#include "bechordal_t.h"
#include "bespill.h"
#include "belive_t.h"
+#include "beabi.h"
#include "irgwalk.h"
#include "besched.h"
#include "beutil.h"
#include "bespillbelady.h"
#include "beverify.h"
+#include "benodesets.h"
#define DBG_LIVE 1
#define DBG_LOOPANA 2
#define DBG_PRESSURE 4
+#define DBG_SPILLS 8
+#define DBG_CHOOSE 16
DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
typedef struct morgan_env {
}
static INLINE int loop_attr_hash(const loop_attr_t *a) {
+#ifdef DEBUG_libfirm
+ return a->loop->loop_nr;
+#else
return HASH_PTR(a->loop);
+#endif
}
static INLINE int block_attr_hash(const block_attr_t *b) {
- return HASH_PTR(b->block);
+ return nodeset_hash(b->block);
}
static INLINE int loop_edge_hash(const loop_edge_t *e) {
- return HASH_PTR(e->block) ^ (e->pos * 31);
+ return nodeset_hash(e->block) ^ (e->pos * 31);
}
static INLINE loop_attr_t *get_loop_attr(morgan_env_t *env, const ir_loop *loop) {
res = set_find(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
if(res == NULL) {
- b_attr.livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
+ b_attr.livethrough_unused = NULL;
res = set_insert(env->block_attr_set, &b_attr, sizeof(b_attr), hash);
}
* Determine edges going out of a loop (= edges that go to a block that is not inside
* the loop or one of its subloops)
*/
-static INLINE void construct_loop_edges(ir_node* block, void* e) {
- morgan_env_t *env = (morgan_env_t*) e;
+static INLINE void construct_loop_edges(ir_node* block, void* data) {
+ morgan_env_t *env = data;
int n_cfgpreds = get_Block_n_cfgpreds(block);
int i;
ir_loop* loop = get_irn_loop(block);
- loop_attr_t *loop_attr = get_loop_attr(env, loop);
DBG((dbg, DBG_LOOPANA, "Loop for %+F: %d (depth %d)\n", block, loop->loop_nr, loop->depth));
for(i = 0; i < n_cfgpreds; ++i) {
+ loop_edge_t edge;
+ int hash;
ir_node* cfgpred = get_Block_cfgpred(block, i);
ir_node* cfgpred_block = get_nodes_block(cfgpred);
ir_loop* cfgpred_loop = get_irn_loop(cfgpred_block);
- loop_attr_t *outedges = get_loop_attr(env, cfgpred_loop);
if(cfgpred_loop == loop)
continue;
- // is it an edge into the loop?
- if(get_loop_depth(loop) > get_loop_depth(cfgpred_loop)) {
- loop_edge_t edge;
- edge.block = block;
- edge.pos = i;
- DBG((dbg, DBG_LOOPANA, "Loop in edge from %+F (loop %d) to %+F (loop %d)\n", cfgpred_block, get_loop_loop_nr(cfgpred_loop), block, get_loop_loop_nr(loop)));
- set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
- } else {
- ir_loop *p_loop = cfgpred_loop;
- while(get_loop_depth(p_loop) > get_loop_depth(loop)) {
- p_loop = get_loop_outer_loop(p_loop);
- }
- if(p_loop != loop) {
- loop_edge_t edge;
- edge.block = block;
- edge.pos = i;
- DBG((dbg, DBG_LOOPANA, "Loop in edge from %+F (loop %d) to %+F (loop %d)\n", cfgpred_block, get_loop_loop_nr(cfgpred_loop), block, get_loop_loop_nr(loop)));
- set_insert(loop_attr->in_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
- }
- }
+ assert(get_loop_depth(cfgpred_loop) != get_loop_depth(loop));
+
+ edge.block = block;
+ edge.pos = i;
+ hash = loop_edge_hash(&edge);
- // an edge out of the loop?
- if(get_loop_depth(cfgpred_loop) >= get_loop_depth(loop)) {
- loop_edge_t edge;
- edge.block = block;
- edge.pos = i;
- DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F\n", cfgpred_block, cfgpred_loop->loop_nr, block));
- set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
+ // edge out of a loop?
+ if(get_loop_depth(cfgpred_loop) > get_loop_depth(loop)) {
+ ir_loop *l;
+
+ DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (loop %d)\n", block, get_loop_loop_nr(loop),
+ cfgpred_block, get_loop_loop_nr(cfgpred_loop)));
+
+ /* this might be a jump out of multiple loops, so add this to all
+ * needed outedge sets */
+ l = cfgpred_loop;
+ do {
+ loop_attr_t *l_attr = get_loop_attr(env, l);
+ set_insert(l_attr->out_edges, &edge, sizeof(edge), hash);
+
+ l = get_loop_outer_loop(l);
+ assert(l != NULL);
+ } while(l != loop);
} else {
- ir_loop *o_loop = loop;
+ ir_loop *l;
- // we might jump in the middle of another inner loop which is not inside
- // our loop (happens for irreducible graphs). This would be a
- // real out edge then.
- while(get_loop_depth(o_loop) > get_loop_depth(cfgpred_loop)) {
- o_loop = get_loop_outer_loop(o_loop);
- }
+ // edge into a loop
+ DBG((dbg, DBG_LOOPANA, "Loop in edge from %+F (loop %d) to %+F (loop %d)\n", block, get_loop_loop_nr(loop),
+ cfgpred_block, get_loop_loop_nr(cfgpred_loop)));
- if(cfgpred_loop != o_loop) {
- loop_edge_t edge;
- edge.block = block;
- edge.pos = i;
- DBG((dbg, DBG_LOOPANA, "Loop out edge from %+F (loop %d) to %+F (into jump)\n", cfgpred_block, cfgpred_loop->loop_nr, block));
- set_insert(outedges->out_edges, &edge, sizeof(edge), loop_edge_hash(&edge));
- }
+ l = loop;
+ do {
+ loop_attr_t *l_attr = get_loop_attr(env, l);
+ set_insert(l_attr->in_edges, &edge, sizeof(edge), hash);
+
+ l = get_loop_outer_loop(l);
+ } while(l != cfgpred_loop);
}
}
}
}
}
+#if 0
/**
* Debugging help, shows all nodes in a (node-)bitset
*/
bitset_foreach(bitset, i) {
ir_node* node = get_idx_irn(irg, i);
- DBG((dbg, DBG_LIVE, "\t%+F\n", node));
+ ir_fprintf(stderr, " %+F", node);
}
+ fprintf(stderr, "\n");
}
+#endif
-/**
- * Construct the livethrough unused set for a block
- */
-static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, const ir_node* block) {
- block_attr_t *block_attr = get_block_attr(env, block);
- ir_node *node;
+static INLINE void init_livethrough_unuseds(block_attr_t *attr, morgan_env_t *env) {
+ const ir_node *block;
int i;
- DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
+ if(attr->livethrough_unused != NULL)
+ return;
+
+ block = attr->block;
+
+ attr->livethrough_unused = bitset_obstack_alloc(&env->obst, get_irg_last_idx(env->irg));
+
// copy all live-outs into the livethrough_unused set
be_lv_foreach(env->cenv->lv, block, be_lv_state_in | be_lv_state_out, i) {
ir_node *irn = be_lv_get_irn(env->cenv->lv, block, i);
int node_idx;
- /*
- if(!live_is_in(li) || !live_is_out(li))
- continue;
- */
if(!consider_for_spilling(env->arch, env->cls, irn))
continue;
node_idx = get_irn_idx(irn);
- bitset_set(block_attr->livethrough_unused, node_idx);
+ bitset_set(attr->livethrough_unused, node_idx);
+ }
+}
+
+/**
+ * Construct the livethrough unused set for a block
+ */
+static void construct_block_livethrough_unused(ir_node *block, void *data) {
+ morgan_env_t* env = data;
+ block_attr_t *block_attr = get_block_attr(env, block);
+ ir_node *node;
+ int n_cfgpreds;
+ block_attr_t **pred_attrs = NULL;
+ int i;
+
+ init_livethrough_unuseds(block_attr, env);
+
+ DBG((dbg, DBG_LIVE, "Processing block %d\n", get_irn_node_nr(block)));
+
+ n_cfgpreds = get_Block_n_cfgpreds(block);
+ if(n_cfgpreds > 1) {
+ pred_attrs = alloca(sizeof(pred_attrs[0]) * n_cfgpreds);
+ for(i = 0; i < n_cfgpreds; ++i) {
+ ir_node *pred_block = get_Block_cfgpred_block(block, i);
+ pred_attrs[i] = get_block_attr(env, pred_block);
+ init_livethrough_unuseds(pred_attrs[i], env);
+ }
}
/*
sched_foreach(block, node) {
int i, arity;
- for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
- int idx = get_irn_idx(get_irn_n(node, i));
- bitset_clear(block_attr->livethrough_unused, idx);
+ // phis are really uses in the pred block
+ if(is_Phi(node)) {
+ int j;
+ for(j = 0; j < n_cfgpreds; ++j) {
+ ir_node *used_value = get_Phi_pred(node, j);
+ int idx = get_irn_idx(used_value);
+ block_attr_t *pred_attr = pred_attrs[j];
+
+ bitset_clear(pred_attr->livethrough_unused, idx);
+ }
+ } else {
+ // mark all used values as used
+ for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
+ int idx = get_irn_idx(get_irn_n(node, i));
+ bitset_clear(block_attr->livethrough_unused, idx);
+ }
}
}
-
- show_nodebitset(env->irg, block_attr->livethrough_unused);
- return block_attr->livethrough_unused;
}
/**
loop_element elem = get_loop_element(loop, i);
switch (*elem.kind) {
case k_ir_node: {
- bitset_t *livethrough_block_unused;
+ ir_node *block = elem.node;
+ block_attr_t *block_attr = get_block_attr(env, block);
+ bitset_t *livethrough_block_unused = block_attr->livethrough_unused;
+
assert(is_Block(elem.node));
- livethrough_block_unused = construct_block_livethrough_unused(env, elem.node);
+ assert(livethrough_block_unused != NULL);
+
if(i == 0) {
bitset_copy(loop_attr->livethrough_unused, livethrough_block_unused);
} else {
bitset_andnot(son_attr->livethrough_unused, loop_attr->livethrough_unused);
DBG((dbg, DBG_LIVE, "Livethroughs for loop %d:\n", loop->loop_nr));
- show_nodebitset(env->irg, son_attr->livethrough_unused);
} else if(*elem.kind == k_ir_node) {
block_attr_t *block_attr = get_block_attr(env, elem.node);
bitset_andnot(block_attr->livethrough_unused, loop_attr->livethrough_unused);
DBG((dbg, DBG_LIVE, "Livethroughs for block %+F\n", elem.node));
- show_nodebitset(env->irg, block_attr->livethrough_unused);
} else {
assert(0);
}
/*---------------------------------------------------------------------------*/
+typedef struct _spillcandidate_t {
+ ir_node *node;
+ int cost;
+} spillcandidate_t;
+
+static int compare_spillcandidates(const void *d1, const void *d2) {
+ const spillcandidate_t *cand1 = d1;
+ const spillcandidate_t *cand2 = d2;
+
+ return cand1->cost - cand2->cost;
+}
+
+static void spill_values(morgan_env_t *env, const loop_attr_t *loop_attr, int spills) {
+ const bitset_t *cand_bitset = loop_attr->livethrough_unused;
+ int candidatecount = bitset_popcnt(cand_bitset);
+ spillcandidate_t *candidates;
+ int i, c;
+ loop_edge_t *edge;
+
+ assert(spills <= candidatecount);
+
+ candidates = alloca(sizeof(candidates[0]) * candidatecount);
+
+ DBG((dbg, DBG_CHOOSE, "Candidates for loop %d\n", get_loop_loop_nr(loop_attr->loop)));
+ // build candidiatelist
+ c = 0;
+ bitset_foreach(cand_bitset, i) {
+ ir_node *node = get_idx_irn(env->irg, i);
+ candidates[c].node = node;
+ candidates[c].cost = 0;
+
+ for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
+ candidates[c].cost += be_get_reload_costs_on_edge(env->senv, node, edge->block, edge->pos);
+ }
+ DBG((dbg, DBG_CHOOSE, "%+F has costs %d\n", node, candidates[c].cost));
+
+ c++;
+ }
+ assert(c == candidatecount);
+
+ // sort list
+ qsort(candidates, candidatecount, sizeof(candidates[0]), compare_spillcandidates);
+
+ // spill values
+ for(i = 0; i < spills; ++i) {
+ ir_node *to_spill = candidates[i].node;
+ DBG((dbg, DBG_CHOOSE, "Spilling %+F ", to_spill));
+
+ for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
+ be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
+ }
+ }
+}
+
static int reduce_register_pressure_in_block(morgan_env_t *env, const ir_node* block, int loop_unused_spills_possible) {
- int pressure;
- ir_node *irn;
- int max_pressure = 0;
- int spills_needed;
+ ir_node *node;
+ int max_pressure;
int loop_unused_spills_needed;
- block_attr_t *block_attr = get_block_attr(env, block);
- int block_unused_spills_possible = bitset_popcnt(block_attr->livethrough_unused);
- int unused_spills_possible = loop_unused_spills_possible + block_unused_spills_possible;
pset *live_nodes = pset_new_ptr_default();
be_liveness_end_of_block(env->cenv->lv, env->arch, env->cls, block, live_nodes);
- pressure = pset_count(live_nodes);
+ max_pressure = pset_count(live_nodes);
DBG((dbg, DBG_LIVE, "Reduce pressure to %d In Block %+F:\n", env->registers_available, block));
/**
- * Walk over all irns in the schedule and check register pressure for each of them
+ * Determine register pressure in block
*/
- sched_foreach_reverse(block, irn) {
- // do we need more spills than possible with unused libethroughs?
- int spills_needed = pressure - env->registers_available - unused_spills_possible;
- if(spills_needed > 0) {
- DBG((dbg, DBG_PRESSURE, "\tWARNING %d more spills needed at %+F\n", spills_needed, irn));
- // TODO further spills needed
- //assert(0);
- }
- if(pressure > max_pressure) {
- max_pressure = pressure;
- }
+ sched_foreach_reverse(block, node) {
+ int pressure;
- /* Register pressure is only important until we reach the first phi (the rest of the block
- * will only be phis.)
- */
- if(is_Phi(irn))
+ if(is_Phi(node))
break;
- // update pressure
- be_liveness_transfer(env->arch, env->cls, irn, live_nodes);
+ be_liveness_transfer(env->arch, env->cls, node, live_nodes);
pressure = pset_count(live_nodes);
+ if(pressure > max_pressure)
+ max_pressure = pressure;
}
+ del_pset(live_nodes);
- DBG((dbg, DBG_PRESSURE, "\tMax Pressure in %+F: %d\n", block, max_pressure));
+ loop_unused_spills_needed = max_pressure - env->registers_available;
- /*
- * Calculate number of spills from loop_unused_spills_possible that we want to use,
- * and spill unused livethroughs from the block if we still don't have enough registers
- */
- spills_needed = max_pressure - env->registers_available;
- if(spills_needed < 0) {
+ if(loop_unused_spills_needed < 0) {
loop_unused_spills_needed = 0;
- } else if(spills_needed > loop_unused_spills_possible) {
- int i, spills;
- int block_unused_spills_needed;
-
+ } else if(loop_unused_spills_needed > loop_unused_spills_possible) {
loop_unused_spills_needed = loop_unused_spills_possible;
- block_unused_spills_needed = spills_needed - loop_unused_spills_possible;
- if(block_unused_spills_needed > block_unused_spills_possible) {
- block_unused_spills_needed = block_unused_spills_possible;
- }
-
- spills = 0;
- /*
- * Spill/Reload unused livethroughs from the block
- */
- bitset_foreach(block_attr->livethrough_unused, i) {
- ir_node *to_spill;
- const ir_edge_t *edge;
-
- if(spills >= block_unused_spills_needed)
- break;
-
- to_spill = get_idx_irn(env->irg, i);
- foreach_block_succ(block, edge) {
- DBG((dbg, DBG_PRESSURE, "Spilling node %+F around block %+F\n", to_spill, block));
- be_add_reload_on_edge(env->senv, to_spill, edge->src, edge->pos);
- }
- spills++;
- }
- } else {
- loop_unused_spills_needed = spills_needed;
}
- del_pset(live_nodes);
-
- DBG((dbg, DBG_PRESSURE, "Unused spills for Block %+F needed: %d\n", block, loop_unused_spills_needed));
+ DBG((dbg, DBG_PRESSURE, "Block %+F: max-pressure %d spills possible: %d spills used: %d\n",
+ block, max_pressure, loop_unused_spills_possible, loop_unused_spills_needed));
return loop_unused_spills_needed;
}
int needed;
assert(is_Block(elem.node));
needed = reduce_register_pressure_in_block(env, elem.node, spills_possible);
+ assert(needed >= 0);
assert(needed <= spills_possible);
if(needed > spills_needed)
spills_needed = needed;
}
case k_ir_loop: {
int needed = reduce_register_pressure_in_loop(env, elem.son, spills_possible);
+ assert(needed >= 0);
assert(needed <= spills_possible);
if(needed > spills_needed)
spills_needed = needed;
spills_to_place = spills_needed;
- bitset_foreach(loop_attr->livethrough_unused, i) {
- loop_edge_t *edge;
- ir_node *to_spill = get_idx_irn(env->irg, i);
-
- for(edge = set_first(loop_attr->out_edges); edge != NULL; edge = set_next(loop_attr->out_edges)) {
- DBG((dbg, DBG_PRESSURE, "Spilling node %+F around loop %d\n", to_spill, loop->loop_nr));
- be_add_reload_on_edge(env->senv, to_spill, edge->block, edge->pos);
- }
+ DBG((dbg, DBG_SPILLS, "%d values unused in loop %d, spilling %d\n",
+ spills_possible - outer_spills_possible, loop->loop_nr, spills_to_place));
- spills_to_place--;
- if(spills_to_place <= 0) {
- break;
- }
- }
+ spill_values(env, loop_attr, spills_to_place);
} else {
outer_spills_needed = spills_needed;
}
return outer_spills_needed;
}
-void be_spill_morgan(const be_chordal_env_t *chordal_env) {
+void be_spill_morgan(be_chordal_env_t *chordal_env) {
+ ir_graph *irg = chordal_env->irg;
morgan_env_t env;
FIRM_DBG_REGISTER(dbg, "ir.be.spillmorgan");
- //firm_dbg_set_mask(dbg, DBG_LOOPANA | DBG_PRESSURE);
+ //firm_dbg_set_mask(dbg, DBG_SPILLS | DBG_LOOPANA);
env.cenv = chordal_env;
env.arch = chordal_env->birg->main_env->arch_env;
obstack_init(&env.obst);
- env.registers_available = arch_count_non_ignore_regs(env.arch, env.cls);
+ env.registers_available = env.cls->n_regs - be_put_ignore_regs(chordal_env->birg, env.cls, NULL);
env.loop_attr_set = new_set(loop_attr_cmp, 5);
env.block_attr_set = new_set(block_attr_cmp, 20);
/*-- Part1: Analysis --*/
- be_liveness_recompute(chordal_env->lv);
+ //Matze: I hope liveness information is up to date at this point...
+ //be_liveness_recompute(chordal_env->lv);
/* construct control flow loop tree */
- construct_cf_backedges(chordal_env->irg);
+ if(! (get_irg_loopinfo_state(irg) & loopinfo_cf_consistent)) {
+ construct_cf_backedges(irg);
+ }
/* construct loop out edges and livethrough_unused sets for loops and blocks */
- irg_block_walk_graph(chordal_env->irg, NULL, construct_loop_edges, &env);
- construct_loop_livethrough_unused(&env, get_irg_loop(env.irg));
+ irg_block_walk_graph(irg, construct_block_livethrough_unused, construct_loop_edges, &env);
+ construct_loop_livethrough_unused(&env, get_irg_loop(irg));
/*-- Part2: Transformation --*/
/* spill unused livethrough values around loops and blocks where
* the pressure is too high
*/
- reduce_register_pressure_in_loop(&env, get_irg_loop(env.irg), 0);
+ reduce_register_pressure_in_loop(&env, get_irg_loop(irg), 0);
- /* Place copies for spilled phis */
- be_place_copies(env.senv);
/* Insert real spill/reload nodes and fix usages */
be_insert_spills_reloads(env.senv);
/* Verify the result */
if (chordal_env->opts->vrfy_option == BE_CH_VRFY_WARN) {
- be_verify_schedule(env.irg);
+ be_verify_schedule(irg);
} else if (chordal_env->opts->vrfy_option == BE_CH_VRFY_ASSERT) {
- assert(be_verify_schedule(env.irg));
+ assert(be_verify_schedule(irg));
}
if (chordal_env->opts->dump_flags & BE_CH_DUMP_SPILL)
- be_dump(env.irg, "-spillmorgan", dump_ir_block_graph_sched);
+ be_dump(irg, "-spillmorgan", dump_ir_block_graph_sched);
/* cleanup */
free_loop_edges(&env);
del_set(env.block_attr_set);
/* fix the remaining places with too high register pressure with beladies algorithm */
-
- /* we have to remove dead nodes from schedule to not confuse liveness calculation */
- be_remove_dead_nodes_from_schedule(env.irg);
- be_liveness_recompute(chordal_env->lv);
-
be_spill_belady_spill_env(chordal_env, env.senv);
be_delete_spill_env(env.senv);