-/**
- * Author: Daniel Grund, Matthias Braun
- * Date: 20.09.2005
- * Copyright: (c) Universitaet Karlsruhe
- * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
+/*
+ * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
*
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
+/**
+ * @file
+ * @brief Beladys spillalgorithm.
+ * @author Daniel Grund, Matthias Braun
+ * @date 20.09.2005
+ * @version $Id$
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
-#ifdef HAVE_ALLOCA_H
-#include <alloca.h>
-#endif
-
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
#include "obst.h"
#include "set.h"
#include "pset.h"
#include "iredges_t.h"
#include "ircons_t.h"
#include "irprintf.h"
+#include "xmalloc.h"
#include "beutil.h"
-#include "bearch.h"
+#include "bearch_t.h"
#include "bespillbelady.h"
-#include "beuses_t.h"
+#include "beuses.h"
#include "besched_t.h"
#include "beirgmod.h"
#include "belive_t.h"
#include "benode_t.h"
#include "bechordal_t.h"
+#include "bespilloptions.h"
+#include "beloopana.h"
+#include "beirg_t.h"
+#include "bemodule.h"
#define DBG_SPILL 1
#define DBG_WSETS 2
* An association between a node and a point in time.
*/
typedef struct _loc_t {
- ir_node *irn; /**< A node. */
- unsigned time; /**< A use time (see beuses.h). */
+ ir_node *irn; /**< A node. */
+ unsigned time; /**< A use time (see beuses.h). */
+ int reloaded_value; /**< the value is a reloaded value */
} loc_t;
typedef struct _workset_t {
typedef struct _belady_env_t {
struct obstack ob;
- const be_chordal_env_t *cenv;
const arch_env_t *arch;
const arch_register_class_t *cls;
+ be_lv_t *lv;
+ be_loopana_t *loop_ana;
int n_regs; /** number of regs in this reg-class */
workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
return p->time - q->time;
}
-void workset_print(const workset_t *w)
+/* debug helper */
+static void workset_print(const workset_t *w)
{
int i;
for(i = 0; i < w->len; ++i) {
- ir_fprintf(stderr, "%+F %d\n", w->vals[i].irn, w->vals[i].time);
+ ir_fprintf(stderr, "%+F %d (%d)\n", w->vals[i].irn, w->vals[i].time, w->vals[i].reloaded_value);
}
+ /* avoid unused warning */
+ (void) workset_print;
}
/**
* 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(belady_env_t *env, workset_t *ws, ir_node *val) {
+static INLINE void workset_insert(belady_env_t *env, workset_t *ws,
+ ir_node *val, int reloaded_value)
+{
int i;
/* check for current regclass */
if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, val)) {
- DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
+ //DBG((dbg, DBG_WORKSET, "Skipped %+F\n", val));
return;
}
/* check if val is already contained */
- for(i=0; i<ws->len; ++i)
- if (ws->vals[i].irn == val)
+ for(i=0; i<ws->len; ++i) {
+ if (ws->vals[i].irn == val) {
+ if(!ws->vals[i].reloaded_value)
+ ws->vals[i].reloaded_value = reloaded_value;
return;
+ }
+ }
/* insert val */
assert(ws->len < env->n_regs && "Workset already full!");
- ws->vals[ws->len++].irn = val;
+ ws->vals[ws->len].irn = val;
+ ws->vals[ws->len].reloaded_value = reloaded_value;
+ ws->len++;
}
/**
return use.time;
}
-#if 0
-/**
- * Fix to remove dead nodes (especially don't spill nodes) from workset.
- */
-static void fix_dead_values(workset_t *ws, ir_node *irn) {
- int idx;
- ir_node *node;
- ir_node *block = get_nodes_block(irn);
-
- DBG((dbg, DBG_DECIDE, "fixing dead values at %+F:\n", irn));
-
- workset_foreach(ws, node, idx) {
- const ir_edge_t *edge;
- int fixme = 1;
-
- /* skip already fixed nodes */
- if (workset_get_time(ws, idx) == INT_MAX)
- continue;
-
- /* check all users */
- foreach_out_edge(node, edge) {
- ir_node *user = get_edge_src_irn(edge);
-
- if ((get_nodes_block(user) != block) || /* user is in a different block */
- (sched_is_scheduled(user) && sched_comes_after(irn, user)) || /* user is scheduled after irn */
- user == irn) /* irn is the user */
- { /* => don't fix distance */
- fixme = 0;
- break;
- }
- }
-
- /* all users scheduled prior to current irn in in same block as irn -> fix */
- if (fixme) {
- workset_set_time(ws, idx, INT_MAX);
- DBG((dbg, DBG_DECIDE, "\tfixing time for %+F to INT_MAX\n", node));
- }
- }
-
-}
-#endif
-
/**
* Performs the actions necessary to grant the request that:
* - new_vals can be held in registers
if (! workset_contains(ws, val)) {
DBG((dbg, DBG_DECIDE, " insert %+F\n", val));
+
to_insert[demand++] = val;
if (is_usage) {
DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr));
- be_add_reload(env->senv, val, env->instr);
+ be_add_reload(env->senv, val, env->instr, env->cls, 1);
}
- }
- else {
- assert(is_usage || "Defined value already in workset?!?");
- DBG((dbg, DBG_DECIDE, " skip %+F\n", val));
+ } else {
+ DBG((dbg, DBG_DECIDE, " %+F already in workset\n", val));
+ assert(is_usage);
}
}
- DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
+ //DBG((dbg, DBG_DECIDE, " demand = %d\n", demand));
/*
2. Make room for at least 'demand' slots
len = workset_get_length(ws);
max_allowed = env->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) {
+ DBG((dbg, DBG_DECIDE, " disposing %d values\n", ws->len - max_allowed));
+
/* get current next-use distance */
for (i = 0; i < ws->len; ++i) {
unsigned dist = get_distance(env, env->instr, env->instr_nr, workset_get_val(ws, i), !is_usage);
workset_set_time(ws, i, dist);
}
-#if 0
- /*
- FIX for don't spill nodes:
- Problem is that get_distance always returns 0 for those nodes even if they are not
- needed anymore (all their usages have already been visited).
- Even if we change this behavior, get_distance doesn't distinguish between not
- used anymore (dead) and live out of block.
- Solution: Set distances of all nodes having all their usages in schedule prior to
- current instruction to MAX_INT.
- */
- fix_dead_values(ws, env->instr);
-#endif
-
/* sort entries by increasing nextuse-distance*/
workset_sort(ws);
for (i = max_allowed; i < ws->len; ++i) {
ir_node *irn = ws->vals[i].irn;
- DBG((dbg, DBG_DECIDE, " disposing %+F (%u)\n", irn, workset_get_time(ws, i)));
+ DBG((dbg, DBG_DECIDE, " disposing %+F (%u)\n", irn,
+ workset_get_time(ws, i)));
+
+ if(!USES_IS_INFINITE(ws->vals[i].time)
+ && !ws->vals[i].reloaded_value) {
+ be_add_spill(env->senv, irn, env->instr);
+ }
if (is_Phi(irn))
continue;
/*
3. Insert the new values into the workset
*/
- for (i = 0; i < demand; ++i)
- workset_insert(env, env->ws, to_insert[i]);
+ for (i = 0; i < demand; ++i) {
+ workset_insert(env, env->ws, to_insert[i], 1);
+ }
}
static void belady(ir_node *block, void *env);
-static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first, ir_node *node, ir_node *block, ir_loop *loop) {
+/** Decides whether a specific node should be in the start workset or not
+ *
+ * @param env belady environment
+ * @param first
+ * @param node the node to test
+ * @param block the block of the node
+ * @param loop the loop of the node
+ */
+static loc_t to_take_or_not_to_take(belady_env_t *env, ir_node* first,
+ ir_node *node, ir_node *block,
+ ir_loop *loop)
+{
be_next_use_t next_use;
loc_t loc;
loc.time = USES_INFINITY;
+ loc.irn = node;
+ (void) block;
if (!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node)) {
loc.time = USES_INFINITY;
return loc;
}
- loc.irn = node;
-
/* We have to keep nonspillable nodes in the workingset */
if(arch_irn_get_flags(env->arch, node) & arch_irn_flags_dont_spill) {
loc.time = 0;
next_use = be_get_next_use(env->uses, first, 0, node, 0);
if(USES_IS_INFINITE(next_use.time)) {
// the nodes marked as live in shouldn't be dead, so it must be a phi
+ assert(is_Phi(node));
loc.time = USES_INFINITY;
DBG((dbg, DBG_START, " %+F not taken (dead)\n", node));
- assert(is_Phi(node));
+ if(is_Phi(node)) {
+ be_spill_phi(env->senv, node);
+ }
return loc;
}
if(next_use.outermost_loop >= get_loop_depth(loop)) {
DBG((dbg, DBG_START, " %+F taken (%u, loop %d)\n", node, loc.time, next_use.outermost_loop));
- return loc;
- // ARR_APP1(loc_t, starters, loc);
} else {
- loc.time = USES_INFINITY;
- DBG((dbg, DBG_START, " %+F not taken (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
- return loc;
+ loc.time = USES_PENDING;
+ DBG((dbg, DBG_START, " %+F delayed (outerloopdepth %d < loopdetph %d)\n", node, next_use.outermost_loop, get_loop_depth(loop)));
}
+ return loc;
}
/*
* and notifies spill algorithm which phis need to be spilled
*/
static void compute_live_ins(ir_node *block, void *data) {
- belady_env_t *env = data;
- block_info_t *block_info;
- ir_node *first, *irn;
- loc_t loc, *starters;
- int i, len, ws_count;
- ir_loop *loop = get_irn_loop(block);
- be_lv_t *lv = env->cenv->birg->lv;
-
- if(get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
+ belady_env_t *env = data;
+ ir_loop *loop = get_irn_loop(block);
+ const be_lv_t *lv = env->lv;
+ block_info_t *block_info;
+ ir_node *first, *irn;
+ loc_t loc, *starters, *delayed;
+ int i, len, ws_count;
+ int free_slots, free_pressure_slots;
+ unsigned pressure;
+
+ if (get_Block_n_cfgpreds(block) == 1 && get_irg_start_block(get_irn_irg(block)) != block)
return;
block_info = new_block_info(&env->ob);
/* Collect all values living at start of block */
starters = NEW_ARR_F(loc_t, 0);
-
- /* rebuild schedule time information, because it seems to be broken */
- // Matze: is this still true?
- //sched_renumber(block);
+ delayed = NEW_ARR_F(loc_t, 0);
DBG((dbg, DBG_START, "Living at start of %+F:\n", block));
first = sched_first(block);
+ /* check all Phis first */
sched_foreach(block, irn) {
- if(!is_Phi(irn))
+ if (! is_Phi(irn))
break;
loc = to_take_or_not_to_take(env, first, irn, block, loop);
- if(!USES_IS_INFINITE(loc.time)) {
- ARR_APP1(loc_t, starters, loc);
- } else {
- be_spill_phi(env->senv, irn);
+ 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);
}
}
+ /* check all Live-Ins */
be_lv_foreach(lv, block, be_lv_state_in, i) {
ir_node *node = be_lv_get_irn(lv, block, i);
loc = to_take_or_not_to_take(env, first, node, block, loop);
- if(!USES_IS_INFINITE(loc.time)) {
- ARR_APP1(loc_t, starters, loc);
+ if (! USES_IS_INFINITE(loc.time)) {
+ if (USES_IS_PENDING(loc.time))
+ ARR_APP1(loc_t, delayed, loc);
+ else
+ ARR_APP1(loc_t, starters, loc);
+ }
+ }
+
+ pressure = be_get_loop_pressure(env->loop_ana, env->cls, loop);
+ assert(ARR_LEN(delayed) <= (signed)pressure);
+ free_slots = env->n_regs - ARR_LEN(starters);
+ free_pressure_slots = env->n_regs - (pressure - ARR_LEN(delayed));
+ free_slots = MIN(free_slots, free_pressure_slots);
+ /* append nodes delayed due to loop structure until start set is full */
+ for (i = 0; i < ARR_LEN(delayed) && i < free_slots; ++i) {
+ DBG((dbg, DBG_START, " delayed %+F taken\n", delayed[i].irn));
+ ARR_APP1(loc_t, starters, delayed[i]);
+ delayed[i].irn = NULL;
+ }
+
+ /* spill all delayed phis which didn't make it into start workset */
+ for ( ; i < ARR_LEN(delayed); ++i) {
+ ir_node *irn = delayed[i].irn;
+ if (irn && is_Phi(irn) && get_nodes_block(irn) == block) {
+ DBG((dbg, DBG_START, " spilling delayed phi %+F\n", irn));
+ be_spill_phi(env->senv, irn);
}
}
+ DEL_ARR_F(delayed);
- // Sort start values by first use
+ /* Sort start values by first use */
qsort(starters, ARR_LEN(starters), sizeof(starters[0]), loc_compare);
/* Copy the best ones from starters to start workset */
- ws_count = MIN(ARR_LEN(starters), env->n_regs);
+ ws_count = MIN(ARR_LEN(starters), env->n_regs);
block_info->ws_start = new_workset(env, &env->ob);
workset_bulk_fill(block_info->ws_start, ws_count, starters);
len = ARR_LEN(starters);
for (i = ws_count; i < len; ++i) {
irn = starters[i].irn;
- if (!is_Phi(irn) || get_nodes_block(irn) != block)
+ if (! is_Phi(irn) || get_nodes_block(irn) != block)
continue;
be_spill_phi(env->senv, irn);
/* allocate all values _used_ by this instruction */
workset_clear(new_vals);
for(i = 0, arity = get_irn_arity(irn); i < arity; ++i) {
- workset_insert(env, new_vals, get_irn_n(irn, i));
+ /* (note that reloaded_value is not interesting here) */
+ workset_insert(env, new_vals, get_irn_n(irn, i), 0);
}
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(env, new_vals, proj);
+ const ir_edge_t *edge;
+
+ foreach_out_edge(irn, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ workset_insert(env, new_vals, proj, 0);
+ }
} else {
- workset_insert(env, new_vals, irn);
+ workset_insert(env, new_vals, irn, 0);
}
displace(env, new_vals, 0);
* 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 *block, void *data) {
- belady_env_t *env = data;
- workset_t *wsb;
- ir_graph *irg = get_irn_irg(block);
- ir_node *startblock = get_irg_start_block(irg);
- int i, max, iter, iter2;
-
- if(block == startblock)
- return;
+static void fix_block_borders(ir_node *block, void *data)
+{
+ ir_graph *irg = get_irn_irg(block);
+ ir_node *startblock = get_irg_start_block(irg);
+ belady_env_t *env = data;
+ workset_t *start_workset;
+ int arity;
+ int i;
+ int iter;
+
+ if(block == startblock)
+ return;
DBG((dbg, DBG_FIX, "\n"));
DBG((dbg, DBG_FIX, "Fixing %+F\n", block));
- wsb = get_block_info(block)->ws_start;
+ start_workset = get_block_info(block)->ws_start;
/* process all pred blocks */
- for (i=0, max=get_irn_arity(block); i<max; ++i) {
- ir_node *irnb, *irnp, *pred = get_Block_cfgpred_block(block, i);
- workset_t *wsp = get_block_info(pred)->ws_end;
+ arity = get_irn_arity(block);
+ for (i = 0; i < arity; ++i) {
+ ir_node *pred = get_Block_cfgpred_block(block, i);
+ workset_t *workset_pred_end = get_block_info(pred)->ws_end;
+ ir_node *node;
DBG((dbg, DBG_FIX, " Pred %+F\n", pred));
- 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) && block == get_nodes_block(irnb)) {
- irnb = get_irn_n(irnb, i);
+ /* spill all values not used anymore */
+ workset_foreach(workset_pred_end, node, iter) {
+ ir_node *n2;
+ int iter2;
+ int found = 0;
+ workset_foreach(start_workset, n2, iter2) {
+ if(n2 == node) {
+ found = 1;
+ break;
+ }
+ /* note that we do not look at phi inputs, becuase the values
+ * will be either live-end and need no spill or
+ * they have other users in which must be somewhere else in the
+ * workset */
+ }
- // we might have unknowns as argument for the phi
- if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, irnb))
- continue;
+ if(!found && be_is_live_out(env->lv, pred, node)
+ && !workset_pred_end->vals[iter].reloaded_value) {
+ ir_node *insert_point
+ = be_get_end_of_block_insertion_point(pred);
+ DBG((dbg, DBG_SPILL, "Spill %+F before %+F\n", node,
+ insert_point));
+ be_add_spill(env->senv, node, insert_point);
}
+ }
- /* Unknowns are available everywhere */
- if(get_irn_opcode(irnb) == iro_Unknown)
- continue;
+ /* reload missing values in predecessors */
+ workset_foreach(start_workset, node, iter) {
+ /* if node is a phi of the current block we reload
+ * the corresponding argument, else node itself */
+ if(is_Phi(node) && block == get_nodes_block(node)) {
+ node = get_irn_n(node, i);
- /* check if irnb is in a register at end of pred */
- workset_foreach(wsp, irnp, iter2) {
- if (irnb == irnp)
- goto next_value;
+ /* we might have unknowns as argument for the phi */
+ if(!arch_irn_consider_in_reg_alloc(env->arch, env->cls, node))
+ continue;
}
- /* irnb is not in memory at the end of pred, so we have to reload it */
- DBG((dbg, DBG_FIX, " reload %+F\n", irnb));
- DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", irnb, block, i));
- be_add_reload_on_edge(env->senv, irnb, block, i);
+ /* check if node is in a register at end of pred */
+ if(workset_contains(workset_pred_end, node))
+ continue;
-next_value:
- /*epsilon statement :)*/;
+ /* node is not in memory at the end of pred -> reload it */
+ DBG((dbg, DBG_FIX, " reload %+F\n", node));
+ DBG((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
+ be_add_reload_on_edge(env->senv, node, block, i, env->cls, 1);
}
}
}
-void be_spill_belady(const be_chordal_env_t *chordal_env) {
- be_spill_belady_spill_env(chordal_env, NULL);
+/**
+ * Do spilling for a register class on a graph using the belady heuristic.
+ * In the transformed graph, the register pressure never exceeds the number
+ * of available registers.
+ *
+ * @param birg The backend graph
+ * @param cls The register class to spill
+ */
+static void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls) {
+ be_spill_belady_spill_env(birg, cls, NULL);
}
-void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
+void be_spill_belady_spill_env(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) {
belady_env_t env;
+ ir_graph *irg = be_get_birg_irg(birg);
+ int n_regs;
- FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
- //firm_dbg_set_mask(dbg, DBG_SPILL);
+ n_regs = cls->n_regs - be_put_ignore_regs(birg, cls, NULL);
+ be_liveness_assure_sets(be_assure_liveness(birg));
+
+ /* construct control flow loop tree */
+ if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
+ construct_cf_backedges(irg);
+ }
+
+ be_clear_links(irg);
/* init belady env */
obstack_init(&env.ob);
- env.cenv = chordal_env;
- env.arch = chordal_env->birg->main_env->arch_env;
- env.cls = chordal_env->cls;
- env.n_regs = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, chordal_env->cls, NULL);
+ env.arch = birg->main_env->arch_env;
+ env.cls = cls;
+ env.lv = be_get_birg_liveness(birg);
+ env.n_regs = n_regs;
env.ws = new_workset(&env, &env.ob);
- env.uses = be_begin_uses(chordal_env->irg, chordal_env->birg->lv);
+ env.uses = be_begin_uses(irg, env.lv);
+ env.loop_ana = be_new_loop_pressure(birg);
if(spill_env == NULL) {
- env.senv = be_new_spill_env(chordal_env);
+ env.senv = be_new_spill_env(birg);
} else {
env.senv = spill_env;
}
- DEBUG_ONLY(be_set_spill_env_dbg_module(env.senv, dbg);)
- be_clear_links(chordal_env->irg);
/* Decide which phi nodes will be spilled and place copies for them into the graph */
- irg_block_walk_graph(chordal_env->irg, compute_live_ins, NULL, &env);
+ irg_block_walk_graph(irg, compute_live_ins, NULL, &env);
/* Fix high register pressure with belady algorithm */
- irg_block_walk_graph(chordal_env->irg, NULL, belady, &env);
+ irg_block_walk_graph(irg, NULL, belady, &env);
/* belady was block-local, fix the global flow by adding reloads on the edges */
- irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &env);
+ irg_block_walk_graph(irg, fix_block_borders, NULL, &env);
+
+ be_end_uses(env.uses);
+ be_free_loop_pressure(env.loop_ana);
+ obstack_free(&env.ob, NULL);
+
/* Insert spill/reload nodes into the graph and fix usages */
be_insert_spills_reloads(env.senv);
/* clean up */
if(spill_env == NULL)
be_delete_spill_env(env.senv);
- be_end_uses(env.uses);
- obstack_free(&env.ob, NULL);
}
+
+void be_init_spillbelady(void)
+{
+ static be_spiller_t belady_spiller = {
+ be_spill_belady
+ };
+
+ be_register_spiller("belady", &belady_spiller);
+ FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);