#include "config.h"
#endif
+#include <stdbool.h>
+
#include "obst.h"
#include "irprintf_t.h"
#include "irgraph.h"
(see bespill.h be_get_reload_costs_no_weight) */
#define RELOAD_COST_FACTOR 10
-typedef enum {
- value_not_reloaded, /* the value has not been reloaded */
- value_partially_reloaded, /* the value has been reloaded on some paths */
- value_reloaded /* the value has been reloaded on all paths */
-} reloaded_state_t;
+#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). */
- reloaded_state_t reloaded; /**< the value is a reloaded value */
+ bool spilled; /**< the value was already spilled on this path */
} loc_t;
typedef struct _workset_t {
* Inserts the value @p val into the workset, iff it is not
* already contained. The workset must not be full.
*/
-static void workset_insert(workset_t *workset, ir_node *val, int reloaded)
+static void workset_insert(workset_t *workset, ir_node *val, bool spilled)
{
loc_t *loc;
int i;
for (i = 0; i < workset->len; ++i) {
loc = &workset->vals[i];
if (loc->node == val) {
- if(!loc->reloaded) {
- loc->reloaded = reloaded;
+ if (spilled) {
+ loc->spilled = true;
}
return;
}
assert(workset->len < n_regs && "Workset already full!");
loc = &workset->vals[workset->len];
loc->node = val;
- loc->reloaded = reloaded;
- loc->time = 6666; /* undefined yet */
+ loc->spilled = spilled;
+ loc->time = TIME_UNDEFINED;
workset->len++;
}
}
}
-static INLINE int workset_contains(const workset_t *ws, const ir_node *val)
+static INLINE const loc_t *workset_contains(const workset_t *ws,
+ const ir_node *val)
{
int i;
- for(i=0; i<ws->len; ++i) {
+ for (i = 0; i < ws->len; ++i) {
if (ws->vals[i].node == val)
- return 1;
+ return &ws->vals[i];
}
- return 0;
+ return NULL;
}
/**
assert(! (flags & arch_irn_flags_ignore));
use = be_get_next_use(uses, from, from_step, def, skip_from_uses);
- if(USES_IS_INFINITE(use.time))
+ if (USES_IS_INFINITE(use.time))
return USES_INFINITY;
/* We have to keep nonspillable nodes in the workingset */
- if(flags & arch_irn_flags_dont_spill)
+ if (flags & arch_irn_flags_dont_spill)
return 0;
costs = be_get_reload_costs_no_weight(senv, def, use.before);
static void displace(workset_t *new_vals, int is_usage)
{
ir_node **to_insert = alloca(n_regs * sizeof(to_insert[0]));
+ bool *spilled = alloca(n_regs * sizeof(spilled[0]));
ir_node *val;
int i;
int len;
/* 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 (is_usage) {
DB((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, instr));
be_add_reload(senv, val, instr, cls, 1);
+ reloaded = true;
}
} else {
DB((dbg, DBG_DECIDE, " %+F already in workset\n", val));
* spilled */
workset_remove(ws, val);
}
- to_insert[demand++] = val;
+ spilled[demand] = reloaded;
+ to_insert[demand] = val;
+ ++demand;
}
/* 2. Make room for at least 'demand' slots */
/* 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));
/* sort entries by increasing nextuse-distance*/
workset_sort(ws);
- /* Logic for not needed live-ins: If a value is disposed
- * before its first usage, remove it from start workset
- * We don't do this for phis though */
for (i = len - spills_needed; i < len; ++i) {
ir_node *val = ws->vals[i].node;
DB((dbg, DBG_DECIDE, " disposing node %+F (%u)\n", val,
workset_get_time(ws, i)));
- if(!USES_IS_INFINITE(ws->vals[i].time)
- && !ws->vals[i].reloaded) {
- //be_add_spill(senv, val, instr);
+#ifdef PLACE_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
+#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));
}
+#endif
}
/* kill the last 'demand' entries in the array */
for (i = 0; i < demand; ++i) {
ir_node *val = to_insert[i];
- workset_insert(ws, val, 1);
+ workset_insert(ws, val, spilled[i]);
+ }
+}
+
+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;
}
}
* @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;
- loc.time = USES_INFINITY;
- loc.node = node;
- //loc.reloaded = rand() % 2; /* provoke a bug... */
- loc.reloaded = 0;
+ loc.time = USES_INFINITY;
+ loc.node = node;
+ loc.spilled = false;
if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node)) {
loc.time = USES_INFINITY;
}
/* We have to keep nonspillable nodes in the workingset */
- if(arch_irn_get_flags(arch_env, node) & arch_irn_flags_dont_spill) {
+ if (arch_irn_get_flags(arch_env, 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)) {
+ 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 (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 (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 {
* beginning of a loop. We try to reload as much values as possible now so they
* don't get reloaded inside the loop.
*/
-static void compute_live_ins(const ir_node *block)
+static void decide_start_workset(const ir_node *block)
{
ir_loop *loop = get_irn_loop(block);
ir_node *first;
int i, len, ws_count;
int free_slots, free_pressure_slots;
unsigned pressure;
- //int arity;
- //int n_pred_worksets;
- //workset_t **pred_worksets;
+ int arity;
+ workset_t **pred_worksets;
+ bool all_preds_known;
+
+ /* check predecessors */
+ arity = get_irn_arity(block);
+ pred_worksets = alloca(sizeof(pred_worksets[0]) * arity);
+ all_preds_known = true;
+ 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;
+ all_preds_known = false;
+ } else {
+ pred_worksets[i] = pred_info->end_workset;
+ }
+ }
/* Collect all values living at start of block */
starters = NEW_ARR_F(loc_t, 0);
/* check all Phis first */
sched_foreach(block, node) {
+ unsigned available;
+
if (! is_Phi(node))
break;
+ if (!arch_irn_consider_in_reg_alloc(arch_env, cls, node))
+ continue;
- loc = to_take_or_not_to_take(first, node, loop);
+ 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, 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);
+ unsigned available;
- loc = to_take_or_not_to_take(first, node, loop);
+ 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, available);
if (! USES_IS_INFINITE(loc.time)) {
if (USES_IS_PENDING(loc.time))
/* 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) {
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];
DB((dbg, DBG_START, " delayed %+F taken\n", loc->node));
ARR_APP1(loc_t, starters, *loc);
loc->node = NULL;
+ --free_slots;
skip_delayed:
;
}
* but not in the start workset */
for (i = ARR_LEN(delayed) - 1; i >= 0; --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));
DEL_ARR_F(starters);
-#if 0
- /* determine reloaded status of the values: If there's 1 pred block (which
- * is no backedge) where the value is reloaded then we must set it to
- * reloaded here. We place spills in all pred where the value was not yet
- * reloaded to be sure we have a spill on each path */
- n_pred_worksets = 0;
- 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)
- continue;
-
- pred_worksets[n_pred_worksets] = pred_info->end_workset;
- ++n_pred_worksets;
- }
-
+ /* 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. */
for(i = 0; i < ws_count; ++i) {
- loc_t *loc = &ws->vals[i];
- ir_node *value = loc->node;
- int reloaded;
+ loc_t *loc = &ws->vals[i];
+ ir_node *value = loc->node;
+ bool spilled;
int n;
- /* phis from this block aren't reloaded */
- if(get_nodes_block(value) == block) {
+ /* phis from this block aren't spilled */
+ if (get_nodes_block(value) == block) {
assert(is_Phi(value));
- loc->reloaded = value_not_reloaded;
+ loc->spilled = false;
continue;
}
- /* was the value reloaded on any of the other inputs */
- reloaded = 0;
- arity = get_Block_n_cfgpreds(block);
- for(n = 0; n < n_pred_worksets; ++n) {
+ /* determine if value was spilled on any predecessor */
+ spilled = false;
+ for(n = 0; n < arity; ++n) {
workset_t *pred_workset = pred_worksets[n];
- int p_len = workset_get_length(pred_workset);
+ int p_len;
int p;
+ if (pred_workset == NULL)
+ continue;
+
+ p_len = workset_get_length(pred_workset);
for(p = 0; p < p_len; ++p) {
loc_t *l = &pred_workset->vals[p];
- if(l->node == value) {
- if(l->reloaded) {
- reloaded = 1;
- }
- break;
+
+ if (l->node != value)
+ continue;
+
+ if (l->spilled) {
+ spilled = true;
}
+ break;
}
- if(p >= p_len) {
- reloaded = 1;
+ }
+
+ loc->spilled = spilled;
+ }
+}
+
+#if 0
+static void decide_start_workset2(const ir_node *block)
+{
+ int arity;
+ workset_t **pred_worksets;
+ int p;
+ int i;
+ 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;
+ }
+
+ /* we construct a new workset */
+ workset_clear(ws);
+
+ /* take values live in all pred blocks */
+ len = workset_get_length(pred_worksets[0]);
+ for (p = 0; p < len; ++p) {
+ const loc_t *l = &pred_worksets[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 = 1; 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 */
+ if (i >= arity) {
+ workset_insert(ws, value, spilled);
}
}
-#endif
+
+ /* Copy the best ones from starters to start workset */
+ ws_count = MIN(ARR_LEN(starters), n_regs);
+ workset_bulk_fill(ws, ws_count, starters);
}
+#endif
/**
* For the given block @p block, decide for each values
const ir_edge_t *edge;
/* no need to process a block twice */
- if(get_block_info(block) != NULL) {
+ if (get_block_info(block) != NULL) {
return;
}
ir_node *pred_block = get_Block_cfgpred_block(block, i);
block_info_t *pred_info = get_block_info(pred_block);
- if(pred_info == NULL) {
+ if (pred_info == NULL) {
/* process predecessor first (it will be in the queue already) */
- if(!is_backedge(block, i)) {
+ if (!is_backedge(block, i)) {
return;
}
has_backedges = 1;
}
}
(void) has_backedges;
- if(arity == 0) {
+ if (arity == 0) {
workset_clear(ws);
- } else if(arity == 1) {
+ } else if (arity == 1) {
ir_node *pred_block = get_Block_cfgpred_block(block, 0);
block_info_t *pred_info = get_block_info(pred_block);
/* 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) */
- compute_live_ins(block);
+ decide_start_workset(block);
}
DB((dbg, DBG_DECIDE, "\n"));
if (!arch_irn_consider_in_reg_alloc(arch_env, cls, in))
continue;
- /* (note that reloaded_value is irrelevant here) */
- workset_insert(new_vals, in, 0);
+ /* (note that "spilled" is irrelevant here) */
+ workset_insert(new_vals, in, false);
}
displace(new_vals, 1);
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, 0);
+ workset_insert(new_vals, proj, false);
}
} else {
if (!arch_irn_consider_in_reg_alloc(arch_env, cls, irn))
continue;
- workset_insert(new_vals, irn, 0);
+ workset_insert(new_vals, irn, false);
}
displace(new_vals, 0);
workset_foreach(pred_end_workset, node, iter) {
ir_node *n2;
int iter2;
- int found = 0;
+ bool found = false;
workset_foreach(start_workset, n2, iter2) {
- if(n2 == node) {
- found = 1;
+ if (n2 == node) {
+ found = true;
break;
}
/* note that we do not look at phi inputs, becuase the values
* workset */
}
-#if 0
- if(!found && be_is_live_out(lv, pred, node)
- && !pred_end_workset->vals[iter].reloaded) {
- ir_node *insert_point
- = be_get_end_of_block_insertion_point(pred);
- DB((dbg, DBG_SPILL, "Spill %+F before %+F\n", node,
+ if (found)
+ continue;
+
+#ifdef PLACE_SPILLS
+ if (be_is_live_in(lv, block, node)
+ && !pred_end_workset->vals[iter].spilled) {
+ ir_node *insert_point;
+ if (arity > 1) {
+ insert_point = be_get_end_of_block_insertion_point(pred);
+ insert_point = sched_prev(insert_point);
+ } else {
+ insert_point = block;
+ }
+ DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
insert_point));
be_add_spill(senv, node, insert_point);
}
#endif
}
- /* reload missing values in predecessors */
+ /* reload missing values in predecessors, add missing spills */
workset_foreach(start_workset, node, iter) {
+ const loc_t *l = &start_workset->vals[iter];
+ const loc_t *pred_loc;
+
/* 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)) {
+ 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(arch_env, cls, node))
continue;
}
/* check if node is in a register at end of pred */
- if(workset_contains(pred_end_workset, node))
- continue;
-
- /* node is not in memory at the end of pred -> reload it */
- DB((dbg, DBG_FIX, " reload %+F\n", node));
- DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
- be_add_reload_on_edge(senv, node, block, i, cls, 1);
+ 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) {
+ ir_node *insert_point
+ = be_get_end_of_block_insertion_point(pred);
+ insert_point = sched_prev(insert_point);
+ DB((dbg, DBG_SPILL, "Spill %+F after %+F\n", node,
+ 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));
+ DB((dbg, DBG_SPILL, "Reload %+F before %+F,%d\n", node, block, i));
+ be_add_reload_on_edge(senv, node, block, i, cls, 1);
+ }
}
}
}
be_liveness_assure_sets(be_assure_liveness(birg));
/* construct control flow loop tree */
- if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
+ if (! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
construct_cf_backedges(irg);
}
/* init belady env */
obstack_init(&obst);
- arch_env = &birg->main_env->arch_env;
+ 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);