X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;ds=inline;f=ir%2Fbe%2Fbespillbelady.c;h=cd9d4b0fd27b4a8d199412531adb750bd251d3d1;hb=e9137de01e1ed5e7a8baee8d23a59e82ad634af5;hp=ed734c0390c631a5a06fa0f8a9629db80d6af569;hpb=712cea35fe399ae20e3a33eb9bf81360b5000cf3;p=libfirm

diff --git a/ir/be/bespillbelady.c b/ir/be/bespillbelady.c
index ed734c039..cd9d4b0fd 100644
--- a/ir/be/bespillbelady.c
+++ b/ir/be/bespillbelady.c
@@ -46,11 +46,10 @@
 #define DBG_START  16
 #define DBG_SLOTS  32
 #define DBG_TRACE  64
+#define DBG_WORKSET 128
 #define DEBUG_LVL 0 //(DBG_START | DBG_DECIDE | DBG_WSETS | DBG_FIX | DBG_SPILL)
 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
 
-#define MIN(a,b) (((a)<(b))?(a):(b))
-
 typedef struct _workset_t workset_t;
 
 typedef struct _belady_env_t {
@@ -64,10 +63,9 @@ typedef struct _belady_env_t {
 	ir_node *instr;		/**< current instruction */
 	unsigned instr_nr;	/**< current instruction number (relative to block start) */
 	pset *used;			/**< holds the values used (so far) in the current BB */
-	pset *copies;		/**< holds all copies placed due to phi-spilling */
+	ir_node **copies;	/**< holds all copies placed due to phi-spilling */
 
-	spill_env_t *senv;	/* see bespill.h */
-	pset *reloads;		/**< all reload nodes placed */
+	spill_env_t *senv;	/**< see bespill.h */
 } belady_env_t;
 
 struct _workset_t {
@@ -134,7 +132,7 @@ static INLINE void workset_insert(workset_t *ws, ir_node *val) {
 	int i;
 	/* check for current regclass */
 	if (arch_get_irn_reg_class(ws->bel->arch, val, -1) != ws->bel->cls) {
-		DBG((dbg, DBG_DECIDE, "Dropped %+F\n", val));
+		DBG((dbg, DBG_WORKSET, "Dropped %+F\n", val));
 		return;
 	}
 
@@ -294,8 +292,9 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
 			to_insert[demand++] = val;
 			if (is_usage)
 				be_add_reload(bel->senv, val, bel->instr);
-		} else
+		} else {
 			DBG((dbg, DBG_DECIDE, "    skip %+F\n", val));
+		}
 	}
 	DBG((dbg, DBG_DECIDE, "    demand = %d\n", demand));
 
@@ -305,6 +304,8 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
 	len = workset_get_length(ws);
 	max_allowed = bel->n_regs - demand;
 
+	DBG((dbg, DBG_DECIDE, "    disposing %d values\n", ws->len - max_allowed));
+
 	/* Only make more free room if we do not have enough */
 	if (len > max_allowed) {
 		/* get current next-use distance */
@@ -325,8 +326,9 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
 				workset_remove(ws_start, irn);
 
 				DBG((dbg, DBG_DECIDE, "    dispose %+F dumb\n", irn));
-			} else
+			} else {
 				DBG((dbg, DBG_DECIDE, "    dispose %+F\n", irn));
+			}
 		}
 
 		/* kill the last 'demand' entries in the array */
@@ -341,6 +343,47 @@ static void displace(belady_env_t *bel, workset_t *new_vals, int is_usage) {
 
 static void belady(ir_node *blk, void *env);
 
+/**
+ * Inserts a spill of a value at the earliest possible location in a block.
+ * That is after the last use of the value or at the beginning of the block if
+ * there is no use
+ */
+static ir_node *insert_copy(belady_env_t *env, ir_node *block, ir_node *value) {
+	ir_node* node;
+	ir_graph *irg = get_irn_irg(block);
+	ir_node *copy = be_new_Copy(env->cls, irg, block, value);
+
+	ARR_APP1(ir_node*, env->copies, copy);
+
+	// walk schedule backwards until we find a usage, or until we have reached the first phi
+	// TODO can we do this faster somehow? This makes insert_copy O(n) in block_size...
+	sched_foreach_reverse(block, node) {
+		int i, arity;
+
+		if(is_Phi(node)) {
+			sched_add_after(node, copy);
+			goto placed;
+		}
+		if(value == node) {
+			sched_add_after(node, copy);
+			goto placed;
+		}
+		for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
+			ir_node *arg = get_irn_n(node, i);
+			if(arg == value) {
+				sched_add_after(node, copy);
+				goto placed;
+			}
+		}
+	}
+	// we didn't find a use or a phi yet, so place the copy at the beginning of the block
+	sched_add_before(sched_first(block), copy);
+
+placed:
+
+	return copy;
+}
+
 /**
  * Collects all values live-in at block @p blk and all phi results in this block.
  * Then it adds the best values (at most n_regs) to the blocks start_workset.
@@ -348,8 +391,8 @@ static void belady(ir_node *blk, void *env);
  *  their args to break interference and make it possible to spill them to the
  *  same spill slot.
  */
-static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
-	belady_env_t *bel = env;
+static block_info_t *compute_block_start_info(ir_node *blk, void *data) {
+	belady_env_t *env = data;
 	ir_node *irn, *first;
 	irn_live_t *li;
 	int i, count, ws_count;
@@ -363,7 +406,7 @@ static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
 		return res;
 
 	/* Create the block info for this block. */
-	res = new_block_info(&bel->ob);
+	res = new_block_info(&env->ob);
 	set_block_info(blk, res);
 
 
@@ -373,24 +416,26 @@ static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
 	DBG((dbg, DBG_START, "Living at start of %+F:\n", blk));
 	first = sched_first(blk);
 	count = 0;
-	sched_foreach(blk, irn)
-		if (is_Phi(irn) && arch_get_irn_reg_class(bel->arch, irn, -1) == bel->cls) {
+	sched_foreach(blk, irn) {
+		if (is_Phi(irn) && arch_get_irn_reg_class(env->arch, irn, -1) == env->cls) {
 			loc.irn = irn;
-			loc.time = get_distance(bel, first, 0, irn, 0);
+			loc.time = get_distance(env, first, 0, irn, 0);
 			obstack_grow(&ob, &loc, sizeof(loc));
 			DBG((dbg, DBG_START, "    %+F:\n", irn));
 			count++;
 		} else
 			break;
+	}
 
-	live_foreach(blk, li)
-		if (live_is_in(li) && arch_get_irn_reg_class(bel->arch, li->irn, -1) == bel->cls) {
+	live_foreach(blk, li) {
+		if (live_is_in(li) && arch_get_irn_reg_class(env->arch, li->irn, -1) == env->cls) {
 			loc.irn = (ir_node *)li->irn;
-			loc.time = get_distance(bel, first, 0, li->irn, 0);
+			loc.time = get_distance(env, first, 0, li->irn, 0);
 			obstack_grow(&ob, &loc, sizeof(loc));
-			DBG((dbg, DBG_START, "    %+F:\n", irn));
+			DBG((dbg, DBG_START, "    %+F:\n", li->irn));
 			count++;
 		}
+	}
 
 	starters = obstack_finish(&ob);
 	qsort(starters, count, sizeof(starters[0]), loc_compare);
@@ -403,25 +448,26 @@ static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
 
 		/* if pred block has not been processed yet, do it now */
 		if (! pred_info) {
-			belady(pred_blk, bel);
+			belady(pred_blk, env);
 			pred_info = get_block_info(pred_blk);
 		}
 
 		/* now we have an end_set of pred */
 		assert(pred_info->ws_end && "The recursive call (above) is supposed to compute an end_set");
-		res->ws_start = workset_clone(&bel->ob, pred_info->ws_end);
+		res->ws_start = workset_clone(&env->ob, pred_info->ws_end);
 
 	} else
 
 	/* Else we want the start_set to be the values used 'the closest' */
 	{
 		/* Copy the best ones from starters to start workset */
-		ws_count = MIN(count, bel->n_regs);
-		res->ws_start = new_workset(&bel->ob, bel);
+		ws_count = MIN(count, env->n_regs);
+		res->ws_start = new_workset(&env->ob, env);
 		workset_bulk_fill(res->ws_start, ws_count, starters);
 	}
 
 
+
 	/* The phis of this block which are not in the start set have to be spilled later.
 	 * Therefore we add temporary copies in the pred_blocks so the spills can spill
 	 * into the same spill slot.
@@ -436,13 +482,11 @@ static block_info_t *compute_block_start_info(ir_node *blk, void *env) {
 		DBG((dbg, DBG_START, "For %+F:\n", irn));
 
 		for (max=get_irn_arity(irn), o=0; o<max; ++o) {
-			ir_node *arg = get_irn_n(irn, o);
 			ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(irn), o);
-			ir_node *cpy = be_new_Copy(bel->cls, irg, pred_block, arg);
-			pset_insert_ptr(bel->copies, cpy);
-			DBG((dbg, DBG_START, "    place a %+F of %+F in %+F\n", cpy, arg, pred_block));
-			sched_add_before(pred_block, cpy);
-			set_irn_n(irn, o, cpy);
+			ir_node *arg = get_irn_n(irn, o);
+			ir_node* copy = insert_copy(env, pred_block, arg);
+
+			set_irn_n(irn, o, copy);
 		}
 	}
 
@@ -559,9 +603,10 @@ static void fix_block_borders(ir_node *blk, void *env) {
 				continue;
 
 			/* check if irnb is in a register at end of pred */
-			workset_foreach(wsp, irnp, iter2)
+			workset_foreach(wsp, irnp, iter2) {
 				if (irnb == irnp)
 					goto next_value;
+			}
 
 			/* irnb is in memory at the end of pred, so we have to reload it */
 			DBG((dbg, DBG_FIX, "    reload %+F\n", irnb));
@@ -573,69 +618,37 @@ next_value:
 	}
 }
 
-/**
- * Removes all used reloads from bel->reloads.
- * The remaining nodes in bel->reloads will be removed from the graph.
- */
-static void rescue_used_reloads(ir_node *irn, void *env) {
-	pset *rlds = (pset *)env;
-	if (pset_find_ptr(rlds, irn))
-		pset_remove_ptr(rlds, irn);
-}
-
 /**
  * Removes all copies introduced for phi-spills
  */
-static void remove_copies(belady_env_t *bel) {
-	ir_node *irn;
-
-	for (irn = pset_first(bel->copies); irn; irn = pset_next(bel->copies)) {
-		ir_node *src, *user;
-
-		assert(be_is_Copy(irn));
-		assert(get_irn_n_edges(irn) == 1 && "This is not a copy introduced in 'compute_block_start_info()'. Who created it?");
-
-		user = get_irn_edge(get_irn_irg(irn), irn, 0)->src;
-
-		src = be_get_Copy_op(irn);
-		set_irn_n(user, 0, src);
-	}
-}
-
-/**
- * Finds all unused reloads and remove them from the schedule
- * Also removes spills if they are not used anymore after removing reloads
- */
-static void remove_unused_reloads(ir_graph *irg, belady_env_t *bel) {
-	ir_node *irn;
+static void remove_copies(belady_env_t *env) {
+	int i;
 
-	irg_walk_graph(irg, rescue_used_reloads, NULL, bel->reloads);
-	for(irn = pset_first(bel->reloads); irn; irn = pset_next(bel->reloads)) {
-		ir_node *spill;
-		DBG((dbg, DBG_SPILL, "Removing %+F before %+F in %+F\n", irn, sched_next(irn), get_nodes_block(irn)));
+	for(i = 0; i < ARR_LEN(env->copies); ++i) {
+		ir_node *node = env->copies[i];
+		ir_node *src;
+		const ir_edge_t *edge, *ne;
 
-		if (be_is_Reload(irn))
-			spill = get_irn_n(irn, be_pos_Reload_mem);
+		assert(be_is_Copy(node));
 
-		/* remove reload */
-		set_irn_n(irn, 0, new_Bad());
-		sched_remove(irn);
+		src = be_get_Copy_op(node);
+		foreach_out_edge_safe(node, edge, ne) {
+			ir_node *user = get_edge_src_irn(edge);
+			int user_pos = get_edge_src_pos(edge);
 
-		if (be_is_Reload(irn)) {
-			/* if spill not used anymore, remove it too
-			 * test of regclass is necessary since spill may be a phi-M */
-			if (get_irn_n_edges(spill) == 0 && bel->cls == arch_get_irn_reg_class(bel->arch, spill, -1)) {
-				set_irn_n(spill, 0, new_Bad());
-				sched_remove(spill);
-			}
+			set_irn_n(user, user_pos, src);
 		}
 	}
 }
 
 void be_spill_belady(const be_chordal_env_t *chordal_env) {
+	be_spill_belady_spill_env(chordal_env, NULL);
+}
+
+void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t *spill_env) {
 	belady_env_t bel;
 
-	FIRM_DBG_REGISTER(dbg, "ir.be.spillbelady");
+	FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
 
 	/* init belady env */
 	obstack_init(&bel.ob);
@@ -644,10 +657,14 @@ void be_spill_belady(const be_chordal_env_t *chordal_env) {
 	bel.n_regs    = arch_register_class_n_regs(bel.cls);
 	bel.ws        = new_workset(&bel.ob, &bel);
 	bel.uses      = be_begin_uses(chordal_env->irg, chordal_env->birg->main_env->arch_env, bel.cls);
-	bel.senv      = be_new_spill_env(chordal_env, is_mem_phi, NULL);
+	if(spill_env == NULL) {
+		bel.senv = be_new_spill_env(chordal_env, is_mem_phi, NULL);
+	} else {
+		bel.senv = spill_env;
+		be_set_is_spilled_phi(bel.senv, is_mem_phi, NULL);
+	}
 	DEBUG_ONLY(be_set_spill_env_dbg_module(bel.senv, dbg);)
-	bel.reloads   = pset_new_ptr_default();
-	bel.copies    = pset_new_ptr_default();
+	bel.copies    = NEW_ARR_F(ir_node*, 0);
 
 	DBG((dbg, LEVEL_1, "running on register class: %s\n", bel.cls->name));
 
@@ -655,13 +672,15 @@ void be_spill_belady(const be_chordal_env_t *chordal_env) {
 	be_clear_links(chordal_env->irg);
 	irg_block_walk_graph(chordal_env->irg, NULL, belady, &bel);
 	irg_block_walk_graph(chordal_env->irg, fix_block_borders, NULL, &bel);
-	be_insert_spills_reloads(bel.senv, bel.reloads);
-	remove_unused_reloads(chordal_env->irg, &bel);
+	be_insert_spills_reloads(bel.senv);
 	remove_copies(&bel);
+	DEL_ARR_F(bel.copies);
+
+	be_remove_dead_nodes_from_schedule(chordal_env->irg);
 
 	/* clean up */
-	del_pset(bel.reloads);
-	be_delete_spill_env(bel.senv);
+	if(spill_env == NULL)
+		be_delete_spill_env(bel.senv);
 	be_end_uses(bel.uses);
 	obstack_free(&bel.ob, NULL);
 }