#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;
- pset *pre_colored; /**< Set of precolored nodes. */
- bitset_t *live; /**< A liveness bitset. */
- bitset_t *colors; /**< The color mask. */
- bitset_t *in_colors; /**< Colors used by live in values. */
- int colors_n; /**< The number of colors. */
+ 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 *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;
#include "fourcc.h"
#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_or(colors, alloc_env->ignore_regs);
+ return bitset_next_clear(colors, 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 {
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_env_t *env, ir_node *irn, struct obstack *obst)
{
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)) {
- o.carrier = p;
- o.irn = irn;
- o.pos = -(get_Proj_proj(p) + 1);
- o.partner = NULL;
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->has_constraints |= arch_register_req_is(&o.req, limited);
+ insn->out_constraints |= o.has_constraints;
+ pre_colored += arch_get_irn_register(arch_env, p) != NULL;
}
}
}
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;
- arch_get_register_req(arch_env, &o.req, irn, -1);
+ o.has_constraints = arch_register_req_is(&o.req, limited);
obstack_grow(obst, &o, sizeof(o));
insn->n_ops++;
- insn->has_constraints |= arch_register_req_is(&o.req, limited);
+ insn->out_constraints |= o.has_constraints;
+ pre_colored += arch_get_irn_register(arch_env, irn) != NULL;
}
- insn->use_start = insn->n_ops;
+ 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;
- arch_get_register_req(arch_env, &o.req, irn, i);
+ o.has_constraints = arch_register_req_is(&o.req, limited);
obstack_grow(obst, &o, sizeof(o));
insn->n_ops++;
- insn->has_constraints |= arch_register_req_is(&o.req, limited);
+ 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 operand_t *find_unpaired_use(insn_t *insn, const operand_t *op, int can_be_constrained)
+static bitset_t *get_decisive_partner_regs(bitset_t *bs, const operand_t *o1, const operand_t *o2)
{
- int i;
- operand_t *res = NULL;
+ bitset_t *res = bs;
- for(i = insn->use_start; i < insn->n_ops; ++i) {
- operand_t *op = &insn->ops[i];
- int has_constraint = arch_register_req_is(&op->req, limited);
+ if(!o1) {
+ bitset_copy(bs, o2->regs);
+ return bs;
+ }
- if(!values_interfere(op->carrier, op->irn) && !op->partner && (!has_constraint || can_be_constrained)) {
- if(arch_register_req_is(&op->req, should_be_same) && op->req.other_same == op->carrier)
- return op;
- else
- res = op;
- }
+ 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(insn_t *insn)
+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;
+ 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 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)
{
- firm_dbg_module_t *dbg = firm_dbg_register("firm.be.chordal.constr");
+ 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];
- int has_constraint = arch_register_req_is(&op->req, limited);
- operand_t *partner = find_unpaired_use(insn, op, !has_constraint);
+ operand_t *op = &insn->ops[i];
+ if(op->has_constraints)
+ bitset_or(out_constr, op->regs);
+ }
- if(partner) {
- op->partner = partner;
- partner->partner = op;
+ /*
+ 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)) {
+ 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(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)
insn_t *insn = scan_insn(env, irn, &env->obst);
ir_node *res = insn->next_insn;
+ 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(be_is_Perm(irn) || be_is_RegParams(irn) || (be_is_Barrier(irn) && !insn->in_constraints))
+ 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 = firm_dbg_register("firm.be.chordal.constr");
+ 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);
int i, n_alloc;
long col;
const ir_edge_t *edge;
- ir_node *perm = insert_Perm_after(aenv, env->cls, env->dom_front, sched_prev(irn));
-
- arch_put_non_ignore_regs(aenv, env->cls, non_ignore);
-
- /* Registers are propagated by insert_Perm_after(). Clean them here! */
- if(perm) {
- foreach_out_edge(perm, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- arch_set_irn_register(aenv, proj, NULL);
- }
- }
-
-
- be_liveness(env->irg);
- insn = scan_insn(env, irn, &env->obst);
-
- DBG((dbg, LEVEL_1, "handling constraints for %+F\n", irn));
+ ir_node *perm = NULL;
/*
- * If there was no Perm made, nothing was alive in this register class.
- * This means, that the node has no operands, thus no input constraints.
- * so it had output constraints. The other results then can be assigned freely.
- */
+ 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);
- pair_up_operands(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(arch_register_req_is(&op->req, limited) && arch_irn_consider_in_reg_alloc(aenv, env->cls, op->carrier)) {
- const arch_register_t *reg = arch_get_irn_register(aenv, 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)) {
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, pmap_get(partners, op->carrier)));
-
- bitset_clear_all(bs);
- op->req.limited(op->req.limited_env, bs);
- assert((!reg || bitset_is_set(bs, reg->index)) && "color of pre-colored node is not in its allowed colors");
- bitset_and(bs, non_ignore);
+ DBG((dbg, LEVEL_2, "\tassociating %+F and %+F\n", op->carrier, op->partner ? op->partner->carrier : NULL));
- /* if the node is pre-colored, explicitly allow the color with which it is pre-colored. */
- if(reg) {
- bitset_set(bs, reg->index);
- }
+ 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));
}
}
+ /*
+ Put all nodes which live by the constrained instruction also to the
+ allocation bipartite graph. They are considered unconstrained.
+ */
if(perm) {
- /* Make the input constraints of the node to output constraints of the Perm's Projs */
- for(i = insn->use_start; i < insn->n_ops; ++i) {
- operand_t *op = &insn->ops[i];
-
- /*
- If the operand is an "in" operand, constrained and the carrier is a Proj to the Perm,
- then copy the in constraint to the Perm's out constraint
- */
- if(arch_register_req_is(&op->req, limited) && is_Proj(op->carrier) && perm == get_Proj_pred(op->carrier))
- be_set_constr_limited(perm, BE_OUT_POS(get_Proj_proj(op->carrier)), &op->req);
- }
-
foreach_out_edge(perm, edge) {
ir_node *proj = get_edge_src_irn(edge);
pmap_insert(partners, proj, NULL);
bitset_clear_all(bs);
- arch_get_allocatable_regs(aenv, proj, -1, bs);
- bitset_and(bs, non_ignore);
+ arch_put_non_ignore_regs(aenv, env->cls, bs);
bitset_foreach(bs, col)
bipartite_add(bp, n_alloc, col);
}
}
+ /* Compute a valid register allocation. */
bipartite_matching(bp, assignment);
/* Assign colors obtained from the matching. */
for(i = 0; i < n_alloc; ++i) {
- int j;
- ir_node *nodes[2];
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]);
/* Allocate the non-constrained Projs of the Perm. */
if(perm) {
+
bitset_clear_all(bs);
/* Put the colors of all Projs in a bitset. */
DBG((dbg, LEVEL_2, "\tchecking reg of %+F: %s\n", proj, reg ? reg->name : "<none>"));
if(reg == NULL) {
- col = bitset_next_clear(bs, 0);
+ 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);
pmap_destroy(partners);
}
+end:
obstack_free(&env->obst, base);
return res;
}
}
else {
- col = bitset_next_clear(colors, 0);
+ 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);
+
+ assert(!arch_register_type_is(reg, ignore) && "Must not assign ignore register");
arch_set_irn_register(arch_env, irn, reg);
DBG((dbg, LEVEL_1, "\tassigning register %s(%d) to %+F\n",
{
be_chordal_alloc_env_t env;
char buf[256];
+ int i;
int colors_n = arch_register_class_n_regs(chordal_env->cls);
ir_graph *irg = chordal_env->irg;
env.colors_n = colors_n;
env.colors = bitset_malloc(colors_n);
env.in_colors = bitset_malloc(colors_n);
+ env.ignore_regs = bitset_malloc(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);
if(chordal_env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
snprintf(buf, sizeof(buf), "-%s-constr", chordal_env->cls->name);
- dump_ir_block_graph_sched(chordal_env->irg, buf);
+ be_dump(chordal_env->irg, buf, dump_ir_block_graph_sched);
}
be_numbering(irg);
free(env.live);
free(env.colors);
free(env.in_colors);
+ free(env.ignore_regs);
del_pset(env.pre_colored);
}