* @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 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;
}
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);
} while(n != NULL);
}
+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;
int i, arity;
ir_node *new_flags_needed = NULL;
+ if(is_Phi(node))
+ break;
+
if(node == flags_needed) {
/* all ok */
flags_needed = NULL;
}
/* test wether node destroys the flags */
- if(flags_needed != NULL && arch_irn_is(arch_env, node, modify_flags)) {
+ if(flags_needed != NULL && is_modify_flags(node)) {
/* rematerialize */
rematerialize_or_move(flags_needed, node, flag_consumers, pn);
flags_needed = NULL;
if(new_flags_needed == NULL)
continue;
+ /* spiller can't (correctly) remat flag consumers at the moment */
+ assert(!arch_irn_is(arch_env, node, rematerializable));
if(new_flags_needed != flags_needed) {
if(flags_needed != NULL) {
/* rematerialize node */
rematerialize_or_move(flags_needed, node, flag_consumers, pn);
- flags_needed = NULL;
+ 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);
+ flags_needed = NULL;
+ flag_consumers = NULL;
+ }
+
assert(flags_needed == NULL);
assert(flag_consumers == NULL);
}
projects / libfirm / blobdiff