#include "iredges_t.h"
#include "irgwalk.h"
#include "irprintf.h"
+#include "irgmod.h"
#include "error.h"
#include "beirg_t.h"
#include "belive_t.h"
#include "bearch_t.h"
+#include "benode_t.h"
#include "besched_t.h"
#include "bemodule.h"
static void process_block(ir_node *block, void *data)
{
- arch_isa_t *isa = arch_env->isa;
unsigned n_classes;
unsigned i;
int l;
(void) data;
/* construct initial register assignment */
- n_classes = arch_isa_get_n_reg_class(isa);
+ n_classes = arch_env_get_n_reg_class(arch_env);
for(i = 0; i < n_classes; ++i) {
- const arch_register_class_t *cls = arch_isa_get_reg_class(isa, 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);
}
}
}
+/**
+ * Walk through the block schedule and skip all barrier nodes.
+ */
+static void skip_barrier(ir_node *ret_blk, ir_graph *irg) {
+ ir_node *irn;
+
+ sched_foreach_reverse(ret_blk, irn) {
+ if (be_is_Barrier(irn)) {
+ const ir_edge_t *edge, *next;
+
+ 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);
+
+ edges_reroute_kind(proj, pred, EDGE_KIND_NORMAL, irg);
+ edges_reroute_kind(proj, pred, EDGE_KIND_DEP, irg);
+ }
+ sched_remove(irn);
+ kill_node(irn);
+ break;
+ }
+ }
+}
+
+/**
+ * Kill the Barrier nodes for better peephole optimization.
+ */
+static void kill_barriers(ir_graph *irg) {
+ ir_node *end_blk = get_irg_end_block(irg);
+ ir_node *start_blk;
+ int i;
+
+ /* skip the barrier on all return blocks */
+ for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
+ ir_node *be_ret = get_Block_cfgpred(end_blk, i);
+ ir_node *ret_blk = get_nodes_block(be_ret);
+
+ skip_barrier(ret_blk, irg);
+ }
+
+ /* skip the barrier on the start block */
+ start_blk = get_irg_start_block(irg);
+ skip_barrier(start_blk, irg);
+}
+
+/*
+ * Tries to optimize a beIncSp node with it's previous IncSP node.
+ * Must be run from a be_peephole_opt() context.
+ */
+void be_peephole_IncSP_IncSP(ir_node *node)
+{
+ int pred_offs;
+ int curr_offs;
+ int offs;
+ ir_node *pred = be_get_IncSP_pred(node);
+ ir_node *predpred;
+
+ if (!be_is_IncSP(pred))
+ return;
+
+ if (get_irn_n_edges(pred) > 1)
+ return;
+
+ 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;
+ }
+ offs = 0;
+ } else if (pred_offs == BE_STACK_FRAME_SIZE_SHRINK) {
+ if (curr_offs != BE_STACK_FRAME_SIZE_EXPAND) {
+ return;
+ }
+ offs = 0;
+ } else if (curr_offs == BE_STACK_FRAME_SIZE_EXPAND ||
+ curr_offs == BE_STACK_FRAME_SIZE_SHRINK) {
+ return;
+ } else {
+ offs = curr_offs + pred_offs;
+ }
+
+ /* add pred offset to ours and remove pred IncSP */
+ be_set_IncSP_offset(node, offs);
+
+ predpred = be_get_IncSP_pred(pred);
+ be_peephole_before_exchange(pred, predpred);
+
+ /* rewire dependency edges */
+ edges_reroute_kind(pred, predpred, EDGE_KIND_DEP, current_ir_graph);
+ be_set_IncSP_pred(node, predpred);
+ sched_remove(pred);
+ be_kill_node(pred);
+
+ be_peephole_after_exchange(predpred);
+}
+
void be_peephole_opt(be_irg_t *birg)
{
- arch_isa_t *isa;
ir_graph *irg = be_get_birg_irg(birg);
unsigned n_classes;
unsigned i;
+ /* barrier nodes are used for register allocations. They hinders
+ * peephole optimizations, so remove them here. */
+ kill_barriers(irg);
+
/* we sometimes find BadE nodes in float apps like optest_float.c or
* kahansum.c for example... */
be_liveness_invalidate(birg->lv);
arch_env = be_get_birg_arch_env(birg);
lv = be_get_birg_liveness(birg);
- isa = arch_env->isa;
- n_classes = arch_isa_get_n_reg_class(isa);
+ 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) {
- const arch_register_class_t *cls = arch_isa_get_reg_class(isa, 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);
}