#include "irgwalk.h"
#include "irnode_t.h"
#include "irprintf.h"
+#include "irdump.h"
#include "obst.h"
#include "raw_bitset.h"
#include "unionfind.h"
#include "bespillutil.h"
#include "beverify.h"
#include "beutil.h"
+#include "bestack.h"
#define USE_FACTOR 1.0f
#define DEF_FACTOR 1.0f
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
static struct obstack obst;
-static be_irg_t *birg;
static ir_graph *irg;
static const arch_register_class_t *cls;
-static const arch_register_req_t *default_cls_req;
static be_lv_t *lv;
static const ir_exec_freq *execfreqs;
static unsigned n_regs;
* the information is per firm-node.
*/
struct allocation_info_t {
- unsigned last_uses; /**< bitset indicating last uses (input pos) */
+ unsigned last_uses[2]; /**< bitset indicating last uses (input pos) */
ir_node *current_value; /**< copy of the value that should be used */
ir_node *original_value; /**< for copies point to original value */
float prefs[0]; /**< register preferences */
return info;
}
-/**
- * Get default register requirement for the current register class
- */
-static const arch_register_req_t *get_default_req_current_cls(void)
-{
- if (default_cls_req == NULL) {
- struct obstack *obst = get_irg_obstack(irg);
- arch_register_req_t *req = OALLOCZ(obst, arch_register_req_t);
-
- req->type = arch_register_req_type_normal;
- req->cls = cls;
-
- default_cls_req = req;
- }
- return default_cls_req;
-}
-
/**
* Link the allocation info of a node to a copy.
* Afterwards, both nodes uses the same allocation info.
return;
penalty *= NEIGHBOR_FACTOR;
- n_allowed = rbitset_popcnt(limited, n_regs);
+ n_allowed = rbitset_popcount(limited, n_regs);
if (n_allowed > 1) {
/* only create a very weak penalty if multiple regs are allowed */
penalty = (penalty * 0.8f) / n_allowed;
static void check_defs(const ir_nodeset_t *live_nodes, float weight,
ir_node *node)
{
- const arch_register_req_t *req;
-
- if (get_irn_mode(node) == mode_T) {
- const ir_edge_t *edge;
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- check_defs(live_nodes, weight, proj);
- }
- return;
- }
-
- if (!arch_irn_consider_in_reg_alloc(cls, node))
- return;
-
- req = arch_get_register_req_out(node);
+ const arch_register_req_t *req = arch_get_register_req_out(node);
if (req->type & arch_register_req_type_limited) {
const unsigned *limited = req->limited;
float penalty = weight * DEF_FACTOR;
int arity = get_irn_arity(insn);
int i;
- float factor = 1.0f / rbitset_popcnt(&req->other_same, arity);
+ float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
for (i = 0; i < arity; ++i) {
ir_node *op;
unsigned r;
if (is_Phi(node))
break;
- if (create_preferences)
- check_defs(&live_nodes, weight, node);
+ if (create_preferences) {
+ ir_node *value;
+ be_foreach_definition(node, cls, value,
+ check_defs(&live_nodes, weight, value);
+ );
+ }
/* mark last uses */
arity = get_irn_arity(node);
/* the allocation info node currently only uses 1 unsigned value
to mark last used inputs. So we will fail for a node with more than
32 inputs. */
- if (arity >= (int) sizeof(unsigned) * 8) {
+ if (arity >= (int) sizeof(info->last_uses) * 8) {
panic("Node with more than %d inputs not supported yet",
- (int) sizeof(unsigned) * 8);
+ (int) sizeof(info->last_uses) * 8);
}
info = get_allocation_info(node);
for (i = 0; i < arity; ++i) {
- ir_node *op = get_irn_n(node, i);
- if (!arch_irn_consider_in_reg_alloc(cls, op))
+ ir_node *op = get_irn_n(node, i);
+ const arch_register_req_t *req = arch_get_register_req_out(op);
+ if (req->cls != cls)
continue;
/* last usage of a value? */
if (!ir_nodeset_contains(&live_nodes, op)) {
- rbitset_set(&info->last_uses, i);
+ rbitset_set(info->last_uses, i);
}
}
ir_nodeset_destroy(&live_nodes);
}
-static void congruence_def(ir_nodeset_t *live_nodes, ir_node *node)
+static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
{
- const arch_register_req_t *req;
-
- if (get_irn_mode(node) == mode_T) {
- const ir_edge_t *edge;
- foreach_out_edge(node, edge) {
- ir_node *def = get_edge_src_irn(edge);
- congruence_def(live_nodes, def);
- }
- return;
- }
-
- if (!arch_irn_consider_in_reg_alloc(cls, node))
- return;
+ const arch_register_req_t *req = arch_get_register_req_out(node);
/* should be same constraint? */
- req = arch_get_register_req_out(node);
if (req->type & arch_register_req_type_should_be_same) {
- ir_node *insn = skip_Proj(node);
+ const ir_node *insn = skip_Proj_const(node);
int arity = get_irn_arity(insn);
int i;
unsigned node_idx = get_irn_idx(node);
/* check should be same constraints */
sched_foreach_reverse(block, node) {
+ ir_node *value;
if (is_Phi(node))
break;
- congruence_def(&live_nodes, node);
+ be_foreach_definition(node, cls, value,
+ congruence_def(&live_nodes, value);
+ );
be_liveness_transfer(cls, node, &live_nodes);
}
-
-
/**
* Assign register reg to the given node.
*
apref_delta = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
apref_delta += pref_delta - split_threshold;
- /* our source register isn't a usefull destination for recursive
+ /* our source register isn't a useful destination for recursive
splits */
old_source_state = rbitset_is_set(forbidden_regs, from_r);
rbitset_set(forbidden_regs, from_r);
unsigned r;
assert(!is_Phi(node));
- assert(arch_irn_consider_in_reg_alloc(cls, node));
-
/* preassigned register? */
reg = arch_get_irn_register(node);
if (reg != NULL) {
return;
}
- /* give should_be_same boni */
- info = get_allocation_info(node);
- req = arch_get_register_req_out(node);
+ req = arch_get_register_req_out(node);
+ /* ignore reqs must be preassigned */
+ assert (! (req->type & arch_register_req_type_ignore));
+ /* give should_be_same boni */
+ info = get_allocation_info(node);
in_node = skip_Proj(node);
if (req->type & arch_register_req_type_should_be_same) {
float weight = (float)get_block_execfreq(execfreqs, block);
break;
}
if (i >= n_regs) {
+ /* the common reason to hit this panic is when 1 of your nodes is not
+ * register pressure faithful */
panic("No register left for %+F\n", node);
}
DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
perm, in[0], in[1], before));
- proj0 = new_r_Proj(block, perm, get_irn_mode(in[0]), 0);
+ proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
mark_as_copy_of(proj0, in[0]);
reg = arch_register_for_index(cls, old_r);
use_reg(proj0, reg);
- proj1 = new_r_Proj(block, perm, get_irn_mode(in[1]), 1);
+ proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
mark_as_copy_of(proj1, in[1]);
reg = arch_register_for_index(cls, r2);
use_reg(proj1, reg);
*/
static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
{
- allocation_info_t *info = get_allocation_info(node);
- const unsigned *last_uses = &info->last_uses;
- int arity = get_irn_arity(node);
- int i;
+ allocation_info_t *info = get_allocation_info(node);
+ const unsigned *last_uses = info->last_uses;
+ int arity = get_irn_arity(node);
+ int i;
for (i = 0; i < arity; ++i) {
ir_node *op;
ir_node *op;
const arch_register_t *reg;
- if (!rbitset_is_set(&info->last_uses, i))
+ if (!rbitset_is_set(info->last_uses, i))
continue;
op = get_irn_n(node, i);
hungarian_problem_t *bp;
unsigned l, r;
unsigned *assignment;
+ ir_node *value;
/* construct a list of register occupied by live-through values */
unsigned *live_through_regs = NULL;
}
/* is any of the live-throughs using a constrained output register? */
- if (get_irn_mode(node) == mode_T) {
- const ir_edge_t *edge;
-
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- const arch_register_req_t *req;
-
- if (!arch_irn_consider_in_reg_alloc(cls, proj))
- continue;
-
- req = arch_get_register_req_out(proj);
- if (!(req->type & arch_register_req_type_limited))
- continue;
-
- if (live_through_regs == NULL) {
- rbitset_alloca(live_through_regs, n_regs);
- determine_live_through_regs(live_through_regs, node);
- }
-
- rbitset_or(forbidden_regs, req->limited, n_regs);
- if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
- good = false;
- }
- }
- } else {
- if (arch_irn_consider_in_reg_alloc(cls, node)) {
- const arch_register_req_t *req = arch_get_register_req_out(node);
- if (req->type & arch_register_req_type_limited) {
- rbitset_alloca(live_through_regs, n_regs);
- determine_live_through_regs(live_through_regs, node);
- if (rbitsets_have_common(req->limited, live_through_regs, n_regs)) {
- good = false;
- rbitset_or(forbidden_regs, req->limited, n_regs);
- }
- }
+ be_foreach_definition(node, cls, value,
+ if (! (req_->type & arch_register_req_type_limited))
+ continue;
+ if (live_through_regs == NULL) {
+ rbitset_alloca(live_through_regs, n_regs);
+ determine_live_through_regs(live_through_regs, node);
}
- }
+ rbitset_or(forbidden_regs, req_->limited, n_regs);
+ if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
+ good = false;
+ );
if (good)
return;
for (r = 0; r < n_regs; ++r) {
if (rbitset_is_set(limited, r))
continue;
- hungarian_remv(bp, r, current_reg);
+ hungarian_remove(bp, r, current_reg);
}
}
hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
assignment = ALLOCAN(unsigned, n_regs);
- res = hungarian_solve(bp, (int*) assignment, NULL, 0);
+ res = hungarian_solve(bp, assignment, NULL, 0);
assert(res == 0);
#if 0
unsigned *permutation;
ir_node **old_assignments;
bool need_permutation;
- ir_node *node;
+ ir_node *phi;
ir_node *pred = get_Block_cfgpred_block(block, p);
block_info_t *pred_info = get_block_info(pred);
/* check phi nodes */
need_permutation = false;
- node = sched_first(block);
- for ( ; is_Phi(node); node = sched_next(node)) {
+ phi = sched_first(block);
+ for ( ; is_Phi(phi); phi = sched_next(phi)) {
const arch_register_t *reg;
+ const arch_register_t *op_reg;
int regn;
int a;
ir_node *op;
- if (!arch_irn_consider_in_reg_alloc(cls, node))
+ if (!arch_irn_consider_in_reg_alloc(cls, phi))
continue;
- op = get_Phi_pred(node, p);
- if (!arch_irn_consider_in_reg_alloc(cls, op))
- continue;
-
- a = find_value_in_block_info(pred_info, op);
+ op = get_Phi_pred(phi, p);
+ a = find_value_in_block_info(pred_info, op);
assert(a >= 0);
- reg = arch_get_irn_register(node);
+ reg = arch_get_irn_register(phi);
regn = arch_register_get_index(reg);
- if (regn != a) {
- permutation[regn] = a;
- need_permutation = true;
- }
+ /* same register? nothing to do */
+ if (regn == a)
+ continue;
+
+ op = pred_info->assignments[a];
+ op_reg = arch_get_irn_register(op);
+ /* virtual or joker registers are ok too */
+ if ((op_reg->type & arch_register_type_joker)
+ || (op_reg->type & arch_register_type_virtual))
+ continue;
+
+ permutation[regn] = a;
+ need_permutation = true;
}
if (need_permutation) {
old_assignments = assignments;
assignments = pred_info->assignments;
permute_values(NULL, be_get_end_of_block_insertion_point(pred),
- permutation);
+ permutation);
assignments = old_assignments;
}
/* change phi nodes to use the copied values */
- node = sched_first(block);
- for ( ; is_Phi(node); node = sched_next(node)) {
+ phi = sched_first(block);
+ for ( ; is_Phi(phi); phi = sched_next(phi)) {
int a;
ir_node *op;
- if (!arch_irn_consider_in_reg_alloc(cls, node))
+ if (!arch_irn_consider_in_reg_alloc(cls, phi))
continue;
- op = get_Phi_pred(node, p);
- /* no need to do anything for Unknown inputs */
- if (!arch_irn_consider_in_reg_alloc(cls, op))
- continue;
+ op = get_Phi_pred(phi, p);
/* we have permuted all values into the correct registers so we can
simply query which value occupies the phis register in the
predecessor */
- a = arch_register_get_index(arch_get_irn_register(node));
+ a = arch_register_get_index(arch_get_irn_register(phi));
op = pred_info->assignments[a];
- set_Phi_pred(node, p, op);
+ set_Phi_pred(phi, p, op);
}
}
int n_phis = 0;
int n;
int res;
- int *assignment;
+ unsigned *assignment;
ir_node *node;
hungarian_problem_t *bp;
//hungarian_print_cost_matrix(bp, 7);
hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
- assignment = ALLOCAN(int, n_regs);
+ assignment = ALLOCAN(unsigned, n_regs);
res = hungarian_solve(bp, assignment, NULL, 0);
assert(res == 0);
if (!arch_irn_consider_in_reg_alloc(cls, node))
continue;
- r = assignment[n++];
+ r = assignment[n++];
assert(rbitset_is_set(normal_regs, r));
reg = arch_register_for_index(cls, r);
DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
*/
static void allocate_coalesce_block(ir_node *block, void *data)
{
- int i;
- ir_nodeset_t live_nodes;
- ir_node *node;
- int n_preds;
- block_info_t *block_info;
- block_info_t **pred_block_infos;
- ir_node **phi_ins;
- unsigned *forbidden_regs; /**< collects registers which must
+ int i;
+ ir_nodeset_t live_nodes;
+ ir_node *node;
+ int n_preds;
+ block_info_t *block_info;
+ block_info_t **pred_block_infos;
+ ir_node **phi_ins;
+ unsigned *forbidden_regs; /**< collects registers which must
not be used for optimistic splits */
(void) data;
ir_nodeset_init(&live_nodes);
/* gather regalloc infos of predecessor blocks */
- n_preds = get_Block_n_cfgpreds(block);
- pred_block_infos = ALLOCAN(block_info_t*, n_preds);
+ n_preds = get_Block_n_cfgpreds(block);
+ pred_block_infos = ALLOCAN(block_info_t*, n_preds);
for (i = 0; i < n_preds; ++i) {
ir_node *pred = get_Block_cfgpred_block(block, i);
block_info_t *pred_info = get_block_info(pred);
/* collect live-in nodes and preassigned values */
be_lv_foreach(lv, block, be_lv_state_in, i) {
- const arch_register_t *reg;
- int p;
- bool need_phi = false;
+ bool need_phi = false;
+ const arch_register_req_t *req;
+ const arch_register_t *reg;
+ int p;
node = be_lv_get_irn(lv, block, i);
- if (!arch_irn_consider_in_reg_alloc(cls, node))
+ req = arch_get_register_req_out(node);
+ if (req->cls != cls)
continue;
+ if (req->type & arch_register_req_type_ignore) {
+ allocation_info_t *info = get_allocation_info(node);
+ info->current_value = node;
+
+ reg = arch_get_irn_register(node);
+ assert(reg != NULL); /* ignore values must be preassigned */
+ use_reg(node, reg);
+ continue;
+ }
+
/* check all predecessors for this value, if it is not everywhere the
same or unknown then we have to construct a phi
(we collect the potential phi inputs here) */
}
if (need_phi) {
- ir_mode *mode = get_irn_mode(node);
- const arch_register_req_t *req = get_default_req_current_cls();
- ir_node *phi;
-
- phi = new_r_Phi(block, n_preds, phi_ins, mode);
- be_set_phi_reg_req(phi, req);
+ ir_mode *mode = get_irn_mode(node);
+ ir_node *phi = be_new_Phi(block, n_preds, phi_ins, mode, cls);
DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
#ifdef DEBUG_libfirm
sched_foreach(block, node) {
int i;
int arity;
+ ir_node *value;
/* phis are already assigned */
if (is_Phi(node))
free_last_uses(&live_nodes, node);
/* assign output registers */
- /* TODO: 2 phases: first: pre-assigned ones, 2nd real regs */
- if (get_irn_mode(node) == mode_T) {
- const ir_edge_t *edge;
- foreach_out_edge(node, edge) {
- ir_node *proj = get_edge_src_irn(edge);
- if (!arch_irn_consider_in_reg_alloc(cls, proj))
- continue;
- assign_reg(block, proj, forbidden_regs);
- }
- } else if (arch_irn_consider_in_reg_alloc(cls, node)) {
- assign_reg(block, node, forbidden_regs);
- }
+ be_foreach_definition_(node, cls, value,
+ assign_reg(block, value, forbidden_regs);
+ );
}
ir_nodeset_destroy(&live_nodes);
{
int i;
- lv = be_assure_liveness(birg);
+ lv = be_assure_liveness(irg);
be_liveness_assure_sets(lv);
ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
}
-static void dump(int mask, ir_graph *irg, const char *suffix,
- void (*dumper)(ir_graph *, const char *))
+static void dump(int mask, ir_graph *irg, const char *suffix)
{
- if (birg->main_env->options->dump_flags & mask)
- be_dump(irg, suffix, dumper);
+ if (be_get_irg_options(irg)->dump_flags & mask)
+ dump_ir_graph(irg, suffix);
}
/**
{
/* make sure all nodes show their real register pressure */
be_timer_push(T_RA_CONSTR);
- be_pre_spill_prepare_constr(birg, cls);
+ be_pre_spill_prepare_constr(irg, cls);
be_timer_pop(T_RA_CONSTR);
- dump(DUMP_RA, irg, "-spillprepare", dump_ir_block_graph_sched);
+ dump(DUMP_RA, irg, "-spillprepare");
/* spill */
be_timer_push(T_RA_SPILL);
- be_do_spill(birg, cls);
+ be_do_spill(irg, cls);
be_timer_pop(T_RA_SPILL);
be_timer_push(T_RA_SPILL_APPLY);
check_for_memory_operands(irg);
be_timer_pop(T_RA_SPILL_APPLY);
- dump(DUMP_RA, irg, "-spill", dump_ir_block_graph_sched);
+ dump(DUMP_RA, irg, "-spill");
}
/**
* The pref register allocator for a whole procedure.
*/
-static void be_pref_alloc(be_irg_t *new_birg)
+static void be_pref_alloc(ir_graph *new_irg)
{
- const arch_env_t *arch_env = new_birg->main_env->arch_env;
- int n_cls = arch_env_get_n_reg_class(arch_env);
+ const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
+ int n_cls = arch_env->n_register_classes;
int c;
obstack_init(&obst);
- birg = new_birg;
- irg = be_get_birg_irg(birg);
- execfreqs = birg->exec_freq;
+ irg = new_irg;
+ execfreqs = be_get_irg_exec_freq(irg);
/* determine a good coloring order */
determine_block_order();
for (c = 0; c < n_cls; ++c) {
- cls = arch_env_get_reg_class(arch_env, c);
- default_cls_req = NULL;
+ cls = &arch_env->register_classes[c];
if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
continue;
n_regs = arch_register_class_n_regs(cls);
normal_regs = rbitset_malloc(n_regs);
- be_abi_set_non_ignore_regs(birg->abi, cls, normal_regs);
+ be_set_allocatable_regs(irg, cls, normal_regs);
spill();
/* verify schedule and register pressure */
be_timer_push(T_VERIFY);
- if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
- be_verify_schedule(birg);
- be_verify_register_pressure(birg, cls, irg);
- } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
- assert(be_verify_schedule(birg) && "Schedule verification failed");
- assert(be_verify_register_pressure(birg, cls, irg)
+ if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
+ be_verify_schedule(irg);
+ be_verify_register_pressure(irg, cls);
+ } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
+ assert(be_verify_schedule(irg) && "Schedule verification failed");
+ assert(be_verify_register_pressure(irg, cls)
&& "Register pressure verification failed");
}
be_timer_pop(T_VERIFY);
}
be_timer_push(T_RA_SPILL_APPLY);
- be_abi_fix_stack_nodes(birg->abi);
+ be_abi_fix_stack_nodes(irg);
be_timer_pop(T_RA_SPILL_APPLY);
be_timer_push(T_VERIFY);
- if (birg->main_env->options->vrfy_option == BE_VRFY_WARN) {
- be_verify_register_allocation(birg);
- } else if (birg->main_env->options->vrfy_option == BE_VRFY_ASSERT) {
- assert(be_verify_register_allocation(birg)
- && "Register allocation invalid");
+ if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
+ be_verify_register_allocation(irg);
+ } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
+ assert(be_verify_register_allocation(irg)
+ && "Register allocation invalid");
}
be_timer_pop(T_VERIFY);
obstack_free(&obst, NULL);
}
-/**
- * Initializes this module.
- */
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc);
void be_init_pref_alloc(void)
{
static be_ra_t be_ra_pref = {
be_register_allocator("pref", &be_ra_pref);
FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");
}
-
-BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc);
projects / libfirm / blobdiff