/**
* @file
* @brief Beladys spillalgorithm version 2.
- * @author Daniel Grund, Matthias Braun, Sebastian Hack
+ * @author Sebastian Hack, Matthias Braun, Daniel Grund
* @date 01.08.2007
* @version $Id$
*
#include "ircons_t.h"
#include "irprintf.h"
#include "execfreq.h"
+#include "dfs_t.h"
#include "xmalloc.h"
#include "beutil.h"
#include "beirg_t.h"
#include "bemodule.h"
+#include <libcore/lc_opts.h>
+#include <libcore/lc_opts_enum.h>
+#include <libcore/lc_timing.h>
+
#define DBG_SPILL 1
#define DBG_WSETS 2
#define DBG_FIX 4
#define DBG_WORKSET 128
#define DBG_GLOBAL 256
-#define DEAD UINT_MAX
+#define ALREADY_SPILLED_FACTOR 2
+
+#define DEAD UINT_MAX
+#define LIVE_END (DEAD-1)
+#define REMAT_DIST (DEAD-2)
+
+static int already_spilled_factor = 2;
+static int remat_live_range_ext = 1;
+static int global_pass_enabled = 1;
+
+static const lc_opt_table_entry_t options[] = {
+ LC_OPT_ENT_INT ("asf", "already spilled factor", &already_spilled_factor),
+ LC_OPT_ENT_BOOL ("remat", "rematerializable ops get infinite long live ranges", &remat_live_range_ext),
+ LC_OPT_ENT_BOOL ("global", "enable/disable the global pass", &global_pass_enabled),
+ LC_OPT_LAST
+};
+
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
/**
*/
typedef struct _loc_t {
ir_node *irn; /**< A node. */
- unsigned time; /**< A use time (see beuses.h). */
- unsigned version; /**< That is used in the global pass below.
- For usage see the comments below.
- In the local belady pass, this is not
- important. */
+ unsigned time; /**< A use time.
+ In the global pass this is used
+ as the version number and not as a time.
+ Only to save space...
+ */
} loc_t;
typedef struct _workset_t {
typedef struct _belady_env_t {
struct obstack ob;
ir_graph *irg;
+ const dfs_t *dfs;
const arch_env_t *arch;
const arch_register_class_t *cls;
be_lv_t *lv;
int n_regs; /**< number of regs in this reg-class */
workset_t *ws; /**< the main workset used while processing a block. ob-allocated */
ir_node *instr; /**< current instruction */
- unsigned instr_nr; /**< current instruction number (relative to block start) */
+ int instr_nr; /**< current instruction number (relative to block start) */
spill_env_t *senv; /**< see bespill.h */
+ bitset_t *spilled; /**< bitset to keep all the irns which have already been spilled. */
} belady_env_t;
{
const loc_t *p = a;
const loc_t *q = b;
- return (int) p->time - (int) q->time;
+ return (p->time > q->time) - (p->time < q->time);
}
static INLINE void workset_print(const workset_t *w)
const ir_node *bl;
workset_t *ws_start, *ws_end;
ir_phase next_uses;
+ int id;
ir_node *first_non_in; /**< First node in block which is not a phi. */
ir_node *last_ins; /**< The instruction before which end of
double exec_freq; /**< The execution frequency of this block. */
} block_info_t;
-static INLINE void *new_block_info(belady_env_t *bel, ir_node *bl) {
+static INLINE void *new_block_info(belady_env_t *bel, int id)
+{
+ ir_node *bl = bel->blocks[id];
block_info_t *res = obstack_alloc(&bel->ob, sizeof(*res));
memset(res, 0, sizeof(res[0]));
res->first_non_in = NULL;
res->last_ins = NULL;
res->bel = bel;
res->bl = bl;
+ res->id = id;
res->entrance_reg = new_workset(bel, &bel->ob);
res->exec_freq = get_block_execfreq(bel->ef, bl);
set_irn_link(bl, res);
transport in values for the global
pass. */
int step; /**< The time step of the use. */
+ ir_node *irn;
struct _next_use_t *next; /**< The next use int this block
or NULL. */
} next_use_t;
{
ir_node *irn;
+ sched_renumber(bi->bl);
+
phase_init(&bi->next_uses, "next uses", bi->bel->irg, PHASE_DEFAULT_GROWTH, next_use_init, NULL);
sched_foreach_reverse(bi->bl, irn) {
- int i, step = sched_get_time_step(irn);
+ int i;
if (is_Phi(irn))
break;
next_use_t *curr = phase_get_irn_data(&bi->next_uses, op);
next_use_t *use = phase_alloc(&bi->next_uses, sizeof(use[0]));
- assert(step >= 0);
use->is_first_use = 1;
- use->step = step;
+ use->step = sched_get_time_step(irn);
use->next = curr;
+ use->irn = irn;
- if (curr)
+ if (curr) {
curr->is_first_use = 0;
+ assert(curr->step >= use->step);
+ }
phase_set_irn_data(&bi->next_uses, op, use);
}
phase_set_irn_data(&bi->next_uses, irn, use->next);
}
+static INLINE unsigned get_curr_distance(block_info_t *bi, const ir_node *irn, int is_usage)
+{
+ belady_env_t *env = bi->bel;
+ int curr_step = sched_get_time_step(env->instr);
+ next_use_t *use = get_current_use(bi, irn);
+ int flags = arch_irn_get_flags(env->arch, irn);
+
+ assert(!(flags & arch_irn_flags_ignore));
+
+ /* We have to keep nonspillable nodes in the workingset */
+ if(flags & arch_irn_flags_dont_spill)
+ return 0;
+
+ if (!is_usage && use && use->step == curr_step)
+ use = use->next;
+
+ if (use) {
+ unsigned res = use->step - curr_step;
+
+ assert(use->step >= curr_step);
+
+ if (res != 0) {
+ if (remat_live_range_ext && be_is_rematerializable(env->senv, irn, use->irn))
+ res = REMAT_DIST;
+ else if (bitset_contains_irn(env->spilled, irn))
+ res *= already_spilled_factor;
+ }
+
+ return res;
+ }
+
+ return be_is_live_end(env->lv, bi->bl, irn) ? LIVE_END : DEAD;
+}
static INLINE int is_local_phi(const ir_node *bl, const ir_node *irn)
{
}
if (insert_reload) {
+ bitset_add_irn(env->spilled, val);
DBG((dbg, DBG_SPILL, "Reload %+F before %+F\n", val, env->instr));
be_add_reload(env->senv, val, env->instr, env->cls, 1);
}
/* Only make more free room if we do not have enough */
if (len > max_allowed) {
- int curr_step = sched_get_time_step(env->instr);
-
DBG((dbg, DBG_DECIDE, " disposing %d values\n", len - max_allowed));
/* get current next-use distance */
for (i = 0; i < ws->len; ++i) {
- ir_node *val = workset_get_val(ws, i);
- next_use_t *use = phase_get_irn_data(&bi->next_uses, val);
- assert(use == NULL || use->step >= curr_step);
-
- if (!is_usage && use)
- use = use->next;
-
- workset_set_time(ws, i, use ? (unsigned) (use->step - curr_step) : DEAD);
+ ir_node *val = workset_get_val(ws, i);
+ unsigned dist = get_curr_distance(bi, val, is_usage);
+ workset_set_time(ws, i, dist);
}
/* sort entries by increasing nextuse-distance*/
workset_insert(env, env->ws, to_insert[i]);
bi->pressure = MAX(bi->pressure, workset_get_length(env->ws));
-
}
/**
* whether it is used from a register or is reloaded
* before the use.
*/
-static void belady(ir_node *block, void *data) {
- belady_env_t *env = data;
- block_info_t *block_info = new_block_info(env, block);
+static void belady(belady_env_t *env, int id) {
+ block_info_t *block_info = new_block_info(env, id);
+ const ir_node *block = block_info->bl;
+ void *obst_state = obstack_base(&env->ob);
workset_t *new_vals;
ir_node *irn;
}
phase_free(&block_info->next_uses);
+ obstack_free(&env->ob, obst_state);
/* Remember end-workset for this block */
block_info->ws_end = workset_clone(env, &env->ob, env->ws);
*/
+#define workset_set_version(ws, i, t) ((ws)->vals[(i)].time = (t))
+#define workset_get_version(ws, i) ((ws)->vals[(i)].time)
+
+#define ver_oldest (0)
+#define ver_youngest ((unsigned) -1)
+#define ver_make_newer(v) ((v) + 1)
+#define ver_is_older(v, w) ((v) < (w))
+#define ver_is_younger(v, w) ((v) > (w))
+
static int block_freq_gt(const void *a, const void *b)
{
const ir_node * const *p = a;
return (diff > 0) - (diff < 0);
}
-typedef struct _block_end_state_t {
- ir_node *bl;
- ir_node *irn;
- double costs;
+static int block_order(const void *a, const void *b)
+{
+ const ir_node * const *p = a;
+ const ir_node * const *q = b;
+ block_info_t *pi = get_block_info(*p);
+ block_info_t *qi = get_block_info(*q);
+ double diff;
+
+ if (pi->exec_freq > 1.0 && qi->exec_freq > 1.0) {
+ const dfs_t *dfs = pi->bel->dfs;
+ int pp = dfs_get_post_num(dfs, pi->bl);
+ int pq = dfs_get_post_num(dfs, qi->bl);
+ return pq - pp;
+ }
+
+ diff = qi->exec_freq - pi->exec_freq;
+ return (diff > 0) - (diff < 0);
+}
+
+enum {
+ irn_act_none = 0,
+ irn_act_reload,
+ irn_act_live_through
+};
+
+typedef struct _block_state_t {
+ struct _block_state_t *next;
+ struct _block_state_t *next_intern;
+ block_info_t *bi;
+ unsigned pressure;
workset_t *end_state;
- unsigned reload_at_end : 1;
- unsigned live_through : 1;
-} block_end_state_t;
+} block_state_t;
+
+typedef struct _irn_action_t {
+ struct _irn_action_t *next;
+ ir_node *irn;
+ const ir_node *bl;
+ int act;
+} irn_action_t;
typedef struct _global_end_state_t {
belady_env_t *env;
bitset_t *succ_phis;
+ bitset_t *committed;
struct obstack obst;
- block_end_state_t *end_info;
- unsigned gauge;
+ void *reset_level;
unsigned version;
+
+ unsigned *bs_tops_vers;
+ block_state_t **bs_tops;
+ block_state_t *bs_top;
+ irn_action_t *ia_top;
} global_end_state_t;
-static block_end_state_t *get_block_end_state(global_end_state_t *state, ir_node *bl, ir_node *irn)
+typedef struct {
+ void *obst_level;
+ block_state_t *bs_top;
+ irn_action_t *ia_top;
+} rollback_info_t;
+
+static INLINE block_state_t *get_block_state(global_end_state_t *ges, block_info_t *bi)
+{
+ int id = bi->id;
+ assert(!ver_is_younger(ges->bs_tops_vers[id], ges->version));
+ return ver_is_older(ges->bs_tops_vers[id], ges->version) ? NULL : ges->bs_tops[bi->id];
+}
+
+static INLINE const workset_t *get_end_state(global_end_state_t *ges, block_info_t *bi)
+{
+ block_state_t *bs = get_block_state(ges, bi);
+ return bs ? bs->end_state : bi->ws_end;
+}
+
+static block_state_t *new_block_state(global_end_state_t *ges, block_info_t *bi)
{
- unsigned i;
+ block_state_t *bs = get_block_state(ges, bi);
+ block_state_t *nw = obstack_alloc(&ges->obst, sizeof(nw[0]));
- for (i = 0; i < state->gauge; ++i) {
- block_end_state_t *bei = &state->end_info[i];
- if (bei->bl == bl && bei->irn == irn)
- return bei;
+ nw->next_intern = bs;
+ nw->next = ges->bs_top;
+ nw->bi = bi;
+
+ if (bs) {
+ nw->pressure = bs->pressure;
+ nw->end_state = workset_clone(ges->env, &ges->obst, bs->end_state);
+ }
+ else {
+ nw->pressure = bi->pressure;
+ nw->end_state = workset_clone(ges->env, &ges->obst, bi->ws_end);
}
- {
- block_info_t *bi = get_block_info(bl);
- block_end_state_t *curr;
+ ges->bs_top = nw;
+ ges->bs_tops[bi->id] = nw;
+ ges->bs_tops_vers[bi->id] = ges->version;
+ return nw;
+}
- /* make sure we have room in the array */
- ARR_EXTO(block_end_state_t, state->end_info, (int) state->gauge);
+static irn_action_t *new_irn_action(global_end_state_t *ges, ir_node *irn, const ir_node *bl)
+{
+ irn_action_t *ia = obstack_alloc(&ges->obst, sizeof(ia[0]));
+
+ ia->irn = irn;
+ ia->bl = bl;
+ ia->act = irn_act_none;
+ ia->next = ges->ia_top;
+ ges->ia_top = ia;
+ return ia;
+}
- curr = &state->end_info[state->gauge];
+static INLINE rollback_info_t trans_begin(global_end_state_t *ges)
+{
+ rollback_info_t rb;
+ rb.obst_level = obstack_base(&ges->obst);
+ rb.bs_top = ges->bs_top;
+ rb.ia_top = ges->ia_top;
+ return rb;
+}
- memset(curr, 0, sizeof(curr[0]));
- curr->bl = bl;
- curr->irn = irn;
- curr->end_state = workset_clone(state->env, &state->obst, bi->ws_end);
- curr->costs = -1.0;
- ++state->gauge;
- return curr;
+static INLINE void trans_rollback(global_end_state_t *ges, rollback_info_t *rb)
+{
+ block_state_t *bs;
+
+ /* unwind all the stacks indiced with the block number */
+ for (bs = ges->bs_top; bs != rb->bs_top; bs = bs->next) {
+ unsigned id = bs->bi->id;
+ ges->bs_tops[id] = bs->next_intern;
}
+
+ ges->ia_top = rb->ia_top;
+ ges->bs_top = rb->bs_top;
+ obstack_free(&ges->obst, rb->obst_level);
}
-static double can_bring_in(global_end_state_t *ges, ir_node *bl, ir_node *irn, int level);
-static double can_make_available_at_end(global_end_state_t *ges, ir_node *bl, ir_node *irn, int level)
+static double can_bring_in(global_end_state_t *ges, ir_node *bl, ir_node *irn, double limit, int level);
+
+static double can_make_available_at_end(global_end_state_t *ges, ir_node *bl, ir_node *irn, double limit, int level)
{
- block_end_state_t *bes = get_block_end_state(ges, bl, irn);
- workset_t *end = bes->end_state;
- block_info_t *bi = get_block_info(bl);
- int n_regs = bi->bel->n_regs;
+ block_info_t *bi = get_block_info(bl);
+ const workset_t *end = get_end_state(ges, bi);
+ double res;
int index;
- DBG((dbg, DBG_GLOBAL, "\t%2Dcan make avail %+F at end of %+F (pressure %d)\n",
- level, irn, bl, bi->pressure));
+ DBG((dbg, DBG_GLOBAL, "\t%2Dcan make avail %+F at end of %+F\n", level, irn, bl));
/*
* to make the value available at end,
* cheaper, do the other thing. If not, reload it here.
*/
- /*
- * we have been here before and already figured out some costs.
- * so we can exit safely.
- */
- if (bes->costs >= 0.0) {
- DBG((dbg, DBG_GLOBAL, "\t%2Dwe\'ve been here before\n", level));
- goto end;
- }
-
/* if the end set contains it already, it is in a reg and it costs nothing
* to load it to one. */
index = workset_get_index(end, irn);
if (index >= 0) {
- unsigned ver = end->vals[index].version;
+ unsigned ver = workset_get_version(end, index);
DBG((dbg, DBG_GLOBAL, "\t%2Dnode is in the end set and is %s fixed\n",
- level, ver > ges->version ? "already" : "not yet"));
+ level, ver_is_older(ver, ges->version) ? "already" : "not yet"));
/*
* if the version is older, the value is already fixed
- * and cannot be removed from the end set. If not,
- * we fix it here by giving it our version.
+ * and cannot be removed from the end set.
+ *
+ * If not, we will create a new block state for that block since
+ * we modify it by giving the end state a new version.
*/
- if (ver < ges->version)
- end->vals[index].version = ges->version;
+ if (ver_is_younger(ver, ges->version)) {
+ block_state_t *bs = new_block_state(ges, bi);
+ workset_set_version(bs->end_state, index, ges->version);
+ }
- bes->costs = 0.0;
+ res = 0.0;
goto end;
}
*/
{
- int len = workset_get_length(end);
+ int n_regs = bi->bel->n_regs;
+ int len = workset_get_length(end);
int slot = -1;
int i;
- bes->costs = HUGE_VAL;
+ res = HUGE_VAL;
/*
* look if there is room in the end array
* for the variable. Note that this does not
- * means that the var is living through the block.
+ * mean that the var can live through the block.
* There is just room at the *end*
*/
if (len < n_regs) {
- DBG((dbg, DBG_GLOBAL, "\t%2Dthe end set has %d free slots\n",
- level, n_regs - len));
+ DBG((dbg, DBG_GLOBAL, "\t%2Dthe end set has %d free slots\n", level, n_regs - len));
slot = len;
} else {
- for (i = 0; i < len; ++i)
- if (end->vals[i].version < ges->version)
+ for (i = 0; i < len; ++i) {
+ unsigned ver = workset_get_version(end, i);
+ if (ver_is_younger(ver, ges->version))
break;
+ }
if (i < len) {
DBG((dbg, DBG_GLOBAL, "\t%2D%+F (slot %d) can be erased from the end set\n",
}
}
+ /*
+ * finally there is some room. we can at least reload the value.
+ * but we will try to let ot live through anyhow.
+ */
if (slot >= 0) {
- int gauge = ges->gauge;
+ irn_action_t *vs = new_irn_action(ges, irn, bi->bl);
+ block_state_t *bs = new_block_state(ges, bi);
+ workset_t *end = bs->end_state;
ir_node *ins_before = block_info_get_last_ins(bi);
double reload_here = be_get_reload_costs(bi->bel->senv, irn, ins_before);
- double bring_in = bi->pressure < n_regs ? can_bring_in(ges, bl, irn, level + 1) : HUGE_VAL;
-
- DBG((dbg, DBG_GLOBAL, "\t%2Dthere is a free slot. capacity=%d, reload here=%f, bring in=%f\n",
- level, n_regs - bi->pressure, reload_here, bring_in));
+ int pressure_ok = bs->pressure < (unsigned) n_regs;
/*
- * reloading here pays off; bringing the value in from elsewhere
- * is too expensive, hence we drop that search by resetting
- * the gauge.
+ * No matter what we do, the value will be in the end set
+ * if the block from now on (of course only regarding the
+ * current state). Enter it and set the new length
+ * appropriately.
*/
- if (reload_here <= bring_in) {
- ges->gauge = gauge;
- bes->costs = reload_here;
- bes->reload_at_end = 1;
- } else {
- bes->live_through = 1;
- bes->costs = bring_in;
+ end->vals[slot].irn = irn;
+ workset_set_version(end, slot, ges->version);
+ workset_set_length(end, MAX(workset_get_length(end), slot + 1));
+
+ vs->act = irn_act_reload;
+ res = reload_here;
+
+ DBG((dbg, DBG_GLOBAL, "\t%2Dthere is a free slot. capacity=%d, reload here=%f, pressure %s\n",
+ level, n_regs - bs->pressure, reload_here, pressure_ok ? "ok" : "insufficient"));
+
+ /* look if we can bring the value in. */
+ if (pressure_ok) {
+ rollback_info_t rb = trans_begin(ges);
+ double new_limit = MIN(reload_here, limit);
+
+ vs->act = irn_act_live_through;
+ bs->pressure += 1;
+ res = can_bring_in(ges, bl, irn, new_limit, level + 1);
+
+ /*
+ * if bring in is too expensive re-adjust the pressure
+ * and roll back the state
+ */
+ if (res >= reload_here) {
+ bs->pressure -= 1;
+ vs->act = irn_act_reload;
+ trans_rollback(ges, &rb);
+ res = reload_here;
+ }
}
- end->vals[slot].irn = irn;
- end->vals[slot].version = ges->version;
- end->len = MAX(end->len, slot + 1);
- } else {
- ges->gauge -= 1;
+ DBG((dbg, DBG_GLOBAL, "\t%2D%s\n", level,
+ vs->act == irn_act_reload ? "reloading" : "bringing in"));
}
}
end:
- DBG((dbg, DBG_GLOBAL, "\t%2D-> %f\n", level, bes->costs));
- return bes->costs;
+ DBG((dbg, DBG_GLOBAL, "\t%2D-> %f\n", level, res));
+ return res;
}
-static double can_bring_in(global_end_state_t *ges, ir_node *bl, ir_node *irn, int level)
+static double can_bring_in(global_end_state_t *ges, ir_node *bl, ir_node *irn, double limit, int level)
{
+ belady_env_t *env = ges->env;
double glob_costs = HUGE_VAL;
- DBG((dbg, DBG_GLOBAL, "\t%2Dcan bring in for %+F at block %+F\n", level, irn, bl));
+ DBG((dbg, DBG_GLOBAL, "\t%2Dcan bring in (max %f) for %+F at block %+F\n", level, limit, irn, bl));
if (is_transport_in(bl, irn)) {
int i, n = get_irn_arity(bl);
ir_node **nodes = alloca(get_irn_arity(bl) * sizeof(nodes[0]));
+ rollback_info_t rb = trans_begin(ges);
- int gauge_begin = ges->gauge;
glob_costs = 0.0;
for (i = 0; i < n; ++i) {
ir_node *pr = get_Block_cfgpred_block(bl, i);
ir_node *op = is_local_phi(bl, irn) ? get_irn_n(irn, i) : irn;
- double c = can_make_available_at_end(ges, pr, op, level + 1);
+ double c;
+
+ /*
+ * there might by unknwons as operands of phis in that case
+ * we set the costs to zero, since they won't get spilled.
+ */
+ if (arch_irn_consider_in_reg_alloc(env->arch, env->cls, op))
+ c = can_make_available_at_end(ges, pr, op, limit - glob_costs, level + 1);
+ else
+ c = 0.0;
+
+ glob_costs += c;
- if (c >= HUGE_VAL) {
- ges->gauge = gauge_begin;
+ if (glob_costs >= limit) {
glob_costs = HUGE_VAL;
+ trans_rollback(ges, &rb);
goto end;
}
-
- glob_costs += c;
}
}
static void materialize_and_commit_end_state(global_end_state_t *ges)
{
belady_env_t *env = ges->env;
- unsigned i;
+ irn_action_t *ia;
+ block_state_t *bs;
DBG((dbg, DBG_GLOBAL, "\tmaterializing\n"));
- for (i = 0; i < ges->gauge; ++i) {
- block_end_state_t *bes = &ges->end_info[i];
- block_info_t *bi = get_block_info(bes->bl);
- int idx, end_pressure;
-
- DBG((dbg, DBG_GLOBAL, "\t\t%+F in %+F, cost %f through: %d, rel: %d\n",
- bes->irn, bes->bl, bes->costs, bes->live_through, bes->reload_at_end));
-
- /* insert the reload if the val was reloaded at the block's end */
- if (bes->reload_at_end) {
- be_add_reload_at_end(env->senv, bes->irn, bes->bl, env->cls, 1);
- DBG((dbg, DBG_GLOBAL, "\t\tadding reload of %+F at end of %+F\n", bes->irn, bes->bl));
- }
-
- end_pressure = 0;
- for (idx = workset_get_length(bes->end_state) - 1; idx >= 0; --idx)
- if (bes->end_state->vals[idx].version >= ges->version)
- end_pressure += 1;
-
- /*
- * if the variable is live through the block,
- * update the pressure indicator.
- */
- DBG((dbg, DBG_GLOBAL, "\t\told pressure %d, ", bi->pressure));
-
- bi->pressure = MAX(bi->pressure + bes->live_through, end_pressure);
-
- DBG((dbg, DBG_GLOBAL, "new pressure: %d, end pressure: %d, end length: %d\n",
- bi->pressure, end_pressure, workset_get_length(bes->end_state)));
-// workset_print(bes->end_state);
- idx = workset_get_index(bes->end_state, bes->irn);
-
- if (is_local_phi(bes->bl, bes->irn) && bes->live_through)
- bitset_add_irn(ges->succ_phis, bes->irn);
+ /*
+ * Perform all the variable actions.
+ */
+ for (ia = ges->ia_top; ia != NULL; ia = ia->next) {
+ switch (ia->act) {
+ case irn_act_live_through:
+ if (is_local_phi(ia->bl, ia->irn)) {
+ bitset_add_irn(ges->succ_phis, ia->irn);
+ DBG((dbg, DBG_GLOBAL, "\t\tlive through phi kept alive: %+F\n", ia->irn));
+ }
+ break;
+ case irn_act_reload:
+ be_add_reload_at_end(env->senv, ia->irn, ia->bl, env->cls, 1);
+ DBG((dbg, DBG_GLOBAL, "\t\tadding reload of %+F at end of %+F\n", ia->irn, ia->bl));
+ break;
+ default:
+ DBG((dbg, DBG_GLOBAL, "\t\t%+F is in the end set of %+F\n", ia->irn, ia->bl));
+ }
+ }
- /*
- * set the version number in the workset.
- * That will mark this value as fixed in the end set
- * and will prevent further investigations from removing
- * it from there.
- * Also "commit" the workset;
- * by copying it back to the block's end workset.
- */
- if (idx >= 0) {
- DBG((dbg, DBG_GLOBAL, "\t\tcommiting workset of %+F with version %x\n", bes->bl, ges->version));
- bes->end_state->vals[idx].version = ges->version;
- workset_copy(env, bi->ws_end, bes->end_state);
+ /*
+ * Commit the block end states
+ */
+ for (bs = ges->bs_top; bs != NULL; bs = bs->next) {
+ block_info_t *bi = bs->bi;
+
+ if (!bitset_is_set(ges->committed, bi->id)) {
+ DBG((dbg, DBG_GLOBAL, "\t\tcommiting workset of %+F with version %x\n", bi->bl, ges->version));
+ // bes->bs->end_state->vals[idx].version = ges->version;
+ workset_copy(env, bi->ws_end, bs->end_state);
+ DBG((dbg, DBG_GLOBAL, "\t\told pressure: %d, new pressure: %d, end length: %d\n",
+ bi->pressure, bs->pressure, workset_get_length(bs->end_state)));
+ bi->pressure = bs->pressure;
+ bitset_set(ges->committed, bi->id);
}
}
+
+ /* clear the committed bitset. the next call is expecting it. */
+ bitset_clear_all(ges->committed);
}
/**
static void fix_block_borders(global_end_state_t *ges, ir_node *block) {
block_info_t *bi = get_block_info(block);
belady_env_t *env = ges->env;
+ void *reset_level = obstack_base(&ges->obst);
ir_node *irn;
int i;
+ for (i = workset_get_length(bi->ws_end) - 1; i >= 0; --i)
+ workset_set_version(bi->ws_end, i, ver_youngest);
+
DBG((dbg, DBG_GLOBAL, "fixing block borders at %+F (%f)\n", block, bi->exec_freq));
/* process all variables which shall be in a reg at
double local_costs = be_get_reload_costs(env->senv, irn, bi->first_non_in);
double bring_in_costs;
- /* reset the gauge and create a new version. */
- ges->gauge = 0;
- ges->version -= 1;
+ /* reset the lists */
+ ges->bs_top = NULL;
+ ges->ia_top = NULL;
DBG((dbg, DBG_GLOBAL, "\ttrans in var %+F, version %x\n", irn, ges->version));
- bring_in_costs = can_bring_in(ges, block, irn, 1);
+ bring_in_costs = can_bring_in(ges, block, irn, local_costs, 1);
DBG((dbg, DBG_GLOBAL, "\tbring in: %f, local: %f", bring_in_costs, local_costs));
* in a register at the entrance of the block
* or it is too costly, so we have to do the reload locally
*/
- if (bring_in_costs > local_costs) {
-
+ if (bring_in_costs >= local_costs) {
DBG((dbg, DBG_GLOBAL, " -> do local reload\n"));
be_add_reload(env->senv, irn, bi->first_non_in, env->cls, 1);
} else {
if (is_local_phi(block, irn))
bitset_add_irn(ges->succ_phis, irn);
- DBG((dbg, DBG_GLOBAL, " -> do remote reload\n"));
+ DBG((dbg, DBG_GLOBAL, " -> bring it in\n"));
materialize_and_commit_end_state(ges);
}
DBG((dbg, DBG_GLOBAL, "\n"));
+
+ /* reset the obstack and create a new version. */
+ obstack_free(&ges->obst, reset_level);
+ ges->version = ver_make_newer(ges->version);
}
}
global_end_state_t ges;
int i;
- obstack_init(&ges.obst);
- ges.gauge = 0;
- ges.env = env;
- ges.version = -1;
- ges.end_info = NEW_ARR_F(block_end_state_t, env->n_blocks);
- ges.succ_phis = bitset_irg_obstack_alloc(&env->ob, env->irg);
-
/*
* sort the blocks according to execution frequency.
* That's not necessary for belady() but for the global pass later on.
*/
- qsort(env->blocks, env->n_blocks, sizeof(env->blocks[0]), block_freq_gt);
+ qsort(env->blocks, env->n_blocks, sizeof(env->blocks[0]), block_order);
+
+ memset(&ges, 0, sizeof(ges));
+ obstack_init(&ges.obst);
+ ges.env = env;
+ ges.version = ver_make_newer(ver_oldest);
+ ges.succ_phis = bitset_irg_obstack_alloc(&ges.obst, env->irg);
+ ges.committed = bitset_obstack_alloc(&ges.obst, env->n_blocks);
+ ges.bs_tops = obstack_alloc(&ges.obst, sizeof(ges.bs_tops[0]) * env->n_blocks);
+ ges.bs_tops_vers = obstack_alloc(&ges.obst, sizeof(ges.bs_tops_vers[0]) * env->n_blocks);
+
+ for (i = 0; i < env->n_blocks; ++i)
+ ges.bs_tops_vers[i] = ver_oldest;
for (i = 0; i < env->n_blocks; ++i)
fix_block_borders(&ges, env->blocks[i]);
be_spill_phi(env->senv, irn);
}
}
-
}
static void collect_blocks(ir_node *bl, void *data)
obstack_ptr_grow(&env->ob, bl);
}
-void be_spill_belady_spill_env2(be_irg_t *birg, const arch_register_class_t *cls, spill_env_t *spill_env) {
+/**
+ * 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
+ */
+void be_spill_belady(be_irg_t *birg, const arch_register_class_t *cls)
+{
ir_graph *irg = be_get_birg_irg(birg);
belady_env_t env;
int i, n_regs;
/* init belady env */
obstack_init(&env.ob);
env.irg = irg;
- env.arch = birg->main_env->arch_env;
+ env.arch = be_get_birg_arch_env(birg);
env.cls = cls;
env.lv = be_get_birg_liveness(birg);
+ env.dfs = env.lv->dfs;
env.n_regs = n_regs;
env.ws = new_workset(&env, &env.ob);
- env.senv = spill_env ? spill_env : be_new_spill_env(birg);
+ env.senv = be_new_spill_env(birg);
env.ef = be_get_birg_exec_freq(birg);
+ env.spilled = bitset_irg_obstack_alloc(&env.ob, irg);
env.n_blocks = 0;
irg_block_walk_graph(irg, NULL, collect_blocks, &env);
/* Fix high register pressure in blocks with belady algorithm */
for (i = 0; i < env.n_blocks; ++i)
- belady(env.blocks[i], &env);
+ belady(&env, i);
global_assign(&env);
be_insert_spills_reloads(env.senv);
/* clean up */
- if(spill_env == NULL)
- be_delete_spill_env(env.senv);
+ be_delete_spill_env(env.senv);
obstack_free(&env.ob, 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_env2(birg, cls, NULL);
-}
-
-
void be_init_spillbelady2(void)
{
+ lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
+ lc_opt_entry_t *spill_grp = lc_opt_get_grp(be_grp, "spill");
+ lc_opt_entry_t *bel2_grp = lc_opt_get_grp(spill_grp, "belady2");
+
static be_spiller_t belady_spiller = {
be_spill_belady
};
+ lc_opt_add_table(bel2_grp, options);
be_register_spiller("belady2", &belady_spiller);
FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady2");
}
projects / libfirm / blobdiff