-/**
- * Chordal register allocation.
- * @author Sebastian Hack
- * @date 8.12.2004
+/*
+ * 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.
*
- * Copyright (C) Universitaet Karlsruhe
- * Released under the GPL
+ * 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 Chordal register allocation.
+ * @author Sebastian Hack
+ * @date 08.12.2004
+ * @version $Id$
+ */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-
-#ifdef HAVE_ALLOCA_H
-#include <alloca.h>
-#endif
-
#include <ctype.h>
#include "obst.h"
#include "pset.h"
#include "list.h"
#include "bitset.h"
+#include "raw_bitset.h"
#include "iterator.h"
#include "bipartite.h"
+#include "hungarian.h"
#include "irmode_t.h"
#include "irgraph_t.h"
#include "irgwalk.h"
#include "irdump.h"
#include "irdom.h"
+#include "irtools.h"
+#include "irbitset.h"
#include "debug.h"
#include "xmalloc.h"
+#include "iredges.h"
#include "beutil.h"
#include "besched.h"
-#include "benumb_t.h"
#include "besched_t.h"
#include "belive_t.h"
#include "benode_t.h"
-#include "bearch.h"
+#include "bearch_t.h"
#include "beirgmod.h"
#include "beifg.h"
-
+#include "beinsn_t.h"
+#include "bestatevent.h"
+#include "beirg_t.h"
+#include "beintlive_t.h"
+#include "bera.h"
#include "bechordal_t.h"
#include "bechordal_draw.h"
+#include "bemodule.h"
-#define DBG_LEVEL SET_LEVEL_0
-#define DBG_LEVEL_CHECK SET_LEVEL_0
+DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
#define NO_COLOR (-1)
-#define MAX(x, y) ((x) > (y) ? (x) : (y))
-#define MIN(x, y) ((x) < (y) ? (x) : (y))
-
#define DUMP_INTERVALS
typedef struct _be_chordal_alloc_env_t {
be_chordal_env_t *chordal_env;
- firm_dbg_module_t *constr_dbg; /**< Debug output for the constraint handler. */
pset *pre_colored; /**< Set of precolored nodes. */
bitset_t *live; /**< A liveness bitset. */
bitset_t *tmp_colors; /**< An auxiliary bitset which is as long as the number of colors in the class. */
bitset_t *colors; /**< The color mask. */
bitset_t *in_colors; /**< Colors used by live in values. */
- bitset_t *ignore_regs; /**< A bitset of all ignore registers in the current class. */
int colors_n; /**< The number of colors. */
} be_chordal_alloc_env_t;
/* Make a fourcc for border checking. */
#define BORDER_FOURCC FOURCC('B', 'O', 'R', 'D')
+#if 0
static void check_border_list(struct list_head *head)
{
border_t *x;
check_border_list(ent->value);
}
}
-
+#endif
/**
* Add an interval border to the list of a block's list
if(!is_def) {
border_t *def;
- b = obstack_alloc(&env->obst, sizeof(*b));
+ b = obstack_alloc(env->obst, sizeof(*b));
/* also allocate the def and tie it to the use. */
- def = obstack_alloc(&env->obst, sizeof(*def));
+ def = obstack_alloc(env->obst, sizeof(*def));
memset(def, 0, sizeof(*def));
b->other_end = def;
def->other_end = b;
*/
set_irn_link(irn, def);
- b->magic = BORDER_FOURCC;
- def->magic = BORDER_FOURCC;
+ DEBUG_ONLY(b->magic = BORDER_FOURCC);
+ DEBUG_ONLY(def->magic = BORDER_FOURCC);
}
/*
b->irn = irn;
b->step = step;
list_add_tail(&b->list, head);
- DBG((env->dbg, LEVEL_5, "\t\t%s adding %+F, step: %d\n", is_def ? "def" : "use", irn, step));
+ DBG((dbg, LEVEL_5, "\t\t%s adding %+F, step: %d\n", is_def ? "def" : "use", irn, step));
return b;
*/
static INLINE int has_reg_class(const be_chordal_env_t *env, const ir_node *irn)
{
- // return arch_irn_has_reg_class(env->main_env->arch_env, irn, -1, env->cls);
return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn);
}
{
bitset_t *tmp = alloc_env->tmp_colors;
bitset_copy(tmp, colors);
- bitset_or(tmp, alloc_env->ignore_regs);
+ bitset_or(tmp, alloc_env->chordal_env->ignore_colors);
return bitset_next_clear(tmp, 0);
}
-typedef struct _operand_t operand_t;
-
-struct _operand_t {
- ir_node *irn;
- ir_node *carrier;
- operand_t *partner;
- bitset_t *regs;
- int pos;
- arch_register_req_t req;
- unsigned has_constraints : 1;
-};
-
-typedef struct {
- operand_t *ops;
- int n_ops;
- int use_start;
- ir_node *next_insn;
- ir_node *irn;
- unsigned in_constraints : 1;
- unsigned out_constraints : 1;
- unsigned has_constraints : 1;
- unsigned pre_colored : 1;
-} insn_t;
-
-#define insn_n_defs(insn) ((insn)->use_start)
-#define insn_n_uses(insn) ((insn)->n_ops - (insn)->use_start)
-
-static insn_t *scan_insn(be_chordal_alloc_env_t *alloc_env, ir_node *irn, struct obstack *obst)
+static bitset_t *get_decisive_partner_regs(bitset_t *bs, const be_operand_t *o1, const be_operand_t *o2)
{
- const be_chordal_env_t *env = alloc_env->chordal_env;
- const arch_env_t *arch_env = env->birg->main_env->arch_env;
- operand_t o;
- insn_t *insn;
- int i, n;
- int pre_colored = 0;
-
- insn = obstack_alloc(obst, sizeof(insn[0]));
- memset(insn, 0, sizeof(insn[0]));
-
- insn->irn = irn;
- insn->next_insn = sched_next(irn);
- if(get_irn_mode(irn) == mode_T) {
- ir_node *p;
-
- for(p = sched_next(irn); is_Proj(p); p = sched_next(p)) {
- if(arch_irn_consider_in_reg_alloc(arch_env, env->cls, p)) {
- arch_get_register_req(arch_env, &o.req, p, -1);
- o.carrier = p;
- o.irn = irn;
- o.pos = -(get_Proj_proj(p) + 1);
- o.partner = NULL;
- o.has_constraints = arch_register_req_is(&o.req, limited);
- obstack_grow(obst, &o, sizeof(o));
- insn->n_ops++;
- insn->out_constraints |= o.has_constraints;
- pre_colored += arch_get_irn_register(arch_env, p) != NULL;
- }
- }
+ bitset_t *res = bs;
- insn->next_insn = p;
+ if(!o1) {
+ bitset_copy(bs, o2->regs);
+ return bs;
}
- else if(arch_irn_consider_in_reg_alloc(arch_env, env->cls, irn)) {
- arch_get_register_req(arch_env, &o.req, irn, -1);
- o.carrier = irn;
- o.irn = irn;
- o.pos = -1;
- o.partner = NULL;
- o.has_constraints = arch_register_req_is(&o.req, limited);
- obstack_grow(obst, &o, sizeof(o));
- insn->n_ops++;
- insn->out_constraints |= o.has_constraints;
- pre_colored += arch_get_irn_register(arch_env, irn) != NULL;
+ if(!o2) {
+ bitset_copy(bs, o1->regs);
+ return bs;
}
- insn->pre_colored = pre_colored == insn->n_ops && insn->n_ops > 0;
- insn->use_start = insn->n_ops;
+ assert(o1->req->cls == o2->req->cls || ! o1->req->cls || ! o2->req->cls);
+
+ if(bitset_contains(o1->regs, o2->regs))
+ bitset_copy(bs, o1->regs);
+ else if(bitset_contains(o2->regs, o1->regs))
+ bitset_copy(bs, o2->regs);
+ else
+ res = NULL;
+
+ return res;
+}
- for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
+static be_insn_t *chordal_scan_insn(be_chordal_env_t *env, ir_node *irn)
+{
+ be_insn_env_t ie;
+
+ ie.ignore_colors = env->ignore_colors;
+ ie.aenv = env->birg->main_env->arch_env;
+ ie.obst = env->obst;
+ ie.cls = env->cls;
+ return be_scan_insn(&ie, irn);
+}
+
+static ir_node *prepare_constr_insn(be_chordal_env_t *env, ir_node *irn)
+{
+ const be_irg_t *birg = env->birg;
+ const arch_env_t *aenv = birg->main_env->arch_env;
+ bitset_t *tmp = bitset_alloca(env->cls->n_regs);
+ bitset_t *def_constr = bitset_alloca(env->cls->n_regs);
+ ir_node *bl = get_nodes_block(irn);
+ be_lv_t *lv = env->birg->lv;
+
+ be_insn_t *insn;
+ int i, j;
+
+ for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
ir_node *op = get_irn_n(irn, i);
+ ir_node *copy;
+ const arch_register_t *reg;
+ const arch_register_req_t *req;
- if(arch_irn_consider_in_reg_alloc(arch_env, env->cls, op)) {
- arch_get_register_req(arch_env, &o.req, irn, i);
- o.carrier = op;
- o.irn = irn;
- o.pos = i;
- o.partner = NULL;
- o.has_constraints = arch_register_req_is(&o.req, limited);
- obstack_grow(obst, &o, sizeof(o));
- insn->n_ops++;
- insn->in_constraints |= o.has_constraints;
- }
- }
+ if (arch_get_irn_reg_class(aenv, irn, i) != env->cls)
+ continue;
+
+ reg = arch_get_irn_register(aenv, op);
- insn->has_constraints = insn->in_constraints | insn->out_constraints;
- insn->ops = obstack_finish(obst);
+ if (reg == NULL || !arch_register_type_is(reg, ignore))
+ continue;
+ if(arch_register_type_is(reg, joker))
+ continue;
- /* Compute the admissible registers bitsets. */
- for(i = 0; i < insn->n_ops; ++i) {
- operand_t *op = &insn->ops[i];
+ req = arch_get_register_req(aenv, irn, i);
+ if (!arch_register_req_is(req, limited))
+ continue;
- assert(op->req.cls == env->cls);
- op->regs = bitset_obstack_alloc(obst, env->cls->n_regs);
+ if (rbitset_is_set(req->limited, reg->index))
+ continue;
- if(arch_register_req_is(&op->req, limited))
- op->req.limited(op->req.limited_env, op->regs);
- else
- arch_put_non_ignore_regs(env->birg->main_env->arch_env, env->cls, op->regs);
+ copy = be_new_Copy(env->cls, env->irg, bl, op);
+ be_stat_ev("constr_copy", 1);
+
+ sched_add_before(irn, copy);
+ set_irn_n(irn, i, copy);
+ DBG((dbg, LEVEL_3, "inserting ignore arg copy %+F for %+F pos %d\n", copy, irn, i));
}
- return insn;
-}
+ insn = chordal_scan_insn(env, irn);
-static bitset_t *get_decisive_partner_regs(bitset_t *bs, const operand_t *o1, const operand_t *o2)
-{
- bitset_t *res = bs;
+ if(!insn->has_constraints)
+ goto end;
- if(!o1) {
- bitset_copy(bs, o2->regs);
- return bs;
+ /* insert copies for nodes that occur constrained more than once. */
+ for(i = insn->use_start; i < insn->n_ops; ++i) {
+ be_operand_t *op = &insn->ops[i];
+
+ if(!op->has_constraints)
+ continue;
+
+ for(j = i + 1; j < insn->n_ops; ++j) {
+ ir_node *copy;
+ be_operand_t *a_op = &insn->ops[j];
+
+ if(a_op->carrier != op->carrier || !a_op->has_constraints)
+ continue;
+
+ if (be_is_Copy(get_irn_n(insn->irn, a_op->pos)))
+ continue;
+
+ copy = be_new_Copy(env->cls, env->irg, bl, op->carrier);
+ be_stat_ev("constr_copy", 1);
+
+ sched_add_before(insn->irn, copy);
+ set_irn_n(insn->irn, a_op->pos, copy);
+ DBG((dbg, LEVEL_3, "inserting multiple constr copy %+F for %+F pos %d\n", copy, insn->irn, a_op->pos));
+ }
}
- if(!o2) {
- bitset_copy(bs, o1->regs);
- return bs;
+ /* collect all registers occuring in out constraints. */
+ for(i = 0; i < insn->use_start; ++i) {
+ be_operand_t *op = &insn->ops[i];
+ if(op->has_constraints)
+ bitset_or(def_constr, op->regs);
}
- assert(o1->req.cls == o2->req.cls);
+ /*
+ insert copies for all constrained arguments living through the node
+ and being constrained to a register which also occurs in out constraints.
+ */
+ for(i = insn->use_start; i < insn->n_ops; ++i) {
+ ir_node *copy;
+ be_operand_t *op = &insn->ops[i];
- if(bitset_contains(o1->regs, o2->regs))
- bitset_copy(bs, o1->regs);
- else if(bitset_contains(o2->regs, o1->regs))
- bitset_copy(bs, o2->regs);
- else
- res = NULL;
+ bitset_copy(tmp, op->regs);
+ bitset_and(tmp, def_constr);
- return res;
+ /*
+ Check, if
+ 1) the operand is constrained.
+ 2) lives through the node.
+ 3) is constrained to a register occuring in out constraints.
+ */
+ if(!op->has_constraints ||
+ !values_interfere(birg, insn->irn, op->carrier) ||
+ bitset_popcnt(tmp) == 0)
+ continue;
+
+ /*
+ only create the copy if the operand is no copy.
+ this is necessary since the assure constraints phase inserts
+ Copies and Keeps for operands which must be different from the
+ results. Additional copies here would destroy this.
+ */
+ if (be_is_Copy(get_irn_n(insn->irn, op->pos)))
+ continue;
+
+ copy = be_new_Copy(env->cls, env->irg, bl, op->carrier);
+
+ sched_add_before(insn->irn, copy);
+ set_irn_n(insn->irn, op->pos, copy);
+ DBG((dbg, LEVEL_3, "inserting constr copy %+F for %+F pos %d\n", copy, insn->irn, op->pos));
+ be_liveness_update(lv, op->carrier);
+ }
+
+end:
+ obstack_free(env->obst, insn);
+ return insn->next_insn;
}
-static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, insn_t *insn)
+static void pre_spill_prepare_constr_walker(ir_node *bl, void *data)
+{
+ be_chordal_env_t *env = data;
+ ir_node *irn;
+ for(irn = sched_first(bl); !sched_is_end(irn);) {
+ irn = prepare_constr_insn(env, irn);
+ }
+}
+
+void be_pre_spill_prepare_constr(be_chordal_env_t *cenv) {
+ irg_block_walk_graph(cenv->irg, pre_spill_prepare_constr_walker, NULL, (void *) cenv);
+}
+
+static void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, be_insn_t *insn)
{
const be_chordal_env_t *env = alloc_env->chordal_env;
- const arch_env_t *aenv = env->birg->main_env->arch_env;
- firm_dbg_module_t *dbg = alloc_env->constr_dbg;
- int n_uses = insn_n_uses(insn);
- int n_defs = insn_n_defs(insn);
- int max_pairs = MIN(n_uses, n_defs);
- bitset_t *bs = bitset_alloca(env->cls->n_regs);
- bipartite_t *bp = bipartite_new(n_defs, n_uses);
- int *pairing = alloca(MAX(n_defs, n_uses) * sizeof(pairing[0]));
+ int n_uses = be_insn_n_uses(insn);
+ int n_defs = be_insn_n_defs(insn);
+ bitset_t *bs = bitset_alloca(env->cls->n_regs);
+ int *pairing = alloca(MAX(n_defs, n_uses) * sizeof(pairing[0]));
int i, j;
For each out operand, try to find an in operand which can be assigned the
same register as the out operand.
*/
- for(j = 0; j < insn->use_start; ++j) {
- operand_t *out_op = &insn->ops[j];
+ for (j = 0; j < insn->use_start; ++j) {
+ int smallest = -1;
+ int smallest_n_regs = 2 * env->cls->n_regs + 1;
+ be_operand_t *out_op = &insn->ops[j];
/* Try to find an in operand which has ... */
for(i = insn->use_start; i < insn->n_ops; ++i) {
- const operand_t *op = &insn->ops[i];
-
- /*
- The in operand can only be paired with a def, if the node defining the
- operand's value does not interfere with the instruction itself. That
- would mean, that it is live at the instruction, so no result of the instruction
- can have the same register as the operand.
-
- Furthermore, tow operands can be paired, if the admissible registers
- of one are a subset of the other's. We record the operand whose constraints
- count in the decisive array.
- */
- if(!values_interfere(op->irn, op->carrier)) {
- if(get_decisive_partner_regs(bs, out_op, op))
- bipartite_add(bp, j, i - insn->use_start);
+ int n_total;
+ const be_operand_t *op = &insn->ops[i];
+
+ if (op->partner != NULL)
+ continue;
+ if (values_interfere(env->birg, op->irn, op->carrier))
+ continue;
+
+ bitset_clear_all(bs);
+ bitset_copy(bs, op->regs);
+ bitset_and(bs, out_op->regs);
+ n_total = bitset_popcnt(op->regs) + bitset_popcnt(out_op->regs);
+
+ if (bitset_popcnt(bs) > 0 && n_total < smallest_n_regs) {
+ smallest = i;
+ smallest_n_regs = n_total;
}
}
- }
- /* Compute the pairing. */
- bipartite_matching(bp, pairing);
- for(i = 0; i < insn->use_start; ++i) {
- int p = pairing[i] + insn->use_start;
+ if (smallest >= 0) {
+ be_operand_t *partner = &insn->ops[smallest];
+ for(i = insn->use_start; i < insn->n_ops; ++i) {
+ if(insn->ops[i].carrier == partner->carrier)
+ insn->ops[i].partner = out_op;
+ }
- if(p >= insn->use_start) {
- insn->ops[i].partner = &insn->ops[p];
- insn->ops[p].partner = &insn->ops[i];
+ out_op->partner = partner;
+ partner->partner = out_op;
}
}
-
- bipartite_free(bp);
}
-static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, insn_t **the_insn)
+static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env,
+ be_insn_t **the_insn)
{
be_chordal_env_t *env = alloc_env->chordal_env;
const arch_env_t *aenv = env->birg->main_env->arch_env;
- firm_dbg_module_t *dbg = alloc_env->constr_dbg;
- insn_t *insn = *the_insn;
- ir_node *bl = get_nodes_block(insn->irn);
- ir_node *copy = NULL;
+ be_insn_t *insn = *the_insn;
ir_node *perm = NULL;
bitset_t *out_constr = bitset_alloca(env->cls->n_regs);
- bitset_t *bs = bitset_alloca(env->cls->n_regs);
-
+ const ir_edge_t *edge;
int i;
assert(insn->has_constraints && "only do this for constrained nodes");
be copied.
*/
for(i = 0; i < insn->use_start; ++i) {
- operand_t *op = &insn->ops[i];
+ be_operand_t *op = &insn->ops[i];
if(op->has_constraints)
bitset_or(out_constr, op->regs);
}
- /*
- Now, figure out which input operand must be copied since it has input
- constraints which are also output constraints.
- */
- for(i = insn->use_start; i < insn->n_ops; ++i) {
- operand_t *op = &insn->ops[i];
- if(op->has_constraints && (values_interfere(op->carrier, insn->irn) || arch_irn_is(aenv, op->carrier, ignore))) {
- bitset_copy(bs, op->regs);
- bitset_and(bs, out_constr);
-
- /*
- The operand (interfering with the node) has input constraints
- which also occur as output constraints, so insert a copy.
- */
- if(bitset_popcnt(bs) > 0) {
- copy = be_new_Copy(op->req.cls, env->irg, bl, op->carrier);
- insn->ops[i].carrier = copy;
- sched_add_before(insn->irn, copy);
-
- DBG((dbg, LEVEL_2, "adding copy for interfering and constrained op %+F\n", op->carrier));
- }
- }
- }
-
/*
Make the Perm, recompute liveness and re-scan the insn since the
in operands are now the Projs of the Perm.
*/
- perm = insert_Perm_after(aenv, env->cls, env->dom_front, sched_prev(insn->irn));
+ perm = insert_Perm_after(env->birg, env->cls, sched_prev(insn->irn));
/* Registers are propagated by insert_Perm_after(). Clean them here! */
- if(perm) {
- const ir_edge_t *edge;
+ if(perm == NULL)
+ return NULL;
- foreach_out_edge(perm, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- arch_set_irn_register(aenv, proj, NULL);
- }
+ be_stat_ev("constr_perm", get_irn_arity(perm));
+ foreach_out_edge(perm, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ arch_set_irn_register(aenv, proj, NULL);
+ }
- /*
- We also have to re-build the insn since the input operands are now the Projs of
- the Perm. Recomputing liveness is also a good idea if a Perm is inserted, since
- the live sets may change.
- */
- be_liveness(env->irg);
- obstack_free(&env->obst, insn);
- *the_insn = insn = scan_insn(alloc_env, insn->irn, &env->obst);
+ /*
+ We also have to re-build the insn since the input operands are now the Projs of
+ the Perm. Recomputing liveness is also a good idea if a Perm is inserted, since
+ the live sets may change.
+ */
+ obstack_free(env->obst, insn);
+ *the_insn = insn = chordal_scan_insn(env, insn->irn);
+ /*
+ Copy the input constraints of the insn to the Perm as output
+ constraints. Succeeding phases (coalescing) will need that.
+ */
+ for(i = insn->use_start; i < insn->n_ops; ++i) {
+ be_operand_t *op = &insn->ops[i];
+ ir_node *proj = op->carrier;
/*
- Copy the input constraints of the insn to the Perm as output
- constraints. Succeeding phases (coalescing will need that).
+ Note that the predecessor must not be a Proj of the Perm,
+ since ignore-nodes are not Perm'ed.
*/
- for(i = insn->use_start; i < insn->n_ops; ++i) {
- operand_t *op = &insn->ops[i];
- ir_node *proj = op->carrier;
- /*
- Note that the predecessor must not be a Proj of the Perm,
- since ignore-nodes are not Perm'ed.
- */
- if(op->has_constraints && is_Proj(proj) && get_Proj_pred(proj) == perm) {
- be_set_constr_limited(perm, BE_OUT_POS(get_Proj_proj(proj)), &op->req);
- }
+ if(op->has_constraints && is_Proj(proj) && get_Proj_pred(proj) == perm) {
+ be_set_constr_limited(perm, BE_OUT_POS(get_Proj_proj(proj)), op->req);
}
}
return perm;
}
-static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env, ir_node *irn, int *silent)
+static ir_node *handle_constraints(be_chordal_alloc_env_t *alloc_env,
+ ir_node *irn, int *silent)
{
+ const arch_env_t *aenv;
+ int n_regs;
+ bitset_t *bs;
+ ir_node **alloc_nodes;
+ //hungarian_problem_t *bp;
+ int *assignment;
+ pmap *partners;
+ int i, n_alloc;
+ bitset_pos_t col;
+ const ir_edge_t *edge;
+ ir_node *perm = NULL;
+ //int match_res, cost;
be_chordal_env_t *env = alloc_env->chordal_env;
- void *base = obstack_base(&env->obst);
- insn_t *insn = scan_insn(alloc_env, irn, &env->obst);
+ void *base = obstack_base(env->obst);
+ be_insn_t *insn = chordal_scan_insn(env, irn);
ir_node *res = insn->next_insn;
int be_silent = *silent;
+ be_irg_t *birg = env->birg;
+ bipartite_t *bp;
if(insn->pre_colored) {
int i;
Perms inserted before the constraint handling phase are considered to be
correctly precolored. These Perms arise during the ABI handling phase.
*/
- if(insn->has_constraints) {
- firm_dbg_module_t *dbg = alloc_env->constr_dbg;
- const arch_env_t *aenv = env->birg->main_env->arch_env;
- int n_regs = env->cls->n_regs;
- bitset_t *bs = bitset_alloca(n_regs);
- bitset_t *non_ignore = bitset_alloca(n_regs);
- ir_node **alloc_nodes = alloca(n_regs * sizeof(alloc_nodes[0]));
- bipartite_t *bp = bipartite_new(n_regs, n_regs);
- int *assignment = alloca(n_regs * sizeof(assignment[0]));
- pmap *partners = pmap_create();
-
- int i, n_alloc;
- long col;
- const ir_edge_t *edge;
- ir_node *perm = NULL;
+ if(!insn->has_constraints)
+ goto end;
- /*
- prepare the constraint handling of this node.
- Perms are constructed and Copies are created for constrained values
- interfering with the instruction.
- */
- perm = pre_process_constraints(alloc_env, &insn);
+ aenv = env->birg->main_env->arch_env;
+ n_regs = env->cls->n_regs;
+ bs = bitset_alloca(n_regs);
+ alloc_nodes = alloca(n_regs * sizeof(alloc_nodes[0]));
+ //bp = hungarian_new(n_regs, n_regs, 2, HUNGARIAN_MATCH_PERFECT);
+ bp = bipartite_new(n_regs, n_regs);
+ assignment = alloca(n_regs * sizeof(assignment[0]));
+ partners = pmap_create();
- /* find suitable in operands to the out operands of the node. */
- pair_up_operands(alloc_env, insn);
+ /*
+ prepare the constraint handling of this node.
+ Perms are constructed and Copies are created for constrained values
+ interfering with the instruction.
+ */
+ perm = pre_process_constraints(alloc_env, &insn);
- /*
- look at the in/out operands and add each operand (and its possible partner)
- to a bipartite graph (left: nodes with partners, right: admissible colors).
- */
- for(i = 0, n_alloc = 0; i < insn->n_ops; ++i) {
- operand_t *op = &insn->ops[i];
+ /* find suitable in operands to the out operands of the node. */
+ pair_up_operands(alloc_env, insn);
- /*
- If the operand has no partner or the partner has not been marked
- for allocation, determine the admissible registers and mark it
- for allocation by associating the node and its partner with the
- set of admissible registers via a bipartite graph.
- */
- if(!op->partner || !pmap_contains(partners, op->partner->carrier)) {
+ /*
+ look at the in/out operands and add each operand (and its possible partner)
+ to a bipartite graph (left: nodes with partners, right: admissible colors).
+ */
+ for(i = 0, n_alloc = 0; i < insn->n_ops; ++i) {
+ be_operand_t *op = &insn->ops[i];
- pmap_insert(partners, op->carrier, op->partner ? op->partner->carrier : NULL);
- alloc_nodes[n_alloc] = op->carrier;
+ /*
+ If the operand has no partner or the partner has not been marked
+ for allocation, determine the admissible registers and mark it
+ for allocation by associating the node and its partner with the
+ set of admissible registers via a bipartite graph.
+ */
+ if(!op->partner || !pmap_contains(partners, op->partner->carrier)) {
+ ir_node *partner = op->partner ? op->partner->carrier : NULL;
+ int i;
+
+ pmap_insert(partners, op->carrier, partner);
+ if(partner != NULL)
+ pmap_insert(partners, partner, op->carrier);
+
+ /* don't insert a node twice */
+ for(i = 0; i < n_alloc; ++i) {
+ if(alloc_nodes[i] == op->carrier) {
+ break;
+ }
+ }
+ if(i < n_alloc)
+ continue;
- DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, op->partner ? op->partner->carrier : NULL));
+ alloc_nodes[n_alloc] = op->carrier;
- bitset_clear_all(bs);
- get_decisive_partner_regs(bs, op, op->partner);
+ DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier,
+ partner));
- DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier, bs));
+ bitset_clear_all(bs);
+ get_decisive_partner_regs(bs, op, op->partner);
- bitset_foreach(bs, col)
- bipartite_add(bp, n_alloc, col);
+ DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier,
+ bs));
- n_alloc++;
+ bitset_foreach(bs, col) {
+ //hungarian_add(bp, n_alloc, col, 1);
+ bipartite_add(bp, n_alloc, col);
}
- }
- /*
- Put all nodes which live by the constrained instruction also to the
- allocation bipartite graph. They are considered unconstrained.
- */
- if(perm) {
- foreach_out_edge(perm, edge) {
- ir_node *proj = get_edge_src_irn(edge);
+ n_alloc++;
+ }
+ }
- assert(is_Proj(proj));
+ /*
+ Put all nodes which live through the constrained instruction also to the
+ allocation bipartite graph. They are considered unconstrained.
+ */
+ if(perm != NULL) {
+ foreach_out_edge(perm, edge) {
+ int i;
+ ir_node *proj = get_edge_src_irn(edge);
- if(values_interfere(proj, irn)) {
- assert(n_alloc < n_regs);
- alloc_nodes[n_alloc] = proj;
- pmap_insert(partners, proj, NULL);
+ assert(is_Proj(proj));
- bitset_clear_all(bs);
- arch_put_non_ignore_regs(aenv, env->cls, bs);
- bitset_foreach(bs, col)
- bipartite_add(bp, n_alloc, col);
+ if(!values_interfere(birg, proj, irn) || pmap_contains(partners, proj))
+ continue;
- n_alloc++;
+ /* don't insert a node twice */
+ for(i = 0; i < n_alloc; ++i) {
+ if(alloc_nodes[i] == proj) {
+ break;
}
}
- }
-
- /* Compute a valid register allocation. */
- bipartite_matching(bp, assignment);
+ if(i < n_alloc)
+ continue;
- /* Assign colors obtained from the matching. */
- for(i = 0; i < n_alloc; ++i) {
- const arch_register_t *reg;
- ir_node *nodes[2];
- int j;
-
- assert(assignment[i] >= 0 && "there must have been a register assigned");
- reg = arch_register_for_index(env->cls, assignment[i]);
- nodes[0] = alloc_nodes[i];
- nodes[1] = pmap_get(partners, alloc_nodes[i]);
+ assert(n_alloc < n_regs);
- for(j = 0; j < 2; ++j) {
- if(!nodes[j])
- continue;
+ alloc_nodes[n_alloc] = proj;
+ pmap_insert(partners, proj, NULL);
- arch_set_irn_register(aenv, nodes[j], reg);
- pset_hinsert_ptr(alloc_env->pre_colored, nodes[j]);
- DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", nodes[j], reg->name));
+ bitset_clear_all(bs);
+ arch_put_non_ignore_regs(aenv, env->cls, bs);
+ bitset_andnot(bs, env->ignore_colors);
+ bitset_foreach(bs, col) {
+ //hungarian_add(bp, n_alloc, col, 1);
+ bipartite_add(bp, n_alloc, col);
}
+
+ n_alloc++;
}
+ }
+ /* Compute a valid register allocation. */
+#if 0
+ hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
+ match_res = hungarian_solve(bp, assignment, &cost, 1);
+ assert(match_res == 0 && "matching failed");
+#else
+ bipartite_matching(bp, assignment);
+#endif
- /* Allocate the non-constrained Projs of the Perm. */
- if(perm) {
+ /* Assign colors obtained from the matching. */
+ for(i = 0; i < n_alloc; ++i) {
+ const arch_register_t *reg;
+ ir_node *irn;
- bitset_clear_all(bs);
+ assert(assignment[i] >= 0 && "there must have been a register assigned");
+ reg = arch_register_for_index(env->cls, assignment[i]);
+ assert(! (reg->type & arch_register_type_ignore));
- /* Put the colors of all Projs in a bitset. */
- foreach_out_edge(perm, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- const arch_register_t *reg = arch_get_irn_register(aenv, proj);
+ irn = alloc_nodes[i];
+ if (irn != NULL) {
+ arch_set_irn_register(aenv, irn, reg);
+ (void) pset_hinsert_ptr(alloc_env->pre_colored, irn);
+ DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", irn, reg->name));
+ }
- if(reg != NULL)
- bitset_set(bs, reg->index);
- }
+ irn = pmap_get(partners, alloc_nodes[i]);
+ if (irn != NULL) {
+ arch_set_irn_register(aenv, irn, reg);
+ (void) pset_hinsert_ptr(alloc_env->pre_colored, irn);
+ DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", irn, reg->name));
+ }
+ }
- /* Assign the not yet assigned Projs of the Perm a suitable color. */
- foreach_out_edge(perm, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- const arch_register_t *reg = arch_get_irn_register(aenv, proj);
+ /* Allocate the non-constrained Projs of the Perm. */
+ if(perm != NULL) {
+ bitset_clear_all(bs);
- DBG((dbg, LEVEL_2, "\tchecking reg of %+F: %s\n", proj, reg ? reg->name : "<none>"));
+ /* Put the colors of all Projs in a bitset. */
+ foreach_out_edge(perm, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ const arch_register_t *reg = arch_get_irn_register(aenv, proj);
- if(reg == NULL) {
- col = get_next_free_reg(alloc_env, bs);
- reg = arch_register_for_index(env->cls, col);
- bitset_set(bs, reg->index);
- arch_set_irn_register(aenv, proj, reg);
- pset_insert_ptr(alloc_env->pre_colored, proj);
- DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", proj, reg->name));
- }
- }
+ if(reg != NULL)
+ bitset_set(bs, reg->index);
}
- pmap_destroy(partners);
+ /* Assign the not yet assigned Projs of the Perm a suitable color. */
+ foreach_out_edge(perm, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+ const arch_register_t *reg = arch_get_irn_register(aenv, proj);
+
+ DBG((dbg, LEVEL_2, "\tchecking reg of %+F: %s\n", proj, reg ? reg->name : "<none>"));
+
+ if(reg == NULL) {
+ col = get_next_free_reg(alloc_env, bs);
+ reg = arch_register_for_index(env->cls, col);
+ bitset_set(bs, reg->index);
+ arch_set_irn_register(aenv, proj, reg);
+ pset_insert_ptr(alloc_env->pre_colored, proj);
+ DBG((dbg, LEVEL_2, "\tsetting %+F to register %s\n", proj, reg->name));
+ }
+ }
}
+ bipartite_free(bp);
+ //hungarian_free(bp);
+ pmap_destroy(partners);
+
end:
- obstack_free(&env->obst, base);
+ obstack_free(env->obst, base);
return res;
}
static void constraints(ir_node *bl, void *data)
{
be_chordal_alloc_env_t *env = data;
- firm_dbg_module_t *dbg = firm_dbg_register("firm.be.chordal.constr");
- arch_env_t *arch_env = env->chordal_env->birg->main_env->arch_env;
/*
Start silent in the start block.
be_chordal_alloc_env_t *alloc_env = env_ptr;
be_chordal_env_t *env = alloc_env->chordal_env;
- const arch_env_t *arch_env = env->birg->main_env->arch_env;
bitset_t *live = alloc_env->live;
- firm_dbg_module_t *dbg = env->dbg;
ir_node *irn;
+ be_lv_t *lv = env->birg->lv;
int i, n;
+ bitset_pos_t elm;
unsigned step = 0;
unsigned pressure = 0;
struct list_head *head;
- pset *live_in = put_live_in(block, pset_new_ptr_default());
- pset *live_end = put_live_end(block, pset_new_ptr_default());
DBG((dbg, LEVEL_1, "Computing pressure in block %+F\n", block));
bitset_clear_all(live);
/* Set up the border list in the block info */
- head = obstack_alloc(&env->obst, sizeof(*head));
+ head = obstack_alloc(env->obst, sizeof(*head));
INIT_LIST_HEAD(head);
assert(pmap_get(env->border_heads, block) == NULL);
pmap_insert(env->border_heads, block, head);
* Make final uses of all values live out of the block.
* They are necessary to build up real intervals.
*/
- for(irn = pset_first(live_end); irn; irn = pset_next(live_end)) {
+ be_lv_foreach(lv, block, be_lv_state_end, i) {
+ ir_node *irn = be_lv_get_irn(lv, block, i);
if(has_reg_class(env, irn)) {
- DBG((dbg, LEVEL_3, "\tMaking live: %+F/%d\n", irn, get_irn_graph_nr(irn)));
- bitset_set(live, get_irn_graph_nr(irn));
+ DBG((dbg, LEVEL_3, "\tMaking live: %+F/%d\n", irn, get_irn_idx(irn)));
+ bitset_set(live, get_irn_idx(irn));
border_use(irn, step, 0);
}
}
*/
sched_foreach_reverse(block, irn) {
DBG((dbg, LEVEL_1, "\tinsn: %+F, pressure: %d\n", irn, pressure));
- DBG((dbg, LEVEL_2, "\tlive: %b\n", live));
+ DBG((dbg, LEVEL_2, "\tlive: %B\n", live));
+
+ if (get_irn_mode(irn) == mode_T) {
+ const ir_edge_t *edge;
+
+ foreach_out_edge(irn, edge) {
+ ir_node *proj = get_edge_src_irn(edge);
+
+ /*
+ * If the node defines some value, which can put into a
+ * register of the current class, make a border for it.
+ */
+ if(has_reg_class(env, proj)) {
+ int nr = get_irn_idx(proj);
+
+ bitset_clear(live, nr);
+ border_def(proj, step, 1);
+ }
+ }
+ }
/*
* If the node defines some value, which can put into a
* register of the current class, make a border for it.
*/
if(has_reg_class(env, irn)) {
- int nr = get_irn_graph_nr(irn);
+ int nr = get_irn_idx(irn);
bitset_clear(live, nr);
border_def(irn, step, 1);
ir_node *op = get_irn_n(irn, i);
if(has_reg_class(env, op)) {
- int nr = get_irn_graph_nr(op);
-
- DBG((dbg, LEVEL_4, "\t\tpos: %d, use: %+F\n", i, op));
+ int nr = get_irn_idx(op);
+ const char *msg = "-";
if(!bitset_is_set(live, nr)) {
border_use(op, step, 1);
bitset_set(live, nr);
+ msg = "X";
}
+
+ DBG((dbg, LEVEL_4, "\t\t%s pos: %d, use: %+F\n", msg, i, op));
}
}
}
++step;
}
- /*
- * Add initial defs for all values live in.
- */
- for(irn = pset_first(live_in); irn; irn = pset_next(live_in)) {
- if(has_reg_class(env, irn)) {
-
- /* Mark the value live in. */
- bitset_set(live, get_irn_graph_nr(irn));
-
- /* Add the def */
+ bitset_foreach(live, elm) {
+ ir_node *irn = get_idx_irn(env->irg, elm);
+ if (be_is_live_in(lv, block, irn))
border_def(irn, step, 0);
- }
}
-
-
- del_pset(live_in);
- del_pset(live_end);
}
static void assign(ir_node *block, void *env_ptr)
{
be_chordal_alloc_env_t *alloc_env = env_ptr;
be_chordal_env_t *env = alloc_env->chordal_env;
- firm_dbg_module_t *dbg = env->dbg;
bitset_t *live = alloc_env->live;
bitset_t *colors = alloc_env->colors;
bitset_t *in_colors = alloc_env->in_colors;
const arch_env_t *arch_env = env->birg->main_env->arch_env;
+ struct list_head *head = get_block_border_head(env, block);
+ be_lv_t *lv = env->birg->lv;
const ir_node *irn;
border_t *b;
- struct list_head *head = get_block_border_head(env, block);
- pset *live_in = put_live_in(block, pset_new_ptr_default());
+ int idx;
bitset_clear_all(colors);
bitset_clear_all(live);
DBG((dbg, LEVEL_4, "\tusedef chain for block\n"));
list_for_each_entry(border_t, b, head, list) {
DBG((dbg, LEVEL_4, "\t%s %+F/%d\n", b->is_def ? "def" : "use",
- b->irn, get_irn_graph_nr(b->irn)));
+ b->irn, get_irn_idx(b->irn)));
}
/*
* Since their colors have already been assigned (The dominators were
* allocated before), we have to mark their colors as used also.
*/
- for(irn = pset_first(live_in); irn; irn = pset_next(live_in)) {
+ be_lv_foreach(lv, block, be_lv_state_in, idx) {
+ irn = be_lv_get_irn(lv, block, idx);
if(has_reg_class(env, irn)) {
const arch_register_t *reg = arch_get_irn_register(arch_env, irn);
int col;
assert(reg && "Node must have been assigned a register");
col = arch_register_get_index(reg);
+ DBG((dbg, LEVEL_4, "%+F has reg %s\n", irn, reg->name));
+
/* Mark the color of the live in value as used. */
bitset_set(colors, col);
bitset_set(in_colors, col);
/* Mark the value live in. */
- bitset_set(live, get_irn_graph_nr(irn));
+ bitset_set(live, get_irn_idx(irn));
}
}
/*
- * Mind that the sequence
- * of defs from back to front defines a perfect
+ * Mind that the sequence of defs from back to front defines a perfect
* elimination order. So, coloring the definitions from first to last
* will work.
*/
list_for_each_entry_reverse(border_t, b, head, list) {
ir_node *irn = b->irn;
- int nr = get_irn_graph_nr(irn);
+ int nr = get_irn_idx(irn);
+ int ignore = arch_irn_is(arch_env, irn, ignore);
/*
* Assign a color, if it is a local def. Global defs already have a
* color.
*/
- if(b->is_def && !is_live_in(block, irn)) {
+ if(b->is_def && !be_is_live_in(lv, block, irn)) {
const arch_register_t *reg;
int col = NO_COLOR;
- if(pset_find_ptr(alloc_env->pre_colored, irn)) {
+ if(ignore || pset_find_ptr(alloc_env->pre_colored, irn)) {
reg = arch_get_irn_register(arch_env, irn);
col = reg->index;
assert(!bitset_is_set(colors, col) && "pre-colored register must be free");
- }
-
- else {
+ } else {
col = get_next_free_reg(alloc_env, colors);
reg = arch_register_for_index(env->cls, col);
assert(arch_get_irn_register(arch_env, irn) == NULL && "This node must not have been assigned a register yet");
bitset_set(colors, col);
arch_set_irn_register(arch_env, irn, reg);
- DBG((dbg, LEVEL_1, "\tassigning register %s(%d) to %+F\n",
- arch_register_get_name(reg), col, irn));
+ DBG((dbg, LEVEL_1, "\tassigning register %s(%d) to %+F\n", arch_register_get_name(reg), col, irn));
assert(!bitset_is_set(live, nr) && "Value's definition must not have been encountered");
bitset_set(live, nr);
assert(reg && "Register must have been assigned");
col = arch_register_get_index(reg);
- assert(bitset_is_set(live, nr) && "Cannot have a non live use");
+#ifndef NDEBUG
+ if(!arch_register_type_is(reg, ignore)) {
+ assert(bitset_is_set(live, nr) && "Cannot have a non live use");
+ }
+#endif
bitset_clear(colors, col);
bitset_clear(live, nr);
}
}
-
- del_pset(live_in);
}
void be_ra_chordal_color(be_chordal_env_t *chordal_env)
{
be_chordal_alloc_env_t env;
char buf[256];
- int i;
+ be_lv_t *lv;
+ be_irg_t *birg = chordal_env->birg;
+ const arch_register_class_t *cls = chordal_env->cls;
- int colors_n = arch_register_class_n_regs(chordal_env->cls);
+ int colors_n = arch_register_class_n_regs(cls);
ir_graph *irg = chordal_env->irg;
+ int allocatable_regs = colors_n - be_put_ignore_regs(birg, cls, NULL);
+
+ /* some special classes contain only ignore regs, no work to be done */
+ if(allocatable_regs == 0)
+ return;
+ be_assure_dom_front(birg);
+ lv = be_assure_liveness(birg);
+ be_liveness_assure_sets(lv);
+ be_liveness_assure_chk(lv);
- if(get_irg_dom_state(irg) != dom_consistent)
- compute_doms(irg);
+ assure_doms(irg);
env.chordal_env = chordal_env;
env.colors_n = colors_n;
env.colors = bitset_alloca(colors_n);
env.tmp_colors = bitset_alloca(colors_n);
env.in_colors = bitset_alloca(colors_n);
- env.ignore_regs = bitset_alloca(colors_n);
env.pre_colored = pset_new_ptr_default();
- env.constr_dbg = firm_dbg_register("firm.be.chordal.constr");
-
- for(i = 0; i < colors_n; ++i)
- if(arch_register_type_is(&chordal_env->cls->regs[i], ignore))
- bitset_set(env.ignore_regs, i);
/* Handle register targeting constraints */
dom_tree_walk_irg(irg, constraints, NULL, &env);
be_dump(chordal_env->irg, buf, dump_ir_block_graph_sched);
}
- be_numbering(irg);
- env.live = bitset_malloc(get_graph_node_count(chordal_env->irg));
+ env.live = bitset_malloc(get_irg_last_idx(chordal_env->irg));
/* First, determine the pressure */
dom_tree_walk_irg(irg, pressure, NULL, &env);
/* Assign the colors */
dom_tree_walk_irg(irg, assign, NULL, &env);
- be_numbering_done(irg);
-
if(chordal_env->opts->dump_flags & BE_CH_DUMP_TREE_INTV) {
- plotter_t *plotter;
+ plotter_t *plotter;
ir_snprintf(buf, sizeof(buf), "ifg_%s_%F.eps", chordal_env->cls->name, irg);
- plotter = new_plotter_ps(buf);
- draw_interval_tree(&draw_chordal_def_opts, chordal_env, plotter);
- plotter_free(plotter);
+ plotter = new_plotter_ps(buf);
+ draw_interval_tree(&draw_chordal_def_opts, chordal_env, plotter);
+ plotter_free(plotter);
}
+ bitset_free(env.live);
del_pset(env.pre_colored);
}
+
+void be_init_chordal(void)
+{
+ FIRM_DBG_REGISTER(dbg, "firm.be.chordal");
+}
+
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_chordal);