) { aff_chunk_t *c = xmalloc(sizeof(*c)); c->weight = -1; c->weight_consistent = 0; + c->n = NEW_ARR_F(ir_node *, 0); c->nodes = bitset_irg_malloc(env->co->irg); + c->interfere = bitset_irg_malloc(env->co->irg); c->id = last_chunk_id++; - pset_new_insert(&env->chunkset, c); + pset_insert(env->chunkset, c, c->id); return c; } @@ -231,8 +240,11 @@ 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) { - pset_new_remove(&env->chunkset, c); + pset_remove(env->chunkset, c, c->id); bitset_free(c->nodes); + bitset_free(c->interfere); + DEL_ARR_F(c->n); + c->deleted = 1; free(c); } @@ -240,9 +252,21 @@ static INLINE void delete_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { * Adds a node to an affinity chunk */ 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))) + 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)); + } } /** @@ -289,8 +313,10 @@ static void *co_mst_irn_init(ir_phase *ph, ir_node *irn, void *old) { /* build list of interfering neighbours */ len = 0; be_ifg_foreach_neighbour(env->ifg, nodes_it, irn, neigh) { - obstack_ptr_grow(phase_obst(ph), neigh); - ++len; + if (! arch_irn_is(env->aenv, neigh, ignore)) { + obstack_ptr_grow(phase_obst(ph), neigh); + ++len; + } } res->int_neighs = obstack_finish(phase_obst(ph)); res->n_neighs = len; @@ -302,17 +328,8 @@ static void *co_mst_irn_init(ir_phase *ph, 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, ir_node *irn) { - const co_mst_irn_t *node = get_co_mst_irn(env, irn); - const ir_node *neigh; - int i; - - for (i = 0; i < node->n_neighs; ++i) { - neigh = node->int_neighs[i]; - if (! arch_irn_is(env->aenv, neigh, ignore) && bitset_is_set(chunk->nodes, get_irn_idx(neigh))) - return 1; - } - - return 0; + (void) env; + return bitset_is_set(chunk->interfere, get_irn_idx(irn)); } /** @@ -323,20 +340,17 @@ static INLINE int aff_chunk_interferes(co_mst_env_t *env, const aff_chunk_t *chu * @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) { - int idx; + bitset_t *tmp; if (c1 == c2) return 0; /* check if there is a node in c2 having an interfering neighbor in c1 */ - bitset_foreach(c2->nodes, idx) { - ir_node *n = get_idx_irn(env->co->irg, idx); - - if (aff_chunk_interferes(env, c1, n)) - return 1; - } + tmp = bitset_alloca(get_irg_last_idx(env->co->irg)); + tmp = bitset_copy(tmp, c1->interfere); + tmp = bitset_and(tmp, c2->nodes); - return 0; + return bitset_popcnt(tmp) > 0; } /** @@ -404,17 +418,22 @@ static int aff_chunk_absorb(co_mst_env_t *env, ir_node *src, ir_node *tgt) { goto absorbed; } } else if (c1 != c2 && ! aff_chunks_interfere(env, c1, c2)) { - int idx; + int idx, len; - bitset_or(c1->nodes, c2->nodes); - c1->weight_consistent = 0; - - bitset_foreach(c2->nodes, idx) { - ir_node *n = get_idx_irn(env->co->irg, idx); + for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx) { + ir_node *n = c2->n[idx]; co_mst_irn_t *mn = get_co_mst_irn(env, n); + mn->chunk = c1; + + if (! bitset_is_set(c1->nodes, get_irn_idx(n))) + ARR_APP1(ir_node *, c1->n, n); } + bitset_or(c1->nodes, c2->nodes); + bitset_or(c1->interfere, c2->interfere); + c1->weight_consistent = 0; + delete_aff_chunk(env, c2); goto absorbed; } @@ -432,17 +451,17 @@ absorbed: static void aff_chunk_assure_weight(const co_mst_env_t *env, aff_chunk_t *c) { if (! c->weight_consistent) { int w = 0; - int idx; + int idx, len; - bitset_foreach(c->nodes, idx) { - ir_node *n = get_idx_irn(env->co->irg, idx); + for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { + ir_node *n = c->n[idx]; const affinity_node_t *an = get_affinity_info(env->co, n); 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; + const int m_idx = get_irn_idx(m); /* skip ignore nodes */ if (arch_irn_is(env->aenv, m, ignore)) @@ -500,7 +519,6 @@ static void build_affinity_chunks(co_mst_env_t *env) { ir_node *n; int i, len; aff_chunk_t *curr_chunk; - pset_new_iterator_t iter; /* at first we create the affinity edge objects */ be_ifg_foreach_node(env->ifg, nodes_it, n) { @@ -520,6 +538,8 @@ static void build_affinity_chunks(co_mst_env_t *env) { 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) { ir_node *m = neigh->irn; int m_idx = get_irn_idx(m); @@ -558,7 +578,7 @@ static void build_affinity_chunks(co_mst_env_t *env) { } /* now insert all chunks into a priority queue */ - foreach_pset_new(&env->chunkset, curr_chunk, iter) { + foreach_pset(env->chunkset, curr_chunk) { aff_chunk_assure_weight(env, curr_chunk); DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id)); @@ -596,7 +616,7 @@ 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, "\nExpanding new chunk (id %d) from %+F:", chunk->id, node->irn)); + DBG((dbg, LEVEL_1, "\n\tExpanding new chunk (#%d) from %+F, color %d:", chunk->id, node->irn, col)); /* init queue and chunk */ waitq_put(nodes, node); @@ -657,29 +677,31 @@ static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *v */ static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp) { bitset_t *visited = bitset_irg_malloc(env->co->irg); - int idx; + int idx, len; aff_chunk_t *best = NULL; - bitset_foreach(c->nodes, idx) { + for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { ir_node *irn; co_mst_irn_t *node; aff_chunk_t *tmp_chunk; decide_func_t *decider; int check_for_best; - if (bitset_is_set(visited, idx)) + irn = c->n[idx]; + if (bitset_is_set(visited, get_irn_idx(irn))) continue; - irn = get_idx_irn(env->co->irg, idx); node = get_co_mst_irn(env, irn); if (get_mst_irn_col(node) == col) { decider = decider_has_color; check_for_best = 1; + DBG((dbg, LEVEL_4, "\tcolor %d wanted", col)); } else { decider = decider_hasnot_color; check_for_best = 0; + DBG((dbg, LEVEL_4, "\tcolor %d forbidden", col)); } /* create a new chunk starting at current node */ @@ -732,10 +754,13 @@ static INLINE void col_cost_init_single(co_mst_env_t *env, col_cost_t *cost, int * ATTENTION: the queue is empty after calling this function! */ static INLINE void reject_coloring(waitq *nodes) { + DB((dbg, LEVEL_4, "\treject coloring for")); while (! waitq_empty(nodes)) { co_mst_irn_t *n = waitq_get(nodes); + DB((dbg, LEVEL_4, " %+F", n->irn)); n->tmp_fixed = 0; } + DB((dbg, LEVEL_4, "\n")); } /** @@ -744,7 +769,8 @@ static INLINE void reject_coloring(waitq *nodes) { static void determine_color_costs(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs) { affinity_node_t *an = get_affinity_info(env->co, node->irn); neighb_t *aff_neigh; - int idx, i; + bitset_pos_t idx; + int i; col_cost_init(env, costs, 0.0); @@ -878,6 +904,7 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost assert(! node->tmp_fixed && "Node must not have been temporary fixed."); node->tmp_fixed = 1; node->tmp_col = tgt_col; + DBG((dbg, LEVEL_4, "\tTemporary setting %+F to color %d\n", node->irn, tgt_col)); assert(waitq_empty(local_changed) && "Node queue should be empty here."); waitq_put(local_changed, node); @@ -894,6 +921,8 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost continue; nn = get_co_mst_irn(env, neigh); + DB((dbg, LEVEL_4, "\tHandling neighbour %+F, at position %d (fixed: %d, tmp_fixed: %d, tmp_col: %d, col: %d)\n", + neigh, j, nn->fixed, nn->tmp_fixed, nn->tmp_col, nn->col)); /* Try to change the color of the neighbor and record all nodes which @@ -972,7 +1001,7 @@ static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, return res; } - DEBUG_ONLY( +#ifndef NDEBUG if (firm_dbg_get_mask(dbg) & LEVEL_4) { if (node->fixed || node->tmp_fixed) DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col)); @@ -982,7 +1011,7 @@ static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, DB((dbg, LEVEL_4, ")\n")); } } - ) +#endif return 0; } @@ -994,34 +1023,60 @@ static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { aff_chunk_t *best_chunk = NULL; int best_color = -1; + int did_all = 0; waitq *changed_ones = new_waitq(); waitq *tmp_chunks = new_waitq(); + waitq *best_starts = NULL; bitset_t *visited; - int col, idx; + int col, idx, len; DB((dbg, LEVEL_2, "fragmentizing chunk #%d", c->id)); DBG_AFF_CHUNK(env, LEVEL_2, c); DB((dbg, LEVEL_2, "\n")); - /* check which color is the "best" for the given chunk */ - for (col = 0; col < env->k; ++col) { - int reg_col = env->map_regs[col]; - int one_good = 0; + /* 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 did_all flag 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). + */ + for (col = 0; col < env->n_regs && !did_all; ++col) { + int one_good = 0; + waitq *good_starts = new_waitq(); aff_chunk_t *local_best; - DB((dbg, LEVEL_3, "\ttrying color %d\n", reg_col)); + /* skip ignore colors */ + if (bitset_is_set(env->ignore_regs, col)) + continue; + + DB((dbg, LEVEL_3, "\ttrying color %d\n", col)); + + /* suppose we can color all nodes to the same color */ + did_all = 1; /* try to bring all nodes of given chunk to the current color. */ - bitset_foreach(c->nodes, idx) { - ir_node *irn = get_idx_irn(env->co->irg, idx); + for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { + ir_node *irn = c->n[idx]; co_mst_irn_t *node = get_co_mst_irn(env, irn); + int good = 0; 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)); - DB((dbg, LEVEL_4, "\t\tBringing %+F from color %d to color %d ...\n", irn, node->col, reg_col)); - one_good |= change_node_color(env, node, reg_col, changed_ones); - DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, reg_col, one_good ? "succeeded" : "failed")); + /* + The order of the colored nodes is important, so we record the successfully + colored ones in the order they appeared. + */ + good = change_node_color(env, node, col, changed_ones); + if (good) + waitq_put(good_starts, node); + one_good |= good; + did_all &= good; + + DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, one_good ? "succeeded" : "failed")); } /* try next color when failed */ @@ -1029,24 +1084,31 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { continue; /* fragment the chunk according to the coloring */ - local_best = fragment_chunk(env, reg_col, c, tmp_chunks); + 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_4, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, reg_col)); + DB((dbg, LEVEL_4, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, col)); DBG_AFF_CHUNK(env, LEVEL_4, local_best); if (! best_chunk || best_chunk->weight < local_best->weight) { best_chunk = local_best; - best_color = reg_col; + best_color = col; + if (best_starts) + del_waitq(best_starts); + best_starts = good_starts; DB((dbg, LEVEL_4, "\n\t\t... setting global best chunk (id %d), color %d\n", best_chunk->id, best_color)); } else { DB((dbg, LEVEL_4, "\n\t\t... omitting, global best is better\n")); + del_waitq(good_starts); } } + else { + del_waitq(good_starts); + } /* reject the coloring and bring the coloring to the initial state */ reject_coloring(changed_ones); @@ -1063,6 +1125,8 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { /* return if coloring failed */ if (! best_chunk) { del_waitq(changed_ones); + if (best_starts) + del_waitq(best_starts); return; } @@ -1071,13 +1135,29 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { DB((dbg, LEVEL_2, "using color %d\n", best_color)); /* get the best fragment from the best list and color it */ - bitset_foreach(best_chunk->nodes, idx) { - ir_node *irn = get_idx_irn(env->co->irg, idx); + while (! waitq_empty(best_starts)) { + co_mst_irn_t *node = waitq_get(best_starts); + int res; + + if (! bitset_is_set(best_chunk->nodes, get_irn_idx(node->irn))) + continue; + + res = change_node_color(env, node, best_color, changed_ones); + if (! res) + panic("Color manifesting failed for %+F, color %d in chunk %d\n", node->irn, best_color, best_chunk->id); + node->fixed = 1; + node->chunk = best_chunk; + } + /* we colored the successful start nodes, now color the rest of the chunk */ + for (idx = 0, len = ARR_LEN(best_chunk->n); idx < len; ++idx) { + ir_node *irn = best_chunk->n[idx]; co_mst_irn_t *node = get_co_mst_irn(env, irn); int res; res = change_node_color(env, node, best_color, changed_ones); - assert(res && "color manifesting failed"); + if (! res) + panic("Color manifesting failed for %+F, color %d in chunk %d\n", irn, best_color, best_chunk->id); + DB((dbg, LEVEL_4, "\tManifesting color %d for %+F, chunk #%d\n", best_color, irn, best_chunk->id)); node->fixed = 1; node->chunk = best_chunk; } @@ -1091,10 +1171,21 @@ 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) { + ir_node *irn = c->n[idx]; + + if (bitset_is_set(best_chunk->nodes, get_irn_idx(irn))) { + int last = ARR_LEN(c->n) - 1; + + c->n[idx] = c->n[last]; + ARR_SHRINKLEN(c->n, last); + len--; + } + } /* we have to get the nodes back into the original chunk because they are scattered over temporary chunks */ - bitset_foreach(c->nodes, idx) { - ir_node *n = get_idx_irn(env->co->irg, idx); + for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { + ir_node *n = c->n[idx]; co_mst_irn_t *nn = get_co_mst_irn(env, n); nn->chunk = c; } @@ -1102,10 +1193,10 @@ 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); - bitset_foreach(c->nodes, idx) { - if (! bitset_is_set(visited, idx)) { + for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { + ir_node *irn = c->n[idx]; + if (! bitset_is_set(visited, get_irn_idx(irn))) { aff_chunk_t *new_chunk = new_aff_chunk(env); - ir_node *irn = get_idx_irn(env->co->irg, idx); co_mst_irn_t *node = get_co_mst_irn(env, irn); expand_chunk_from(env, node, visited, new_chunk, c, decider_always_yes, 0); @@ -1118,16 +1209,17 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { delete_aff_chunk(env, best_chunk); bitset_free(visited); del_waitq(changed_ones); + if (best_starts) + del_waitq(best_starts); } /** * Main driver for mst safe coalescing algorithm. */ -int co_solve_heuristic_mst(copy_opt_t *co) -{ +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 k, idx, num; + unsigned k; ir_node *irn; co_mst_env_t mst_env; @@ -1137,16 +1229,6 @@ int co_solve_heuristic_mst(copy_opt_t *co) k = be_put_ignore_regs(co->cenv->birg, co->cls, ignore_regs); k = n_regs - k; - /* Create a color to register number map. In some architectures registers are ignore "in the middle" - of the register set. */ - mst_env.map_regs = NEW_ARR_D(int, phase_obst(&mst_env.ph), k); - for (idx = num = 0; idx < n_regs; ++idx) { - if (bitset_is_set(ignore_regs, idx)) - continue; - mst_env.map_regs[num++] = idx; - } - assert(num == k); - mst_env.n_regs = n_regs; mst_env.k = k; mst_env.chunks = new_pqueue(); @@ -1154,7 +1236,7 @@ int co_solve_heuristic_mst(copy_opt_t *co) mst_env.ignore_regs = ignore_regs; mst_env.ifg = co->cenv->ifg; mst_env.aenv = co->aenv; - pset_new_init(&mst_env.chunkset); + mst_env.chunkset = pset_new_ptr(512); DBG((dbg, LEVEL_1, "==== Coloring %+F, class %s ====\n", co->irg, co->cls->name)); @@ -1192,7 +1274,7 @@ int co_solve_heuristic_mst(copy_opt_t *co) /* free allocated memory */ del_pqueue(mst_env.chunks); phase_free(&mst_env.ph); - pset_new_destroy(&mst_env.chunkset); + del_pset(mst_env.chunkset); return 0; }