/*
- * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
*
* This file is part of libFirm.
*
* @brief modifies schedule so flags dependencies are respected.
* @author Matthias Braun, Christoph Mallon
* @version $Id: besched.h 14693 2007-06-21 15:35:49Z beck $
+ *
+ * Fixup schedule to respect flag constraints by moving and rematerialisation of
+ * nodes.
+ *
+ * Flags are modeled as register classes with ignore registers. However to avoid
+ * bloating the graph, only flag-consumer -> producer dependencies are
+ * explicitely modeled in the graph. Nodes that just change the flags are only
+ * marked with the arch_irn_flags_modify_flags flag.
+ *
+ * Flags are usually a limited resource that can't (or at least shouldn't) be
+ * spilled. So in some situations (for example 2 adc-nodes that use the flags of
+ * a single add node on x86) operations have to be repeated to work correctly.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#include "bearch_t.h"
#include "beirg_t.h"
#include "besched_t.h"
+#include "benode_t.h"
static const arch_env_t *arch_env = NULL;
static const arch_register_class_t *flag_class = NULL;
static const arch_register_t *flags_reg = NULL;
static func_rematerialize remat = NULL;
+static int changed;
static ir_node *default_remat(ir_node *node, ir_node *after)
{
- ir_node *copy = exact_copy(node);
+ ir_node *block, *copy;
+ if(is_Block(after))
+ block = after;
+ else
+ block = get_nodes_block(after);
+
+ copy = exact_copy(node);
+ set_nodes_block(copy, block);
sched_add_after(after, copy);
return copy;
/* phi represents a usage at block end */
if(is_Phi(out2))
continue;
- if(sched_get_time_step(out2) <= sched_get_time_step(after)) {
+ if(is_Sync(out2)) {
+ const ir_edge_t *edge3;
+ foreach_out_edge(out2, edge3) {
+ ir_node *out3 = get_edge_src_irn(edge3);
+ /* phi represents a usage at block end */
+ if(is_Phi(out3))
+ continue;
+ assert(!is_Sync(out3));
+ if(sched_get_time_step(out3) <= sched_get_time_step(after)) {
+ return 0;
+ }
+ }
+ } else if(sched_get_time_step(out2) <= sched_get_time_step(after)) {
return 0;
}
}
}
}
- ir_fprintf(stderr, "Can move node %+F after node %+F\n", node, after);
return 1;
}
static void rematerialize_or_move(ir_node *flags_needed, ir_node *node,
- ir_node *flag_consumers, int pn)
+ ir_node *flag_consumers, int pn, be_lv_t *lv)
{
ir_node *n;
ir_node *copy;
ir_node *value;
- if(can_move(flags_needed, node)) {
+ if(!is_Block(node) &&
+ get_nodes_block(flags_needed) == get_nodes_block(node) &&
+ can_move(flags_needed, node)) {
/* move it */
sched_remove(flags_needed);
sched_add_after(node, flags_needed);
+ /* No need to update liveness, because the node stays in the same block */
return;
}
- copy = remat(flags_needed, node);
+ changed = 1;
+ copy = remat(flags_needed, node);
if(get_irn_mode(copy) == mode_T) {
ir_node *block = get_nodes_block(copy);
}
n = get_irn_link(n);
} while(n != NULL);
+
+ /* No need to introduce the copy, because it only lives in this block, but
+ * we have to update the liveness of all operands */
+ if (is_Block(node) ||
+ get_nodes_block(node) != get_nodes_block(flags_needed)) {
+ int i;
+
+ for (i = get_irn_arity(copy) - 1; i >= 0; --i) {
+ be_liveness_update(lv, get_irn_n(copy, i));
+ }
+ }
+}
+
+static int is_modify_flags(ir_node *node) {
+ int i, arity;
+
+ if(arch_irn_is(arch_env, node, modify_flags))
+ return 1;
+ if(!be_is_Keep(node))
+ return 0;
+
+ arity = get_irn_arity(node);
+ for(i = 0; i < arity; ++i) {
+ ir_node *in = get_irn_n(node, i);
+ in = skip_Proj(in);
+ if(arch_irn_is(arch_env, in, modify_flags))
+ return 1;
+ }
+
+ return 0;
}
+/**
+ * walks up the schedule and makes sure there are no flag-destroying nodes
+ * between a flag-consumer -> flag-producer chain. Fixes problematic situations
+ * by moving and/or rematerialisation of the flag-producers.
+ * (This can be extended in the future to do some register allocation on targets
+ * like ppc32 where we conceptually have 8 flag registers)
+ */
static void fix_flags_walker(ir_node *block, void *env)
{
ir_node *node;
ir_node *flags_needed = NULL;
ir_node *flag_consumers = NULL;
- int pn;
- (void) env;
+ int pn = -1;
sched_foreach_reverse(block, node) {
int i, arity;
ir_node *new_flags_needed = NULL;
+ if(is_Phi(node))
+ break;
+
if(node == flags_needed) {
/* all ok */
flags_needed = NULL;
flag_consumers = NULL;
}
- /* test wether node destroys the flags */
- if(flags_needed != NULL && arch_irn_is(arch_env, node, modify_flags)) {
+ /* test whether node destroys the flags */
+ if(flags_needed != NULL && is_modify_flags(node)) {
/* rematerialize */
- rematerialize_or_move(flags_needed, node, flag_consumers, pn);
+ rematerialize_or_move(flags_needed, node, flag_consumers, pn, env);
flags_needed = NULL;
flag_consumers = NULL;
}
- /* test wether the current node needs flags */
+ /* test whether the current node needs flags */
arity = get_irn_arity(node);
for(i = 0; i < arity; ++i) {
- //ir_node *in = get_irn_n(node, i);
const arch_register_class_t *cls
= arch_get_irn_reg_class(arch_env, node, i);
if(cls == flag_class) {
if(new_flags_needed == NULL)
continue;
- if(new_flags_needed != flags_needed) {
+ /* spiller can't (correctly) remat flag consumers at the moment */
+ assert(!arch_irn_is(arch_env, node, rematerializable));
+
+ if(skip_Proj(new_flags_needed) != flags_needed) {
if(flags_needed != NULL) {
/* rematerialize node */
- rematerialize_or_move(flags_needed, node, flag_consumers, pn);
- flags_needed = NULL;
+ rematerialize_or_move(flags_needed, node, flag_consumers, pn, env);
+ flags_needed = NULL;
flag_consumers = NULL;
}
- if(get_nodes_block(new_flags_needed) != block) {
- panic("remat across blocks not implemented yet");
- flags_needed = NULL;
- flag_consumers = NULL;
- } else {
- flags_needed = new_flags_needed;
- arch_set_irn_register(arch_env, flags_needed, flags_reg);
- if(is_Proj(flags_needed)) {
- pn = get_Proj_proj(flags_needed);
- flags_needed = get_Proj_pred(flags_needed);
- }
- flag_consumers = node;
- set_irn_link(flag_consumers, NULL);
- assert(arch_irn_is(arch_env, flags_needed, rematerializable));
+ flags_needed = new_flags_needed;
+ arch_set_irn_register(arch_env, flags_needed, flags_reg);
+ if(is_Proj(flags_needed)) {
+ pn = get_Proj_proj(flags_needed);
+ flags_needed = get_Proj_pred(flags_needed);
}
+ flag_consumers = node;
+ set_irn_link(flag_consumers, NULL);
+ assert(arch_irn_is(arch_env, flags_needed, rematerializable));
} else {
/* link all consumers in a list */
- set_irn_link(flag_consumers, node);
+ set_irn_link(node, flag_consumers);
flag_consumers = node;
}
}
+ if(flags_needed != NULL) {
+ assert(get_nodes_block(flags_needed) != block);
+ rematerialize_or_move(flags_needed, node, flag_consumers, pn, env);
+ flags_needed = NULL;
+ flag_consumers = NULL;
+ }
+
assert(flags_needed == NULL);
assert(flag_consumers == NULL);
}
flag_class = flag_cls;
flags_reg = & flag_class->regs[0];
remat = remat_func;
+ changed = 0;
if(remat == NULL)
remat = &default_remat;
- set_using_irn_link(irg);
- irg_block_walk_graph(irg, fix_flags_walker, NULL, NULL);
- clear_using_irn_link(irg);
+ ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
+ irg_block_walk_graph(irg, fix_flags_walker, NULL, birg->lv);
+ ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+
+ if(changed) {
+ be_remove_dead_nodes_from_schedule(birg);
+ }
}
projects / libfirm / blobdiff