* @author Matthias Braun
* @version $Id$
*/
-#ifdef HAVE_CONFIG_H
#include "config.h"
-#endif
#include "bepeephole.h"
#include "iredges_t.h"
#include "irgwalk.h"
#include "irprintf.h"
+#include "ircons.h"
#include "irgmod.h"
#include "error.h"
-#include "beirg_t.h"
+#include "beirg.h"
#include "belive_t.h"
-#include "bearch_t.h"
-#include "benode_t.h"
-#include "besched_t.h"
+#include "bearch.h"
+#include "benode.h"
+#include "besched.h"
#include "bemodule.h"
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
unsigned reg_idx;
unsigned cls_idx;
- if(!mode_is_data(get_irn_mode(node)))
+ if (!mode_is_data(get_irn_mode(node)))
return;
- reg = arch_get_irn_register(arch_env, node);
- if(reg == NULL) {
- panic("No register assigned at %+F\n", node);
+ reg = arch_get_irn_register(node);
+ if (reg == NULL) {
+ panic("No register assigned at %+F", node);
}
- if(arch_register_type_is(reg, virtual))
+ if (arch_register_type_is(reg, virtual))
return;
cls = arch_register_get_class(reg);
reg_idx = arch_register_get_index(reg);
cls_idx = arch_register_class_index(cls);
//assert(register_values[cls_idx][reg_idx] != NULL);
- DBG((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
+ DB((dbg, LEVEL_1, "Clear Register %s\n", reg->name));
register_values[cls_idx][reg_idx] = NULL;
}
unsigned reg_idx;
unsigned cls_idx;
- if(!mode_is_data(get_irn_mode(node)))
+ if (!mode_is_data(get_irn_mode(node)))
return;
- reg = arch_get_irn_register(arch_env, node);
- if(reg == NULL) {
- panic("No register assigned at %+F\n", node);
+ reg = arch_get_irn_register(node);
+ if (reg == NULL) {
+ panic("No register assigned at %+F", node);
}
- if(arch_register_type_is(reg, virtual))
+ if (arch_register_type_is(reg, virtual))
return;
cls = arch_register_get_class(reg);
reg_idx = arch_register_get_index(reg);
cls_idx = arch_register_class_index(cls);
- DBG((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
+ DB((dbg, LEVEL_1, "Set Register %s: %+F\n", reg->name, node));
register_values[cls_idx][reg_idx] = node;
}
static void clear_defs(ir_node *node)
{
/* clear values defined */
- if(get_irn_mode(node) == mode_T) {
+ 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);
/* set values used */
arity = get_irn_arity(node);
- for(i = 0; i < arity; ++i) {
+ for (i = 0; i < arity; ++i) {
ir_node *in = get_irn_n(node, i);
set_reg_value(in);
}
}
-void be_peephole_before_exchange(const ir_node *old_node, ir_node *new_node)
+void be_peephole_new_node(ir_node * nw)
+{
+ be_liveness_introduce(lv, nw);
+}
+
+/**
+ * must be called from peephole optimisations before a node will be killed
+ * and its users will be redirected to new_node.
+ * so bepeephole can update it's internal state.
+ *
+ * Note: killing a node and rewiring os only allowed if new_node produces
+ * the same registers as old_node.
+ */
+static void be_peephole_before_exchange(const ir_node *old_node,
+ ir_node *new_node)
{
const arch_register_t *reg;
const arch_register_class_t *cls;
unsigned reg_idx;
unsigned cls_idx;
- DBG((dbg, LEVEL_1, "About to exchange %+F with %+F\n", old_node, new_node));
+ DB((dbg, LEVEL_1, "About to exchange and kill %+F with %+F\n", old_node, new_node));
- if(old_node == current_node) {
- if(is_Proj(new_node)) {
- current_node = get_Proj_pred(new_node);
- } else {
- current_node = new_node;
- }
+ if (current_node == old_node) {
+ /* next node to be processed will be killed. Its scheduling predecessor
+ * must be processed next. */
+ current_node = sched_next(current_node);
+ assert (!is_Bad(current_node));
}
- if(!mode_is_data(get_irn_mode(old_node)))
+ if (!mode_is_data(get_irn_mode(old_node)))
return;
- reg = arch_get_irn_register(arch_env, old_node);
- if(reg == NULL) {
- panic("No register assigned at %+F\n", old_node);
+ reg = arch_get_irn_register(old_node);
+ if (reg == NULL) {
+ panic("No register assigned at %+F", old_node);
}
+ assert(reg == arch_get_irn_register(new_node) &&
+ "KILLING a node and replacing by different register is not allowed");
+
cls = arch_register_get_class(reg);
reg_idx = arch_register_get_index(reg);
cls_idx = arch_register_class_index(cls);
- if(register_values[cls_idx][reg_idx] == old_node) {
+ if (register_values[cls_idx][reg_idx] == old_node) {
register_values[cls_idx][reg_idx] = new_node;
}
be_liveness_remove(lv, old_node);
}
-void be_peephole_after_exchange(ir_node *new_node)
+void be_peephole_exchange(ir_node *old, ir_node *nw)
{
- be_liveness_introduce(lv, new_node);
+ be_peephole_before_exchange(old, nw);
+ sched_remove(old);
+ exchange(old, nw);
+ be_peephole_new_node(nw);
}
+/**
+ * block-walker: run peephole optimization on the given block.
+ */
static void process_block(ir_node *block, void *data)
{
unsigned n_classes;
/* construct initial register assignment */
n_classes = arch_env_get_n_reg_class(arch_env);
- for(i = 0; i < n_classes; ++i) {
+ for (i = 0; i < n_classes; ++i) {
const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
unsigned n_regs = arch_register_class_n_regs(cls);
memset(register_values[i], 0, sizeof(ir_node*) * n_regs);
}
assert(lv->nodes && "live sets must be computed");
- DBG((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
+ DB((dbg, LEVEL_1, "\nProcessing block %+F (from end)\n", block));
be_lv_foreach(lv, block, be_lv_state_end, l) {
ir_node *node = be_lv_get_irn(lv, block, l);
set_reg_value(node);
}
- DBG((dbg, LEVEL_1, "\nstart processing\n"));
+ DB((dbg, LEVEL_1, "\nstart processing\n"));
/* walk the block from last insn to the first */
current_node = sched_last(block);
- for( ; !sched_is_begin(current_node);
+ for ( ; !sched_is_begin(current_node);
current_node = sched_prev(current_node)) {
ir_op *op;
- ir_node *last;
- peephole_opt_func func;
+ peephole_opt_func peephole_node;
- if(is_Phi(current_node))
+ assert(!is_Bad(current_node));
+ if (is_Phi(current_node))
break;
clear_defs(current_node);
set_uses(current_node);
- op = get_irn_op(current_node);
- func = (peephole_opt_func) op->ops.generic;
- if(func == NULL)
+ op = get_irn_op(current_node);
+ peephole_node = (peephole_opt_func)op->ops.generic;
+ if (peephole_node == NULL)
continue;
- last = current_node;
- func(current_node);
- /* was the current node replaced? */
- if(current_node != last) {
- set_uses(current_node);
- }
+ DB((dbg, LEVEL_2, "optimize %+F\n", current_node));
+ peephole_node(current_node);
+ assert(!is_Bad(current_node));
}
}
+static void kill_node_and_preds(ir_node *node)
+{
+ int arity, i;
+
+ arity = get_irn_arity(node);
+ for (i = 0; i < arity; ++i) {
+ ir_node *pred = get_irn_n(node, i);
+
+ set_irn_n(node, i, new_Bad());
+ if (get_irn_n_edges(pred) != 0)
+ continue;
+
+ kill_node_and_preds(pred);
+ }
+
+ if (!is_Proj(node))
+ sched_remove(node);
+ kill_node(node);
+}
+
/**
* Walk through the block schedule and skip all barrier nodes.
*/
-static void skip_barrier(ir_node *ret_blk, ir_graph *irg) {
+static void skip_barrier(ir_node *block, ir_graph *irg)
+{
ir_node *irn;
- sched_foreach_reverse(ret_blk, irn) {
- if (be_is_Barrier(irn)) {
- const ir_edge_t *edge, *next;
+ sched_foreach_reverse(block, irn) {
+ int arity;
+ unsigned *used;
+ unsigned n_used;
+ const ir_edge_t *edge, *next;
+
+ if (!be_is_Barrier(irn))
+ continue;
+
+ /* track which outputs are actually used, as we have to create
+ * keep nodes for unused outputs */
+ arity = get_irn_arity(irn);
+ rbitset_alloca(used, arity);
+
+ foreach_out_edge_safe(irn, edge, next) {
+ ir_node *proj = get_edge_src_irn(edge);
+ int pn;
+ ir_node *pred;
+
+ if (is_Anchor(proj))
+ continue;
+
+ pn = (int) get_Proj_proj(proj);
+ pred = get_irn_n(irn, pn);
- foreach_out_edge_safe(irn, edge, next) {
- ir_node *proj = get_edge_src_irn(edge);
- int pn = (int)get_Proj_proj(proj);
- ir_node *pred = get_irn_n(irn, pn);
+ rbitset_set(used, pn);
- edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
- edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
+ edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
+ edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
+ }
+
+ /* the barrier also had the effect of a Keep for unused inputs.
+ * we now have to create an explicit Keep for them */
+ n_used = rbitset_popcount(used, arity);
+ if (n_used < (unsigned) arity) {
+ int n_in = arity - (int) n_used;
+ ir_node **in = ALLOCAN(ir_node*, n_in);
+ int i = 0;
+ int n = 0;
+ ir_node *keep;
+
+ for (i = 0; i < arity; ++i) {
+ if (rbitset_is_set(used, i))
+ continue;
+ assert(n < n_in);
+ in[n++] = get_irn_n(irn, i);
}
- sched_remove(irn);
- kill_node(irn);
- break;
+ keep = be_new_Barrier(get_nodes_block(irn), n_in, in);
+ keep_alive(keep);
+ sched_add_before(irn, keep);
}
+
+ kill_node_and_preds(irn);
+ break;
}
}
/**
* Kill the Barrier nodes for better peephole optimization.
*/
-static void kill_barriers(ir_graph *irg) {
+static void kill_barriers(ir_graph *irg)
+{
ir_node *end_blk = get_irg_end_block(irg);
- ir_node *start_blk;
+ ir_node *start_blk = get_irg_start_block(irg);
int i;
/* skip the barrier on all return blocks */
ir_node *be_ret = get_Block_cfgpred(end_blk, i);
ir_node *ret_blk = get_nodes_block(be_ret);
+ if (ret_blk == start_blk)
+ continue;
+
skip_barrier(ret_blk, irg);
}
skip_barrier(start_blk, irg);
}
+/**
+ * Check whether the node has only one user. Explicitly ignore the anchor.
+ */
+static int has_only_one_user(ir_node *node)
+{
+ int n = get_irn_n_edges(node);
+ const ir_edge_t *edge;
+
+ if (n <= 1)
+ return 1;
+
+ if (n > 2)
+ return 0;
+
+ foreach_out_edge(node, edge) {
+ ir_node *src = get_edge_src_irn(edge);
+ if (is_Anchor(src))
+ return 1;
+ }
+
+ return 0;
+}
+
/*
- * Tries to optimize a beIncSp node with it's previous IncSP node.
+ * Tries to optimize a beIncSP node with its previous IncSP node.
* Must be run from a be_peephole_opt() context.
*/
-void be_peephole_IncSP_IncSP(ir_node *node)
+ir_node *be_peephole_IncSP_IncSP(ir_node *node)
{
int pred_offs;
int curr_offs;
ir_node *pred = be_get_IncSP_pred(node);
if (!be_is_IncSP(pred))
- return;
+ return node;
- if (get_irn_n_edges(pred) > 1)
- return;
+ if (!has_only_one_user(pred))
+ return node;
pred_offs = be_get_IncSP_offset(pred);
curr_offs = be_get_IncSP_offset(node);
if (pred_offs == BE_STACK_FRAME_SIZE_EXPAND) {
if (curr_offs != BE_STACK_FRAME_SIZE_SHRINK) {
- return;
+ return node;
}
offs = 0;
} else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
- return;
+ return node;
}
offs = 0;
} else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND ||
curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
- return;
+ return node;
} else {
offs = curr_offs + pred_offs;
}
/* add node offset to pred and remove our IncSP */
be_set_IncSP_offset(pred, offs);
- be_peephole_before_exchange(node, pred);
-
- /* rewire dependency/data edges */
- edges_reroute_kind(node, pred, EDGE_KIND_DEP, current_ir_graph);
- edges_reroute(node, pred, current_ir_graph);
- sched_remove(node);
- be_kill_node(node);
-
- be_peephole_after_exchange(pred);
+ be_peephole_exchange(node, pred);
+ return pred;
}
-void be_peephole_opt(be_irg_t *birg)
+void be_peephole_opt(ir_graph *irg)
{
- ir_graph *irg = be_get_birg_irg(birg);
- unsigned n_classes;
- unsigned i;
+ unsigned n_classes;
+ unsigned i;
/* barrier nodes are used for register allocations. They hinders
* peephole optimizations, so remove them here. */
/* we sometimes find BadE nodes in float apps like optest_float.c or
* kahansum.c for example... */
- be_liveness_invalidate(birg->lv);
- be_liveness_assure_sets(be_assure_liveness(birg));
+ be_liveness_invalidate(be_get_irg_liveness(irg));
+ be_liveness_assure_sets(be_assure_liveness(irg));
- arch_env = be_get_birg_arch_env(birg);
- lv = be_get_birg_liveness(birg);
+ arch_env = be_get_irg_arch_env(irg);
+ lv = be_get_irg_liveness(irg);
n_classes = arch_env_get_n_reg_class(arch_env);
- register_values = alloca(sizeof(register_values[0]) * n_classes);
- for(i = 0; i < n_classes; ++i) {
+ register_values = XMALLOCN(ir_node**, n_classes);
+ for (i = 0; i < n_classes; ++i) {
const arch_register_class_t *cls = arch_env_get_reg_class(arch_env, i);
unsigned n_regs = arch_register_class_n_regs(cls);
- register_values[i] = alloca(sizeof(ir_node*) * n_regs);
+ register_values[i] = XMALLOCN(ir_node*, n_regs);
}
irg_block_walk_graph(irg, process_block, NULL, NULL);
-}
-void be_peephole_init(void)
-{
- clear_irp_opcodes_generic_func();
+ for (i = 0; i < n_classes; ++i) {
+ xfree(register_values[i]);
+ }
+ xfree(register_values);
}
+BE_REGISTER_MODULE_CONSTRUCTOR(be_init_peephole);
void be_init_peephole(void)
{
FIRM_DBG_REGISTER(dbg, "firm.be.peephole");
}
-
-BE_REGISTER_MODULE_CONSTRUCTOR(be_init_spillbelady);