-/**
- * Author: Christian Wuerdig
- * Date: 2005/12/14
- * Copyright: (c) Universitaet Karlsruhe
- * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
- * CVS-Id: $Id$
+/*
+ * Copyright (C) 1995-2008 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.
*
- * Performs lowering of perm nodes and spill/reload optimization.
+ * 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 Performs lowering of perm nodes. Inserts copies to assure register constraints.
+ * @author Christian Wuerdig
+ * @date 14.12.2005
+ * @version $Id$
*/
-#ifdef HAVE_CONFIG_H
#include "config.h"
-#endif
#include <stdlib.h>
#include "ircons.h"
#include "debug.h"
#include "irhooks.h"
-
-#include "bearch.h"
-#include "belower.h"
-#include "benode_t.h"
-#include "bechordal_t.h"
-#include "besched_t.h"
-
+#include "xmalloc.h"
+#include "irnodeset.h"
#include "irgmod.h"
#include "iredges_t.h"
#include "irgwalk.h"
+#include "array_t.h"
-#ifdef HAVE_MALLOC_H
- #include <malloc.h>
-#endif
-#ifdef HAVE_ALLOCA_H
- #include <alloca.h>
-#endif
-
-#undef is_Perm
-#define is_Perm(arch_env, irn) (arch_irn_classify(arch_env, irn) == arch_irn_class_perm)
-
-/* collect static data about perms */
-typedef struct _perm_stat_t {
- const arch_register_class_t *cls; /**< the current register class */
- int *perm_size_ar; /**< the sizes of all perms in an irg */
- int *real_perm_size_ar; /**< the sizes of all perms in an irg */
- int *chain_len_ar; /**< the sizes of all chains for all perms */
- int *cycle_len_ar; /**< the siyes of all cycles for all perms */
- int num_perms; /**< number of all perms */
- int num_real_perms; /**< number of all perms */
- int num_chains; /**< the number of all chains */
- int num_cycles; /**< the number of all cycles */
-} perm_stat_t;
-
-/* lowering walker environment */
+#include "bearch_t.h"
+#include "belower.h"
+#include "benode_t.h"
+#include "besched_t.h"
+#include "bestat.h"
+#include "bessaconstr.h"
+#include "beintlive_t.h"
+
+#undef KEEP_ALIVE_COPYKEEP_HACK
+
+/** Associates an op with it's copy and CopyKeep. */
+typedef struct {
+ ir_node *op; /**< an irn which must be different */
+ ir_nodeset_t copies; /**< all non-spillable copies of this irn */
+ const arch_register_class_t *cls;
+} op_copy_assoc_t;
+
+/** Environment for constraints. */
+typedef struct {
+ be_irg_t *birg;
+ pset *op_set;
+ struct obstack obst;
+ DEBUG_ONLY(firm_dbg_module_t *dbg;)
+} constraint_env_t;
+
+/** Lowering walker environment. */
typedef struct _lower_env_t {
- be_chordal_env_t *chord_env;
- unsigned do_copy:1;
- unsigned do_stat:1;
- unsigned pstat_n:30;
- perm_stat_t **pstat;
+ be_irg_t *birg;
+ unsigned do_copy : 1;
DEBUG_ONLY(firm_dbg_module_t *dbg_module;)
} lower_env_t;
-/* holds a perm register pair */
+/** Holds a Perm register pair. */
typedef struct _reg_pair_t {
const arch_register_t *in_reg; /**< a perm IN register */
ir_node *in_node; /**< the in node to which the register belongs */
PERM_COPY
} perm_type_t;
-/* structure to represent cycles or chains in a perm */
+/** Structure to represent cycles or chains in a Perm. */
typedef struct _perm_cycle_t {
const arch_register_t **elems; /**< the registers in the cycle */
int n_elems; /**< number of elements in the cycle */
perm_type_t type; /**< type (CHAIN or CYCLE) */
} perm_cycle_t;
-/* Compare the in registers of two register pairs */
+/** Compare the two operands. */
+static int cmp_op_copy_assoc(const void *a, const void *b) {
+ const op_copy_assoc_t *op1 = a;
+ const op_copy_assoc_t *op2 = b;
+
+ return op1->op != op2->op;
+}
+
+/** Compare the in registers of two register pairs. */
static int compare_reg_pair(const void *a, const void *b) {
const reg_pair_t *pair_a = a;
const reg_pair_t *pair_b = b;
return -1;
}
-/* returns the number register pairs marked as checked */
+/** returns the number register pairs marked as checked. */
static int get_n_checked_pairs(reg_pair_t *pairs, int n) {
int i, n_checked = 0;
if (in_out) {
for (i = 0; i < n; i++) {
/* out register matches */
- if (pairs[i].out_reg->index == reg_idx)
+ if ((int) pairs[i].out_reg->index == reg_idx)
return i;
}
}
else {
for (i = 0; i < n; i++) {
/* in register matches */
- if (pairs[i].in_reg->index == reg_idx)
+ if ((int) pairs[i].in_reg->index == reg_idx)
return i;
}
}
}
/* assume worst case: all remaining pairs build a cycle or chain */
- cycle->elems = xcalloc((n - n_pairs_done) * 2, sizeof(cycle->elems[0]));
+ cycle->elems = XMALLOCNZ(const arch_register_t*, (n - n_pairs_done) * 2);
cycle->n_elems = 2; /* initial number of elements is 2 */
cycle->elems[0] = pairs[start].in_reg;
cycle->elems[1] = pairs[start].out_reg;
* @param walk_env The environment
*/
static void lower_perm_node(ir_node *irn, void *walk_env) {
+ ir_graph *irg = get_irn_irg(irn);
const arch_register_class_t *reg_class;
- const arch_env_t *arch_env;
lower_env_t *env = walk_env;
- perm_stat_t **pstat = env->pstat;
- int pstat_idx = -1;
- int real_size = 0;
- int n, i, pn, do_copy, j, n_ops;
+ int real_size = 0;
+ int keep_perm = 0;
+ int n, i, pn, do_copy, j, n_ops;
reg_pair_t *pairs;
const ir_edge_t *edge;
- perm_cycle_t *cycle;
ir_node *sched_point, *block, *in[2];
ir_node *arg1, *arg2, *res1, *res2;
ir_node *cpyxchg = NULL;
DEBUG_ONLY(firm_dbg_module_t *mod;)
- arch_env = env->chord_env->birg->main_env->arch_env;
do_copy = env->do_copy;
DEBUG_ONLY(mod = env->dbg_module;)
block = get_nodes_block(irn);
should be ok.
*/
sched_point = sched_prev(irn);
+ DBG((mod, LEVEL_1, "perm: %+F\n", irn));
DBG((mod, LEVEL_1, "sched point is %+F\n", sched_point));
assert(sched_point && "Perm is not scheduled or has no predecessor");
n = get_irn_arity(irn);
assert(n == get_irn_n_edges(irn) && "perm's in and out numbers different");
- reg_class = arch_get_irn_register(arch_env, get_irn_n(irn, 0))->reg_class;
+ reg_class = arch_get_irn_register(get_irn_n(irn, 0))->reg_class;
pairs = alloca(n * sizeof(pairs[0]));
- if (env->do_stat) {
- unsigned i;
-
- /* determine index in statistics */
- for (i = 0; i < env->pstat_n; i++) {
- if (strcmp(pstat[i]->cls->name, reg_class->name) == 0) {
- pstat_idx = i;
- break;
- }
- }
- assert(pstat_idx >= 0 && "could not determine class index for statistics");
-
- pstat[pstat_idx]->num_perms++;
- pstat[pstat_idx]->perm_size_ar[n - 1]++;
- }
-
/* build the list of register pairs (in, out) */
i = 0;
foreach_out_edge(irn, edge) {
pn = get_Proj_proj(pairs[i].out_node);
pairs[i].in_node = get_irn_n(irn, pn);
- pairs[i].in_reg = arch_get_irn_register(arch_env, pairs[i].in_node);
- pairs[i].out_reg = arch_get_irn_register(arch_env, pairs[i].out_node);
+ pairs[i].in_reg = arch_get_irn_register(pairs[i].in_node);
+ pairs[i].out_reg = arch_get_irn_register(pairs[i].out_node);
pairs[i].checked = 0;
i++;
/* We have to check for a special case:
The in-node could be a Proj from a Perm. In this case,
we need to correct the projnum */
- if (is_Perm(arch_env, pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
+ if (be_is_Perm(pairs[i].in_node) && is_Proj(pairs[i].in_node)) {
set_Proj_proj(pairs[i].out_node, get_Proj_proj(pairs[i].in_node));
}
- /* remove the proj from the schedule */
- sched_remove(pairs[i].out_node);
-
/* reroute the edges from the proj to the argument */
- edges_reroute(pairs[i].out_node, pairs[i].in_node, env->chord_env->irg);
+ exchange(pairs[i].out_node, pairs[i].in_node);
+ //edges_reroute(pairs[i].out_node, pairs[i].in_node, env->birg->irg);
+ //set_irn_n(pairs[i].out_node, 0, new_Bad());
pairs[i].checked = 1;
}
do_copy = 0;
}
- if (env->do_stat && get_n_checked_pairs(pairs, n) < n) {
- pstat[pstat_idx]->num_real_perms++;
- pstat[pstat_idx]->real_perm_size_ar[n - 1]++;
- real_size = n - get_n_checked_pairs(pairs, n);
- }
-
- hook_be_block_stat_perm(reg_class->name, reg_class->n_regs, irn, block, n, real_size);
+ real_size = n - get_n_checked_pairs(pairs, n);
/* check for cycles and chains */
while (get_n_checked_pairs(pairs, n) < n) {
+ perm_cycle_t *cycle;
+
i = n_ops = 0;
/* go to the first not-checked pair */
while (pairs[i].checked) i++;
- cycle = xcalloc(1, sizeof(*cycle));
+ cycle = XMALLOCZ(perm_cycle_t);
cycle = get_perm_cycle(cycle, pairs, n, i);
DB((mod, LEVEL_1, "%+F: following %s created:\n ", irn, cycle->type == PERM_CHAIN ? "chain" : "cycle"));
}
DB((mod, LEVEL_1, "\n"));
- /* statistics */
- if (env->do_stat) {
- int n_idx = cycle->n_elems - 1;
- if (cycle->type == PERM_CHAIN) {
- pstat[pstat_idx]->num_chains++;
- pstat[pstat_idx]->chain_len_ar[n_idx]++;
- }
- else {
- pstat[pstat_idx]->num_cycles++;
- pstat[pstat_idx]->cycle_len_ar[n_idx]++;
- }
- }
-
- /* We don't need to do anything if we have a Perm with two
+ /*
+ We don't need to do anything if we have a Perm with two
elements which represents a cycle, because those nodes
- already represent exchange nodes */
+ already represent exchange nodes
+ */
if (n == 2 && cycle->type == PERM_CYCLE) {
free(cycle);
+ keep_perm = 1;
continue;
}
DBG((mod, LEVEL_1, "%+F (%+F, %s) and (%+F, %s)\n",
irn, res1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
- cpyxchg = be_new_Perm(reg_class, env->chord_env->irg, block, 2, in);
+ cpyxchg = be_new_Perm(reg_class, irg, block, 2, in);
n_ops++;
if (i > 0) {
int pidx = get_pairidx_for_regidx(pairs, n, cycle->elems[i]->index, 0);
/* create intermediate proj */
- res1 = new_r_Proj(get_irn_irg(irn), block, cpyxchg, get_irn_mode(res1), 0);
+ res1 = new_r_Proj(irg, block, cpyxchg, get_irn_mode(res1), 0);
/* set as in for next Perm */
pairs[pidx].in_node = res1;
}
- else {
- sched_remove(res1);
- }
-
- sched_remove(res2);
set_Proj_pred(res2, cpyxchg);
set_Proj_proj(res2, 0);
set_Proj_pred(res1, cpyxchg);
set_Proj_proj(res1, 1);
- sched_add_after(sched_point, res1);
- sched_add_after(sched_point, res2);
-
- arch_set_irn_register(arch_env, res2, cycle->elems[i + 1]);
- arch_set_irn_register(arch_env, res1, cycle->elems[i]);
+ arch_set_irn_register(res2, cycle->elems[i + 1]);
+ arch_set_irn_register(res1, cycle->elems[i]);
/* insert the copy/exchange node in schedule after the magic schedule node (see above) */
sched_add_after(sched_point, cpyxchg);
DBG((mod, LEVEL_1, "%+F creating copy node (%+F, %s) -> (%+F, %s)\n",
irn, arg1, cycle->elems[i]->name, res2, cycle->elems[i + 1]->name));
- cpyxchg = be_new_Copy(reg_class, env->chord_env->irg, block, arg1);
- arch_set_irn_register(arch_env, cpyxchg, cycle->elems[i + 1]);
+ cpyxchg = be_new_Copy(reg_class, irg, block, arg1);
+ arch_set_irn_register(cpyxchg, cycle->elems[i + 1]);
n_ops++;
- /* remove the proj from the schedule */
- sched_remove(res2);
-
/* exchange copy node and proj */
exchange(res2, cpyxchg);
}
}
- if (env->do_stat) {
- hook_be_block_stat_permcycle(reg_class->name, irn, block, \
- cycle->type == PERM_CHAIN, cycle->n_elems, n_ops);
- }
-
free((void *) cycle->elems);
free(cycle);
}
-
-
/* remove the perm from schedule */
- sched_remove(irn);
+ if (! keep_perm) {
+ sched_remove(irn);
+ kill_node(irn);
+ }
}
-static int get_n_out_edges(const ir_node *irn) {
- const ir_edge_t *edge;
- int cnt = 0;
-
- foreach_out_edge(irn, edge) {
- cnt++;
- }
-
- return cnt;
+static int has_irn_users(const ir_node *irn) {
+ return get_irn_out_edge_first_kind(irn, EDGE_KIND_NORMAL) != 0;
}
-static ir_node *belower_skip_proj(ir_node *irn) {
+/**
+ * Skip all Proj nodes.
+ */
+static INLINE ir_node *belower_skip_proj(ir_node *irn) {
while(is_Proj(irn))
irn = get_Proj_pred(irn);
return irn;
}
-static void fix_in(ir_node *irn, ir_node *old, ir_node *nw) {
- int i, n;
-
- irn = belower_skip_proj(irn);
- n = get_irn_arity(irn);
-
- for (i = 0; i < n; i++) {
- if (get_irn_n(irn, i) == old) {
- set_irn_n(irn, i, nw);
- break;
- }
+static ir_node *find_copy(ir_node *irn, ir_node *op)
+{
+ ir_node *block = get_nodes_block(irn);
+ ir_node *cur_node;
+
+ for (cur_node = sched_prev(irn);
+ ! is_Block(cur_node) && be_is_Copy(cur_node) && get_nodes_block(cur_node) == block;
+ cur_node = sched_prev(cur_node))
+ {
+ if (be_get_Copy_op(cur_node) == op && arch_irn_is(cur_node, dont_spill))
+ return cur_node;
}
-}
-static void gen_assure_different_pattern(ir_node *irn, be_irg_t *birg, ir_node *other_different) {
- const arch_env_t *arch_env = birg->main_env->arch_env;
- ir_node *in[2], *keep, *cpy, *temp;
- ir_node *block = get_nodes_block(irn);
- const arch_register_class_t *cls = arch_get_irn_reg_class(arch_env, other_different, -1);
- FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower");
+ return NULL;
+}
- if (arch_irn_is(arch_env, other_different, ignore) || ! mode_is_datab(get_irn_mode(other_different))) {
+static void gen_assure_different_pattern(ir_node *irn, ir_node *other_different, constraint_env_t *env) {
+ be_irg_t *birg = env->birg;
+ ir_graph *irg = be_get_birg_irg(birg);
+ pset *op_set = env->op_set;
+ ir_node *block = get_nodes_block(irn);
+ const arch_register_class_t *cls = arch_get_irn_reg_class(other_different, -1);
+ ir_node *in[2], *keep, *cpy;
+ op_copy_assoc_t key, *entry;
+ DEBUG_ONLY(firm_dbg_module_t *mod = env->dbg;)
+
+ if (arch_irn_is(other_different, ignore) ||
+ !mode_is_datab(get_irn_mode(other_different))) {
DBG((mod, LEVEL_1, "ignore constraint for %+F because other_irn is ignore or not a datab node\n", irn));
return;
}
/* in block far far away */
/* The copy is optimized later if not needed */
- temp = new_rd_Unknown(birg->irg, get_irn_mode(other_different));
- cpy = be_new_Copy(cls, birg->irg, block, temp);
- be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
+ /* check if already exists such a copy in the schedule immediately before */
+ cpy = find_copy(belower_skip_proj(irn), other_different);
+ if (! cpy) {
+ cpy = be_new_Copy(cls, irg, block, other_different);
+ be_node_set_flags(cpy, BE_OUT_POS(0), arch_irn_flags_dont_spill);
+ DBG((mod, LEVEL_1, "created non-spillable %+F for value %+F\n", cpy, other_different));
+ }
+ else {
+ DBG((mod, LEVEL_1, "using already existing %+F for value %+F\n", cpy, other_different));
+ }
in[0] = irn;
in[1] = cpy;
- /* Let the irn use the copy instead of the old other_different */
- fix_in(irn, other_different, cpy);
-
/* Add the Keep resp. CopyKeep and reroute the users */
/* of the other_different irn in case of CopyKeep. */
- if (get_n_out_edges(other_different) == 0) {
- keep = be_new_Keep(cls, birg->irg, block, 2, in);
+ if (has_irn_users(other_different)) {
+ keep = be_new_CopyKeep_single(cls, irg, block, cpy, irn, get_irn_mode(other_different));
+ be_node_set_reg_class(keep, 1, cls);
}
else {
- keep = be_new_CopyKeep_single(cls, birg->irg, block, cpy, irn, get_irn_mode(other_different));
- be_node_set_reg_class(keep, 1, cls);
- edges_reroute(other_different, keep, birg->irg);
+ keep = be_new_Keep(cls, irg, block, 2, in);
}
- /* after rerouting: let the copy point to the other_different irn */
- set_irn_n(cpy, 0, other_different);
+ DBG((mod, LEVEL_1, "created %+F(%+F, %+F)\n\n", keep, irn, cpy));
+
+ /* insert copy and keep into schedule */
+ assert(sched_is_scheduled(irn) && "need schedule to assure constraints");
+ if (! sched_is_scheduled(cpy))
+ sched_add_before(belower_skip_proj(irn), cpy);
+ sched_add_after(irn, keep);
- DBG((mod, LEVEL_1, "created %+F for %+F to assure should_be_different\n", keep, irn));
+ /* insert the other different and it's copies into the set */
+ key.op = other_different;
+ entry = pset_find(op_set, &key, hash_irn(other_different));
+
+ if (! entry) {
+ entry = obstack_alloc(&env->obst, sizeof(*entry));
+ ir_nodeset_init(&entry->copies);
+ entry->op = other_different;
+ entry->cls = cls;
+ }
+
+ /* insert copy */
+ ir_nodeset_insert(&entry->copies, cpy);
+
+ /* insert keep in case of CopyKeep */
+ if (be_is_CopyKeep(keep)) {
+ ir_nodeset_insert(&entry->copies, keep);
+ }
+
+ pset_insert(op_set, entry, hash_irn(other_different));
}
/**
- * Checks if node has a should_be_different constraint in output
- * and adds a Keep then to assure the constraint.
+ * Checks if node has a must_be_different constraint in output and adds a Keep
+ * then to assure the constraint.
*/
-static void assure_different_constraints(ir_node *irn, be_irg_t *birg) {
- const arch_env_t *arch_env = birg->main_env->arch_env;
- const arch_register_req_t *req;
- arch_register_req_t req_temp;
- int i, n;
-
- req = arch_get_register_req(arch_env, &req_temp, irn, -1);
-
- if (req) {
- if (arch_register_req_is(req, should_be_different)) {
- gen_assure_different_pattern(irn, birg, req->other_different);
+static void assure_different_constraints(ir_node *irn, constraint_env_t *env) {
+ ir_node *skipped_irn = belower_skip_proj(irn);
+ const arch_register_req_t *req = arch_get_register_req(irn, -1);
+
+ if (arch_register_req_is(req, must_be_different)) {
+ const unsigned other = req->other_different;
+ int i;
+
+ if (arch_register_req_is(req, should_be_same)) {
+ const unsigned same = req->other_same;
+
+ if (is_po2(other) && is_po2(same)) {
+ int idx_other = ntz(other);
+ int idx_same = ntz(same);
+
+ /*
+ * We can safely ignore a should_be_same x must_be_different y
+ * IFF both inputs are equal!
+ */
+ if (get_irn_n(skipped_irn, idx_other) == get_irn_n(skipped_irn, idx_same)) {
+ return;
+ }
+ }
}
- else if (arch_register_req_is(req, should_be_different_from_all)) {
- n = get_irn_arity(belower_skip_proj(irn));
- for (i = 0; i < n; i++) {
- gen_assure_different_pattern(irn, birg, get_irn_n(belower_skip_proj(irn), i));
+ for (i = 0; 1U << i <= other; ++i) {
+ if (other & (1U << i)) {
+ ir_node *different_from = get_irn_n(skipped_irn, i);
+ gen_assure_different_pattern(irn, different_from, env);
}
}
}
}
-
-
/**
* Calls the functions to assure register constraints.
*
if (is_Block(irn))
return;
- if (mode_is_datab(get_irn_mode(irn)))
+ if (sched_is_scheduled(irn) && mode_is_datab(get_irn_mode(irn)))
assure_different_constraints(irn, walk_env);
-
- return;
}
+/**
+ * Melt all copykeeps pointing to the same node
+ * (or Projs of the same node), copying the same operand.
+ */
+static void melt_copykeeps(constraint_env_t *cenv) {
+ be_irg_t *birg = cenv->birg;
+ ir_graph *irg = be_get_birg_irg(birg);
+ op_copy_assoc_t *entry;
+
+ /* for all */
+ foreach_pset(cenv->op_set, entry) {
+ int idx, num_ck;
+ ir_node *cp;
+ struct obstack obst;
+ ir_nodeset_iterator_t iter;
+ ir_node **ck_arr, **melt_arr;
+
+ obstack_init(&obst);
+
+ /* collect all copykeeps */
+ num_ck = idx = 0;
+ foreach_ir_nodeset(&entry->copies, cp, iter) {
+ if (be_is_CopyKeep(cp)) {
+ obstack_grow(&obst, &cp, sizeof(cp));
+ ++num_ck;
+ }
+#ifdef KEEP_ALIVE_COPYKEEP_HACK
+ else {
+ set_irn_mode(cp, mode_ANY);
+ keep_alive(cp);
+ }
+#endif /* KEEP_ALIVE_COPYKEEP_HACK */
+ }
+ /* compare each copykeep with all other copykeeps */
+ ck_arr = (ir_node **)obstack_finish(&obst);
+ for (idx = 0; idx < num_ck; ++idx) {
+ ir_node *ref, *ref_mode_T;
+
+ if (ck_arr[idx]) {
+ int j, n_melt;
+ ir_node **new_ck_in;
+ ir_node *new_ck;
+ ir_node *sched_pt = NULL;
+
+ n_melt = 1;
+ ref = ck_arr[idx];
+ ref_mode_T = skip_Proj(get_irn_n(ref, 1));
+ obstack_grow(&obst, &ref, sizeof(ref));
+
+ DBG((cenv->dbg, LEVEL_1, "Trying to melt %+F:\n", ref));
+
+ /* check for copykeeps pointing to the same mode_T node as the reference copykeep */
+ for (j = 0; j < num_ck; ++j) {
+ ir_node *cur_ck = ck_arr[j];
+
+ if (j != idx && cur_ck && skip_Proj(get_irn_n(cur_ck, 1)) == ref_mode_T) {
+ obstack_grow(&obst, &cur_ck, sizeof(cur_ck));
+ ir_nodeset_remove(&entry->copies, cur_ck);
+ DBG((cenv->dbg, LEVEL_1, "\t%+F\n", cur_ck));
+ ck_arr[j] = NULL;
+ ++n_melt;
+ sched_remove(cur_ck);
+ }
+ }
+ ck_arr[idx] = NULL;
+
+ /* check, if we found some candidates for melting */
+ if (n_melt == 1) {
+ DBG((cenv->dbg, LEVEL_1, "\tno candidate found\n"));
+ continue;
+ }
+
+ ir_nodeset_remove(&entry->copies, ref);
+ sched_remove(ref);
+
+ melt_arr = (ir_node **)obstack_finish(&obst);
+ /* melt all found copykeeps */
+ NEW_ARR_A(ir_node *, new_ck_in, n_melt);
+ for (j = 0; j < n_melt; ++j) {
+ new_ck_in[j] = get_irn_n(melt_arr[j], 1);
+
+ /* now, we can kill the melted keep, except the */
+ /* ref one, we still need some information */
+ if (melt_arr[j] != ref)
+ kill_node(melt_arr[j]);
+ }
+
+#ifdef KEEP_ALIVE_COPYKEEP_HACK
+ new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, mode_ANY);
+ keep_alive(new_ck);
+#else
+ new_ck = be_new_CopyKeep(entry->cls, irg, get_nodes_block(ref), be_get_CopyKeep_op(ref), n_melt, new_ck_in, get_irn_mode(ref));
+#endif /* KEEP_ALIVE_COPYKEEP_HACK */
+
+ /* set register class for all kept inputs */
+ for (j = 1; j <= n_melt; ++j)
+ be_node_set_reg_class(new_ck, j, entry->cls);
+
+ ir_nodeset_insert(&entry->copies, new_ck);
+
+ /* find scheduling point */
+ sched_pt = ref_mode_T;
+ do {
+ /* just walk along the schedule until a non-Keep/CopyKeep node is found */
+ sched_pt = sched_next(sched_pt);
+ } while (be_is_Keep(sched_pt) || be_is_CopyKeep(sched_pt));
+
+ sched_add_before(sched_pt, new_ck);
+ DBG((cenv->dbg, LEVEL_1, "created %+F, scheduled before %+F\n", new_ck, sched_pt));
+
+ /* finally: kill the reference copykeep */
+ kill_node(ref);
+ }
+ }
+
+ obstack_free(&obst, NULL);
+ }
+}
/**
* Walks over all nodes to assure register constraints.
* @param birg The birg structure containing the irg
*/
void assure_constraints(be_irg_t *birg) {
- irg_walk_blkwise_graph(birg->irg, NULL, assure_constraints_walker, birg);
-}
+ ir_graph *irg = be_get_birg_irg(birg);
+ constraint_env_t cenv;
+ op_copy_assoc_t *entry;
+ ir_node **nodes;
+ FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.constr");
+
+ DEBUG_ONLY(cenv.dbg = mod;)
+ cenv.birg = birg;
+ cenv.op_set = new_pset(cmp_op_copy_assoc, 16);
+ obstack_init(&cenv.obst);
+
+ irg_walk_blkwise_graph(irg, NULL, assure_constraints_walker, &cenv);
+
+ /* melt copykeeps, pointing to projs of */
+ /* the same mode_T node and keeping the */
+ /* same operand */
+ melt_copykeeps(&cenv);
+
+ /* for all */
+ foreach_pset(cenv.op_set, entry) {
+ int n;
+ ir_node *cp;
+ ir_nodeset_iterator_t iter;
+ be_ssa_construction_env_t senv;
+
+ n = ir_nodeset_size(&entry->copies);
+ nodes = alloca(n * sizeof(nodes[0]));
+
+ /* put the node in an array */
+ DBG((mod, LEVEL_1, "introduce copies for %+F ", entry->op));
+
+ /* collect all copies */
+ n = 0;
+ foreach_ir_nodeset(&entry->copies, cp, iter) {
+ nodes[n++] = cp;
+ DB((mod, LEVEL_1, ", %+F ", cp));
+ }
+ DB((mod, LEVEL_1, "\n"));
+
+ /* introduce the copies for the operand and it's copies */
+ be_ssa_construction_init(&senv, birg);
+ be_ssa_construction_add_copy(&senv, entry->op);
+ be_ssa_construction_add_copies(&senv, nodes, n);
+ be_ssa_construction_fix_users(&senv, entry->op);
+ be_ssa_construction_destroy(&senv);
+
+ /* Could be that not all CopyKeeps are really needed, */
+ /* so we transform unnecessary ones into Keeps. */
+ foreach_ir_nodeset(&entry->copies, cp, iter) {
+ if (be_is_CopyKeep(cp) && get_irn_n_edges(cp) < 1) {
+ ir_node *keep;
+ int n = get_irn_arity(cp);
+
+ keep = be_new_Keep(arch_get_irn_reg_class(cp, -1),
+ irg, get_nodes_block(cp), n, get_irn_in(cp) + 1);
+ sched_add_before(cp, keep);
+
+ /* Set all ins (including the block) of the CopyKeep BAD to keep the verifier happy. */
+ sched_remove(cp);
+ kill_node(cp);
+ }
+ }
+
+ ir_nodeset_destroy(&entry->copies);
+ }
+
+ del_pset(cenv.op_set);
+ obstack_free(&cenv.obst, NULL);
+ be_liveness_invalidate(be_get_birg_liveness(birg));
+}
/**
- * Calls the corresponding lowering function for the node.
- *
- * @param irn The node to be checked for lowering
- * @param walk_env The walker environment
+ * Push nodes that do not need to be permed through the Perm.
+ * This is commonly a reload cascade at block ends.
+ * @note This routine needs interference.
+ * @note Probably, we can implement it a little more efficient.
+ * Especially searching the frontier lazily might be better.
+ * @param perm The perm.
+ * @param data The walker data (lower_env_t).
+ * @return 1, if there is something left to perm over.
+ * 0, if removed the complete perm.
*/
-static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
- lower_env_t *env = walk_env;
- const arch_env_t *arch_env = env->chord_env->birg->main_env->arch_env;
+static int push_through_perm(ir_node *perm, void *data)
+{
+ lower_env_t *env = data;
+
+ ir_graph *irg = get_irn_irg(perm);
+ ir_node *bl = get_nodes_block(perm);
+ ir_node *node;
+ int arity = get_irn_arity(perm);
+ int *map;
+ int *proj_map;
+ bitset_t *moved = bitset_alloca(arity);
+ int n_moved;
+ int new_size;
+ ir_node *frontier = bl;
+ FIRM_DBG_REGISTER(firm_dbg_module_t *mod, "firm.be.lower.permmove");
- if (!is_Block(irn) && !is_Proj(irn)) {
- if (is_Perm(arch_env, irn)) {
- lower_perm_node(irn, walk_env);
+ int i, n;
+ const ir_edge_t *edge;
+ ir_node *one_proj = NULL, *irn;
+ const arch_register_class_t *cls = NULL;
+
+ DBG((mod, LEVEL_1, "perm move %+F irg %+F\n", perm, irg));
+
+ /* get some Proj and find out the register class of that Proj. */
+ edge = get_irn_out_edge_first_kind(perm, EDGE_KIND_NORMAL);
+ one_proj = get_edge_src_irn(edge);
+ assert(is_Proj(one_proj));
+ cls = arch_get_irn_reg_class(one_proj, -1);
+
+ /* Find the point in the schedule after which the
+ * potentially movable nodes must be defined.
+ * A Perm will only be pushed up to first instruction
+ * which lets an operand of itself die.
+ * If we would allow to move the Perm above this instruction,
+ * the former dead operand would be live now at the point of
+ * the Perm, increasing the register pressure by one.
+ */
+ sched_foreach_reverse_from (sched_prev(perm), irn) {
+ for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
+ ir_node *op = get_irn_n(irn, i);
+ if (arch_irn_consider_in_reg_alloc(cls, op) &&
+ !values_interfere(env->birg, op, one_proj)) {
+ frontier = irn;
+ goto found_front;
+ }
}
}
+found_front:
+
+ DBG((mod, LEVEL_2, "\tfrontier: %+F\n", frontier));
+
+ node = sched_prev(perm);
+ n_moved = 0;
+ while(!sched_is_begin(node)) {
+ const arch_register_req_t *req;
+ int input = -1;
+ ir_node *proj;
+
+ /* search if node is a INPUT of Perm */
+ foreach_out_edge(perm, edge) {
+ ir_node *out = get_edge_src_irn(edge);
+ int pn = get_Proj_proj(out);
+ ir_node *in = get_irn_n(perm, pn);
+ if (node == in) {
+ proj = out;
+ input = pn;
+ break;
+ }
+ }
+ /* it wasn't an input to the perm, we can't do anything more */
+ if(input < 0)
+ break;
+ if(!sched_comes_after(frontier, node))
+ break;
+ if (arch_irn_is(node, modify_flags))
+ break;
+ if(is_Proj(node)) {
+ req = arch_get_register_req(get_Proj_pred(node),
+ -1 - get_Proj_proj(node));
+ } else {
+ req = arch_get_register_req(node, -1);
+ }
+ if(req->type != arch_register_req_type_normal)
+ break;
+ for(i = get_irn_arity(node) - 1; i >= 0; --i) {
+ ir_node *opop = get_irn_n(node, i);
+ if (arch_irn_consider_in_reg_alloc(cls, opop)) {
+ break;
+ }
+ }
+ if(i >= 0)
+ break;
- return;
-}
+ DBG((mod, LEVEL_2, "\tmoving %+F after %+F, killing %+F\n", node, perm, proj));
-static void lower_print_perm_stat(lower_env_t *env) {
- int j, total_len_chain, total_len_cycle, total_size_perm, total_size_real_perm;
- unsigned i;
+ /* move the movable node in front of the Perm */
+ sched_remove(node);
+ sched_add_after(perm, node);
- printf("=== IRG: %s ===\n", get_entity_name(get_irg_entity(env->chord_env->irg)));
- for (i = 0; i < env->pstat_n; i++) {
- if (env->pstat[i]->num_perms == 0)
- continue;
+ /* give it the proj's register */
+ arch_set_irn_register(node, arch_get_irn_register(proj));
- printf("CLASS: %s\n", env->pstat[i]->cls->name);
- printf("# total perms: %d (size:num -> 1:%d", env->pstat[i]->num_perms, env->pstat[i]->perm_size_ar[0]);
+ /* reroute all users of the proj to the moved node. */
+ edges_reroute(proj, node, irg);
- total_size_perm = env->pstat[i]->perm_size_ar[0];
- for (j = 1; j < env->pstat[i]->cls->n_regs; j++) {
- total_size_perm += (j + 1) * env->pstat[i]->perm_size_ar[j];
- printf(", %d:%d", j + 1, env->pstat[i]->perm_size_ar[j]);
- }
- printf(")\n");
- printf("avg perm size: %.2f\n", env->pstat[i]->num_perms ? (float)total_size_perm / (float)env->pstat[i]->num_perms : 0);
+ /* and kill it */
+ set_Proj_pred(proj, new_Bad());
+ kill_node(proj);
- printf("# real perms: %d (size:num -> 1:%d", env->pstat[i]->num_real_perms, env->pstat[i]->real_perm_size_ar[0]);
+ bitset_set(moved, input);
+ n_moved++;
- total_size_real_perm = env->pstat[i]->real_perm_size_ar[0];
- for (j = 1; j < env->pstat[i]->cls->n_regs; j++) {
- total_size_real_perm += (j + 1) * env->pstat[i]->real_perm_size_ar[j];
- printf(", %d:%d", j + 1, env->pstat[i]->real_perm_size_ar[j]);
- }
- printf(")\n");
- printf("avg real perm size: %.2f\n", env->pstat[i]->num_real_perms ? (float)total_size_real_perm / (float)env->pstat[i]->num_real_perms : 0);
+ node = sched_prev(node);
+ }
- printf("# total chains: %d (lenght:num -> 1:%d", env->pstat[i]->num_chains, env->pstat[i]->chain_len_ar[0]);
+ /* well, we could not push anything through the perm */
+ if(n_moved == 0)
+ return 1;
- total_len_chain = env->pstat[i]->chain_len_ar[0];
+ new_size = arity - n_moved;
+ if(new_size == 0) {
+ return 0;
+ }
- for (j = 1; j < env->pstat[i]->cls->n_regs; j++) {
- total_len_chain += (j + 1) * env->pstat[i]->chain_len_ar[j];
- printf(", %d:%d", j + 1, env->pstat[i]->chain_len_ar[j]);
- }
- printf(")\n");
- printf("avg chain length: %.2f\n", env->pstat[i]->num_chains ? (float)total_len_chain / (float)env->pstat[i]->num_chains : 0);
- printf("avg chains/perm: %.2f\n", env->pstat[i]->num_real_perms ? (float)env->pstat[i]->num_chains / (float)env->pstat[i]->num_real_perms : 0);
+ map = alloca(new_size * sizeof(map[0]));
+ proj_map = alloca(arity * sizeof(proj_map[0]));
+ memset(proj_map, -1, sizeof(proj_map[0]));
+ n = 0;
+ for(i = 0; i < arity; ++i) {
+ if(bitset_is_set(moved, i))
+ continue;
+ map[n] = i;
+ proj_map[i] = n;
+ n++;
+ }
+ assert(n == new_size);
+ foreach_out_edge(perm, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ int pn = get_Proj_proj(proj);
+ pn = proj_map[pn];
+ assert(pn >= 0);
+ set_Proj_proj(proj, pn);
+ }
- printf("# total cycles: %d (length:num -> 1:%d", env->pstat[i]->num_cycles, env->pstat[i]->cycle_len_ar[0]);
+ be_Perm_reduce(perm, new_size, map);
+ return 1;
+}
- total_len_cycle = env->pstat[i]->cycle_len_ar[0];
- for (j = 1; j < env->pstat[i]->cls->n_regs; j++) {
- total_len_cycle += (j + 1) * env->pstat[i]->cycle_len_ar[j];
- printf(", %d:%d", j + 1, env->pstat[i]->cycle_len_ar[j]);
- }
- printf(")\n");
- printf("avg cycle length: %.2f\n", env->pstat[i]->num_cycles ? (float)total_len_cycle / (float)env->pstat[i]->num_cycles : 0);
- printf("avg cycles/perm: %.2f\n", env->pstat[i]->num_real_perms ? (float)env->pstat[i]->num_cycles / (float)env->pstat[i]->num_real_perms : 0);
- }
+/**
+ * Calls the corresponding lowering function for the node.
+ *
+ * @param irn The node to be checked for lowering
+ * @param walk_env The walker environment
+ */
+static void lower_nodes_after_ra_walker(ir_node *irn, void *walk_env) {
+ int perm_stayed;
+
+ if (is_Block(irn) || is_Proj(irn))
+ return;
+ if (!be_is_Perm(irn))
+ return;
+
+ perm_stayed = push_through_perm(irn, walk_env);
+ if (!perm_stayed)
+ return;
+
+ lower_perm_node(irn, walk_env);
}
/**
* Walks over all blocks in an irg and performs lowering need to be
* done after register allocation (e.g. perm lowering).
*
- * @param chord_env The chordal environment containing the irg
+ * @param birg The birg object
* @param do_copy 1 == resolve cycles with a free reg if available
*/
-void lower_nodes_after_ra(be_chordal_env_t *chord_env, int do_copy, int do_stat) {
+void lower_nodes_after_ra(be_irg_t *birg, int do_copy) {
lower_env_t env;
+ ir_graph *irg = be_get_birg_irg(birg);
- env.chord_env = chord_env;
- env.do_copy = do_copy;
- env.do_stat = do_stat;
+ env.birg = birg;
+ env.do_copy = do_copy;
FIRM_DBG_REGISTER(env.dbg_module, "firm.be.lower");
- /* if we want statistics: allocate memory for the data and initialize with 0 */
- if (do_stat) {
- const arch_isa_t *isa = chord_env->birg->main_env->arch_env->isa;
- int i, n = arch_isa_get_n_reg_class(isa);
+ /* we will need interference */
+ be_liveness_assure_chk(be_get_birg_liveness(birg));
- env.pstat = alloca(n * sizeof(env.pstat[0]));
- env.pstat_n = n;
-
- for (i = 0; i < n; i++) {
- const arch_register_class_t *cls = arch_isa_get_reg_class(isa, i);
- int n_regs = cls->n_regs;
-
- env.pstat[i] = alloca(sizeof(*(env.pstat[0])));
- memset(env.pstat[i], 0, sizeof(*(env.pstat[0])));
-
- env.pstat[i]->perm_size_ar = alloca(n_regs * sizeof(env.pstat[i]->perm_size_ar[0]));
- env.pstat[i]->real_perm_size_ar = alloca(n_regs * sizeof(env.pstat[i]->real_perm_size_ar[0]));
- env.pstat[i]->chain_len_ar = alloca(n_regs * sizeof(env.pstat[i]->chain_len_ar[0]));
- env.pstat[i]->cycle_len_ar = alloca(n_regs * sizeof(env.pstat[i]->cycle_len_ar[0]));
-
- memset(env.pstat[i]->perm_size_ar, 0, n_regs * sizeof(env.pstat[i]->perm_size_ar[0]));
- memset(env.pstat[i]->real_perm_size_ar, 0, n_regs * sizeof(env.pstat[i]->real_perm_size_ar[0]));
- memset(env.pstat[i]->chain_len_ar, 0, n_regs * sizeof(env.pstat[i]->chain_len_ar[0]));
- memset(env.pstat[i]->cycle_len_ar, 0, n_regs * sizeof(env.pstat[i]->cycle_len_ar[0]));
-
- env.pstat[i]->cls = cls;
- }
- }
-
- irg_walk_blkwise_graph(chord_env->irg, NULL, lower_nodes_after_ra_walker, &env);
-
- if (do_stat) {
- lower_print_perm_stat(&env);
- }
+ irg_walk_blkwise_graph(irg, NULL, lower_nodes_after_ra_walker, &env);
}
-
-#undef is_Perm
projects / libfirm / blobdiff