X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=948bbe46d0048a120e6209b475ad1821830519ba;hb=501a13c6429ecb1de1cbb274782e67fa0030b970;hp=e8fdf8a143e3b955c03975c4f7b82b7722df8456;hpb=05413b49d91e03ba18a82a0ee55b358184e1afc9;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index e8fdf8a14..948bbe46d 100644 --- a/ir/be/bespillbelady.c +++ b/ir/be/bespillbelady.c @@ -1,22 +1,33 @@ -/** - * Author: Daniel Grund - * Date: 20.09.2005 - * Copyright: (c) Universitaet Karlsruhe - * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. +/* + * 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. + */ + +/** + * @file + * @brief Beladys spillalgorithm. + * @author Daniel Grund, Matthias Braun + * @date 20.09.2005 + * @version $Id$ */ #ifdef HAVE_CONFIG_H #include "config.h" #endif -#ifdef HAVE_ALLOCA_H -#include -#endif - -#ifdef HAVE_MALLOC_H -#include -#endif - #include "obst.h" #include "set.h" #include "pset.h" @@ -25,19 +36,25 @@ #include "irnode.h" #include "irmode.h" #include "irgwalk.h" +#include "irloop.h" #include "iredges_t.h" #include "ircons_t.h" #include "irprintf.h" +#include "xmalloc.h" #include "beutil.h" -#include "bearch.h" +#include "bearch_t.h" #include "bespillbelady.h" -#include "beuses_t.h" +#include "beuses.h" #include "besched_t.h" #include "beirgmod.h" #include "belive_t.h" #include "benode_t.h" #include "bechordal_t.h" +#include "bespilloptions.h" +#include "beloopana.h" +#include "beirg_t.h" +#include "bemodule.h" #define DBG_SPILL 1 #define DBG_WSETS 2 @@ -47,16 +64,27 @@ #define DBG_SLOTS 32 #define DBG_TRACE 64 #define DBG_WORKSET 128 -#define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL) DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) -typedef struct _workset_t workset_t; +/** + * An association between a node and a point in time. + */ +typedef struct _loc_t { + ir_node *irn; /**< A node. */ + unsigned time; /**< A use time (see beuses.h). */ +} loc_t; + +typedef struct _workset_t { + int len; /**< current length */ + loc_t vals[0]; /**< inlined array of the values/distances in this working set */ +} workset_t; typedef struct _belady_env_t { struct obstack ob; - const be_chordal_env_t *cenv; const arch_env_t *arch; const arch_register_class_t *cls; + const be_lv_t *lv; + be_loopana_t *loop_ana; int n_regs; /** number of regs in this reg-class */ workset_t *ws; /**< the main workset used while processing a block. ob-allocated */ @@ -68,12 +96,14 @@ typedef struct _belady_env_t { spill_env_t *senv; /**< see bespill.h */ } belady_env_t; -struct _workset_t { - int len; /**< current length */ - loc_t vals[0]; /**< inlined array of the values/distances in this working set */ -}; +static int loc_compare(const void *a, const void *b) +{ + const loc_t *p = a; + const loc_t *q = b; + return p->time - q->time; +} -void workset_print(const workset_t *w) +static INLINE void workset_print(const workset_t *w) { int i; @@ -132,7 +162,7 @@ static INLINE void workset_insert(belady_env_t *env, workset_t *ws, ir_node *val int i; /* check for current regclass */ if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) { - DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val)); + DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val)); return; } @@ -208,64 +238,28 @@ static INLINE void *new_block_info(struct obstack *ob) { return res; } -#define get_block_info(blk) ((block_info_t *)get_irn_link(blk)) -#define set_block_info(blk, info) set_irn_link(blk, info) +#define get_block_info(block) ((block_info_t *)get_irn_link(block)) +#define set_block_info(block, info) set_irn_link(block, info) /** * @return The distance to the next use or 0 if irn has dont_spill flag set */ -static INLINE unsigned get_distance(belady_env_t *env, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses) +static INLINE unsigned get_distance(belady_env_t *env, ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses) { + be_next_use_t use; int flags = arch_irn_get_flags(env->arch, def); - unsigned dist = be_get_next_use(env->uses, from, from_step, def, skip_from_uses); assert(! (flags & arch_irn_flags_ignore)); - /* we have to keep nonspillable nodes in the workingset */ + use = be_get_next_use(env->uses, from, from_step, def, skip_from_uses); + if(USES_IS_INFINITE(use.time)) + return USES_INFINITY; + + /* We have to keep nonspillable nodes in the workingset */ if(flags & arch_irn_flags_dont_spill) return 0; - return dist; -} - -/** - * Fix to remove dead nodes (especially don't spill nodes) from workset. - */ -static void fix_dead_values(workset_t *ws, ir_node *irn) { - int idx; - ir_node *node; - ir_node *block = get_nodes_block(irn); - - DBG((dbg, DBG_DECIDE, "fixing dead values at %+F:\n", irn)); - - workset_foreach(ws, node, idx) { - const ir_edge_t *edge; - int fixme = 1; - - /* skip already fixed nodes */ - if (workset_get_time(ws, idx) == INT_MAX) - continue; - - /* check all users */ - foreach_out_edge(node, edge) { - ir_node *user = get_edge_src_irn(edge); - - if ((get_nodes_block(user) != block) || /* user is in a different block */ - (sched_is_scheduled(user) && sched_comes_after(irn, user)) || /* user is scheduled after irn */ - user == irn) /* irn is the user */ - { /* => don't fix distance */ - fixme = 0; - break; - } - } - - /* all users scheduled prior to current irn in in same block as irn -> fix */ - if (fixme) { - workset_set_time(ws, idx, INT_MAX); - DBG((dbg, DBG_DECIDE, "\tfixing time for %+F to INT_MAX\n", node)); - } - } - + return use.time; } /** @@ -296,8 +290,10 @@ static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) { if (! workset_contains(ws, val)) { DBG((dbg, DBG_DECIDE, " insert %+F\n", val)); to_insert[demand++] = val; - if (is_usage) - be_add_reload(env->senv, val, env->instr); + if (is_usage) { + DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr)); + be_add_reload(env->senv, val, env->instr, env->cls, 1); + } } else { assert(is_usage || "Defined value already in workset?!?"); @@ -317,19 +313,10 @@ static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) { /* Only make more free room if we do not have enough */ if (len > max_allowed) { /* get current next-use distance */ - for (i = 0; i < ws->len; ++i) - workset_set_time(ws, i, get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage)); - - /* - FIX for don't spill nodes: - Problem is that get_distance always returns 0 for those nodes even if they are not - needed anymore (all their usages have already been visited). - Even if we change this behavior, get_distance doesn't distinguish between not - used anymore (dead) and live out of block. - Solution: Set distances of all nodes having all their usages in schedule prior to - current instruction to MAX_INT. - */ - fix_dead_values(ws, env->instr); + for (i = 0; i < ws->len; ++i) { + unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage); + workset_set_time(ws, i, dist); + } /* sort entries by increasing nextuse-distance*/ workset_sort(ws); @@ -342,6 +329,8 @@ static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) { for (i = max_allowed; i < ws->len; ++i) { ir_node *irn = ws->vals[i].irn; + DBG((dbg, DBG_DECIDE, " disposing %+F (%u)\n", irn, workset_get_time(ws, i))); + if (is_Phi(irn)) continue; @@ -350,10 +339,7 @@ static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) { workset_t *ws_start = get_block_info(curr_bb)->ws_start; workset_remove(ws_start, irn); - DBG((dbg, DBG_DECIDE, " dispose %+F dumb\n", irn)); - } - else { - DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn)); + DBG((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", irn)); } } @@ -368,20 +354,76 @@ static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) { workset_insert(env, env->ws, to_insert[i]); } -static void belady(ir_node *blk, void *env); +static void belady(ir_node *block, void *env); + +/** Decides whether a specific node should be in the start workset or not + * + * @param env belady environment + * @param first + * @param node the node to test + * @param block the block of the node + * @param loop the loop of the node + */ +static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first, + ir_node *node, ir_node *block, + ir_loop *loop) +{ + be_next_use_t next_use; + loc_t loc; + loc.time = USES_INFINITY; + loc.irn = node; + + if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) { + loc.time = USES_INFINITY; + return loc; + } + + /* We have to keep nonspillable nodes in the workingset */ + if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) { + loc.time = 0; + DBG((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time)); + return loc; + } + + next_use = be_get_next_use(env->uses, first, 0, node, 0); + if(USES_IS_INFINITE(next_use.time)) { + // the nodes marked as live in shouldn't be dead, so it must be a phi + assert(is_Phi(node)); + loc.time = USES_INFINITY; + DBG((dbg, DBG_START, " %+F not taken (dead)\n", node)); + if(is_Phi(node)) { + be_spill_phi(env->senv, node); + } + return loc; + } + + loc.time = next_use.time; + + if(next_use.outermost_loop >= get_loop_depth(loop)) { + DBG((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop)); + } else { + loc.time = USES_PENDING; + DBG((dbg, DBG_START, " %+F delayed (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop))); + } + return loc; +} /* * Computes set of live-ins for each block with multiple predecessors * and notifies spill algorithm which phis need to be spilled */ -static void spill_phi_walker(ir_node *block, void *data) { - belady_env_t *env = data; - block_info_t *block_info; - ir_node *first, *irn; - loc_t loc, *starters; - int i, len, ws_count; - - if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block) +static void compute_live_ins(ir_node *block, void *data) { + belady_env_t *env = data; + ir_loop *loop = get_irn_loop(block); + const be_lv_t *lv = env->lv; + block_info_t *block_info; + ir_node *first, *irn; + loc_t loc, *starters, *delayed; + int i, len, ws_count; + int free_slots, free_pressure_slots; + unsigned pressure; + + if (get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block) return; block_info = new_block_info(&env->ob); @@ -389,47 +431,75 @@ static void spill_phi_walker(ir_node *block, void *data) { /* Collect all values living at start of block */ starters = NEW_ARR_F(loc_t, 0); - - /* rebuild schedule time information, because it seems to be broken */ - sched_renumber(block); + delayed = NEW_ARR_F(loc_t, 0); DBG((dbg, DBG_START, "Living at start of %+F:\n", block)); first = sched_first(block); + + /* check all Phis first */ sched_foreach(block, irn) { - if(!is_Phi(irn)) + if (! is_Phi(irn)) break; - if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn)) - continue; - loc.irn = irn; - loc.time = get_distance(env, first, 0, irn, 0); - ARR_APP1(loc_t, starters, loc); - DBG((dbg, DBG_START, " %+F:\n", irn)); + loc = to_take_or_not_to_take(env, first, irn, block, loop); + + if (! USES_IS_INFINITE(loc.time)) { + if (USES_IS_PENDING(loc.time)) + ARR_APP1(loc_t, delayed, loc); + else + ARR_APP1(loc_t, starters, loc); + } } - be_lv_foreach(env->cenv->lv, block, be_lv_state_in, i) { - ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i); - if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn)) - continue; + /* check all Live-Ins */ + be_lv_foreach(lv, block, be_lv_state_in, i) { + ir_node *node = be_lv_get_irn(lv, block, i); - loc.irn = irn; - loc.time = get_distance(env, first, 0, irn, 0); - ARR_APP1(loc_t, starters, loc); - DBG((dbg, DBG_START, " %+F:\n", irn)); + loc = to_take_or_not_to_take(env, first, node, block, loop); + + if (! USES_IS_INFINITE(loc.time)) { + if (USES_IS_PENDING(loc.time)) + ARR_APP1(loc_t, delayed, loc); + else + ARR_APP1(loc_t, starters, loc); + } + } + + pressure = be_get_loop_pressure(env->loop_ana, env->cls, loop); + assert(ARR_LEN(delayed) <= (signed)pressure); + free_slots = env->n_regs - ARR_LEN(starters); + free_pressure_slots = env->n_regs - (pressure - ARR_LEN(delayed)); + free_slots = MIN(free_slots, free_pressure_slots); + /* append nodes delayed due to loop structure until start set is full */ + for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) { + DBG((dbg, DBG_START, " delayed %+F taken\n", delayed[i].irn)); + ARR_APP1(loc_t, starters, delayed[i]); + delayed[i].irn = NULL; + } + + /* spill all delayed phis which didn't make it into start workset */ + for (i = ARR_LEN(delayed) - 1; i >= 0; --i) { + ir_node *irn = delayed[i].irn; + if (irn && is_Phi(irn)) { + DBG((dbg, DBG_START, " spilling delayed phi %+F\n", irn)); + be_spill_phi(env->senv, irn); + } } + DEL_ARR_F(delayed); - // Sort start values by first use + /* Sort start values by first use */ qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare); /* Copy the best ones from starters to start workset */ - ws_count = MIN(ARR_LEN(starters), env->n_regs); + ws_count = MIN(ARR_LEN(starters), env->n_regs); block_info->ws_start = new_workset(env, &env->ob); workset_bulk_fill(block_info->ws_start, ws_count, starters); /* The phis of this block which are not in the start set have to be spilled later. */ - for (i = ws_count, len = ARR_LEN(starters); i < len; ++i) { + len = ARR_LEN(starters); + for (i = ws_count; i < len; ++i) { irn = starters[i].irn; - if (!is_Phi(irn) || get_nodes_block(irn) != block) + if (! is_Phi(irn) || get_nodes_block(irn) != block) continue; be_spill_phi(env->senv, irn); @@ -439,7 +509,7 @@ static void spill_phi_walker(ir_node *block, void *data) { } /** - * Collects all values live-in at block @p blk and all phi results in this block. + * Collects all values live-in at block @p block and all phi results in this block. * Then it adds the best values (at most n_regs) to the blocks start_workset. * The phis among the remaining values get spilled: Introduce psudo-copies of * their args to break interference and make it possible to spill them to the @@ -478,7 +548,7 @@ static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) /** - * For the given block @p blk, decide for each values + * For the given block @p block, decide for each values * whether it is used from a register or is reloaded * before the use. */ @@ -505,7 +575,7 @@ static void belady(ir_node *block, void *data) { workset_copy(env, env->ws, block_info->ws_start); DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", block)); workset_foreach(env->ws, irn, iter) - DBG((dbg, DBG_WSETS, " %+F\n", irn)); + DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(env->ws, iter))); /* process the block from start to end */ DBG((dbg, DBG_WSETS, "Processing...\n")); @@ -555,7 +625,7 @@ static void belady(ir_node *block, void *data) { block_info->processed = 1; DBG((dbg, DBG_WSETS, "End workset for %+F:\n", block)); workset_foreach(block_info->ws_end, irn, iter) - DBG((dbg, DBG_WSETS, " %+F\n", irn)); + DBG((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(block_info->ws_end, iter))); } /** @@ -563,19 +633,24 @@ static void belady(ir_node *block, void *data) { * about the set of live-ins. Thus we must adapt the * live-outs to the live-ins at each block-border. */ -static void fix_block_borders(ir_node *blk, void *data) { +static void fix_block_borders(ir_node *block, void *data) { belady_env_t *env = data; workset_t *wsb; + ir_graph *irg = get_irn_irg(block); + ir_node *startblock = get_irg_start_block(irg); int i, max, iter, iter2; + if(block == startblock) + return; + DBG((dbg, DBG_FIX, "\n")); - DBG((dbg, DBG_FIX, "Fixing %+F\n", blk)); + DBG((dbg, DBG_FIX, "Fixing %+F\n", block)); - wsb = get_block_info(blk)->ws_start; + wsb = get_block_info(block)->ws_start; /* process all pred blocks */ - for (i=0, max=get_irn_arity(blk); iws_end; DBG((dbg, DBG_FIX, " Pred %+F\n", pred)); @@ -583,7 +658,7 @@ static void fix_block_borders(ir_node *blk, void *data) { workset_foreach(wsb, irnb, iter) { /* if irnb is a phi of the current block we reload * the corresponding argument, else irnb itself */ - if(is_Phi(irnb) && blk == get_nodes_block(irnb)) { + if(is_Phi(irnb) && block == get_nodes_block(irnb)) { irnb = get_irn_n(irnb, i); // we might have unknowns as argument for the phi @@ -603,7 +678,8 @@ static void fix_block_borders(ir_node *blk, void *data) { /* irnb is not in memory at the end of pred, so we have to reload it */ DBG((dbg, DBG_FIX, " reload %+F\n", irnb)); - be_add_reload_on_edge(env->senv, irnb, blk, i); + DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", irnb, block, i)); + be_add_reload_on_edge(env->senv, irnb, block, i, env->cls, 1); next_value: /*epsilon statement :)*/; @@ -611,40 +687,58 @@ next_value: } } -void be_spill_belady(const be_chordal_env_t *chordal_env) { - be_spill_belady_spill_env(chordal_env, NULL); +/** + * Do spilling for a register class on a graph using the belady heuristic. + * In the transformed graph, the register pressure never exceeds the number + * of available registers. + * + * @param birg The backend graph + * @param cls The register class to spill + */ +static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) { + be_spill_belady_spill_env(birg, cls, NULL); } -void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) { +void be_spill_belady_spill_env(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) { belady_env_t env; + ir_graph *irg = be_get_birg_irg(birg); + int n_regs; - FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady"); - //firm_dbg_set_mask(dbg, DBG_START); + /* some special classes contain only ignore regs, nothing to do then */ + n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL); + if(n_regs == 0) + return; + + be_invalidate_liveness(birg); + be_assure_liveness(birg); + /* construct control flow loop tree */ + if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) { + construct_cf_backedges(irg); + } + + be_clear_links(irg); /* init belady env */ obstack_init(&env.ob); - env.cenv = chordal_env; - env.arch = chordal_env->birg->main_env->arch_env; - env.cls = chordal_env->cls; - env.n_regs = arch_count_non_ignore_regs(env.arch, env.cls); + env.arch = birg->main_env->arch_env; + env.cls = cls; + env.lv = be_get_birg_liveness(birg); + env.n_regs = n_regs; env.ws = new_workset(&env, &env.ob); - env.uses = be_begin_uses(chordal_env->irg, chordal_env->lv, chordal_env->birg->main_env->arch_env, env.cls); + env.uses = be_begin_uses(irg, env.lv); + env.loop_ana = be_new_loop_pressure(birg); if(spill_env == NULL) { - env.senv = be_new_spill_env(chordal_env); + env.senv = be_new_spill_env(birg); } else { env.senv = spill_env; } - DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);) - - DBG((dbg, LEVEL_1, "running on register class: %s\n", env.cls->name)); - be_clear_links(chordal_env->irg); /* Decide which phi nodes will be spilled and place copies for them into the graph */ - irg_block_walk_graph(chordal_env->irg, spill_phi_walker, NULL, &env); + irg_block_walk_graph(irg, compute_live_ins, NULL, &env); /* Fix high register pressure with belady algorithm */ - irg_block_walk_graph(chordal_env->irg, NULL, belady, &env); + irg_block_walk_graph(irg, NULL, belady, &env); /* belady was block-local, fix the global flow by adding reloads on the edges */ - irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &env); + irg_block_walk_graph(irg, fix_block_borders, NULL, &env); /* Insert spill/reload nodes into the graph and fix usages */ be_insert_spills_reloads(env.senv); @@ -652,5 +746,18 @@ void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t if(spill_env == NULL) be_delete_spill_env(env.senv); be_end_uses(env.uses); + be_free_loop_pressure(env.loop_ana); obstack_free(&env.ob, NULL); } + +void be_init_spillbelady(void) +{ + static be_spiller_t belady_spiller = { + be_spill_belady + }; + + be_register_spiller("belady", &belady_spiller); + FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady"); +} + +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);