X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=cd9d4b0fd27b4a8d199412531adb750bd251d3d1;hb=e9137de01e1ed5e7a8baee8d23a59e82ad634af5;hp=04cf88710825c6e87ba164a01cbab59eca6f697a;hpb=08f75f5670b62b8f76da7e2bc38648372f422356;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index 04cf88710..cd9d4b0fd 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" @@ -46,11 +46,9 @@ #define DBG_START 16 #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) -static firm_dbg_module_t *dbg = NULL; - -#define MIN(a,b) (((a)<(b))?(a):(b)) -#undef SINGLE_START_PROJS +DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) typedef struct _workset_t workset_t; @@ -65,22 +63,25 @@ typedef struct _belady_env_t { 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 *copies; /**< holds all copies placed due to phi-spilling */ + ir_node **copies; /**< holds all copies placed due to phi-spilling */ - spill_env_t *senv; /* see bespill.h */ - pset *reloads; /**< all reload nodes placed */ + spill_env_t *senv; /**< see bespill.h */ } 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 @@ -131,7 +132,7 @@ static INLINE void workset_insert(workset_t *ws, ir_node *val) { int i; /* check for current regclass */ if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) { - DBG((dbg, DBG_DECIDE, "Dropped %+F\n", val)); + DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val)); return; } @@ -204,9 +205,15 @@ 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 +221,50 @@ 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); - /* 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; +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); - 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); + if(!USES_IS_INIFINITE(dist) && (fl & (arch_irn_flags_ignore | arch_irn_flags_dont_spill)) != 0) + return 0; - /* 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 +274,7 @@ 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 +282,7 @@ 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); @@ -332,23 +292,25 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) { to_insert[demand++] = val; if (is_usage) be_add_reload(bel->senv, val, bel->instr); - } else + } else { DBG((dbg, DBG_DECIDE, " skip %+F\n", val)); + } } DBG((dbg, DBG_DECIDE, " demand = %d\n", demand)); - /* * 2. Make room for at least 'demand' slots */ len = workset_get_length(ws); max_allowed = bel->n_regs - demand; + DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed)); + /* Only make more free room if we do not have enough */ 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); @@ -357,14 +319,16 @@ 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)); - } else + } else { DBG((dbg, DBG_DECIDE, " dispose %+F\n", irn)); + } } /* kill the last 'demand' entries in the array */ @@ -377,6 +341,160 @@ 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); + +/** + * Inserts a spill of a value at the earliest possible location in a block. + * That is after the last use of the value or at the beginning of the block if + * there is no use + */ +static ir_node *insert_copy(belady_env_t *env, ir_node *block, ir_node *value) { + ir_node* node; + ir_graph *irg = get_irn_irg(block); + ir_node *copy = be_new_Copy(env->cls, irg, block, value); + + ARR_APP1(ir_node*, env->copies, copy); + + // walk schedule backwards until we find a usage, or until we have reached the first phi + // TODO can we do this faster somehow? This makes insert_copy O(n) in block_size... + sched_foreach_reverse(block, node) { + int i, arity; + + if(is_Phi(node)) { + sched_add_after(node, copy); + goto placed; + } + if(value == node) { + sched_add_after(node, copy); + goto placed; + } + for(i = 0, arity = get_irn_arity(node); i < arity; ++i) { + ir_node *arg = get_irn_n(node, i); + if(arg == value) { + sched_add_after(node, copy); + goto placed; + } + } + } + // we didn't find a use or a phi yet, so place the copy at the beginning of the block + sched_add_before(sched_first(block), copy); + +placed: + + return copy; +} + +/** + * 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 *data) { + belady_env_t *env = data; + 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(&env->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(env->arch, irn, -1) == env->cls) { + loc.irn = irn; + loc.time = get_distance(env, 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(env->arch, li->irn, -1) == env->cls) { + loc.irn = (ir_node *)li->irn; + loc.time = get_distance(env, first, 0, li->irn, 0); + obstack_grow(&ob, &loc, sizeof(loc)); + DBG((dbg, DBG_START, " %+F:\n", li->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, env); + 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(&env->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, env->n_regs); + res->ws_start = new_workset(&env->ob, env); + 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); iws, 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 +531,13 @@ 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 +565,8 @@ 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 +578,21 @@ 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)) @@ -492,9 +603,10 @@ 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; + } /* irnb is in memory at the end of pred, so we have to reload it */ DBG((dbg, DBG_FIX, " reload %+F\n", irnb)); @@ -506,94 +618,69 @@ next_value: } } -/** - * Removes all used reloads from bel->reloads. - * The remaining nodes in bel->reloads will be removed from the graph. - */ -static void rescue_used_reloads(ir_node *irn, void *env) { - pset *rlds = (pset *)env; - if (pset_find_ptr(rlds, irn)) - pset_remove_ptr(rlds, irn); -} - /** * Removes all copies introduced for phi-spills */ -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; - - 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?"); - - src = get_irn_n(irn, 0); - set_irn_n(spill, 0, src); - } -} - -/** - * Finds all unused reloads and remove them from the schedule - * Also removes spills if they are not used anymore after removing reloads - */ -static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) { - ir_node *irn; +static void remove_copies(belady_env_t *env) { + int i; - irg_walk_graph(irg, rescue_used_reloads, NULL, bel->reloads); - for(irn = pset_first(bel->reloads); irn; irn = pset_next(bel->reloads)) { - ir_node *spill; - DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn))); + for(i = 0; i < ARR_LEN(env->copies); ++i) { + ir_node *node = env->copies[i]; + ir_node *src; + const ir_edge_t *edge, *ne; - spill = get_irn_n(irn, 0); + assert(be_is_Copy(node)); - /* remove reload */ - set_irn_n(irn, 0, new_Bad()); - sched_remove(irn); + src = be_get_Copy_op(node); + foreach_out_edge_safe(node, edge, ne) { + ir_node *user = get_edge_src_irn(edge); + int user_pos = get_edge_src_pos(edge); - /* 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); + set_irn_n(user, user_pos, src); } } } void be_spill_belady(const be_chordal_env_t *chordal_env) { + be_spill_belady_spill_env(chordal_env, NULL); +} + +void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) { belady_env_t bel; - dbg = firm_dbg_register("ir.be.spillbelady"); + FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady"); /* 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); + if(spill_env == NULL) { + bel.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL); + } else { + bel.senv = spill_env; + be_set_is_spilled_phi(bel.senv, is_mem_phi, NULL); + } + DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);) + bel.copies = NEW_ARR_F(ir_node*, 0); 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); + be_insert_spills_reloads(bel.senv); remove_copies(&bel); + DEL_ARR_F(bel.copies); - + be_remove_dead_nodes_from_schedule(chordal_env->irg); /* clean up */ - del_pset(bel.reloads); - be_delete_spill_env(bel.senv); + if(spill_env == NULL) + be_delete_spill_env(bel.senv); be_end_uses(bel.uses); obstack_free(&bel.ob, NULL); }