X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbeblocksched.c;h=d1b58e500383d8b92e85c459a6fba1af79cc0849;hb=2be6077d3c0d1cdc8fad965f1857ac3a08c632aa;hp=4e76c02eecf0eed3b59fe1e0bded38e76316a2e6;hpb=3bb47464dc990c9def630a208a12da1bdd035bbb;p=libfirm diff --git a/ir/be/beblocksched.c b/ir/be/beblocksched.c index 4e76c02ee..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" @@ -120,13 +106,17 @@ struct edge_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; }; +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. * Also determines edge with highest frequency. @@ -141,7 +131,7 @@ static void collect_egde_frequency(ir_node *block, void *data) memset(&edge, 0, sizeof(edge)); - entry = OALLOCZ(env->obst, blocksched_entry_t); + entry = OALLOCZ(&env->obst, blocksched_entry_t); entry->block = block; set_irn_link(block, entry); @@ -159,11 +149,11 @@ static void collect_egde_frequency(ir_node *block, void *data) } 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(env->execfreqs, 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(env->execfreqs, pred_block); + float pred_freq = (float)get_block_execfreq(pred_block); edge.outedge_penalty_freq = -(pred_freq - freq); } @@ -182,7 +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; @@ -200,20 +190,53 @@ static void collect_egde_frequency(ir_node *block, void *data) } } +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 = (const edge_t*)d1; const edge_t *e2 = (const edge_t*)d2; - - return QSORT_CMP(e2->execfreq, e1->execfreq); + 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); + } } 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; - /* reverse sorting as penalties are negative */ - return QSORT_CMP(e1->outedge_penalty_freq, e2->outedge_penalty_freq); + 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) @@ -256,8 +279,8 @@ static void coalesce_blocks(blocksched_env_t *env) continue; pred_block = get_Block_cfgpred_block(block, pos); - entry = (blocksched_entry_t*)get_irn_link(block); - pred_entry = (blocksched_entry_t*)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; @@ -295,8 +318,8 @@ static void coalesce_blocks(blocksched_env_t *env) continue; pred_block = get_Block_cfgpred_block(block, pos); - entry = (blocksched_entry_t*)get_irn_link(block); - pred_entry = (blocksched_entry_t*)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; @@ -337,8 +360,8 @@ static void coalesce_blocks(blocksched_env_t *env) continue; pred_block = get_Block_cfgpred_block(block, pos); - entry = (blocksched_entry_t*)get_irn_link(block); - pred_entry = (blocksched_entry_t*)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) @@ -357,8 +380,7 @@ 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; @@ -376,7 +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 = (blocksched_entry_t*)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) { @@ -407,16 +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 = (blocksched_entry_t*)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; @@ -435,7 +456,7 @@ static void pick_block_successor(blocksched_entry_t *entry, blocksched_env_t *en } while (irn_visited(succ)); } - succ_entry = (blocksched_entry_t*)get_irn_link(succ); + succ_entry = get_blocksched_entry(succ); entry->next = succ_entry; succ_entry->prev = entry; @@ -446,7 +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 = (blocksched_entry_t*)get_irn_link(startblock); + blocksched_entry_t *entry = get_blocksched_entry(startblock); ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); inc_irg_visited(irg); @@ -482,21 +503,17 @@ static ir_node **create_blocksched_array(blocksched_env_t *env, blocksched_entry 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); @@ -513,7 +530,7 @@ static ir_node **create_block_schedule_greedy(ir_graph *irg, ir_exec_freq *execf be_get_be_obst(irg)); DEL_ARR_F(env.edges); - obstack_free(&obst, NULL); + obstack_free(&env.obst, NULL); return block_list; } @@ -565,11 +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; @@ -577,7 +593,7 @@ 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 = OALLOC(env->env.obst, blocksched_ilp_entry_t); + entry = OALLOC(&env->env.obst, blocksched_ilp_entry_t); entry->block = block; entry->next = NULL; entry->prev = NULL; @@ -589,14 +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_equal, arity - 1); + cst_idx = lpp_add_cst_uniq(env->lpp, name, lpp_greater_equal, arity - 1); for (i = 0; i < arity; ++i) { double execfreq; @@ -604,19 +621,22 @@ 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 edge_count = ARR_LEN(env->ilpedges); - be_options_t *options = be_get_irg_options(env->env.irg); int i; /* complete out constraints */ @@ -631,14 +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); 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); } - lpp_solve_net(env->lpp, options->ilp_server, options->ilp_solver); + lpp_solve_net(env->lpp, be_options.ilp_server, be_options.ilp_solver); assert(lpp_is_sol_valid(env->lpp)); /* Apply results to edges */ @@ -659,8 +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; @@ -668,21 +688,17 @@ 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 = lpp_new("blockschedule", lpp_minimize); lpp_set_time_limit(env.lpp, 20); @@ -700,7 +716,7 @@ static ir_node **create_block_schedule_ilp(ir_graph *irg, ir_exec_freq *execfreq DEL_ARR_F(env.ilpedges); lpp_free(env.lpp); - obstack_free(&obst, NULL); + obstack_free(&env.env.obst, NULL); return block_list; } @@ -725,14 +741,12 @@ void be_init_blocksched(void) ir_node **be_create_block_schedule(ir_graph *irg) { - ir_exec_freq *execfreqs = be_get_irg_exec_freq(irg); - switch (algo) { case BLOCKSCHED_GREEDY: case BLOCKSCHED_NAIV: - return create_block_schedule_greedy(irg, execfreqs); + return create_block_schedule_greedy(irg); case BLOCKSCHED_ILP: - return create_block_schedule_ilp(irg, execfreqs); + return create_block_schedule_ilp(irg); } panic("unknown blocksched algo");