X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespillbelady.c;h=ad654a7fb63c39d49becf2d704af6b774f620f08;hb=29087feac9466883c278e53eec325d5e3099df1d;hp=eaf93d036edc159510a929782e370ff48b2e366c;hpb=44d51ff21a4a48494fd8e42dd10f5b489b4fca78;p=libfirm diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c index eaf93d036..ad654a7fb 100644 --- a/ir/be/bespillbelady.c +++ b/ir/be/bespillbelady.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -22,16 +22,12 @@ * @brief Beladys spillalgorithm. * @author Daniel Grund, Matthias Braun * @date 20.09.2005 - * @version $Id$ */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include #include "obst.h" -#include "irprintf_t.h" #include "irgraph.h" #include "irnode.h" #include "irmode.h" @@ -41,21 +37,22 @@ #include "ircons_t.h" #include "irprintf.h" #include "irnodeset.h" -#include "xmalloc.h" -#include "pdeq.h" +#include "irtools.h" +#include "statev_t.h" +#include "util.h" #include "beutil.h" -#include "bearch_t.h" +#include "bearch.h" #include "beuses.h" -#include "besched_t.h" +#include "besched.h" #include "beirgmod.h" #include "belive_t.h" -#include "benode_t.h" +#include "benode.h" #include "bechordal_t.h" -#include "bespilloptions.h" -#include "beloopana.h" -#include "beirg_t.h" #include "bespill.h" +#include "beloopana.h" +#include "beirg.h" +#include "bespillutil.h" #include "bemodule.h" #define DBG_SPILL 1 @@ -68,71 +65,54 @@ #define DBG_WORKSET 128 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;) -/* factor to weight the different costs of reloading/rematerializing a node - (see bespill.h be_get_reload_costs_no_weight) */ -#define RELOAD_COST_FACTOR 10 - #define TIME_UNDEFINED 6666 -#define PLACE_SPILLS - /** * An association between a node and a point in time. */ typedef struct loc_t { ir_node *node; unsigned time; /**< A use time (see beuses.h). */ - bool spilled; /**< the value was already spilled on this path */ + bool spilled; /**< value was already spilled on this path */ } loc_t; -typedef struct _workset_t { - int len; /**< current length */ - loc_t vals[0]; /**< inlined array of the values/distances in this working set */ +typedef struct workset_t { + unsigned len; /**< current length */ + loc_t vals[]; /**< array of the values/distances in this working set */ } workset_t; static struct obstack obst; -static const arch_env_t *arch_env; static const arch_register_class_t *cls; static const be_lv_t *lv; static be_loopana_t *loop_ana; -static int n_regs; +static unsigned n_regs; static workset_t *ws; /**< the main workset used while processing a block. */ static be_uses_t *uses; /**< env for the next-use magic */ -static ir_node *instr; /**< current instruction */ -static unsigned instr_nr; /**< current instruction number - (relative to block start) */ -static ir_nodeset_t used; static spill_env_t *senv; /**< see bespill.h */ -static pdeq *worklist; +static ir_node **blocklist; -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) -{ - int i; +static int move_spills = true; +static int respectloopdepth = true; +static int improve_known_preds = true; +/* factor to weight the different costs of reloading/rematerializing a node + (see bespill.h be_get_reload_costs_no_weight) */ +static int remat_bonus = 10; - for(i = 0; i < w->len; ++i) { - ir_fprintf(stderr, "%+F %d\n", w->vals[i].node, w->vals[i].time); - } -} +static const lc_opt_table_entry_t options[] = { + LC_OPT_ENT_BOOL ("movespills", "try to move spills out of loops", &move_spills), + LC_OPT_ENT_BOOL ("respectloopdepth", "outermost loop cutting", &respectloopdepth), + LC_OPT_ENT_BOOL ("improveknownpreds", "known preds cutting", &improve_known_preds), + LC_OPT_ENT_INT ("rematbonus", "give bonus to rematerialisable nodes", &remat_bonus), + LC_OPT_LAST +}; /** * Alloc a new workset on obstack @p ob with maximum size @p max */ static workset_t *new_workset(void) { - workset_t *res; - size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]); - - res = obstack_alloc(&obst, size); - memset(res, 0, size); - return res; + return OALLOCFZ(&obst, workset_t, vals, n_regs); } /** @@ -140,10 +120,8 @@ static workset_t *new_workset(void) */ static workset_t *workset_clone(workset_t *workset) { - workset_t *res; - size_t size = sizeof(*res) + n_regs * sizeof(res->vals[0]); - res = obstack_alloc(&obst, size); - memcpy(res, workset, size); + workset_t *res = OALLOCF(&obst, workset_t, vals, n_regs); + memcpy(res, workset, sizeof(*res) + n_regs * sizeof(res->vals[0])); return res; } @@ -173,10 +151,10 @@ static void workset_bulk_fill(workset_t *workset, int count, const loc_t *locs) */ static void workset_insert(workset_t *workset, ir_node *val, bool spilled) { - loc_t *loc; - int i; + loc_t *loc; + unsigned i; /* check for current regclass */ - assert(arch_irn_consider_in_reg_alloc(arch_env, cls, val)); + assert(arch_irn_consider_in_reg_alloc(cls, val)); /* check if val is already contained */ for (i = 0; i < workset->len; ++i) { @@ -209,10 +187,10 @@ static void workset_clear(workset_t *workset) /** * Removes the value @p val from the workset if present. */ -static INLINE void workset_remove(workset_t *workset, ir_node *val) +static void workset_remove(workset_t *workset, ir_node *val) { - int i; - for(i = 0; i < workset->len; ++i) { + unsigned i; + for (i = 0; i < workset->len; ++i) { if (workset->vals[i].node == val) { workset->vals[i] = workset->vals[--workset->len]; return; @@ -220,11 +198,9 @@ static INLINE void workset_remove(workset_t *workset, ir_node *val) } } -static INLINE const loc_t *workset_contains(const workset_t *ws, - const ir_node *val) +static const loc_t *workset_contains(const workset_t *ws, const ir_node *val) { - int i; - + unsigned i; for (i = 0; i < ws->len; ++i) { if (ws->vals[i].node == val) return &ws->vals[i]; @@ -233,65 +209,107 @@ static INLINE const loc_t *workset_contains(const workset_t *ws, return NULL; } +static int loc_compare(const void *a, const void *b) +{ + const loc_t *p = ((const loc_t*) a); + const loc_t *q = ((const loc_t*) b); + const unsigned pt = p->time; + const unsigned qt = q->time; + + if (pt < qt) + return -1; + if (pt > qt) + return 1; + + return get_irn_node_nr(p->node) - get_irn_node_nr(q->node); +} + +static void workset_sort(workset_t *workset) +{ + qsort(workset->vals, workset->len, sizeof(workset->vals[0]), loc_compare); +} + +static inline unsigned workset_get_time(const workset_t *workset, unsigned idx) +{ + return workset->vals[idx].time; +} + +static inline void workset_set_time(workset_t *workset, unsigned idx, + unsigned time) +{ + workset->vals[idx].time = time; +} + +static inline unsigned workset_get_length(const workset_t *workset) +{ + return workset->len; +} + +static inline void workset_set_length(workset_t *workset, unsigned len) +{ + workset->len = len; +} + +static inline ir_node *workset_get_val(const workset_t *workset, unsigned idx) +{ + return workset->vals[idx].node; +} + /** * 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].node : NULL, i < ws->len; \ - ++i) - -#define workset_set_time(ws, i, t) (ws)->vals[i].time=t -#define workset_get_time(ws, i) (ws)->vals[i].time -#define workset_set_length(ws, length) (ws)->len = length -#define workset_get_length(ws) ((ws)->len) -#define workset_get_val(ws, i) ((ws)->vals[i].node) -#define workset_sort(ws) qsort((ws)->vals, (ws)->len, sizeof((ws)->vals[0]), loc_compare); - -typedef struct _block_info_t -{ +#define workset_foreach(ws, v, i) \ + for (i=0; v=(i < ws->len) ? ws->vals[i].node : NULL, i < ws->len; ++i) + +typedef struct block_info_t { workset_t *start_workset; workset_t *end_workset; } block_info_t; - -static void *new_block_info(void) +static block_info_t *new_block_info(void) { - block_info_t *res = obstack_alloc(&obst, sizeof(res[0])); - memset(res, 0, sizeof(res[0])); + return OALLOCZ(&obst, block_info_t); +} - return res; +static inline block_info_t *get_block_info(const ir_node *block) +{ + return (block_info_t*)get_irn_link(block); } -#define get_block_info(block) ((block_info_t *)get_irn_link(block)) -#define set_block_info(block, info) set_irn_link(block, info) +static inline void set_block_info(ir_node *block, block_info_t *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(ir_node *from, unsigned from_step, - const ir_node *def, int skip_from_uses) +static unsigned get_distance(ir_node *from, const ir_node *def, int skip_from_uses) { be_next_use_t use; - int flags = arch_irn_get_flags(arch_env, def); unsigned costs; unsigned time; - assert(! (flags & arch_irn_flags_ignore)); + assert(!arch_irn_is_ignore(def)); - use = be_get_next_use(uses, from, from_step, def, skip_from_uses); - if(USES_IS_INFINITE(use.time)) + use = be_get_next_use(uses, from, def, skip_from_uses); + time = use.time; + if (USES_IS_INFINITE(time)) return USES_INFINITY; /* We have to keep nonspillable nodes in the workingset */ - if(flags & arch_irn_flags_dont_spill) + if (arch_get_irn_flags(skip_Proj_const(def)) & arch_irn_flags_dont_spill) return 0; - costs = be_get_reload_costs_no_weight(senv, def, use.before); - assert(costs * RELOAD_COST_FACTOR < 1000); - time = use.time + 1000 - (costs * RELOAD_COST_FACTOR); + /* give some bonus to rematerialisable nodes */ + if (remat_bonus > 0) { + costs = be_get_reload_costs_no_weight(senv, def, use.before); + assert(costs * remat_bonus < 1000); + time += 1000 - (costs * remat_bonus); + } return time; } @@ -305,26 +323,22 @@ static INLINE unsigned get_distance(ir_node *from, unsigned from_step, * @p is_usage indicates that the values in new_vals are used (not defined) * In this case reloads must be performed */ -static void displace(workset_t *new_vals, int is_usage) +static void displace(workset_t *const new_vals, int const is_usage, ir_node *const instr) { - ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0])); - bool *spilled = alloca(n_regs * sizeof(spilled[0])); + ir_node **to_insert = ALLOCAN(ir_node*, n_regs); + bool *spilled = ALLOCAN(bool, n_regs); ir_node *val; int i; int len; int spills_needed; int demand; - int iter; + unsigned iter; /* 1. Identify the number of needed slots and the values to reload */ demand = 0; workset_foreach(new_vals, val, iter) { bool reloaded = false; - /* mark value as used */ - if (is_usage) - ir_nodeset_insert(&used, val); - if (! workset_contains(ws, val)) { DB((dbg, DBG_DECIDE, " insert %+F\n", val)); if (is_usage) { @@ -351,17 +365,12 @@ static void displace(workset_t *new_vals, int is_usage) /* Only make more free room if we do not have enough */ if (spills_needed > 0) { -#ifndef PLACE_SPILLS - ir_node *curr_bb = get_nodes_block(instr); - workset_t *ws_start = get_block_info(curr_bb)->start_workset; -#endif - DB((dbg, DBG_DECIDE, " disposing %d values\n", spills_needed)); /* calculate current next-use distance for live values */ for (i = 0; i < len; ++i) { ir_node *val = workset_get_val(ws, i); - unsigned dist = get_distance(instr, instr_nr, val, !is_usage); + unsigned dist = get_distance(instr, val, !is_usage); workset_set_time(ws, i, dist); } @@ -374,22 +383,15 @@ static void displace(workset_t *new_vals, int is_usage) DB((dbg, DBG_DECIDE, " disposing node %+F (%u)\n", val, workset_get_time(ws, i))); -#ifdef PLACE_SPILLS - if(!USES_IS_INFINITE(ws->vals[i].time) && !ws->vals[i].spilled) { - ir_node *after_pos = sched_prev(instr); - be_add_spill(senv, val, after_pos); - } -#endif - -#ifndef PLACE_SPILLS - /* Logic for not needed live-ins: If a value is disposed - * before its first use, remove it from start workset - * We don't do this for phis though */ - if (!is_Phi(val) && ! ir_nodeset_contains(&used, val)) { - workset_remove(ws_start, val); - DB((dbg, DBG_DECIDE, " (and removing %+F from start workset)\n", val)); + if (move_spills) { + if (!USES_IS_INFINITE(ws->vals[i].time) + && !ws->vals[i].spilled) { + ir_node *after_pos = sched_prev(instr); + DB((dbg, DBG_DECIDE, "Spill %+F after node %+F\n", val, + after_pos)); + be_add_spill(senv, val, after_pos); + } } -#endif } /* kill the last 'demand' entries in the array */ @@ -404,6 +406,64 @@ static void displace(workset_t *new_vals, int is_usage) } } +enum { + AVAILABLE_EVERYWHERE, + AVAILABLE_NOWHERE, + AVAILABLE_PARTLY, + AVAILABLE_UNKNOWN +}; + +static unsigned available_in_all_preds(workset_t* const* pred_worksets, + size_t n_pred_worksets, + const ir_node *value, bool is_local_phi) +{ + size_t i; + bool avail_everywhere = true; + bool avail_nowhere = true; + + assert(n_pred_worksets > 0); + + /* value available in all preds? */ + for (i = 0; i < n_pred_worksets; ++i) { + bool found = false; + const workset_t *p_workset = pred_worksets[i]; + int p_len = workset_get_length(p_workset); + int p_i; + const ir_node *l_value; + + if (is_local_phi) { + assert(is_Phi(value)); + l_value = get_irn_n(value, i); + } else { + l_value = value; + } + + for (p_i = 0; p_i < p_len; ++p_i) { + const loc_t *p_l = &p_workset->vals[p_i]; + if (p_l->node != l_value) + continue; + + found = true; + break; + } + + if (found) { + avail_nowhere = false; + } else { + avail_everywhere = false; + } + } + + if (avail_everywhere) { + assert(!avail_nowhere); + return AVAILABLE_EVERYWHERE; + } else if (avail_nowhere) { + return AVAILABLE_NOWHERE; + } else { + return AVAILABLE_PARTLY; + } +} + /** Decides whether a specific node should be in the start workset or not * * @param env belady environment @@ -412,7 +472,7 @@ static void displace(workset_t *new_vals, int is_usage) * @param loop the loop of the node */ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, - ir_loop *loop) + ir_loop *loop, unsigned available) { be_next_use_t next_use; loc_t loc; @@ -421,33 +481,38 @@ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, loc.node = node; loc.spilled = false; - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) { - loc.time = USES_INFINITY; - return loc; - } - /* We have to keep nonspillable nodes in the workingset */ - if(arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) { + if (arch_get_irn_flags(skip_Proj_const(node)) & arch_irn_flags_dont_spill) { loc.time = 0; DB((dbg, DBG_START, " %+F taken (dontspill node)\n", node, loc.time)); return loc; } - next_use = be_get_next_use(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 + next_use = be_get_next_use(uses, first, 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; DB((dbg, DBG_START, " %+F not taken (dead)\n", node)); - if(is_Phi(node)) { - be_spill_phi(senv, node); - } return loc; } loc.time = next_use.time; - if(next_use.outermost_loop >= get_loop_depth(loop)) { + if (improve_known_preds) { + if (available == AVAILABLE_EVERYWHERE) { + DB((dbg, DBG_START, " %+F taken (%u, live in all preds)\n", + node, loc.time)); + return loc; + } else if (available == AVAILABLE_NOWHERE) { + DB((dbg, DBG_START, " %+F not taken (%u, live in no pred)\n", + node, loc.time)); + loc.time = USES_INFINITY; + return loc; + } + } + + if (!respectloopdepth || next_use.outermost_loop >= get_loop_depth(loop)) { DB((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop)); } else { @@ -455,6 +520,7 @@ static loc_t to_take_or_not_to_take(ir_node* first, ir_node *node, DB((dbg, DBG_START, " %+F delayed (outerdepth %d < loopdepth %d)\n", node, next_use.outermost_loop, get_loop_depth(loop))); } + return loc; } @@ -468,15 +534,33 @@ static void decide_start_workset(const ir_node *block) { ir_loop *loop = get_irn_loop(block); ir_node *first; - ir_node *node; loc_t loc; loc_t *starters; loc_t *delayed; - int i, len, ws_count; - int free_slots, free_pressure_slots; + unsigned len; + unsigned i; + unsigned ws_count; + int free_slots, free_pressure_slots; unsigned pressure; int arity; workset_t **pred_worksets; + bool all_preds_known; + + /* check predecessors */ + arity = get_irn_arity(block); + pred_worksets = ALLOCAN(workset_t*, arity); + all_preds_known = true; + for (int in = 0; in < arity; ++in) { + ir_node *pred_block = get_Block_cfgpred_block(block, in); + block_info_t *pred_info = get_block_info(pred_block); + + if (pred_info == NULL) { + pred_worksets[in] = NULL; + all_preds_known = false; + } else { + pred_worksets[in] = pred_info->end_workset; + } + } /* Collect all values living at start of block */ starters = NEW_ARR_F(loc_t, 0); @@ -487,24 +571,41 @@ static void decide_start_workset(const ir_node *block) /* check all Phis first */ sched_foreach(block, node) { + unsigned available; + if (! is_Phi(node)) break; + if (!arch_irn_consider_in_reg_alloc(cls, node)) + continue; + + if (all_preds_known) { + available = available_in_all_preds(pred_worksets, arity, node, true); + } else { + available = AVAILABLE_UNKNOWN; + } - loc = to_take_or_not_to_take(first, node, loop); + loc = to_take_or_not_to_take(first, node, loop, available); 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); + } else { + be_spill_phi(senv, node); } } /* check all Live-Ins */ - be_lv_foreach(lv, block, be_lv_state_in, i) { - ir_node *node = be_lv_get_irn(lv, block, i); + be_lv_foreach_cls(lv, block, be_lv_state_in, cls, node) { + unsigned available; + if (all_preds_known) { + available = available_in_all_preds(pred_worksets, arity, node, false); + } else { + available = AVAILABLE_UNKNOWN; + } - loc = to_take_or_not_to_take(first, node, loop); + loc = to_take_or_not_to_take(first, node, loop, available); if (! USES_IS_INFINITE(loc.time)) { if (USES_IS_PENDING(loc.time)) @@ -515,7 +616,7 @@ static void decide_start_workset(const ir_node *block) } pressure = be_get_loop_pressure(loop_ana, cls, loop); - assert(ARR_LEN(delayed) <= (signed)pressure); + assert(ARR_LEN(delayed) <= pressure); free_slots = n_regs - ARR_LEN(starters); free_pressure_slots = n_regs - (pressure - ARR_LEN(delayed)); free_slots = MIN(free_slots, free_pressure_slots); @@ -523,34 +624,41 @@ static void decide_start_workset(const ir_node *block) /* so far we only put nodes into the starters list that are used inside * the loop. If register pressure in the loop is low then we can take some * values and let them live through the loop */ + DB((dbg, DBG_START, "Loop pressure %d, taking %d delayed vals\n", + pressure, free_slots)); if (free_slots > 0) { + size_t i; + qsort(delayed, ARR_LEN(delayed), sizeof(delayed[0]), loc_compare); - for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) { + for (i = 0; i < ARR_LEN(delayed) && free_slots > 0; ++i) { int p, arity; loc_t *loc = & delayed[i]; - /* don't use values which are dead in a known predecessors - * to not induce unnecessary reloads */ - arity = get_irn_arity(block); - for (p = 0; p < arity; ++p) { - ir_node *pred_block = get_Block_cfgpred_block(block, p); - block_info_t *pred_info = get_block_info(pred_block); - - if (pred_info == NULL) - continue; - - if (!workset_contains(pred_info->end_workset, loc->node)) { - DB((dbg, DBG_START, - " delayed %+F not live at pred %+F\n", loc->node, - pred_block)); - goto skip_delayed; + if (!is_Phi(loc->node)) { + /* don't use values which are dead in a known predecessors + * to not induce unnecessary reloads */ + arity = get_irn_arity(block); + for (p = 0; p < arity; ++p) { + ir_node *pred_block = get_Block_cfgpred_block(block, p); + block_info_t *pred_info = get_block_info(pred_block); + + if (pred_info == NULL) + continue; + + if (!workset_contains(pred_info->end_workset, loc->node)) { + DB((dbg, DBG_START, + " delayed %+F not live at pred %+F\n", loc->node, + pred_block)); + goto skip_delayed; + } } } DB((dbg, DBG_START, " delayed %+F taken\n", loc->node)); ARR_APP1(loc_t, starters, *loc); loc->node = NULL; + --free_slots; skip_delayed: ; } @@ -558,9 +666,10 @@ static void decide_start_workset(const ir_node *block) /* spill phis (the actual phis not just their values) that are in this block * but not in the start workset */ - for (i = ARR_LEN(delayed) - 1; i >= 0; --i) { + len = ARR_LEN(delayed); + for (i = 0; i < len; ++i) { ir_node *node = delayed[i].node; - if(node == NULL || !is_Phi(node) || get_nodes_block(node) != block) + if (node == NULL || !is_Phi(node) || get_nodes_block(node) != block) continue; DB((dbg, DBG_START, " spilling delayed phi %+F\n", node)); @@ -572,7 +681,7 @@ static void decide_start_workset(const ir_node *block) 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), n_regs); + ws_count = MIN((unsigned) ARR_LEN(starters), n_regs); workset_clear(ws); workset_bulk_fill(ws, ws_count, starters); @@ -593,26 +702,14 @@ static void decide_start_workset(const ir_node *block) /* determine spill status of the values: If there's 1 pred block (which * is no backedge) where the value is spilled then we must set it to * spilled here. */ - arity = get_irn_arity(block); - pred_worksets = alloca(sizeof(pred_worksets[0]) * arity); - for(i = 0; i < arity; ++i) { - ir_node *pred_block = get_Block_cfgpred_block(block, i); - block_info_t *pred_info = get_block_info(pred_block); - - if(pred_info == NULL) - pred_worksets[i] = NULL; - else - pred_worksets[i] = pred_info->end_workset; - } - - for(i = 0; i < ws_count; ++i) { + for (i = 0; i < ws_count; ++i) { loc_t *loc = &ws->vals[i]; ir_node *value = loc->node; bool spilled; int n; /* phis from this block aren't spilled */ - if(get_nodes_block(value) == block) { + if (get_nodes_block(value) == block) { assert(is_Phi(value)); loc->spilled = false; continue; @@ -620,7 +717,7 @@ static void decide_start_workset(const ir_node *block) /* determine if value was spilled on any predecessor */ spilled = false; - for(n = 0; n < arity; ++n) { + for (n = 0; n < arity; ++n) { workset_t *pred_workset = pred_worksets[n]; int p_len; int p; @@ -629,7 +726,7 @@ static void decide_start_workset(const ir_node *block) continue; p_len = workset_get_length(pred_workset); - for(p = 0; p < p_len; ++p) { + for (p = 0; p < p_len; ++p) { loc_t *l = &pred_workset->vals[p]; if (l->node != value) @@ -646,127 +743,35 @@ static void decide_start_workset(const ir_node *block) } } -#if 0 -static void decide_start_workset2(const ir_node *block) -{ - int arity; - workset_t **pred_worksets; - int p; - int len; - - /* check if all predecessors are known */ - arity = get_irn_arity(block); - pred_worksets = alloca(sizeof(pred_worksets[0]) * arity); - for (i = 0; i < arity; ++i) { - ir_node *pred_block = get_Block_cfgpred_block(block, i); - block_info_t *pred_info = get_block_info(pred_block); - - if (pred_info == NULL) { - /* not all predecessors known, use decide_start_workset */ - decide_start_workset(block); - return; - } - - pred_worksets[i] = pred_info->end_workset; - } - - /* take values live in all pred blocks */ - len = workset_get_length(pred_workset[0]); - for (p = 0; p < p_len; ++p) { - const loc_t *l = &pred_workset[0]->vals[p]; - ir_node *value; - bool spilled = false; - - if (USES_IS_INFINITE(l->time)) - continue; - - /* value available in all preds? */ - value = l->node; - for (i = 0; i < arity; ++i) { - bool found = false; - workset_t p_workset = &pred_worksets[i]; - int p_len = workset_get_length(p_workset); - int p_i; - - for (p_i = 0; p_i < p_len; ++p_i) { - const loc_t *p_l = &p_workset->vals[p_i]; - if (p_l->node != value) - continue; - - found = true; - if (p_l->spilled) - spilled = true; - break; - } - - if (!found) - break; - } - - /* it was available in all preds, TODO: insert spills... */ - if (i >= arity) { - workset_insert(ws, value, spilled); - } - } - - - /* Copy the best ones from starters to start workset */ - ws_count = MIN(ARR_LEN(starters), n_regs); - workset_clear(ws); - workset_bulk_fill(ws, ws_count, starters); -} - -#endif - /** * For the given block @p block, decide for each values * whether it is used from a register or is reloaded * before the use. */ -static void belady(ir_node *block) +static void process_block(ir_node *block) { - workset_t *new_vals; - ir_node *irn; - int iter; - block_info_t *block_info; - int i, arity; - int has_backedges = 0; - //int first = 0; - const ir_edge_t *edge; + workset_t *new_vals; + unsigned iter; + block_info_t *block_info; + int arity; /* no need to process a block twice */ - if(get_block_info(block) != NULL) { - return; - } + assert(get_block_info(block) == NULL); - /* check if all predecessor blocks are processed yet (though for backedges - * we have to make an exception as we can't process them first) */ + /* construct start workset */ arity = get_Block_n_cfgpreds(block); - for(i = 0; i < arity; ++i) { - ir_node *pred_block = get_Block_cfgpred_block(block, i); - block_info_t *pred_info = get_block_info(pred_block); - - if(pred_info == NULL) { - /* process predecessor first (it will be in the queue already) */ - if(!is_backedge(block, i)) { - return; - } - has_backedges = 1; - } - } - (void) has_backedges; - if(arity == 0) { + if (arity == 0) { + /* no predecessor -> empty set */ workset_clear(ws); - } else if(arity == 1) { + } else if (arity == 1) { + /* one predecessor, copy its end workset */ ir_node *pred_block = get_Block_cfgpred_block(block, 0); block_info_t *pred_info = get_block_info(pred_block); assert(pred_info != NULL); workset_copy(ws, pred_info->end_workset); } else { - /* we need 2 heuristics here, for the case when all predecessor blocks - * are known and when some are backedges (and therefore can't be known - * yet) */ + /* multiple predecessors, do more advanced magic :) */ decide_start_workset(block); } @@ -777,22 +782,21 @@ static void belady(ir_node *block) set_block_info(block, block_info); DB((dbg, DBG_WSETS, "Start workset for %+F:\n", block)); - workset_foreach(ws, irn, iter) { - DB((dbg, DBG_WSETS, " %+F (%u)\n", irn, - workset_get_time(ws, iter))); + { + ir_node *irn; + workset_foreach(ws, irn, iter) { + DB((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(ws, iter))); + } } block_info->start_workset = workset_clone(ws); /* process the block from start to end */ DB((dbg, DBG_WSETS, "Processing...\n")); - ir_nodeset_init(&used); - instr_nr = 0; /* TODO: this leaks (into the obstack)... */ new_vals = new_workset(); sched_foreach(block, irn) { - int i, arity; assert(workset_get_length(ws) <= n_regs); /* Phis are no real instr (see insert_starters()) */ @@ -801,54 +805,30 @@ static void belady(ir_node *block) } DB((dbg, DBG_DECIDE, " ...%+F\n", irn)); - /* set instruction in the workset */ - instr = irn; - /* allocate all values _used_ by this instruction */ workset_clear(new_vals); - for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) { - ir_node *in = get_irn_n(irn, i); - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in)) - continue; - + be_foreach_use(irn, cls, in_req_, in, in_req, /* (note that "spilled" is irrelevant here) */ workset_insert(new_vals, in, false); - } - displace(new_vals, 1); + ); + displace(new_vals, 1, irn); /* allocate all values _defined_ by this instruction */ workset_clear(new_vals); - if (get_irn_mode(irn) == mode_T) { - const ir_edge_t *edge; - - foreach_out_edge(irn, edge) { - ir_node *proj = get_edge_src_irn(edge); - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, proj)) - continue; - workset_insert(new_vals, proj, false); - } - } else { - if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn)) - continue; - workset_insert(new_vals, irn, false); - } - displace(new_vals, 0); - - instr_nr++; + be_foreach_definition(irn, cls, value, req, + assert(req->width == 1); + workset_insert(new_vals, value, false); + ); + displace(new_vals, 0, irn); } - ir_nodeset_destroy(&used); /* Remember end-workset for this block */ block_info->end_workset = workset_clone(ws); DB((dbg, DBG_WSETS, "End workset for %+F:\n", block)); - workset_foreach(ws, irn, iter) - DB((dbg, DBG_WSETS, " %+F (%u)\n", irn, - workset_get_time(ws, iter))); - - /* add successor blocks into worklist */ - foreach_block_succ(block, edge) { - ir_node *succ = get_edge_src_irn(edge); - pdeq_putr(worklist, succ); + { + ir_node *irn; + workset_foreach(ws, irn, iter) + DB((dbg, DBG_WSETS, " %+F (%u)\n", irn, workset_get_time(ws, iter))); } } @@ -859,19 +839,23 @@ static void belady(ir_node *block) */ static void fix_block_borders(ir_node *block, void *data) { - workset_t *start_workset; - int arity; - int i; - int iter; + workset_t *start_workset; + int arity; + int i; + unsigned iter; (void) data; DB((dbg, DBG_FIX, "\n")); DB((dbg, DBG_FIX, "Fixing %+F\n", block)); + arity = get_irn_arity(block); + /* can happen for endless loops */ + if (arity == 0) + return; + start_workset = get_block_info(block)->start_workset; /* process all pred blocks */ - arity = get_irn_arity(block); for (i = 0; i < arity; ++i) { ir_node *pred = get_Block_cfgpred_block(block, i); workset_t *pred_end_workset = get_block_info(pred)->end_workset; @@ -882,10 +866,10 @@ static void fix_block_borders(ir_node *block, void *data) /* spill all values not used anymore */ workset_foreach(pred_end_workset, node, iter) { ir_node *n2; - int iter2; + unsigned iter2; bool found = false; workset_foreach(start_workset, n2, iter2) { - if(n2 == node) { + if (n2 == node) { found = true; break; } @@ -898,8 +882,7 @@ static void fix_block_borders(ir_node *block, void *data) if (found) continue; -#ifdef PLACE_SPILLS - if(be_is_live_in(lv, block, node) + if (move_spills && be_is_live_in(lv, block, node) && !pred_end_workset->vals[iter].spilled) { ir_node *insert_point; if (arity > 1) { @@ -912,7 +895,6 @@ static void fix_block_borders(ir_node *block, void *data) insert_point)); be_add_spill(senv, node, insert_point); } -#endif } /* reload missing values in predecessors, add missing spills */ @@ -922,21 +904,20 @@ static void fix_block_borders(ir_node *block, void *data) /* if node is a phi of the current block we reload * the corresponding argument, else node itself */ - if(is_Phi(node) && get_nodes_block(node) == block) { + if (is_Phi(node) && get_nodes_block(node) == block) { node = get_irn_n(node, i); assert(!l->spilled); /* we might have unknowns as argument for the phi */ - if(!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) + if (!arch_irn_consider_in_reg_alloc(cls, node)) continue; } /* check if node is in a register at end of pred */ pred_loc = workset_contains(pred_end_workset, node); if (pred_loc != NULL) { -#ifdef PLACE_SPILLS /* we might have to spill value on this path */ - if (!pred_loc->spilled && l->spilled) { + if (move_spills && !pred_loc->spilled && l->spilled) { ir_node *insert_point = be_get_end_of_block_insertion_point(pred); insert_point = sched_prev(insert_point); @@ -944,7 +925,6 @@ static void fix_block_borders(ir_node *block, void *data) insert_point)); be_add_spill(senv, node, insert_point); } -#endif } else { /* node is not in register at the end of pred -> reload it */ DB((dbg, DBG_FIX, " reload %+F\n", node)); @@ -955,45 +935,50 @@ static void fix_block_borders(ir_node *block, void *data) } } -static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) +static void be_spill_belady(ir_graph *irg, const arch_register_class_t *rcls) { - ir_graph *irg = be_get_birg_irg(birg); + int i; - be_liveness_assure_sets(be_assure_liveness(birg)); + be_assure_live_sets(irg); - /* construct control flow loop tree */ - if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) { - construct_cf_backedges(irg); - } + stat_ev_tim_push(); + assure_loopinfo(irg); + stat_ev_tim_pop("belady_time_backedges"); + stat_ev_tim_push(); be_clear_links(irg); + stat_ev_tim_pop("belady_time_clear_links"); + + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); /* init belady env */ + stat_ev_tim_push(); obstack_init(&obst); - arch_env = birg->main_env->arch_env; - cls = rcls; - lv = be_get_birg_liveness(birg); - n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL); - ws = new_workset(); - uses = be_begin_uses(irg, lv); - loop_ana = be_new_loop_pressure(birg); - senv = be_new_spill_env(birg); - worklist = new_pdeq(); - - pdeq_putr(worklist, get_irg_start_block(irg)); - - while(!pdeq_empty(worklist)) { - ir_node *block = pdeq_getl(worklist); - belady(block); + cls = rcls; + lv = be_get_irg_liveness(irg); + n_regs = be_get_n_allocatable_regs(irg, cls); + ws = new_workset(); + uses = be_begin_uses(irg, lv); + loop_ana = be_new_loop_pressure(irg, cls); + senv = be_new_spill_env(irg); + blocklist = be_get_cfgpostorder(irg); + stat_ev_tim_pop("belady_time_init"); + + stat_ev_tim_push(); + /* walk blocks in reverse postorder */ + for (i = ARR_LEN(blocklist) - 1; i >= 0; --i) { + process_block(blocklist[i]); } - /* end block might not be reachable in endless loops */ - belady(get_irg_end_block(irg)); - - del_pdeq(worklist); + DEL_ARR_F(blocklist); + stat_ev_tim_pop("belady_time_belady"); + stat_ev_tim_push(); /* belady was block-local, fix the global flow by adding reloads on the * edges */ irg_block_walk_graph(irg, fix_block_borders, NULL, NULL); + stat_ev_tim_pop("belady_time_fix_borders"); + + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); /* Insert spill/reload nodes into the graph and fix usages */ be_insert_spills_reloads(senv); @@ -1005,14 +990,16 @@ static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *rcls) obstack_free(&obst, NULL); } +BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady) void be_init_spillbelady(void) { static be_spiller_t belady_spiller = { be_spill_belady }; + lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be"); + lc_opt_entry_t *belady_group = lc_opt_get_grp(be_grp, "belady"); + lc_opt_add_table(belady_group, options); be_register_spiller("belady", &belady_spiller); FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady"); } - -BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);