X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=ed734c0390c631a5a06fa0f8a9629db80d6af569;hb=712cea35fe399ae20e3a33eb9bf81360b5000cf3;hp=688172561f1dc7b3b45e7b0846ceb37f3ab7f941;hpb=d9760f58556914ee7a2bc0a42b0a57c2b0658e85;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index 688172561..ed734c039 100644 --- a/ir/be/bespillbelady.c +++ b/ir/be/bespillbelady.c @@ -4,8 +4,6 @@ * Copyright: (c) Universitaet Karlsruhe * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. * - * NOTE: Comments my be (partially) wrong, since there was a major bug - * (spilling of phis, prespill) whose fixing changed a lot. */ #ifdef HAVE_CONFIG_H #include "config.h" @@ -22,12 +20,14 @@ #include "obst.h" #include "set.h" #include "pset.h" +#include "irprintf_t.h" #include "irgraph.h" #include "irnode.h" #include "irmode.h" #include "irgwalk.h" #include "iredges_t.h" #include "ircons_t.h" +#include "irprintf.h" #include "beutil.h" #include "bearch.h" @@ -47,10 +47,9 @@ #define DBG_SLOTS 32 #define DBG_TRACE 64 #define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL) -static firm_dbg_module_t *dbg = NULL; +DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) #define MIN(a,b) (((a)<(b))?(a):(b)) -#undef SINGLE_START_PROJS typedef struct _workset_t workset_t; @@ -73,14 +72,18 @@ typedef struct _belady_env_t { struct _workset_t { belady_env_t *bel; - int i; /**< used for iteration TODO remove this form the struct */ int len; /**< current length */ loc_t vals[1]; /**< inlined array of the values/distances in this working set */ }; -typedef struct _block_info_t { - workset_t *ws_start, *ws_end; -} block_info_t; +void workset_print(const workset_t *w) +{ + int i; + + for(i = 0; i < w->len; ++i) { + ir_printf("%+F %d\n", w->vals[i].irn, w->vals[i].time); + } +} /** * Alloc a new workset on obstack @p ob with maximum size @p max @@ -204,9 +207,15 @@ static INLINE int workset_contains(const workset_t *ws, const ir_node *val) { return 0; } -#define workset_foreach(ws, v) for(ws->i=0; \ - v=(ws->i < ws->len) ? ws->vals[ws->i].irn : NULL, ws->i < ws->len; \ - ws->i++) +/** + * Iterates over all values in the working set. + * @p ws The workset to iterate + * @p v A variable to put the current value in + * @p i An integer for internal use + */ +#define workset_foreach(ws, v, i) for(i=0; \ + v=(i < ws->len) ? ws->vals[i].irn : NULL, i < ws->len; \ + ++i) #define workset_set_time(ws, i, t) (ws)->vals[i].time=t #define workset_set_length(ws, length) (ws)->len = length @@ -214,97 +223,50 @@ static INLINE int workset_contains(const workset_t *ws, const ir_node *val) { #define workset_get_val(ws, i) ((ws)->vals[i].irn) #define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare); +typedef struct _block_info_t { + workset_t *ws_start, *ws_end; +} block_info_t; + + +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; + + 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) static int is_mem_phi(const ir_node *irn, void *data) { workset_t *sws; ir_node *blk = get_nodes_block(irn); DBG((dbg, DBG_SPILL, "Is %+F a mem-phi?\n", irn)); - sws = ((block_info_t *) get_irn_link(blk))->ws_start; + sws = get_block_info(blk)->ws_start; DBG((dbg, DBG_SPILL, " %d\n", !workset_contains(sws, irn))); return !workset_contains(sws, irn); } /** - * 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. + * @return The distance to the next use + * Or 0 if irn is an ignore node */ -static void compute_block_start_info(ir_node *blk, void *env) { - belady_env_t *bel = env; - block_info_t *blk_info; - ir_node *irn, *first; - irn_live_t *li; - int i, count, ws_count; - loc_t loc, *starters; - ir_graph *irg = get_irn_irg(blk); - struct obstack ob; - obstack_init(&ob); +static INLINE unsigned get_distance(belady_env_t *bel, const ir_node *from, unsigned from_step, const ir_node *def, int skip_from_uses) +{ + arch_irn_flags_t fl = arch_irn_get_flags(bel->arch, def); + unsigned dist = be_get_next_use(bel->uses, from, from_step, def, skip_from_uses); - /* Get all values living at the block start */ - 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_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) { - loc.irn = irn; - loc.time = be_get_next_use(bel->uses, first, 0, irn, 0); - obstack_grow(&ob, &loc, sizeof(loc)); - DBG((dbg, DBG_START, " %+F:\n", irn)); - count++; - } else - break; + if(!USES_IS_INIFINITE(dist) && (fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0) + return 0; - live_foreach(blk, li) - if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) { - loc.irn = (ir_node *)li->irn; - loc.time = be_get_next_use(bel->uses, first, 0, li->irn, 0); - obstack_grow(&ob, &loc, sizeof(loc)); - DBG((dbg, DBG_START, " %+F:\n", irn)); - count++; - } - starters = obstack_finish(&ob); - - /* Sort all values */ - qsort(starters, count, sizeof(starters[0]), loc_compare); - - /* Create the start workset for this block. Copy the best ones from starters */ - blk_info = obstack_alloc(&bel->ob, sizeof(*blk_info)); - set_irn_link(blk, blk_info); - - ws_count = MIN(count, bel->n_regs); - blk_info->ws_start = new_workset(&bel->ob, bel); - workset_bulk_fill(blk_info->ws_start, ws_count, starters); - - /* Spill the phis among the remaining values */ - for (i=ws_count; icls, irg, pred_block, arg); - pset_insert_ptr(bel->copies, cpy); - DBG((dbg, DBG_START, " place a %+F of %+F in %+F\n", cpy, arg, pred_block)); - sched_add_before(pred_block, cpy); - set_irn_n(irn, o, cpy); - } - } - - obstack_free(&ob, NULL); + return dist; } /** - * Performs the actions neccessary to grant the request that: + * Performs the actions necessary to grant the request that: * - new_vals can be held in registers * - as few as possible other values are disposed * - the worst values get disposed @@ -314,7 +276,7 @@ static void compute_block_start_info(ir_node *blk, void *env) { */ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { ir_node *val; - int i, len, max_allowed, demand; + int i, len, max_allowed, demand, iter; workset_t *ws = bel->ws; ir_node **to_insert = alloca(bel->n_regs * sizeof(*to_insert)); @@ -322,7 +284,7 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { * 1. Identify the number of needed slots and the values to reload */ demand = 0; - workset_foreach(new_vals, val) { + workset_foreach(new_vals, val, iter) { /* mark value as used */ if (is_usage) pset_insert_ptr(bel->used, val); @@ -337,7 +299,6 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { } DBG((dbg, DBG_DECIDE, " demand = %d\n", demand)); - /* * 2. Make room for at least 'demand' slots */ @@ -348,7 +309,7 @@ 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, be_get_next_use(bel->uses, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage)); + workset_set_time(ws, i, get_distance(bel, bel->instr, bel->instr_nr, workset_get_val(ws, i), !is_usage)); /* sort entries by increasing nextuse-distance*/ workset_sort(ws); @@ -357,9 +318,10 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { before its first usage, remove it from start workset */ 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); - workset_t *ws_start = ((block_info_t *) get_irn_link(curr_bb))->ws_start; + 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)); @@ -377,6 +339,118 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { workset_bulk_insert(bel->ws, demand, to_insert); } +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. + */ +static block_info_t *compute_block_start_info(ir_node *blk, void *env) { + belady_env_t *bel = env; + ir_node *irn, *first; + irn_live_t *li; + int i, count, ws_count; + loc_t loc, *starters; + ir_graph *irg = get_irn_irg(blk); + struct obstack ob; + block_info_t *res = get_block_info(blk); + + /* Have we seen this block before? */ + if (res) + return res; + + /* Create the block info for this block. */ + res = new_block_info(&bel->ob); + set_block_info(blk, res); + + + /* Get all values living at the block start sorted by next use*/ + obstack_init(&ob); + + 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_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) { + loc.irn = irn; + loc.time = get_distance(bel, first, 0, irn, 0); + obstack_grow(&ob, &loc, sizeof(loc)); + DBG((dbg, DBG_START, " %+F:\n", irn)); + count++; + } else + break; + + live_foreach(blk, li) + if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) { + loc.irn = (ir_node *)li->irn; + loc.time = get_distance(bel, first, 0, li->irn, 0); + obstack_grow(&ob, &loc, sizeof(loc)); + DBG((dbg, DBG_START, " %+F:\n", irn)); + count++; + } + + starters = obstack_finish(&ob); + qsort(starters, count, sizeof(starters[0]), loc_compare); + + + /* 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); + + /* if pred block has not been processed yet, do it now */ + if (! pred_info) { + belady(pred_blk, bel); + pred_info = get_block_info(pred_blk); + } + + /* 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(&bel->ob, pred_info->ws_end); + + } else + + /* 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, bel->n_regs); + res->ws_start = new_workset(&bel->ob, bel); + workset_bulk_fill(res->ws_start, ws_count, starters); + } + + + /* 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); icls, irg, pred_block, arg); + pset_insert_ptr(bel->copies, cpy); + DBG((dbg, DBG_START, " place a %+F of %+F in %+F\n", cpy, arg, pred_block)); + sched_add_before(pred_block, cpy); + set_irn_n(irn, o, cpy); + } + } + + obstack_free(&ob, NULL); + return res; +} + + /** * For the given block @p blk, decide for each values * whether it is used from a register or is reloaded @@ -386,17 +460,22 @@ static void belady(ir_node *blk, void *env) { belady_env_t *bel = env; workset_t *new_vals; ir_node *irn; -#ifdef SINGLE_START_PROJS - ir_node *start_blk = get_irg_start_block(get_irn_irg(blk)); -#endif - block_info_t *blk_info = get_irn_link(blk); + int iter; + block_info_t *blk_info; + + /* Don't do a block twice */ + if (get_block_info(blk)) + 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)); workset_copy(bel->ws, blk_info->ws_start); - DBG((dbg, DBG_WSETS, "Initial start workset for %+F:\n", blk)); - workset_foreach(bel->ws, irn) + DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk)); + workset_foreach(bel->ws, irn, iter) DBG((dbg, DBG_WSETS, " %+F\n", irn)); /* process the block from start to end */ @@ -408,22 +487,13 @@ static void belady(ir_node *blk, void *env) { assert(workset_get_length(bel->ws) <= bel->n_regs && "Too much values in workset!"); -#ifdef SINGLE_START_PROJS - if (is_Phi(irn) || - (is_Proj(irn) && blk!=start_blk) || - (get_irn_mode(irn) == mode_T && blk==start_blk)) { - DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn)); - continue; - } -#else /* projs are handled with the tuple value. - * Phis are no real instr (see insert_starters) + * Phis are no real instr (see insert_starters()) * instr_nr does not increase */ if (is_Proj(irn) || is_Phi(irn)) { DBG((dbg, DBG_DECIDE, " ...%+F skipped\n", irn)); continue; } -#endif DBG((dbg, DBG_DECIDE, " ...%+F\n", irn)); /* set instruction in the workset */ @@ -451,11 +521,8 @@ static void belady(ir_node *blk, void *env) { /* Remember end-workset for this block */ blk_info->ws_end = workset_clone(&bel->ob, bel->ws); - DBG((dbg, DBG_WSETS, "Start workset for %+F:\n", blk)); - workset_foreach(blk_info->ws_start, irn) - DBG((dbg, DBG_WSETS, " %+F\n", irn)); DBG((dbg, DBG_WSETS, "End workset for %+F:\n", blk)); - workset_foreach(blk_info->ws_end, irn) + workset_foreach(blk_info->ws_end, irn, iter) DBG((dbg, DBG_WSETS, " %+F\n", irn)); } @@ -467,21 +534,21 @@ static void belady(ir_node *blk, void *env) { static void fix_block_borders(ir_node *blk, void *env) { workset_t *wsb; belady_env_t *bel = env; - int i, max; + int i, max, iter, iter2; DBG((dbg, DBG_FIX, "\n")); DBG((dbg, DBG_FIX, "Fixing %+F\n", blk)); - wsb = ((block_info_t *)get_irn_link(blk))->ws_start; + wsb = get_block_info(blk)->ws_start; /* process all pred blocks */ for (i=0, max=get_irn_arity(blk); iws_end; + workset_t *wsp = get_block_info(pred)->ws_end; DBG((dbg, DBG_FIX, " Pred %+F\n", pred)); - workset_foreach(wsb, irnb) { + 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)) @@ -492,7 +559,7 @@ static void fix_block_borders(ir_node *blk, void *env) { continue; /* check if irnb is in a register at end of pred */ - workset_foreach(wsp, irnp) + workset_foreach(wsp, irnp, iter2) if (irnb == irnp) goto next_value; @@ -523,16 +590,15 @@ static void remove_copies(belady_env_t *bel) { ir_node *irn; for (irn = pset_first(bel->copies); irn; irn = pset_next(bel->copies)) { - ir_node *src, *spill; + ir_node *src, *user; assert(be_is_Copy(irn)); assert(get_irn_n_edges(irn) == 1 && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?"); - spill = get_irn_edge(get_irn_irg(irn), irn, 0)->src; - assert(be_is_Spill(spill) && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?"); + user = get_irn_edge(get_irn_irg(irn), irn, 0)->src; - src = get_irn_n(irn, 0); - set_irn_n(spill, 0, src); + src = be_get_Copy_op(irn); + set_irn_n(user, 0, src); } } @@ -548,17 +614,20 @@ static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) { ir_node *spill; DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn))); - spill = get_irn_n(irn, 0); + if (be_is_Reload(irn)) + spill = get_irn_n(irn, be_pos_Reload_mem); /* remove reload */ set_irn_n(irn, 0, new_Bad()); sched_remove(irn); - /* if spill not used anymore, remove it too - * test of regclass is necessary since spill may be a phi-M */ - if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) { - set_irn_n(spill, 0, new_Bad()); - sched_remove(spill); + if (be_is_Reload(irn)) { + /* if spill not used anymore, remove it too + * test of regclass is necessary since spill may be a phi-M */ + if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) { + set_irn_n(spill, 0, new_Bad()); + sched_remove(spill); + } } } } @@ -566,30 +635,30 @@ static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) { void be_spill_belady(const be_chordal_env_t *chordal_env) { belady_env_t bel; - dbg = firm_dbg_register("ir.be.spillbelady"); - firm_dbg_set_mask(dbg, DEBUG_LVL); + FIRM_DBG_REGISTER(dbg, "ir.be.spillbelady"); /* init belady env */ obstack_init(&bel.ob); - bel.arch = chordal_env->main_env->arch_env; - bel.cls = chordal_env->cls; - bel.n_regs = arch_register_class_n_regs(bel.cls); - bel.ws = new_workset(&bel.ob, &bel); - bel.uses = be_begin_uses(chordal_env->irg, chordal_env->main_env->arch_env, bel.cls); - bel.senv = be_new_spill_env(dbg, chordal_env, is_mem_phi, NULL); - bel.reloads = pset_new_ptr_default(); - bel.copies = pset_new_ptr_default(); + bel.arch = chordal_env->birg->main_env->arch_env; + bel.cls = chordal_env->cls; + bel.n_regs = arch_register_class_n_regs(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); + bel.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL); + DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);) + bel.reloads = pset_new_ptr_default(); + bel.copies = pset_new_ptr_default(); + + DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name)); /* do the work */ - irg_block_walk_graph(chordal_env->irg, compute_block_start_info, NULL, &bel); - irg_block_walk_graph(chordal_env->irg, belady, NULL, &bel); + 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, bel.reloads); remove_unused_reloads(chordal_env->irg, &bel); remove_copies(&bel); - - /* clean up */ del_pset(bel.reloads); be_delete_spill_env(bel.senv);