+ return arch_irn_consider_in_reg_alloc(env->birg->main_env->arch_env, env->cls, irn);
+}
+
+#define has_limited_constr(req, irn) \
+ (arch_get_register_req(arch_env, (req), irn, -1) && (req)->type == arch_register_req_type_limited)
+
+static int get_next_free_reg(const be_chordal_alloc_env_t *alloc_env, bitset_t *colors)
+{
+ bitset_t *tmp = alloc_env->tmp_colors;
+ bitset_copy(tmp, colors);
+ bitset_or(tmp, alloc_env->chordal_env->ignore_colors);
+ return bitset_next_clear(tmp, 0);
+}
+
+static bitset_t *get_decisive_partner_regs(bitset_t *bs, const be_operand_t *o1, const be_operand_t *o2)
+{
+ bitset_t *res = bs;
+
+ if(!o1) {
+ bitset_copy(bs, o2->regs);
+ return bs;
+ }
+
+ if(!o2) {
+ bitset_copy(bs, o1->regs);
+ return bs;
+ }
+
+ 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;
+}
+
+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_get_irn_reg_class(aenv, irn, i) != env->cls)
+ continue;
+
+ reg = arch_get_irn_register(aenv, op);
+
+ if (reg == NULL || !arch_register_type_is(reg, ignore))
+ continue;
+ if(arch_register_type_is(reg, joker))
+ continue;
+
+ req = arch_get_register_req(aenv, irn, i);
+ if (!arch_register_req_is(req, limited))
+ continue;
+
+ if (rbitset_is_set(req->limited, reg->index))
+ continue;
+
+ 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));
+ }
+
+ insn = chordal_scan_insn(env, irn);
+
+ if(!insn->has_constraints)
+ goto end;
+
+ /* 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));
+ }
+ }
+
+ /* 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);
+ }
+
+ /*
+ 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];
+
+ bitset_copy(tmp, op->regs);
+ bitset_and(tmp, def_constr);
+
+ /*
+ 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 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;
+
+ 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) {
+ 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) {
+ 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;
+ }
+ }
+
+ 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;
+ }
+
+ out_op->partner = partner;
+ partner->partner = out_op;
+ }
+ }
+}
+
+
+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;
+ be_insn_t *insn = *the_insn;
+ ir_node *perm = NULL;
+ bitset_t *out_constr = bitset_alloca(env->cls->n_regs);
+ const ir_edge_t *edge;
+ int i;
+
+ assert(insn->has_constraints && "only do this for constrained nodes");
+
+ /*
+ Collect all registers that occur in output constraints.
+ This is necessary, since if the insn has one of these as an input constraint
+ and the corresponding operand interferes with the insn, the operand must
+ be copied.
+ */
+ for(i = 0; i < insn->use_start; ++i) {
+ be_operand_t *op = &insn->ops[i];
+ if(op->has_constraints)
+ bitset_or(out_constr, op->regs);
+ }
+
+ /*
+ 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(env->birg, env->cls, sched_prev(insn->irn));
+
+ /* Registers are propagated by insert_Perm_after(). Clean them here! */
+ if(perm == NULL)
+ return 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.
+ */
+ 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;
+ /*
+ 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);
+ }
+ }
+
+ return perm;
+}
+
+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);
+ 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;
+ for(i = 0; i < insn->use_start; ++i)
+ pset_insert_ptr(alloc_env->pre_colored, insn->ops[i].carrier);
+ }
+
+ /*
+ If the current node is a barrier toggle the silent flag.
+ If we are in the start block, we are ought to be silent at the beginning,
+ so the toggling activates the constraint handling but skips the barrier.
+ If we are in the end block we handle the in requirements of the barrier
+ and set the rest to silent.
+ */
+ if(be_is_Barrier(irn))
+ *silent = !*silent;
+
+ if(be_silent)
+ goto end;
+
+ /*
+ 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)
+ goto end;
+
+ 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();
+
+ /*
+ 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);
+
+ /* find suitable in operands to the out operands of the node. */
+ pair_up_operands(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) {
+ be_operand_t *op = &insn->ops[i];
+
+ /*
+ 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;
+
+ alloc_nodes[n_alloc] = op->carrier;
+
+ DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier,
+ partner));
+
+ bitset_clear_all(bs);
+ get_decisive_partner_regs(bs, op, op->partner);
+
+ DBG((dbg, LEVEL_2, "\tallowed registers for %+F: %B\n", op->carrier,
+ bs));
+
+ bitset_foreach(bs, col) {
+ //hungarian_add(bp, n_alloc, col, 1);
+ bipartite_add(bp, n_alloc, col);
+ }
+
+ n_alloc++;
+ }
+ }
+
+ /*
+ 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);
+
+ assert(is_Proj(proj));
+
+ if(!values_interfere(birg, proj, irn) || pmap_contains(partners, proj))
+ continue;
+
+ /* don't insert a node twice */
+ for(i = 0; i < n_alloc; ++i) {
+ if(alloc_nodes[i] == proj) {
+ break;
+ }
+ }
+ if(i < n_alloc)
+ continue;
+
+
+ assert(n_alloc < n_regs);
+
+ alloc_nodes[n_alloc] = proj;
+ pmap_insert(partners, proj, NULL);
+
+ 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
+
+ /* Assign colors obtained from the matching. */
+ for(i = 0; i < n_alloc; ++i) {
+ const arch_register_t *reg;
+ ir_node *irn;
+
+ 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));
+
+ 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));
+ }
+
+ 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));
+ }
+ }
+
+ /* Allocate the non-constrained Projs of the Perm. */
+ if(perm != NULL) {
+ bitset_clear_all(bs);
+
+ /* 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)
+ bitset_set(bs, reg->index);
+ }
+
+ /* 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);
+ return res;
+}
+
+/**
+ * Handle constraint nodes in each basic block.
+ * handle_constraints() inserts Perm nodes which perm
+ * over all values live at the constrained node right in front
+ * of the constrained node. These Perms signal a constrained node.
+ * For further comments, refer to handle_constraints().
+ */
+static void constraints(ir_node *bl, void *data)
+{
+ be_chordal_alloc_env_t *env = data;
+
+ /*
+ Start silent in the start block.
+ The silence remains until the first barrier is seen.
+ Each other block is begun loud.
+ */
+ int silent = bl == get_irg_start_block(get_irn_irg(bl));
+ ir_node *irn;
+
+ /*
+ If the block is the start block search the barrier and
+ start handling constraints from there.
+ */
+
+ for(irn = sched_first(bl); !sched_is_end(irn);) {
+ irn = handle_constraints(env, irn, &silent);
+ }