X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur4.c;h=dbe5acc87077f0ca241ae239f05d553e7bb8db3d;hb=6ee1fce95429dbf57fda4455ca5f2cf011ac8190;hp=ee2926d99485331f23878514438746148528c070;hpb=89af99b9c92c524c11d3d8f297f4f3277af9b8d5;p=libfirm diff --git a/ir/be/becopyheur4.c b/ir/be/becopyheur4.c index ee2926d99..dbe5acc87 100644 --- a/ir/be/becopyheur4.c +++ b/ir/be/becopyheur4.c @@ -1,20 +1,6 @@ /* - * Copyright (C) 1995-2007 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. + * Copyright (C) 2012 University of Karlsruhe. */ /** @@ -22,16 +8,15 @@ * @brief Simple copy minimization heuristics. * @author Christian Wuerdig * @date 27.04.2007 - * @version $Id$ * * This is the C implementation of the mst algorithm * originally written in Java by Sebastian Hack. * (also known as "heur3" :) * Performs simple copy minimization. */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif /* HAVE_CONFIG_H */ + +#define DISABLE_STATEV #include @@ -39,14 +24,16 @@ #include "irnode_t.h" #include "bitset.h" #include "raw_bitset.h" -#include "irphase_t.h" +#include "irnodemap.h" #include "pqueue.h" #include "xmalloc.h" #include "pdeq.h" -#include "pset.h" #include "irprintf.h" -#include "irbitset.h" +#include "util.h" +#include "irtools.h" #include "error.h" +#include "list.h" +#include "statev_t.h" #include "bearch.h" #include "beifg.h" @@ -55,81 +42,150 @@ #include "bemodule.h" -#define COL_COST_INFEASIBLE DBL_MAX -#define AFF_NEIGHBOUR_FIX_BENEFIT 128.0 -#define NEIGHBOUR_CONSTR_COSTS 64.0 +#ifdef DEBUG_libfirm -#ifdef NDEBUG +#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) -#define DBG_AFF_CHUNK(env, level, chunk) -#define DBG_COL_COST(env, level, cost) +static firm_dbg_module_t *dbg = NULL; #else -static firm_dbg_module_t *dbg = NULL; -#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) +#define DBG_AFF_CHUNK(env, level, chunk) +#define DBG_COL_COST(env, level, cost) #endif -static int last_chunk_id = 0; +typedef float real_t; +#define REAL(C) (C ## f) + +static unsigned last_chunk_id = 0; +static int recolor_limit = 7; +static double dislike_influence = REAL(0.1); -typedef struct _col_cost_t { - int col; - double cost; +typedef struct col_cost_t { + int col; + real_t cost; } 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. */ +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 weight_consistent : 1; /**< Set if the weight is consistent. */ + unsigned deleted : 1; /**< For debugging: Set if the was deleted. */ + unsigned id; /**< An id of this chunk. */ + unsigned visited; + list_head list; + col_cost_t color_affinity[1]; } aff_chunk_t; /** * An affinity edge. */ -typedef struct _aff_edge_t { - ir_node *src; /**< Source node. */ - ir_node *tgt; /**< Target node. */ - double weight; /**< The weight of this 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 { +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 */ - ir_phase ph; /**< phase object holding data for nodes */ - pqueue *chunks; /**< priority queue for chunks */ - pset *chunkset; /**< set holding all chunks */ + 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 */ - const arch_env_t *aenv; /**< the arch environment */ copy_opt_t *co; /**< the copy opt object */ + unsigned chunk_visited; + col_cost_t **single_cols; } co_mst_env_t; /* stores coalescing related information for a node */ -typedef struct _co_mst_irn_t { - 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) */ - int n_neighs; /**< length of the interfering neighbours array. */ - int int_aff_neigh; /**< number of interfering affinity neighbours */ - int col; /**< color currently assigned */ - int init_col; /**< the initial color */ - int tmp_col; /**< a temporary assigned color */ - unsigned fixed : 1; /**< the color is fixed */ - unsigned tmp_fixed : 1; /**< the color is temporary fixed */ +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) */ + int n_neighs; /**< length of the interfering neighbours array. */ + int int_aff_neigh; /**< number of interfering affinity neighbours */ + int col; /**< color currently assigned */ + int init_col; /**< the initial color */ + int tmp_col; /**< a temporary assigned color */ + unsigned fixed : 1; /**< the color is fixed */ + struct list_head list; /**< Queue for coloring undo. */ + 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))) +/** + * 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 *res = OALLOC(&env->obst, co_mst_irn_t); + + const arch_register_req_t *req; + neighbours_iter_t nodes_it; + unsigned len; + + res->irn = irn; + res->chunk = NULL; + res->fixed = 0; + res->tmp_col = -1; + res->int_neighs = NULL; + res->int_aff_neigh = 0; + 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 */ + res->adm_colors = bitset_obstack_alloc(&env->obst, env->n_regs); + + /* Exclude colors not assignable to the irn */ + req = arch_get_irn_register_req(irn); + if (arch_register_req_is(req, limited)) { + rbitset_copy_to_bitset(req->limited, res->adm_colors); + /* exclude global ignore registers as well */ + bitset_and(res->adm_colors, env->allocatable_regs); + } else { + bitset_copy(res->adm_colors, env->allocatable_regs); + } + + /* compute the constraint factor */ + res->constr_factor = (real_t) (1 + env->n_regs - bitset_popcount(res->adm_colors)) / env->n_regs; + + /* set the number of interfering affinity neighbours to -1, they are calculated later */ + res->int_aff_neigh = -1; + + /* build list of interfering neighbours */ + len = 0; + 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 int decide_func_t(const co_mst_irn_t *node, int col); @@ -138,13 +194,15 @@ 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; +static void dbg_aff_chunk(const co_mst_env_t *env, const aff_chunk_t *c) +{ + 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, "}"); @@ -153,64 +211,67 @@ 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) { - bitset_pos_t idx; +static void dbg_admissible_colors(const co_mst_env_t *env, const co_mst_irn_t *node) +{ (void) env; - if (bitset_popcnt(node->adm_colors) < 1) + if (bitset_popcount(node->adm_colors) < 1) fprintf(stderr, "no admissible colors?!?"); else { - bitset_foreach(node->adm_colors, idx) - fprintf(stderr, " %d", idx); + 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) { +static void dbg_col_cost(const co_mst_env_t *env, const col_cost_t *cost) +{ int i; - for (i = 0; i < env->n_regs; ++i) { - if (cost[i].cost == COL_COST_INFEASIBLE) - fprintf(stderr, " (%d, INF)", cost[i].col); - else - fprintf(stderr, " (%d, %.1f)", cost[i].col, cost[i].cost); - } + for (i = 0; i < env->n_regs; ++i) + fprintf(stderr, " (%d, %.4f)", cost[i].col, cost[i].cost); } #endif /* DEBUG_libfirm */ -static INLINE int get_mst_irn_col(const co_mst_irn_t *node) { - return node->tmp_fixed ? node->tmp_col : node->col; +static inline int get_mst_irn_col(const co_mst_irn_t *node) +{ + return node->tmp_col >= 0 ? node->tmp_col : node->col; } /** * @return 1 if node @p node has color @p col, 0 otherwise. */ -static int decider_has_color(const co_mst_irn_t *node, int col) { +static int decider_has_color(const co_mst_irn_t *node, int col) +{ return get_mst_irn_col(node) == col; } /** * @return 1 if node @p node has not color @p col, 0 otherwise. */ -static int decider_hasnot_color(const co_mst_irn_t *node, int col) { +static int decider_hasnot_color(const co_mst_irn_t *node, int col) +{ return get_mst_irn_col(node) != col; } /** * Always returns true. */ -static int decider_always_yes(const co_mst_irn_t *node, int col) { +static int decider_always_yes(const co_mst_irn_t *node, int col) +{ (void) node; (void) col; return 1; } /** compares two affinity edges by its weight */ -static int cmp_aff_edge(const void *a, const void *b) { - const aff_edge_t *e1 = a; - const aff_edge_t *e2 = b; +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) @@ -223,150 +284,177 @@ static int cmp_aff_edge(const void *a, const void *b) { } /** compares to color-cost pairs */ -static int cmp_col_cost(const void *a, const void *b) { - const col_cost_t *c1 = a; - const col_cost_t *c2 = b; +static __attribute__((unused)) int cmp_col_cost_lt(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 = c1->cost - c2->cost; + + if (diff < 0) + return 1; + if (diff > 0) + return -1; - return c1->cost < c2->cost ? -1 : 1; + return QSORT_CMP(c1->col, c2->col); +} + +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 = 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(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_insert(env->chunkset, c, c->id); + c->deleted = 0; + 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(co_mst_env_t *env, aff_chunk_t *c) { - pset_remove(env->chunkset, c, c->id); - bitset_free(c->nodes); - bitset_free(c->interfere); +static inline void delete_aff_chunk(aff_chunk_t *c) +{ + list_del(&c->list); + DEL_ARR_F(c->interfere); DEL_ARR_F(c->n); c->deleted = 1; free(c); } /** - * Adds a node to an affinity chunk + * 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 void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) { - int i; +static inline int nodes_bsearch(const ir_node **arr, const ir_node *n) +{ + int hi = ARR_LEN(arr); + int lo = 0; - if (bitset_is_set(c->nodes, get_irn_idx(node->irn))) - return; + while (lo < hi) { + int md = lo + ((hi - lo) >> 1); - c->weight_consistent = 0; - node->chunk = c; - bitset_set(c->nodes, get_irn_idx(node->irn)); + if (arr[md] == n) + return md; + if (arr[md] < n) + lo = md + 1; + else + hi = md; + } - ARR_APP1(ir_node *, c->n, node->irn); + return ~lo; +} - for (i = node->n_neighs - 1; i >= 0; --i) { - ir_node *neigh = node->int_neighs[i]; - bitset_set(c->interfere, get_irn_idx(neigh)); +/** 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; } /** - * In case there is no phase information for irn, initialize it. + * Adds a node to an affinity chunk */ -static void *co_mst_irn_init(ir_phase *ph, ir_node *irn, void *old) { - co_mst_irn_t *res = old ? old : phase_alloc(ph, sizeof(res[0])); - co_mst_env_t *env = ph->priv; - - if (res != old) { - const arch_register_req_t *req; - void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg); - ir_node *neigh; - unsigned len; - - res->irn = irn; - res->chunk = NULL; - res->fixed = 0; - res->tmp_fixed = 0; - 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->init_col = res->col; - - DB((dbg, LEVEL_4, "Creating phase info for %+F\n", irn)); - - /* set admissible registers */ - 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); - if (arch_register_req_is(req, limited)) - rbitset_copy_to_bitset(req->limited, res->adm_colors); - else - bitset_set_all(res->adm_colors); +static inline void aff_chunk_add_node(aff_chunk_t *c, co_mst_irn_t *node) +{ + int i; - /* exclude global ignore registers as well */ - bitset_andnot(res->adm_colors, env->ignore_regs); + if (! nodes_insert(&c->n, node->irn)) + return; - /* set the number of interfering affinity neighbours to -1, they are calculated later */ - res->int_aff_neigh = -1; + c->weight_consistent = 0; + node->chunk = c; - /* 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)) { - obstack_ptr_grow(phase_obst(ph), neigh); - ++len; - } - } - res->int_neighs = obstack_finish(phase_obst(ph)); - res->n_neighs = len; + for (i = node->n_neighs - 1; i >= 0; --i) { + ir_node *neigh = node->int_neighs[i]; + nodes_insert(&c->interfere, neigh); } - return res; } /** * 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) { - (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) { - bitset_t *tmp; +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 */ - tmp = bitset_alloca(get_irg_last_idx(env->co->irg)); - tmp = bitset_copy(tmp, c1->interfere); - tmp = bitset_and(tmp, c2->nodes); + for (i = ARR_LEN(c2->n) - 1; i >= 0; --i) { + const ir_node *irn = c2->n[i]; - return bitset_popcnt(tmp) > 0; + 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, 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; } @@ -376,18 +464,19 @@ static INLINE aff_chunk_t *get_aff_chunk(co_mst_env_t *env, ir_node *irn) { * are no interference edges from chunk(src) to chunk(tgt)). * @return 1 if successful, 0 if not possible */ -static int aff_chunk_absorb(co_mst_env_t *env, ir_node *src, ir_node *tgt) { +static int aff_chunk_absorb(co_mst_env_t *env, const ir_node *src, const ir_node *tgt) +{ aff_chunk_t *c1 = get_aff_chunk(env, src); aff_chunk_t *c2 = get_aff_chunk(env, tgt); -#ifndef NDEBUG - DB((dbg, LEVEL_4, "Attempt to let c1 (id %d): ", c1 ? c1->id : -1)); +#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 %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 { @@ -415,35 +504,31 @@ static int aff_chunk_absorb(co_mst_env_t *env, ir_node *src, ir_node *tgt) { } } 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) { - ir_node *n = c2->n[idx]; - co_mst_irn_t *mn = get_co_mst_irn(env, n); + for (idx = 0, len = ARR_LEN(c2->n); idx < len; ++idx) + aff_chunk_add_node(c1, get_co_mst_irn(env, c2->n[idx])); - mn->chunk = c1; - - if (! bitset_is_set(c1->nodes, get_irn_idx(n))) - ARR_APP1(ir_node *, c1->n, n); + for (idx = 0, len = ARR_LEN(c2->interfere); idx < len; ++idx) { + const ir_node *irn = c2->interfere[idx]; + nodes_insert(&c1->interfere, irn); } - bitset_or(c1->nodes, c2->nodes); - bitset_or(c1->interfere, c2->interfere); c1->weight_consistent = 0; - delete_aff_chunk(env, c2); + delete_aff_chunk(c2); goto absorbed; } DB((dbg, LEVEL_4, " ... c1 interferes with c2, skipped\n")); @@ -457,31 +542,45 @@ absorbed: /** * Assures that the weight of the given chunk is consistent. */ -static void aff_chunk_assure_weight(const co_mst_env_t *env, aff_chunk_t *c) { +static void aff_chunk_assure_weight(co_mst_env_t *env, aff_chunk_t *c) +{ if (! c->weight_consistent) { int w = 0; - int idx, len; + int idx, len, i; + + for (i = 0; i < env->n_regs; ++i) { + c->color_affinity[i].col = i; + c->color_affinity[i].cost = REAL(0.0); + } 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); + 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) { - 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 *= (REAL(1.0) / ARR_LEN(c->n)); + c->weight = w; + // c->weight = bitset_popcount(c->nodes); c->weight_consistent = 1; } } @@ -489,9 +588,9 @@ static void aff_chunk_assure_weight(const co_mst_env_t *env, aff_chunk_t *c) { /** * Count the number of interfering affinity neighbours */ -static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t *an) { - const neighb_t *neigh; - ir_node *irn = an->irn; +static int 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); int res = 0; @@ -499,8 +598,7 @@ 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 */ @@ -522,36 +620,30 @@ static int count_interfering_aff_neighs(co_mst_env_t *env, const affinity_node_t * merged if there are no interference edges from one * chunk to the other. */ -static void build_affinity_chunks(co_mst_env_t *env) { - void *nodes_it = be_ifg_nodes_iter_alloca(env->ifg); - aff_edge_t *edges = NEW_ARR_F(aff_edge_t, 0); - ir_node *n; - int i, len; - aff_chunk_t *curr_chunk; +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, nodes_it, n) { + be_ifg_foreach_node(env->ifg, n) { int n_idx = get_irn_idx(n); 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); an = get_affinity_info(env->co, n); if (an != NULL) { - neighb_t *neigh; - 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); + const ir_node *m = neigh->irn; + int m_idx = get_irn_idx(m); /* record the edge in only one direction */ if (n_idx < m_idx) { @@ -559,7 +651,7 @@ 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; @@ -574,7 +666,7 @@ static void build_affinity_chunks(co_mst_env_t *env) { * these weights are pure hackery ;-). * It's not chriswue's fault but mine. */ - edge.weight = (double)neigh->costs / (double)(1 + n1->int_aff_neigh + n2->int_aff_neigh); + edge.weight = neigh->costs; ARR_APP1(aff_edge_t, edges, edge); } } @@ -582,45 +674,111 @@ static void build_affinity_chunks(co_mst_env_t *env) { } /* now: sort edges and build the affinity chunks */ - len = ARR_LEN(edges); + size_t const len = ARR_LEN(edges); qsort(edges, len, sizeof(edges[0]), cmp_aff_edge); - for (i = 0; i < len; ++i) { + for (size_t i = 0; 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)); (void)aff_chunk_absorb(env, edges[i].src, edges[i].tgt); } /* now insert all chunks into a priority queue */ - foreach_pset(env->chunkset, curr_chunk) { + list_for_each_entry(aff_chunk_t, curr_chunk, &env->chunklist, list) { 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); - if (mirn->chunk == NULL) { - /* 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); + for (size_t pn = 0; pn < ARR_LEN(env->map.data); ++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); + aff_chunk_assure_weight(env, curr_chunk); - DBG((dbg, LEVEL_1, "entry #%d", curr_chunk->id)); - DBG_AFF_CHUNK(env, LEVEL_1, curr_chunk); - DBG((dbg, LEVEL_1, "\n")); + 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); - } + pqueue_put(env->chunks, curr_chunk, curr_chunk->weight); } DEL_ARR_F(edges); } +static __attribute__((unused)) void chunk_order_nodes(co_mst_env_t *env, aff_chunk_t *chunk) +{ + pqueue_t *grow = new_pqueue(); + ir_node const *max_node = NULL; + int max_weight = 0; + size_t i; + + for (i = ARR_LEN(chunk->n); i != 0;) { + const ir_node *irn = chunk->n[--i]; + affinity_node_t *an = get_affinity_info(env->co, irn); + int w = 0; + + if (arch_irn_is_ignore(irn)) + continue; + + if (an) { + co_gs_foreach_neighb(an, neigh) + w += neigh->costs; + + if (w > max_weight) { + max_weight = w; + max_node = irn; + } + } + } + + if (max_node) { + bitset_t *visited = bitset_malloc(get_irg_last_idx(env->co->irg)); + + for (i = ARR_LEN(chunk->n); i != 0;) + bitset_set(visited, get_irn_idx(chunk->n[--i])); + + pqueue_put(grow, (void *) max_node, max_weight); + bitset_clear(visited, get_irn_idx(max_node)); + i = 0; + while (!pqueue_empty(grow)) { + ir_node *irn = (ir_node*)pqueue_pop_front(grow); + affinity_node_t *an = get_affinity_info(env->co, irn); + + if (arch_irn_is_ignore(irn)) + continue; + + assert(i <= ARR_LEN(chunk->n)); + chunk->n[i++] = irn; + + assert(an); + + /* build the affinity edges */ + co_gs_foreach_neighb(an, neigh) { + co_mst_irn_t *node = get_co_mst_irn(env, neigh->irn); + + if (bitset_is_set(visited, get_irn_idx(node->irn))) { + pqueue_put(grow, (void *) neigh->irn, neigh->costs); + bitset_clear(visited, get_irn_idx(node->irn)); + } + } + } + + del_pqueue(grow); + bitset_free(visited); + } +} + /** * Greedy collect affinity neighbours into thew new chunk @p chunk starting at node @p node. */ @@ -629,7 +787,7 @@ 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); @@ -639,28 +797,26 @@ static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *v /* as long as there are nodes in the queue */ while (! waitq_empty(nodes)) { - co_mst_irn_t *n = waitq_get(nodes); + co_mst_irn_t *n = (co_mst_irn_t*)waitq_get(nodes); affinity_node_t *an = get_affinity_info(env->co, n->irn); /* check all affinity neighbors */ if (an != NULL) { - neighb_t *neigh; co_gs_foreach_neighb(an, neigh) { - ir_node *m = neigh->irn; - int m_idx = get_irn_idx(m); + const ir_node *m = neigh->irn; + 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: @@ -688,13 +844,14 @@ static void expand_chunk_from(co_mst_env_t *env, co_mst_irn_t *node, bitset_t *v /** * 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, int col, aff_chunk_t *c, waitq *tmp) { - bitset_t *visited = bitset_irg_malloc(env->co->irg); +static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, waitq *tmp) +{ + bitset_t *visited = bitset_malloc(get_irg_last_idx(env->co->irg)); int idx, len; aff_chunk_t *best = NULL; for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { - ir_node *irn; + const ir_node *irn; co_mst_irn_t *node; aff_chunk_t *tmp_chunk; decide_func_t *decider; @@ -709,19 +866,19 @@ 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); @@ -734,153 +891,116 @@ static aff_chunk_t *fragment_chunk(co_mst_env_t *env, int col, aff_chunk_t *c, w return best; } -/** - * Initializes an array of color-cost pairs. - * Sets forbidden colors to costs COL_COST_INFEASIBLE and all others to @p c. - */ -static INLINE void col_cost_init(co_mst_env_t *env, col_cost_t *cost, double c) { - int i; - - for (i = 0; i < env->n_regs; ++i) { - cost[i].col = i; - if (bitset_is_set(env->ignore_regs, i)) - cost[i].cost = COL_COST_INFEASIBLE; - else - cost[i].cost = c; - } -} - -/** - * Initializes an array of color-cost pairs. - * Sets all colors except color @p col to COL_COST_INFEASIBLE and @p col to 0.0 - */ -static INLINE void col_cost_init_single(co_mst_env_t *env, col_cost_t *cost, int col) { - assert(! bitset_is_set(env->ignore_regs, col) && "Attempt to use forbidden color."); - col_cost_init(env, cost, COL_COST_INFEASIBLE); - cost[col].col = 0; - cost[0].col = col; - cost[0].cost = 0.0; -} - /** * 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(waitq *nodes) { +static inline void reject_coloring(struct list_head *nodes) +{ DB((dbg, LEVEL_4, "\treject coloring for")); - while (! waitq_empty(nodes)) { - co_mst_irn_t *n = waitq_get(nodes); + list_for_each_entry_safe(co_mst_irn_t, n, temp, nodes, list) { DB((dbg, LEVEL_4, " %+F", n->irn)); - n->tmp_fixed = 0; + 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, int col, struct list_head *changed) +{ + assert(col >= 0); + 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 int is_loose(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) { - affinity_node_t *an = get_affinity_info(env->co, node->irn); - neighb_t *aff_neigh; - bitset_pos_t idx; - int i; - - col_cost_init(env, costs, 0.0); - - /* calculate (negative) costs for affinity neighbours */ - if (an != NULL) { - co_gs_foreach_neighb(an, aff_neigh) { - ir_node *m = aff_neigh->irn; - co_mst_irn_t *neigh; - double c; - - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, m, ignore)) - continue; - - neigh = get_co_mst_irn(env, m); - c = (double)aff_neigh->costs; +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; - /* calculate costs for fixed affinity neighbours */ - if (neigh->tmp_fixed || neigh->fixed) { - int col = get_mst_irn_col(neigh); - costs[col].cost -= c * AFF_NEIGHBOUR_FIX_BENEFIT; - } - } + 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 : REAL(0.0); } - /* calculate (positive) costs for interfering neighbours */ for (i = 0; i < node->n_neighs; ++i) { - co_mst_irn_t *neigh; - int col, col_cnt; - ir_node *int_neigh; - - int_neigh = node->int_neighs[i]; - - /* skip ignore nodes */ - if (arch_irn_is(env->aenv, int_neigh, ignore)) - continue; - - neigh = get_co_mst_irn(env, int_neigh); - col = get_mst_irn_col(neigh); - col_cnt = bitset_popcnt(neigh->adm_colors); - - if (neigh->tmp_fixed || neigh->fixed) { - /* colors of fixed interfering neighbours are infeasible */ - costs[col].cost = COL_COST_INFEASIBLE; - } - else if (col_cnt < env->k) { - /* calculate costs for constrained interfering neighbours */ - double ratio = 1.0 - ((double)col_cnt / (double)env->k); - - bitset_foreach_clear(neigh->adm_colors, idx) { - /* check only explicitly forbidden colors (skip global forbidden ones) */ - if (! bitset_is_set(env->ignore_regs, idx)) { - costs[col].cost += ratio * NEIGHBOUR_CONSTR_COSTS; - } - } - } + co_mst_irn_t *n = get_co_mst_irn(env, node->int_neighs[i]); + int col = get_mst_irn_col(n); + if (is_loose(n)) { + ++n_loose; + ++neigh_cols[col]; + } else + costs[col].cost = REAL(0.0); } - /* set all not admissible colors to COL_COST_INFEASIBLE */ - bitset_foreach_clear(node->adm_colors, idx) - costs[idx].cost = COL_COST_INFEASIBLE; + if (n_loose > 0) { + coeff = REAL(1.0) / n_loose; + for (i = 0; i < env->n_regs; ++i) + costs[i].cost *= REAL(1.0) - coeff * neigh_cols[i]; + } } /* need forward declaration due to recursive call */ -static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, waitq *changed_ones); +static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed_ones, int depth, int *max_depth, int *trip); /** * Tries to change node to a color but @p explude_col. * @return 1 if succeeded, 0 otherwise. */ -static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, waitq *changed_ones) { +static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int exclude_col, struct list_head *changed, int depth, int *max_depth, int *trip) +{ int col = get_mst_irn_col(node); int res = 0; /* neighbours has already a different color -> good, temporary fix it */ if (col != exclude_col) { - node->tmp_fixed = 1; - node->tmp_col = col; - waitq_put(changed_ones, node); + if (is_loose(node)) + set_temp_color(node, col, changed); return 1; } /* The node has the color it should not have _and_ has not been visited yet. */ - if (! (node->tmp_fixed || node->fixed)) { - col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0])); + 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 = COL_COST_INFEASIBLE; + 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); + qsort(costs, env->n_regs, sizeof(costs[0]), cmp_col_cost_gt); /* Try recoloring the node using the color list. */ - res = recolor_nodes(env, node, costs, changed_ones); + res = recolor_nodes(env, node, costs, changed, depth + 1, max_depth, trip); } return res; @@ -891,14 +1011,23 @@ static int change_node_color_excluded(co_mst_env_t *env, co_mst_irn_t *node, int * ATTENTION: Expect @p costs already sorted by increasing costs. * @return 1 if coloring could be applied, 0 otherwise. */ -static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, waitq *changed_ones) { +static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *costs, struct list_head *changed, int depth, int *max_depth, int *trip) +{ int i; - waitq *local_changed = new_waitq(); - waitq *tmp = new_waitq(); + struct list_head local_changed; - DBG((dbg, LEVEL_1, "\tRecoloring %+F with color-costs", node->irn)); - DBG_COL_COST(env, LEVEL_1, costs); - DB((dbg, LEVEL_1, "\n")); + ++*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 >= 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; @@ -906,22 +1035,17 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost 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 == COL_COST_INFEASIBLE) { - node->tmp_fixed = 0; - del_waitq(local_changed); - del_waitq(tmp); + 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_fixed && "Node must not have been temporary fixed."); - node->tmp_fixed = 1; - node->tmp_col = tgt_col; + 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)); - assert(waitq_empty(local_changed) && "Node queue should be empty here."); - waitq_put(local_changed, node); - /* try to color all interfering neighbours with current color forbidden */ for (j = 0; j < node->n_neighs; ++j) { co_mst_irn_t *nn; @@ -929,13 +1053,12 @@ 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); - 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)); + 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 @@ -945,13 +1068,9 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost */ 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, tmp); - - /* join lists of changed nodes */ - while (! waitq_empty(tmp)) - waitq_put(local_changed, waitq_get(tmp)); + neigh_ok = change_node_color_excluded(env, nn, tgt_col, &local_changed, depth + 1, max_depth, trip); - if (! neigh_ok) + if (!neigh_ok) break; } } @@ -962,61 +1081,59 @@ static int recolor_nodes(co_mst_env_t *env, co_mst_irn_t *node, col_cost_t *cost */ if (neigh_ok) { /* append the local_changed ones to global ones */ - while (! waitq_empty(local_changed)) - waitq_put(changed_ones, waitq_get(local_changed)); - del_waitq(local_changed); - del_waitq(tmp); + list_splice(&local_changed, changed); return 1; } else { /* coloring of neighbours failed, so we try next color */ - reject_coloring(local_changed); + reject_coloring(&local_changed); } } - del_waitq(local_changed); - del_waitq(tmp); + DBG((dbg, LEVEL_4, "\tAll colors failed\n")); return 0; } /** - * Tries to bring node @p node and all it's neighbours to color @p tgt_col. + * Tries to bring node @p node and all its neighbours to color @p tgt_col. * @return 1 if color @p col could be applied, 0 otherwise */ -static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, waitq *changed_ones) { +static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, struct list_head *changed) +{ int col = get_mst_irn_col(node); /* if node already has the target color -> good, temporary fix it */ if (col == tgt_col) { DBG((dbg, LEVEL_4, "\t\tCNC: %+F has already color %d, fix temporary\n", node->irn, tgt_col)); - if (! node->tmp_fixed) { - node->tmp_fixed = 1; - node->tmp_col = tgt_col; - waitq_put(changed_ones, node); - } + if (is_loose(node)) + set_temp_color(node, tgt_col, changed); return 1; } /* Node has not yet a fixed color and target color is admissible - -> try to recolor node and it's affinity neighbours + -> try to recolor node and its affinity neighbours */ - if (! (node->fixed || node->tmp_fixed) && bitset_is_set(node->adm_colors, tgt_col)) { - col_cost_t *costs = alloca(env->n_regs * sizeof(costs[0])); - int res; + if (is_loose(node) && bitset_is_set(node->adm_colors, tgt_col)) { + col_cost_t *costs = env->single_cols[tgt_col]; + int res, max_depth, trip; - col_cost_init_single(env, costs, tgt_col); + max_depth = 0; + trip = 0; DBG((dbg, LEVEL_4, "\t\tCNC: Attempt to recolor %+F ===>>\n", node->irn)); - res = recolor_nodes(env, node, costs, changed_ones); + 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; } -#ifndef NDEBUG +#ifdef DEBUG_libfirm if (firm_dbg_get_mask(dbg) & LEVEL_4) { - if (node->fixed || node->tmp_fixed) + if (!is_loose(node)) DB((dbg, LEVEL_4, "\t\tCNC: %+F has already fixed color %d\n", node->irn, col)); else { DB((dbg, LEVEL_4, "\t\tCNC: color %d not admissible for %+F (", tgt_col, node->irn)); @@ -1033,20 +1150,58 @@ static int change_node_color(co_mst_env_t *env, co_mst_irn_t *node, int tgt_col, * 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) { +static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) +{ aff_chunk_t *best_chunk = NULL; + int n_nodes = ARR_LEN(c->n); int best_color = -1; - int did_all = 0; - waitq *changed_ones = new_waitq(); + int n_int_chunks = 0; waitq *tmp_chunks = new_waitq(); waitq *best_starts = NULL; + col_cost_t *order = ALLOCANZ(col_cost_t, env->n_regs); bitset_t *visited; - int col, idx, len; - - DB((dbg, LEVEL_2, "fragmentizing chunk #%d", c->id)); + int i; + size_t idx; + size_t len; + size_t nidx; + size_t pos; + struct list_head changed; + + 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 */ + 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) { + assert(!chunk->deleted); + chunk->visited = env->chunk_visited; + ++n_int_chunks; + + aff_chunk_assure_weight(env, chunk); + for (i = 0; i < env->n_regs; ++i) + order[i].cost += chunk->color_affinity[i].cost; + } + } + + for (i = 0; i < env->n_regs; ++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, sizeof(order[0]), 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 @@ -1056,25 +1211,26 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { * 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(); + for (i = 0; i < env->n_regs; ++i) { + int col = order[i].col; + waitq *good_starts; aff_chunk_t *local_best; + int n_succeeded; /* skip ignore colors */ - if (bitset_is_set(env->ignore_regs, col)) + if (!bitset_is_set(env->allocatable_regs, col)) continue; - DB((dbg, LEVEL_3, "\ttrying color %d\n", col)); + DB((dbg, LEVEL_2, "\ttrying color %d\n", col)); - /* suppose we can color all nodes to the same color */ - did_all = 1; + n_succeeded = 0; + good_starts = new_waitq(); /* try to bring all nodes of given chunk to the current color. */ 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; + const ir_node *irn = c->n[idx]; + co_mst_irn_t *node = get_co_mst_irn(env, irn); + int good; 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)); @@ -1083,18 +1239,34 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { 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) + INIT_LIST_HEAD(&changed); + stat_ev_tim_push(); + good = change_node_color(env, node, col, &changed); + stat_ev_tim_pop("heur4_recolor"); + if (good) { waitq_put(good_starts, node); - one_good |= good; - did_all &= good; + materialize_coloring(&changed); + node->fixed = 1; + } - DB((dbg, LEVEL_4, "\t\t... %+F attempt from %d to %d %s\n", irn, node->col, col, one_good ? "succeeded" : "failed")); + 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 (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 (! one_good) + if (n_succeeded == 0) { + del_waitq(good_starts); continue; + } /* fragment the chunk according to the coloring */ local_best = fragment_chunk(env, col, c, tmp_chunks); @@ -1104,8 +1276,8 @@ 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_4, "\t\tlocal best chunk (id %d) for color %d: ", local_best->id, col)); - DBG_AFF_CHUNK(env, LEVEL_4, 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; @@ -1113,9 +1285,9 @@ 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_4, "\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_4, "\n\t\t... omitting, global best is better\n")); + DB((dbg, LEVEL_3, "\n\t\t... omitting, global best is better\n")); del_waitq(good_starts); } } @@ -1123,91 +1295,84 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { del_waitq(good_starts); } - /* reject the coloring and bring the coloring to the initial state */ - reject_coloring(changed_ones); + /* 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 (! waitq_empty(tmp_chunks)) { - aff_chunk_t *tmp = waitq_get(tmp_chunks); + aff_chunk_t *tmp = (aff_chunk_t*)waitq_get(tmp_chunks); if (tmp != best_chunk) - delete_aff_chunk(env, tmp); + delete_aff_chunk(tmp); } del_waitq(tmp_chunks); /* return if coloring failed */ if (! best_chunk) { - del_waitq(changed_ones); if (best_starts) del_waitq(best_starts); 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)); - /* get the best fragment from the best list and color it */ - 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); - 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; - } - - /* materialize colors on changed nodes */ - while (! waitq_empty(changed_ones)) { - co_mst_irn_t *n = waitq_get(changed_ones); - n->tmp_fixed = 0; - n->col = n->tmp_col; + const ir_node *irn = best_chunk->n[idx]; + co_mst_irn_t *node = get_co_mst_irn(env, irn); + int res; + + /* 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)); + INIT_LIST_HEAD(&changed); + stat_ev_tim_push(); + 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 */ - bitset_andnot(c->nodes, best_chunk->nodes); - for (idx = 0, len = ARR_LEN(c->n); idx < len; ++idx) { - 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) { - ir_node *n = c->n[idx]; - co_mst_irn_t *nn = get_co_mst_irn(env, n); + 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 */ - visited = bitset_irg_malloc(env->co->irg); - bitset_or(visited, best_chunk->nodes); + visited = bitset_malloc(get_irg_last_idx(env->co->irg)); + 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) { - ir_node *irn = c->n[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); @@ -1218,82 +1383,125 @@ static void color_aff_chunk(co_mst_env_t *env, aff_chunk_t *c) { } } + for (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(env, best_chunk); + delete_aff_chunk(best_chunk); bitset_free(visited); - del_waitq(changed_ones); if (best_starts) del_waitq(best_starts); + + stat_ev_ctx_pop("heur4_color_chunk"); } /** * Main driver for mst safe coalescing algorithm. */ -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; - ir_node *irn; - co_mst_env_t mst_env; - - /* init phase */ - phase_init(&mst_env.ph, "co_mst", co->irg, PHASE_DEFAULT_GROWTH, co_mst_irn_init, &mst_env); +static int co_solve_heuristic_mst(copy_opt_t *co) +{ + last_chunk_id = 0; - k = be_put_ignore_regs(co->cenv->birg, co->cls, ignore_regs); - k = n_regs - k; + stat_ev_tim_push(); - mst_env.n_regs = n_regs; - mst_env.k = k; - mst_env.chunks = new_pqueue(); - 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); + /* 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 = pqueue_get(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 (%d) done\n", chunk->id)); - delete_aff_chunk(&mst_env, chunk); + DB((dbg, LEVEL_4, "<<<====== Coloring chunk (%u) done\n", chunk->id)); + delete_aff_chunk(chunk); } /* apply coloring */ - foreach_phase_irn(&mst_env.ph, irn) { - co_mst_irn_t *mirn = get_co_mst_irn(&mst_env, irn); + 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]; const arch_register_t *reg; - - if (arch_irn_is(mst_env.aenv, irn, ignore)) + if (mirn == NULL) + continue; + ir_node *const irn = get_idx_irn(co->irg, pn); + if (arch_irn_is_ignore(irn)) continue; - - assert(mirn->fixed && "Node should have fixed color"); /* skip nodes where color hasn't changed */ if (mirn->init_col == mirn->col) 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)); } /* free allocated memory */ del_pqueue(mst_env.chunks); - phase_free(&mst_env.ph); - del_pset(mst_env.chunkset); + obstack_free(&mst_env.obst, NULL); + ir_nodemap_destroy(&mst_env.map); + + stat_ev_tim_pop("heur4_total"); return 0; } -void be_init_copyheur4(void) { +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, 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);