+/**
+ * Check, if an irn is of the register class currently under processing.
+ * @param env The chordal environment.
+ * @param irn The node.
+ * @return 1, if the node is of that register class, 0 if not.
+ */
+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);
+}
+
+#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->ignore_regs);
+ 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)
+{
+ 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;
+ }
+ }
+
+ insn->next_insn = p;
+ }
+
+ 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;
+ }
+
+ insn->pre_colored = pre_colored == insn->n_ops && insn->n_ops > 0;
+ insn->use_start = insn->n_ops;
+
+ for(i = 0, n = get_irn_arity(irn); i < n; ++i) {
+ ir_node *op = get_irn_n(irn, i);
+
+ 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;
+ }
+ }
+
+ insn->has_constraints = insn->in_constraints | insn->out_constraints;
+ insn->ops = obstack_finish(obst);
+
+ /* Compute the admissible registers bitsets. */
+ for(i = 0; i < insn->n_ops; ++i) {
+ operand_t *op = &insn->ops[i];
+
+ assert(op->req.cls == env->cls);
+ op->regs = bitset_obstack_alloc(obst, env->cls->n_regs);
+
+ 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);
+ }
+
+ return insn;
+}
+
+static bitset_t *get_decisive_partner_regs(bitset_t *bs, const operand_t *o1, const 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);
+
+ 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 void pair_up_operands(const be_chordal_alloc_env_t *alloc_env, insn_t *insn)
+{
+ const be_chordal_env_t *env = alloc_env->chordal_env;
+
+ int n_uses = insn_n_uses(insn);
+ int n_defs = insn_n_defs(insn);
+ 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 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];
+
+ /* 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);
+ }
+ }
+ }
+
+ /* Compute the pairing. */
+ bipartite_matching(bp, pairing);
+ for(i = 0; i < insn->use_start; ++i) {
+ int p = pairing[i] + insn->use_start;
+
+ if(p >= insn->use_start) {
+ insn->ops[i].partner = &insn->ops[p];
+ insn->ops[p].partner = &insn->ops[i];
+ }
+ }
+
+ bipartite_free(bp);
+}
+
+
+static ir_node *pre_process_constraints(be_chordal_alloc_env_t *alloc_env, 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;
+ ir_node *perm = NULL;
+ bitset_t *out_constr = bitset_alloca(env->cls->n_regs);
+ bitset_t *bs = bitset_alloca(env->cls->n_regs);
+
+ 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) {
+ 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));
+
+ /* Registers are propagated by insert_Perm_after(). Clean them here! */
+ if(perm) {
+ const ir_edge_t *edge;
+
+ 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);
+
+ /*
+ 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) {
+ 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)
+{
+ 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);
+ ir_node *res = insn->next_insn;
+ int be_silent = *silent;
+
+ 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) {
+ 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);
+ 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;
+
+ /*
+ 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) {
+ 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)) {
+
+ pmap_insert(partners, op->carrier, op->partner ? op->partner->carrier : NULL);
+ alloc_nodes[n_alloc] = op->carrier;
+
+ DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, op->partner ? op->partner->carrier : NULL));
+
+ 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)
+ bipartite_add(bp, n_alloc, col);
+
+ n_alloc++;
+ }
+ }
+
+ /*
+ 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);
+
+ assert(is_Proj(proj));
+
+ if(values_interfere(proj, irn) && !pmap_contains(partners, proj)) {
+ 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_foreach(bs, col)
+ bipartite_add(bp, n_alloc, col);
+
+ n_alloc++;
+ }
+ }
+ }
+
+ /* Compute a valid register allocation. */
+ bipartite_matching(bp, assignment);
+
+ /* 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]);
+
+ for(j = 0; j < 2; ++j) {
+ if(!nodes[j])
+ continue;
+
+ 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));
+ }
+ }
+
+
+ /* Allocate the non-constrained Projs of the Perm. */
+ if(perm) {
+
+ 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));
+ }
+ }
+ }
+
+ 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);
+ }
+}
+