X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbeblocksched.c;h=d1b58e500383d8b92e85c459a6fba1af79cc0849;hb=945c6c2ceebef5e41c0486c31f49d2319cacb3da;hp=11fb8c0a1211ca8a95097d8342e5acf5a38e2581;hpb=5f854aadb7073a3b9318c965a76234c309858704;p=libfirm diff --git a/ir/be/beblocksched.c b/ir/be/beblocksched.c index 11fb8c0a1..d1b58e500 100644 --- a/ir/be/beblocksched.c +++ b/ir/be/beblocksched.c @@ -1,20 +1,6 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. - * * This file is part of libFirm. - * - * This file may be distributed and/or modified under the terms of the - * GNU General Public License version 2 as published by the Free Software - * Foundation and appearing in the file LICENSE.GPL included in the - * packaging of this file. - * - * Licensees holding valid libFirm Professional Edition licenses may use - * this file in accordance with the libFirm Commercial License. - * Agreement provided with the Software. - * - * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE - * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR - * PURPOSE. + * Copyright (C) 2012 University of Karlsruhe. */ /** @@ -22,7 +8,6 @@ * @brief Block-scheduling strategies. * @author Matthias Braun, Christoph Mallon * @date 27.09.2006 - * @version $Id$ * * The goals of the greedy (and ILP) algorithm here works by assuming that * we want to change as many jumps to fallthroughs as possible (executed jumps @@ -42,7 +27,7 @@ #include "array.h" #include "pdeq.h" - +#include "beirg.h" #include "iredges.h" #include "irgwalk.h" #include "irnode_t.h" @@ -52,6 +37,7 @@ #include "execfreq.h" #include "irdump_t.h" #include "irtools.h" +#include "util.h" #include "debug.h" #include "beirgmod.h" #include "bemodule.h" @@ -61,26 +47,21 @@ #include "lc_opts.h" #include "lc_opts_enum.h" -#ifdef WITH_ILP -#include -#include -#endif /* WITH_ILP */ +#include "lpp.h" +#include "lpp_net.h" DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) -typedef enum _blocksched_algos_t { - BLOCKSCHED_NAIV, BLOCKSCHED_EXTBB, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP +typedef enum blocksched_algos_t { + BLOCKSCHED_NAIV, BLOCKSCHED_GREEDY, BLOCKSCHED_ILP } blocksched_algos_t; static int algo = BLOCKSCHED_GREEDY; static const lc_opt_enum_int_items_t blockschedalgo_items[] = { - { "naiv", BLOCKSCHED_NAIV }, - { "extbb", BLOCKSCHED_EXTBB }, + { "naiv", BLOCKSCHED_NAIV }, { "greedy", BLOCKSCHED_GREEDY }, -#ifdef WITH_ILP { "ilp", BLOCKSCHED_ILP }, -#endif /* WITH_ILP */ { NULL, 0 } }; @@ -89,7 +70,7 @@ static lc_opt_enum_int_var_t algo_var = { }; static const lc_opt_table_entry_t be_blocksched_options[] = { - LC_OPT_ENT_ENUM_INT ("algo", "the block scheduling algorithm", &algo_var), + LC_OPT_ENT_ENUM_INT ("blockscheduler", "the block scheduling algorithm", &algo_var), LC_OPT_LAST }; @@ -102,28 +83,39 @@ static const lc_opt_table_entry_t be_blocksched_options[] = { * |___/ */ -typedef struct _blocksched_entry_t { - ir_node *block; - struct _blocksched_entry_t *next; - struct _blocksched_entry_t *prev; -} blocksched_entry_t; +typedef struct blocksched_entry_t blocksched_entry_t; +struct blocksched_entry_t { + ir_node *block; + blocksched_entry_t *next; + blocksched_entry_t *prev; +}; -typedef struct _edge_t { +typedef struct edge_t edge_t; +struct edge_t { ir_node *block; /**< source block */ int pos; /**< number of cfg predecessor (target) */ double execfreq; /**< the frequency */ - int highest_execfreq; /**< flag that indicates whether this edge is the edge with the highest - execfreq pointing away from this block */ -} edge_t; + double outedge_penalty_freq; /**< for edges leaving the loop this is the + penality when we make them a + fallthrough. */ + int highest_execfreq; /**< flag that indicates whether this edge is + the edge with the highest execfreq pointing + away from this block */ +}; -typedef struct _blocksched_env_t { +typedef struct blocksched_env_t blocksched_env_t; +struct blocksched_env_t { ir_graph *irg; - struct obstack *obst; - ir_exec_freq *execfreqs; + struct obstack obst; edge_t *edges; pdeq *worklist; int blockcount; -} blocksched_env_t; +}; + +static blocksched_entry_t* get_blocksched_entry(const ir_node *block) +{ + return (blocksched_entry_t*)get_irn_link(block); +} /** * Collect cfg frequencies of all edges between blocks. @@ -131,29 +123,43 @@ typedef struct _blocksched_env_t { */ static void collect_egde_frequency(ir_node *block, void *data) { - blocksched_env_t *env = data; + blocksched_env_t *env = (blocksched_env_t*)data; int arity; edge_t edge; blocksched_entry_t *entry; + ir_loop *loop; + + memset(&edge, 0, sizeof(edge)); - entry = obstack_alloc(env->obst, sizeof(entry[0])); + entry = OALLOCZ(&env->obst, blocksched_entry_t); entry->block = block; - entry->next = NULL; - entry->prev = NULL; set_irn_link(block, entry); + loop = get_irn_loop(block); + arity = get_Block_n_cfgpreds(block); if (arity == 0) { + /* must be the start block (or end-block for endless loops), + * everything else is dead code and should be removed by now */ assert(block == get_irg_start_block(env->irg) || block == get_irg_end_block(env->irg)); - /* must be the start block (or end-block for endless loops), nothing to - * do here */ + /* nothing to do here */ return; } else if (arity == 1) { + ir_node *pred_block = get_Block_cfgpred_block(block, 0); + ir_loop *pred_loop = get_irn_loop(pred_block); + float freq = (float)get_block_execfreq(block); + + /* is it an edge leaving a loop */ + if (get_loop_depth(pred_loop) > get_loop_depth(loop)) { + float pred_freq = (float)get_block_execfreq(pred_block); + edge.outedge_penalty_freq = -(pred_freq - freq); + } + edge.block = block; edge.pos = 0; - edge.execfreq = get_block_execfreq(env->execfreqs, block); + edge.execfreq = freq; edge.highest_execfreq = 1; ARR_APP1(edge_t, env->edges, edge); } else { @@ -166,7 +172,7 @@ static void collect_egde_frequency(ir_node *block, void *data) double execfreq; ir_node *pred_block = get_Block_cfgpred_block(block, i); - execfreq = get_block_execfreq(env->execfreqs, pred_block); + execfreq = get_block_execfreq(pred_block); edge.pos = i; edge.execfreq = execfreq; @@ -179,27 +185,86 @@ static void collect_egde_frequency(ir_node *block, void *data) } } - if(highest_edge_num >= 0) + if (highest_edge_num >= 0) env->edges[highest_edge_num].highest_execfreq = 1; } } +static int cmp_edges_base(const edge_t *e1, const edge_t *e2) +{ + long nr1 = get_irn_node_nr(e1->block); + long nr2 = get_irn_node_nr(e2->block); + if (nr1 < nr2) { + return 1; + } else if (nr1 > nr2) { + return -1; + } else { + if (e1->pos < e2->pos) { + return 1; + } else if (e1->pos > e2->pos) { + return -1; + } else { + return 0; + } + } +} + static int cmp_edges(const void *d1, const void *d2) { - const edge_t *e1 = d1; - const edge_t *e2 = d2; + const edge_t *e1 = (const edge_t*)d1; + const edge_t *e2 = (const edge_t*)d2; + double freq1 = e1->execfreq; + double freq2 = e2->execfreq; + if (freq1 < freq2) { + return 1; + } else if (freq1 > freq2) { + return -1; + } else { + return cmp_edges_base(e1, e2); + } +} - return QSORT_CMP(e2->execfreq, e1->execfreq); +static int cmp_edges_outedge_penalty(const void *d1, const void *d2) +{ + const edge_t *e1 = (const edge_t*)d1; + const edge_t *e2 = (const edge_t*)d2; + double pen1 = e1->outedge_penalty_freq; + double pen2 = e2->outedge_penalty_freq; + if (pen1 > pen2) { + return 1; + } else if (pen1 < pen2) { + return -1; + } else { + return cmp_edges_base(e1, e2); + } +} + +static void clear_loop_links(ir_loop *loop) +{ + int i, n; + + set_loop_link(loop, NULL); + n = get_loop_n_elements(loop); + for (i = 0; i < n; ++i) { + loop_element elem = get_loop_element(loop, i); + if (*elem.kind == k_ir_loop) { + clear_loop_links(elem.son); + } + } } static void coalesce_blocks(blocksched_env_t *env) { int i; int edge_count = ARR_LEN(env->edges); + edge_t *edges = env->edges; + + /* sort interblock edges by execution frequency */ + qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges); /* run1: only look at jumps */ for (i = 0; i < edge_count; ++i) { - const edge_t *edge = &env->edges[i]; + const edge_t *edge = &edges[i]; ir_node *block = edge->block; int pos = edge->pos; ir_node *pred_block; @@ -214,8 +279,8 @@ static void coalesce_blocks(blocksched_env_t *env) continue; pred_block = get_Block_cfgpred_block(block, pos); - entry = get_irn_link(block); - pred_entry = get_irn_link(pred_block); + entry = get_blocksched_entry(block); + pred_entry = get_blocksched_entry(pred_block); if (pred_entry->next != NULL || entry->prev != NULL) continue; @@ -225,38 +290,85 @@ static void coalesce_blocks(blocksched_env_t *env) continue; /* schedule the 2 blocks behind each other */ - DBG((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n", + DB((dbg, LEVEL_1, "Coalesce (Jump) %+F -> %+F (%.3g)\n", pred_entry->block, entry->block, edge->execfreq)); pred_entry->next = entry; entry->prev = pred_entry; } - /* run2: remaining edges */ + /* run2: pick loop fallthroughs */ + clear_loop_links(get_irg_loop(env->irg)); + + qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges_outedge_penalty); for (i = 0; i < edge_count; ++i) { - const edge_t *edge = &env->edges[i]; + const edge_t *edge = &edges[i]; ir_node *block = edge->block; int pos = edge->pos; ir_node *pred_block; blocksched_entry_t *entry, *pred_entry; + ir_loop *loop; + ir_loop *outer_loop; + + /* already seen all loop outedges? */ + if (edge->outedge_penalty_freq == 0) + break; /* the block might have been removed already... */ - if (is_Bad(get_Block_cfgpred(block, 0))) + if (is_Bad(get_Block_cfgpred(block, pos))) + continue; + + pred_block = get_Block_cfgpred_block(block, pos); + entry = get_blocksched_entry(block); + pred_entry = get_blocksched_entry(pred_block); + + if (pred_entry->next != NULL || entry->prev != NULL) continue; - /* we can't do fallthroughs in backedges */ - if (is_backedge(block, pos)) + /* we want at most 1 outedge fallthrough per loop */ + loop = get_irn_loop(pred_block); + if (get_loop_link(loop) != NULL) + continue; + + /* schedule the 2 blocks behind each other */ + DB((dbg, LEVEL_1, "Coalesce (Loop Outedge) %+F -> %+F (%.3g)\n", + pred_entry->block, entry->block, edge->execfreq)); + pred_entry->next = entry; + entry->prev = pred_entry; + + /* all loops left have an outedge now */ + outer_loop = get_irn_loop(block); + do { + /* we set loop link to loop to mark it */ + set_loop_link(loop, loop); + loop = get_loop_outer_loop(loop); + } while (loop != outer_loop); + } + + /* sort interblock edges by execution frequency */ + qsort(edges, ARR_LEN(edges), sizeof(edges[0]), cmp_edges); + + /* run3: remaining edges */ + for (i = 0; i < edge_count; ++i) { + const edge_t *edge = &edges[i]; + ir_node *block = edge->block; + int pos = edge->pos; + ir_node *pred_block; + blocksched_entry_t *entry, *pred_entry; + + /* the block might have been removed already... */ + if (is_Bad(get_Block_cfgpred(block, pos))) continue; pred_block = get_Block_cfgpred_block(block, pos); - entry = get_irn_link(block); - pred_entry = get_irn_link(pred_block); + entry = get_blocksched_entry(block); + pred_entry = get_blocksched_entry(pred_block); /* is 1 of the blocks already attached to another block? */ if (pred_entry->next != NULL || entry->prev != NULL) continue; /* schedule the 2 blocks behind each other */ - DBG((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n", + DB((dbg, LEVEL_1, "Coalesce (CondJump) %+F -> %+F (%.3g)\n", pred_entry->block, entry->block, edge->execfreq)); pred_entry->next = entry; entry->prev = pred_entry; @@ -268,15 +380,14 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en ir_node *block = entry->block; ir_node *succ = NULL; blocksched_entry_t *succ_entry; - const ir_edge_t *edge; - double best_succ_execfreq; + double best_succ_execfreq; if (irn_visited_else_mark(block)) return; env->blockcount++; - DBG((dbg, LEVEL_1, "Pick succ of %+F\n", block)); + DB((dbg, LEVEL_1, "Pick succ of %+F\n", block)); /* put all successors into the worklist */ foreach_block_succ(block, edge) { @@ -287,7 +398,7 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en /* we only need to put the first of a series of already connected * blocks into the worklist */ - succ_entry = get_irn_link(succ_block); + succ_entry = get_blocksched_entry(succ_block); while (succ_entry->prev != NULL) { /* break cycles... */ if (succ_entry->prev->block == succ_block) { @@ -296,12 +407,12 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en break; } succ_entry = succ_entry->prev; - }; + } if (irn_visited(succ_entry->block)) continue; - DBG((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block)); + DB((dbg, LEVEL_1, "Put %+F into worklist\n", succ_entry->block)); pdeq_putr(env->worklist, succ_entry->block); } @@ -310,7 +421,7 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en return; } - DBG((dbg, LEVEL_1, "deciding...\n")); + DB((dbg, LEVEL_1, "deciding...\n")); best_succ_execfreq = -1; /* no successor yet: pick the successor block with the highest execution @@ -318,16 +429,15 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en foreach_block_succ(block, edge) { ir_node *succ_block = get_edge_src_irn(edge); - double execfreq; if (irn_visited(succ_block)) continue; - succ_entry = get_irn_link(succ_block); + succ_entry = get_blocksched_entry(succ_block); if (succ_entry->prev != NULL) continue; - execfreq = get_block_execfreq(env->execfreqs, succ_block); + double execfreq = get_block_execfreq(succ_block); if (execfreq > best_succ_execfreq) { best_succ_execfreq = execfreq; succ = succ_block; @@ -335,18 +445,18 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en } if (succ == NULL) { - DBG((dbg, LEVEL_1, "pick from worklist\n")); + DB((dbg, LEVEL_1, "pick from worklist\n")); do { if (pdeq_empty(env->worklist)) { - DBG((dbg, LEVEL_1, "worklist empty\n")); + DB((dbg, LEVEL_1, "worklist empty\n")); return; } - succ = pdeq_getl(env->worklist); + succ = (ir_node*)pdeq_getl(env->worklist); } while (irn_visited(succ)); } - succ_entry = get_irn_link(succ); + succ_entry = get_blocksched_entry(succ); entry->next = succ_entry; succ_entry->prev = entry; @@ -357,7 +467,7 @@ static blocksched_entry_t *finish_block_schedule(blocksched_env_t *env) { ir_graph *irg = env->irg; ir_node *startblock = get_irg_start_block(irg); - blocksched_entry_t *entry = get_irn_link(startblock); + blocksched_entry_t *entry = get_blocksched_entry(startblock); ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); inc_irg_visited(irg); @@ -381,50 +491,46 @@ static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry (void) env; block_list = NEW_ARR_D(ir_node *, obst, count); - DBG((dbg, LEVEL_1, "Blockschedule:\n")); + DB((dbg, LEVEL_1, "Blockschedule:\n")); for (entry = first; entry != NULL; entry = entry->next) { assert(i < count); block_list[i++] = entry->block; - DBG((dbg, LEVEL_1, "\t%+F\n", entry->block)); + DB((dbg, LEVEL_1, "\t%+F\n", entry->block)); } assert(i == count); return block_list; } -static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execfreqs) +static ir_node **create_block_schedule_greedy(ir_graph *irg) { blocksched_env_t env; - struct obstack obst; blocksched_entry_t *start_entry; ir_node **block_list; - obstack_init(&obst); - env.irg = irg; - env.obst = &obst; - env.execfreqs = execfreqs; env.edges = NEW_ARR_F(edge_t, 0); env.worklist = NULL; env.blockcount = 0; + obstack_init(&env.obst); + + assure_loopinfo(irg); // collect edge execution frequencies irg_block_walk_graph(irg, collect_egde_frequency, NULL, &env); - // sort interblock edges by execution frequency - qsort(env.edges, ARR_LEN(env.edges), sizeof(env.edges[0]), cmp_edges); - (void)be_remove_empty_blocks(irg); if (algo != BLOCKSCHED_NAIV) coalesce_blocks(&env); start_entry = finish_block_schedule(&env); - block_list = create_blocksched_array(&env, start_entry, env.blockcount, get_irg_obstack(irg)); + block_list = create_blocksched_array(&env, start_entry, env.blockcount, + be_get_be_obst(irg)); DEL_ARR_F(env.edges); - obstack_free(&obst, NULL); + obstack_free(&env.obst, NULL); return block_list; } @@ -438,23 +544,22 @@ static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execf * */ -#ifdef WITH_ILP -typedef struct _ilp_edge_t { +typedef struct ilp_edge_t { ir_node *block; /**< source block */ int pos; /**< number of cfg predecessor (target) */ int ilpvar; } ilp_edge_t; -typedef struct _blocksched_ilp_env_t { +typedef struct blocksched_ilp_env_t { blocksched_env_t env; ilp_edge_t *ilpedges; lpp_t *lpp; } blocksched_ilp_env_t; -typedef struct _blocksched_ilp_entry_t { +typedef struct blocksched_ilp_entry_t { ir_node *block; - struct _blocksched_entry_t *next; - struct _blocksched_entry_t *prev; + struct blocksched_entry_t *next; + struct blocksched_entry_t *prev; int out_cst; } blocksched_ilp_entry_t; @@ -477,11 +582,10 @@ static int add_ilp_edge(ir_node *block, int pos, double execfreq, blocksched_ilp static void collect_egde_frequency_ilp(ir_node *block, void *data) { - blocksched_ilp_env_t *env = data; + blocksched_ilp_env_t *env = (blocksched_ilp_env_t*)data; ir_graph *irg = env->env.irg; ir_node *startblock = get_irg_start_block(irg); int arity; - lpp_cst_t cst; char name[64]; int out_count; blocksched_ilp_entry_t *entry; @@ -489,11 +593,11 @@ static void collect_egde_frequency_ilp(ir_node *block, void *data) snprintf(name, sizeof(name), "block_out_constr_%ld", get_irn_node_nr(block)); out_count = get_irn_n_edges_kind(block, EDGE_KIND_BLOCK); - entry = obstack_alloc(env->env.obst, sizeof(entry[0])); + entry = OALLOC(&env->env.obst, blocksched_ilp_entry_t); entry->block = block; entry->next = NULL; entry->prev = NULL; - entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, out_count - 1); + entry->out_cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, out_count - 1); set_irn_link(block, entry); if (block == startblock) @@ -501,14 +605,15 @@ static void collect_egde_frequency_ilp(ir_node *block, void *data) arity = get_irn_arity(block); if (arity == 1) { - double execfreq = get_block_execfreq(env->env.execfreqs, block); + double execfreq = get_block_execfreq(block); add_ilp_edge(block, 0, execfreq, env); } else { int i; + int cst_idx; snprintf(name, sizeof(name), "block_in_constr_%ld", get_irn_node_nr(block)); - cst = lpp_add_cst_uniq(env->lpp, name, lpp_greater, arity - 1); + cst_idx = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1); for (i = 0; i < arity; ++i) { double execfreq; @@ -516,22 +621,26 @@ static void collect_egde_frequency_ilp(ir_node *block, void *data) ilp_edge_t *edge; ir_node *pred_block = get_Block_cfgpred_block(block, i); - execfreq = get_block_execfreq(env->env.execfreqs, pred_block); + execfreq = get_block_execfreq(pred_block); edgenum = add_ilp_edge(block, i, execfreq, env); edge = &env->ilpedges[edgenum]; - lpp_set_factor_fast(env->lpp, cst, edge->ilpvar, 1.0); + lpp_set_factor_fast(env->lpp, cst_idx, edge->ilpvar, 1.0); } } } +static blocksched_ilp_entry_t *get_blocksched_ilp_entry(const ir_node *block) +{ + return (blocksched_ilp_entry_t*)get_irn_link(block); +} static void coalesce_blocks_ilp(blocksched_ilp_env_t *env) { - int i; - int edge_count = ARR_LEN(env->ilpedges); + int edge_count = ARR_LEN(env->ilpedges); + int i; /* complete out constraints */ - for(i = 0; i < edge_count; ++i) { + for (i = 0; i < edge_count; ++i) { const ilp_edge_t *edge = &env->ilpedges[i]; ir_node *block = edge->block; ir_node *pred; @@ -542,27 +651,14 @@ static void coalesce_blocks_ilp(blocksched_ilp_env_t *env) continue; pred = get_Block_cfgpred_block(block, edge->pos); - entry = get_irn_link(pred); + entry = get_blocksched_ilp_entry(pred); - DBG((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n", + DB((dbg, LEVEL_1, "Adding out cst to %+F from %+F,%d\n", pred, block, edge->pos)); lpp_set_factor_fast(env->lpp, entry->out_cst, edge->ilpvar, 1.0); } -#if 0 - { - FILE *f; - char fname[256]; - lpp_dump(env->lpp, "lpp.out"); - snprintf(fname, sizeof(fname), "lpp_%s.plain", get_irg_dump_name(env->env.irg)); - f = fopen(fname, "w"); - lpp_dump_plain(env->lpp, f); - fclose(f); - } -#endif - - //lpp_solve_net(env->lpp, main_env->options->ilp_server, main_env->options->ilp_solver); - lpp_solve_net(env->lpp, "i44pc52", "cplex"); + lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver); assert(lpp_is_sol_valid(env->lpp)); /* Apply results to edges */ @@ -583,8 +679,8 @@ static void coalesce_blocks_ilp(blocksched_ilp_env_t *env) continue; pred = get_Block_cfgpred_block(block, edge->pos); - entry = get_irn_link(block); - pred_entry = get_irn_link(pred); + entry = get_blocksched_entry(block); + pred_entry = get_blocksched_entry(pred); assert(entry->prev == NULL && pred_entry->next == NULL); entry->prev = pred_entry; @@ -592,23 +688,19 @@ static void coalesce_blocks_ilp(blocksched_ilp_env_t *env) } } -static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreqs) +static ir_node **create_block_schedule_ilp(ir_graph *irg) { blocksched_ilp_env_t env; - struct obstack obst; blocksched_entry_t *start_entry; ir_node **block_list; - obstack_init(&obst); - env.env.irg = irg; - env.env.obst = &obst; - env.env.execfreqs = execfreqs; env.env.worklist = NULL; env.env.blockcount = 0; env.ilpedges = NEW_ARR_F(ilp_edge_t, 0); + obstack_init(&env.env.obst); - env.lpp = new_lpp("blockschedule", lpp_minimize); + env.lpp = lpp_new("blockschedule", lpp_minimize); lpp_set_time_limit(env.lpp, 20); lpp_set_log(env.lpp, stdout); @@ -618,114 +710,16 @@ static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreq coalesce_blocks_ilp(&env); start_entry = finish_block_schedule(&env.env); - block_list = create_blocksched_array(&env.env, start_entry, env.env.blockcount, get_irg_obstack(irg)); + block_list = create_blocksched_array(&env.env, start_entry, + env.env.blockcount, + be_get_be_obst(irg)); DEL_ARR_F(env.ilpedges); - free_lpp(env.lpp); - obstack_free(&obst, NULL); + lpp_free(env.lpp); + obstack_free(&env.env.obst, NULL); return block_list; } -#endif /* WITH_ILP */ - -/* - * _____ _ ____ ____ - * | ____|_ _| |_| __ )| __ ) - * | _| \ \/ / __| _ \| _ \ - * | |___ > <| |_| |_) | |_) | - * |_____/_/\_\\__|____/|____/ - * - */ - -/** A simple forward single linked list. */ -typedef struct { - ir_node *start; /**< start of the list */ - ir_node *end; /**< last block in the list */ - unsigned n_blks; /**< number of blocks in the list */ -} anchor; - -static void add_block(anchor *list, ir_node *block) { - if (list->start == NULL) { - list->start = block; - list->end = block; - } else { - set_irn_link(list->end, block); - list->end = block; - } - - list->n_blks++; -} - -static void create_block_list(ir_node *leader_block, anchor *list) { - int i; - const ir_edge_t *edge; - ir_node *block = NULL; - ir_extblk *extbb = get_Block_extbb(leader_block); - - if (extbb_visited(extbb)) - return; - mark_extbb_visited(extbb); - - for (i = 0; i < get_extbb_n_blocks(extbb); ++i) { - block = get_extbb_block(extbb, i); - add_block(list, block); - } - - assert(block != NULL); - - /* pick successor extbbs */ - foreach_block_succ(block, edge) { - ir_node *succ = get_edge_src_irn(edge); - create_block_list(succ, list); - } - - for (i = 0; i < get_extbb_n_blocks(extbb) - 1; ++i) { - block = get_extbb_block(extbb, i); - - foreach_block_succ(block, edge) { - ir_node *succ = get_edge_src_irn(edge); - create_block_list(succ, list); - } - } -} - -void compute_extbb_execfreqs(ir_graph *irg, ir_exec_freq *execfreqs); - -/* - * Calculates a block schedule. The schedule is stored as a linked - * list starting at the start_block of the irg. - */ -static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs) -{ - anchor list; - ir_node **blk_list, *b, *n; - unsigned i; - - /* schedule extended basic blocks */ - compute_extbb_execfreqs(irg, execfreqs); - //compute_extbb(irg); - - list.start = NULL; - list.end = NULL; - list.n_blks = 0; - - ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK); - inc_irg_block_visited(irg); - - create_block_list(get_irg_start_block(irg), &list); - - /** create an array, so we can go forward and backward */ - blk_list = NEW_ARR_D(ir_node *, irg->obst,list.n_blks); - - for (i = 0, b = list.start; b; b = n, ++i) { - n = get_irn_link(b); - blk_list[i] = b; - } - - ir_free_resources(irg, IR_RESOURCE_IRN_VISITED | IR_RESOURCE_IRN_LINK); - - return blk_list; -} /* * __ __ _ @@ -735,32 +729,25 @@ static ir_node **create_extbb_block_schedule(ir_graph *irg, ir_exec_freq *execfr * |_| |_|\__,_|_|_| |_| * */ +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched) void be_init_blocksched(void) { lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); - lc_opt_entry_t *blocksched_grp = lc_opt_get_grp(be_grp, "blocksched"); - lc_opt_add_table(blocksched_grp, be_blocksched_options); + lc_opt_add_table(be_grp, be_blocksched_options); FIRM_DBG_REGISTER(dbg, "firm.be.blocksched"); } -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_blocksched); - -ir_node **be_create_block_schedule(ir_graph *irg, ir_exec_freq *execfreqs) +ir_node **be_create_block_schedule(ir_graph *irg) { - switch(algo) { + switch (algo) { case BLOCKSCHED_GREEDY: case BLOCKSCHED_NAIV: - return create_block_schedule_greedy(irg, execfreqs); - case BLOCKSCHED_EXTBB: - return create_extbb_block_schedule(irg, execfreqs); -#ifdef WITH_ILP + return create_block_schedule_greedy(irg); case BLOCKSCHED_ILP: - return create_block_schedule_ilp(irg, execfreqs); -#endif /* WITH_ILP */ + return create_block_schedule_ilp(irg); } panic("unknown blocksched algo"); - return NULL; }