X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=4b5b16ad0a12a6e54247a4f2b1ebd27cff55a027;hb=6866751667f73eeec02ac8275f1011f9e424cce2;hp=c799a50d0cb5ef113cec1dbbfc7ca7ec7ab29f2a;hpb=a220ff4c14309623e11a0d4a6650130baebc5886;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index c799a50d0..4b5b16ad0 100644 --- a/ir/be/bespillbelady.c +++ b/ir/be/bespillbelady.c @@ -62,13 +62,12 @@ typedef struct _belady_env_t { be_uses_t *uses; /**< env for the next-use magic */ ir_node *instr; /**< current instruction */ unsigned instr_nr; /**< current instruction number (relative to block start) */ - pset *used; /**< holds the values used (so far) in the current BB */ + pset *used; spill_env_t *senv; /**< see bespill.h */ } belady_env_t; struct _workset_t { - belady_env_t *bel; int len; /**< current length */ loc_t vals[0]; /**< inlined array of the values/distances in this working set */ }; @@ -85,21 +84,20 @@ void workset_print(const workset_t *w) /** * Alloc a new workset on obstack @p ob with maximum size @p max */ -static INLINE workset_t *new_workset(struct obstack *ob, belady_env_t *bel) { +static INLINE workset_t *new_workset(belady_env_t *env, struct obstack *ob) { workset_t *res; - size_t size = sizeof(*res) + (bel->n_regs)*sizeof(res->vals[0]); + size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]); res = obstack_alloc(ob, size); memset(res, 0, size); - res->bel = bel; return res; } /** * Alloc a new instance on obstack and make it equal to @param ws */ -static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) { +static INLINE workset_t *workset_clone(belady_env_t *env, struct obstack *ob, workset_t *ws) { workset_t *res; - size_t size = sizeof(*res) + (ws->bel->n_regs)*sizeof(res->vals[0]); + size_t size = sizeof(*res) + (env->n_regs)*sizeof(res->vals[0]); res = obstack_alloc(ob, size); memcpy(res, ws, size); return res; @@ -107,10 +105,10 @@ static INLINE workset_t *workset_clone(struct obstack *ob, workset_t *ws) { /** * Do NOT alloc anything. Make @param tgt equal to @param src. - * returns @param tgt for convinience + * returns @param tgt for convenience */ -static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) { - size_t size = sizeof(*src) + (src->bel->n_regs)*sizeof(src->vals[0]); +static INLINE workset_t *workset_copy(belady_env_t *env, workset_t *tgt, workset_t *src) { + size_t size = sizeof(*src) + (env->n_regs)*sizeof(src->vals[0]); memcpy(tgt, src, size); return tgt; } @@ -120,17 +118,19 @@ static INLINE workset_t *workset_copy(workset_t *tgt, workset_t *src) { * @param count locations given at memory @param locs. * Set the length of @param ws to count. */ -#define workset_bulk_fill(ws, count, locs) memcpy(&(ws)->vals[0], locs, ((ws)->len=count)*sizeof(locs[0])); - +static INLINE void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) { + workset->len = count; + memcpy(&(workset->vals[0]), locs, count * sizeof(locs[0])); +} /** * Inserts the value @p val into the workset, iff it is not * already contained. The workset must not be full. */ -static INLINE void workset_insert(workset_t *ws, ir_node *val) { +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(ws->bel->arch, ws->bel->cls, val)) { + if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) { DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val)); return; } @@ -141,14 +141,16 @@ static INLINE void workset_insert(workset_t *ws, ir_node *val) { return; /* insert val */ - assert(ws->len < ws->bel->n_regs && "Workset already full!"); + assert(ws->len < env->n_regs && "Workset already full!"); ws->vals[ws->len++].irn = val; } /** * Removes all entries from this workset */ -#define workset_clear(ws) (ws)->len = 0; +static INLINE void workset_clear(workset_t *ws) { + ws->len = 0; +} /** * Removes the value @p val from the workset if present. @@ -191,6 +193,7 @@ static INLINE int workset_contains(const workset_t *ws, const ir_node *val) { typedef struct _block_info_t { workset_t *ws_start, *ws_end; + int processed; } block_info_t; @@ -198,6 +201,7 @@ static INLINE void *new_block_info(struct obstack *ob) { block_info_t *res = obstack_alloc(ob, sizeof(*res)); res->ws_start = NULL; res->ws_end = NULL; + res->processed = 0; return res; } @@ -230,11 +234,11 @@ static INLINE unsigned get_distance(belady_env_t *env, const ir_node *from, unsi * @p is_usage indicates that the values in new_vals are used (not defined) * In this case reloads must be performed */ -static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { +static void displace(belady_env_t *env, workset_t *new_vals, int is_usage) { ir_node *val; int i, len, max_allowed, demand, iter; - workset_t *ws = bel->ws; - ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert)); + workset_t *ws = env->ws; + ir_node **to_insert = alloca(env->n_regs * sizeof(*to_insert)); /* * 1. Identify the number of needed slots and the values to reload @@ -243,13 +247,13 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { workset_foreach(new_vals, val, iter) { /* mark value as used */ if (is_usage) - pset_insert_ptr(bel->used, val); + pset_insert_ptr(env->used, val); if (!workset_contains(ws, val)) { DBG((dbg, DBG_DECIDE, " insert %+F\n", val)); to_insert[demand++] = val; if (is_usage) - be_add_reload(bel->senv, val, bel->instr); + be_add_reload(env->senv, val, env->instr); } else { assert(is_usage || "Defined value already in workset?!?"); DBG((dbg, DBG_DECIDE, " skip %+F\n", val)); @@ -261,7 +265,7 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { * 2. Make room for at least 'demand' slots */ len = workset_get_length(ws); - max_allowed = bel->n_regs - demand; + max_allowed = env->n_regs - demand; DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed)); @@ -269,18 +273,23 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { if (len > max_allowed) { /* get current next-use distance */ for (i=0; ilen; ++i) - workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage)); + workset_set_time(ws, i, get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage)); /* sort entries by increasing nextuse-distance*/ workset_sort(ws); /* Logic for not needed live-ins: If a value is disposed - before its first usage, remove it from start workset */ + * before its first usage, remove it from start workset + * We don't do this for phis though + */ for (i=max_allowed; ilen; ++i) { ir_node *irn = ws->vals[i].irn; - if (!pset_find_ptr(bel->used, irn)) { - ir_node *curr_bb = get_nodes_block(bel->instr); + if(is_Phi(irn)) + continue; + + if (!pset_find_ptr(env->used, irn)) { + ir_node *curr_bb = get_nodes_block(env->instr); workset_t *ws_start = get_block_info(curr_bb)->ws_start; workset_remove(ws_start, irn); @@ -298,107 +307,111 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { * 3. Insert the new values into the workset */ for(i = 0; i < demand; ++i) - workset_insert(bel->ws, to_insert[i]); + workset_insert(env, env->ws, to_insert[i]); } static void belady(ir_node *blk, void *env); -/** - * Collects all values live-in at block @p blk 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 - * same spill slot. +/* + * Computes set of live-ins for each block with multiple predecessors and + * places copies in the predecessors when phis get spilled */ -static block_info_t *compute_block_start_info(ir_node *blk, void *data) { +static void place_copy_walker(ir_node *block, void *data) { belady_env_t *env = data; - ir_node *irn, *first; + block_info_t *block_info; irn_live_t *li; - int count, ws_count; + ir_node *first, *irn; loc_t loc, *starters; - struct obstack ob; - block_info_t *res = get_block_info(blk); + int i, len, ws_count; - /* Have we seen this block before? */ - if (res) - return res; + if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block) + return; - /* Create the block info for this block. */ - res = new_block_info(&env->ob); - set_block_info(blk, res); + block_info = new_block_info(&env->ob); + set_block_info(block, block_info); - /* Get all values living at the block start sorted by next use*/ - obstack_init(&ob); + /* Collect all values living at start of block */ + starters = NEW_ARR_F(loc_t, 0); - DBG((dbg, DBG_START, "Living at start of %+F:\n", blk)); - first = sched_first(blk); - count = 0; - sched_foreach(blk, irn) { - if(!is_Phi(irn) || !arch_irn_consider_in_reg_alloc(env->arch, env->cls, irn)) + DBG((dbg, DBG_START, "Living at start of %+F:\n", block)); + first = sched_first(block); + sched_foreach(block, 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); - obstack_grow(&ob, &loc, sizeof(loc)); + ARR_APP1(loc_t, starters, loc); DBG((dbg, DBG_START, " %+F:\n", irn)); - count++; } - live_foreach(blk, li) { + live_foreach(block, li) { if (!live_is_in(li) || !arch_irn_consider_in_reg_alloc(env->arch, env->cls, li->irn)) continue; loc.irn = (ir_node *)li->irn; loc.time = get_distance(env, first, 0, li->irn, 0); - obstack_grow(&ob, &loc, sizeof(loc)); + ARR_APP1(loc_t, starters, loc); DBG((dbg, DBG_START, " %+F:\n", li->irn)); - count++; } - starters = obstack_finish(&ob); - qsort(starters, count, sizeof(starters[0]), loc_compare); + // 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); + block_info->ws_start = new_workset(env, &env->ob); + workset_bulk_fill(block_info->ws_start, ws_count, starters); - /* If we have only one predecessor, we want the start_set of blk to be the end_set of pred */ - if (get_Block_n_cfgpreds(blk) == 1 && blk != get_irg_start_block(get_irn_irg(blk))) { - ir_node *pred_blk = get_Block_cfgpred_block(blk, 0); - block_info_t *pred_info = get_block_info(pred_blk); + /* 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) { + irn = starters[i].irn; + if (!is_Phi(irn) || get_nodes_block(irn) != block) + continue; - /* if pred block has not been processed yet, do it now */ - if (! pred_info) { - belady(pred_blk, env); - pred_info = get_block_info(pred_blk); - } + be_spill_phi(env->senv, irn); + } - /* now we have an end_set of pred */ - assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set"); - res->ws_start = workset_clone(&env->ob, pred_info->ws_end); - } else { - //int i; - - /* Else we want the start_set to be the values used 'the closest' */ - /* Copy the best ones from starters to start workset */ - ws_count = MIN(count, env->n_regs); - res->ws_start = new_workset(&env->ob, env); - workset_bulk_fill(res->ws_start, ws_count, starters); - -#if 0 - /* The phis of this block which are not in the start set have to be spilled later. - * Therefore we add temporary copies in the pred_blocks so the spills can spill - * into the same spill slot. - * After spilling these copies get deleted. - */ - for (i=workset_get_length(res->ws_start); isenv, irn); - } -#endif +/** + * Collects all values live-in at block @p blk 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 + * same spill slot. + */ +static block_info_t *compute_block_start_info(belady_env_t *env, ir_node *block) { + ir_node *pred_block; + block_info_t *res, *pred_info; + + /* Have we seen this block before? */ + res = get_block_info(block); + if (res) + return res; + + /* Create the block info for this block. */ + res = new_block_info(&env->ob); + set_block_info(block, res); + + /* Use endset of predecessor block as startset */ + assert(get_Block_n_cfgpreds(block) == 1 && block != get_irg_start_block(get_irn_irg(block))); + pred_block = get_Block_cfgpred_block(block, 0); + pred_info = get_block_info(pred_block); + + /* if pred block has not been processed yet, do it now */ + if (pred_info == NULL || pred_info->processed == 0) { + belady(pred_block, env); + pred_info = get_block_info(pred_block); } - obstack_free(&ob, NULL); + /* now we have an end_set of pred */ + assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set"); + res->ws_start = workset_clone(env, &env->ob, pred_info->ws_end); + return res; } @@ -408,36 +421,39 @@ static block_info_t *compute_block_start_info(ir_node *blk, void *data) { * whether it is used from a register or is reloaded * before the use. */ -static void belady(ir_node *blk, void *env) { - belady_env_t *bel = env; +static void belady(ir_node *block, void *data) { + belady_env_t *env = data; workset_t *new_vals; ir_node *irn; int iter; - block_info_t *blk_info; + block_info_t *block_info; + + /* make sure we have blockinfo (with startset) */ + block_info = get_block_info(block); + if (block_info == NULL) + block_info = compute_block_start_info(env, block); /* Don't do a block twice */ - if (get_block_info(blk)) + if(block_info->processed) return; /* get the starting workset for this block */ - blk_info = compute_block_start_info(blk, bel); - DBG((dbg, DBG_DECIDE, "\n")); - DBG((dbg, DBG_DECIDE, "Decide for %+F\n", blk)); + DBG((dbg, DBG_DECIDE, "Decide for %+F\n", block)); - workset_copy(bel->ws, blk_info->ws_start); - DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk)); - workset_foreach(bel->ws, irn, iter) + 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)); /* process the block from start to end */ DBG((dbg, DBG_WSETS, "Processing...\n")); - bel->used = pset_new_ptr(32); - bel->instr_nr = 0; - new_vals = new_workset(&bel->ob, bel); - sched_foreach(blk, irn) { + env->used = pset_new_ptr_default(); + env->instr_nr = 0; + new_vals = new_workset(env, &env->ob); + sched_foreach(block, irn) { int i, arity; - assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!"); + assert(workset_get_length(env->ws) <= env->n_regs && "Too much values in workset!"); /* projs are handled with the tuple value. * Phis are no real instr (see insert_starters()) @@ -449,34 +465,35 @@ static void belady(ir_node *blk, void *env) { DBG((dbg, DBG_DECIDE, " ...%+F\n", irn)); /* set instruction in the workset */ - bel->instr = irn; + env->instr = irn; /* allocate all values _used_ by this instruction */ workset_clear(new_vals); for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) { - workset_insert(new_vals, get_irn_n(irn, i)); + workset_insert(env, new_vals, get_irn_n(irn, i)); } - displace(bel, new_vals, 1); + displace(env, new_vals, 1); /* allocate all values _defined_ by this instruction */ workset_clear(new_vals); if (get_irn_mode(irn) == mode_T) { /* special handling for tuples and projs */ ir_node *proj; for(proj=sched_next(irn); is_Proj(proj); proj=sched_next(proj)) - workset_insert(new_vals, proj); + workset_insert(env, new_vals, proj); } else { - workset_insert(new_vals, irn); + workset_insert(env, new_vals, irn); } - displace(bel, new_vals, 0); + displace(env, new_vals, 0); - bel->instr_nr++; + env->instr_nr++; } - del_pset(bel->used); + del_pset(env->used); /* Remember end-workset for this block */ - blk_info->ws_end = workset_clone(&bel->ob, bel->ws); - DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk)); - workset_foreach(blk_info->ws_end, irn, iter) + block_info->ws_end = workset_clone(env, &env->ob, env->ws); + 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)); } @@ -485,9 +502,9 @@ static void belady(ir_node *blk, void *env) { * 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 *env) { +static void fix_block_borders(ir_node *blk, void *data) { + belady_env_t *env = data; workset_t *wsb; - belady_env_t *bel = env; int i, max, iter, iter2; DBG((dbg, DBG_FIX, "\n")); @@ -505,9 +522,14 @@ static void fix_block_borders(ir_node *blk, void *env) { 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) && blk == get_nodes_block(irnb)) { irnb = get_irn_n(irnb, i); + // we might have unknowns as argument for the phi + if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb)) + continue; + } + /* Unknowns are available everywhere */ if(get_irn_opcode(irnb) == iro_Unknown) continue; @@ -520,7 +542,7 @@ static void fix_block_borders(ir_node *blk, void *env) { /* 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(bel->senv, irnb, blk, i); + be_add_reload_on_edge(env->senv, irnb, blk, i); next_value: /*epsilon statement :)*/; @@ -533,37 +555,43 @@ void be_spill_belady(const be_chordal_env_t *chordal_env) { } void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) { - belady_env_t bel; + belady_env_t env; FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady"); + //firm_dbg_set_mask(dbg, DBG_WSETS); /* init belady env */ - obstack_init(&bel.ob); - bel.arch = chordal_env->birg->main_env->arch_env; - bel.cls = chordal_env->cls; - bel.n_regs = arch_count_non_ignore_regs(bel.arch, bel.cls); - bel.ws = new_workset(&bel.ob, &bel); - bel.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls); + obstack_init(&env.ob); + 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.ws = new_workset(&env, &env.ob); + env.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, env.cls); if(spill_env == NULL) { - bel.senv = be_new_spill_env(chordal_env); + env.senv = be_new_spill_env(chordal_env); } else { - bel.senv = spill_env; + env.senv = spill_env; } - DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);) + DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);) - DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name)); + DBG((dbg, LEVEL_1, "running on register class: %s\n", env.cls->name)); - /* do the work */ be_clear_links(chordal_env->irg); - irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel); - irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel); - be_insert_spills_reloads(bel.senv); + /* Decide which phi nodes will be spilled and place copies for them into the graph */ + irg_block_walk_graph(chordal_env->irg, place_copy_walker, NULL, &env); + be_place_copies(env.senv); + /* Fix high register pressure with belady algorithm */ + irg_block_walk_graph(chordal_env->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); + /* Insert spill/reload nodes into the graph and fix usages */ + be_insert_spills_reloads(env.senv); be_remove_dead_nodes_from_schedule(chordal_env->irg); /* clean up */ if(spill_env == NULL) - be_delete_spill_env(bel.senv); - be_end_uses(bel.uses); - obstack_free(&bel.ob, NULL); + be_delete_spill_env(env.senv); + be_end_uses(env.uses); + obstack_free(&env.ob, NULL); }