X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbespill.c;h=aabdbfec2a009d245065cbd519af1a3f87948bf5;hb=d920844bd5cd3296b03bb281ce43782967e0ce65;hp=7640de586f385ea1a944def536720d8a1db9b5e6;hpb=44c68b33c3d7ca8ebf83e5b269fdf8fb21ee0c63;p=libfirm

diff --git a/ir/be/bespill.c b/ir/be/bespill.c
index 7640de586..aabdbfec2 100644
--- a/ir/be/bespill.c
+++ b/ir/be/bespill.c
@@ -144,6 +144,26 @@ static spill_ctx_t *be_get_spill_ctx(set *sc, ir_node *to_spill, ir_node *ctx_ir
 	return set_insert(sc, &templ, sizeof(templ), HASH_COMBINE(HASH_PTR(to_spill), HASH_PTR(ctx_irn)));
 }
 
+/**
+ * Schedules a node after an instruction. (That is the place after all projs and phis
+ * that are scheduled after the instruction)
+ */
+static void sched_add_after_insn(ir_node *sched_after, ir_node *node) {
+	ir_node *next = sched_next(sched_after);
+	while(!sched_is_end(next)) {
+		if(!is_Proj(next) && !is_Phi(next))
+			break;
+		next = sched_next(next);
+	}
+
+	if(sched_is_end(next)) {
+		next = sched_last(get_nodes_block(sched_after));
+		sched_add_after(next, node);
+	} else {
+		sched_add_before(next, node);
+	}
+}
+
 /**
  * Creates a spill.
  *
@@ -171,6 +191,7 @@ static ir_node *be_spill_irn(spill_env_t *senv, ir_node *irn, ir_node *ctx_irn)
 	}
 
 	ctx->spill = be_spill(env->arch_env, irn, ctx_irn);
+	sched_add_after_insn(irn, ctx->spill);
 
 	return ctx->spill;
 }
@@ -200,46 +221,44 @@ static void remove_copies(spill_env_t *env) {
 	ARR_SETLEN(ir_node*, env->copies, 0);
 }
 
+static INLINE ir_node *skip_projs(ir_node *node) {
+	while(is_Proj(node)) {
+		node = sched_next(node);
+		assert(!sched_is_end(node));
+	}
+
+	return node;
+}
+
 /**
- * Inserts a copy (needed for spilled phi handling) of a value at the earliest
- * possible location in a block. That is after the last use/def of the value or at
- * the beginning of the block if there is no use/def.
+ * Searchs the schedule backwards until we reach the first use or def of a
+ * value or a phi.
+ * Returns the node before this node (so that you can do sched_add_before)
  */
-static ir_node *insert_copy(spill_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);
+static ir_node *find_last_use_def(spill_env_t *env, ir_node *block, ir_node *value) {
+	ir_node *node, *last;
 
-	ARR_APP1(ir_node*, env->copies, copy);
-
-	// walk schedule backwards until we find a use/def, or until we have reached the first phi
-	// TODO we could also do this by iterating over all uses and checking the
-	// sched_get_time_step value. Need benchmarks to decide this...
+	last = NULL;
 	sched_foreach_reverse(block, node) {
 		int i, arity;
 
 		if(is_Phi(node)) {
-			sched_add_after(node, copy);
-			goto placed;
+			return last;
 		}
 		if(value == node) {
-			sched_add_after(node, copy);
-			goto placed;
+			return skip_projs(last);
 		}
 		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;
+				return skip_projs(node);
 			}
 		}
+		last = node;
 	}
-	// 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;
+	// simply return first node if no def or use found
+	return sched_first(block);
 }
 
 /**
@@ -289,9 +308,12 @@ static ir_node *spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn, set
 			phi_spill_assoc_t *entry;
 
 			if(is_Phi(arg) && pset_find_ptr(senv->mem_phis, arg)) {
-				if (! bitset_is_set(bs, get_irn_idx(arg)))
+				// looping edge?
+				if(arg == phi) {
+					sub_res = phi_spill;
+				} else if (! bitset_is_set(bs, get_irn_idx(arg))) {
 					sub_res = spill_phi(senv, arg, ctx_irn, already_visited_phis, bs);
-				else {
+				} else {
 					/* we already visited the argument phi: get it's spill */
 					key.phi   = arg;
 					key.spill = NULL;
@@ -300,9 +322,9 @@ static ir_node *spill_phi(spill_env_t *senv, ir_node *phi, ir_node *ctx_irn, set
 					sub_res   = entry->spill;
 					assert(sub_res && "spill missing?!?");
 				}
-			}
-			else
+			} else {
 				sub_res = be_spill_irn(senv, arg, ctx_irn);
+			}
 
 			set_irn_n(phi_spill, i, sub_res);
 		}
@@ -455,9 +477,35 @@ static ir_node *do_remat(spill_env_t *senv, ir_node *spilled, ir_node *reloader)
 }
 
 void be_spill_phi(spill_env_t *env, ir_node *node) {
+	int i, arity;
+
 	assert(is_Phi(node));
 
 	pset_insert_ptr(env->mem_phis, node);
+
+	/* We have to place copy nodes in the predecessor blocks to temporarily
+	 * produce new values that get separate spill slots
+	 */
+	for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
+		ir_node *pred_block, *arg, *copy, *insert_point;
+
+		/* Don't do anything for looping edges (there's no need
+		 * and placing copies here breaks stuff as it suddenly
+		 * generates new living values through the whole loop)
+		 */
+		arg = get_irn_n(node, i);
+		if(arg == node)
+			continue;
+
+		pred_block = get_Block_cfgpred_block(get_nodes_block(node), i);
+		copy = be_new_Copy(env->cls, get_irn_irg(arg), pred_block, arg);
+
+		ARR_APP1(ir_node*, env->copies, copy);
+		insert_point = find_last_use_def(env, pred_block, arg);
+		sched_add_before(insert_point, copy);
+
+		set_irn_n(node, i, copy);
+	}
 }
 
 void be_insert_spills_reloads(spill_env_t *env) {
@@ -468,18 +516,8 @@ void be_insert_spills_reloads(spill_env_t *env) {
 	DBG((env->dbg, LEVEL_1, "Reloads for mem-phis:\n"));
 	foreach_pset(env->mem_phis, node) {
 		const ir_edge_t *e;
-		int i, arity;
-
-		/* We have to place copy nodes in the predecessor blocks to temporarily
-		* produce new values that get separate spill slots
-		*/
-		for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
-			ir_node *pred_block = get_Block_cfgpred_block(get_nodes_block(node), i);
-			ir_node *arg = get_irn_n(node, i);
-			ir_node *copy = insert_copy(env, pred_block, arg);
 
-			set_irn_n(node, i, copy);
-		}
+		assert(is_Phi(node));
 
 		/* Add reloads for mem_phis */
 		/* BETTER: These reloads (1) should only be inserted, if they are really needed */
@@ -499,7 +537,6 @@ void be_insert_spills_reloads(spill_env_t *env) {
 	for(si = set_first(env->spills); si; si = set_next(env->spills)) {
 		reloader_t *rld;
 		ir_mode *mode = get_irn_mode(si->spilled_node);
-		//ir_node *value;
 		pset *values = pset_new_ptr(16);
 
 		/* go through all reloads for this spill */
@@ -512,9 +549,7 @@ void be_insert_spills_reloads(spill_env_t *env) {
 #ifdef REMAT
 			if (check_remat_conditions(env, spill, si->spilled_node, rld->reloader)) {
 				new_val = do_remat(env, si->spilled_node, rld->reloader);
-				//pdeq_putl(possibly_dead, spill);
-			}
-			else
+			} else
 #endif
 				/* do a reload */
 				new_val = be_reload(arch_env, env->cls, rld->reloader, mode, spill);
@@ -542,12 +577,9 @@ void be_add_reload(spill_env_t *env, ir_node *to_spill, ir_node *before) {
 	spill_info_t templ, *res;
 	reloader_t *rel;
 
+	assert(sched_is_scheduled(before));
 	assert(arch_irn_consider_in_reg_alloc(env->chordal_env->birg->main_env->arch_env, env->cls, to_spill));
 
-	if(is_Phi(to_spill)) {
-		be_spill_phi(env, to_spill);
-	}
-
 	templ.spilled_node = to_spill;
 	templ.reloaders    = NULL;
 	res = set_insert(env->spills, &templ, sizeof(templ), HASH_PTR(to_spill));
@@ -558,12 +590,32 @@ void be_add_reload(spill_env_t *env, ir_node *to_spill, ir_node *before) {
 	res->reloaders = rel;
 }
 
-void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *bl, int pos) {
-	ir_node *insert_bl = get_irn_arity(bl) == 1 ? sched_first(bl) : get_Block_cfgpred_block(bl, pos);
-	be_add_reload(env, to_spill, insert_bl);
-}
+void be_add_reload_on_edge(spill_env_t *env, ir_node *to_spill, ir_node *block, int pos) {
+	ir_node *predblock, *last;
 
+	/* simply add the reload to the beginning of the block if we only have 1 predecessor
+	 * (we don't need to check for phis as there can't be any in a block with only 1 pred)
+	 */
+	if(get_Block_n_cfgpreds(block) == 1) {
+		assert(!is_Phi(sched_first(block)));
+		be_add_reload(env, to_spill, sched_first(block));
+		return;
+	}
+
+	/* We have to reload the value in pred-block */
+	predblock = get_Block_cfgpred_block(block, pos);
+	last = sched_last(predblock);
+
+	/* we might have projs and keepanys behind the jump... */
+	while(is_Proj(last) || be_is_Keep(last)) {
+		last = sched_prev(last);
+		assert(!sched_is_end(last));
+	}
+	assert(is_cfop(last));
 
+	// add the reload before the (cond-)jump
+	be_add_reload(env, to_spill, last);
+}
 
 /****************************************
 
@@ -741,21 +793,21 @@ interf_detected: /*nothing*/ ;
  * mode but different alignments.
  */
 static ir_type *get_spill_type(pmap *types, spill_slot_t *ss) {
-  pmap_entry *e = pmap_find(types, ss->largest_mode);
-  ir_type *res;
+	pmap_entry *e = pmap_find(types, ss->largest_mode);
+	ir_type *res;
 
-  if (! e) {
+	if (! e) {
 		char buf[64];
-    snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(ss->largest_mode));
-    res = new_type_primitive(new_id_from_str(buf), ss->largest_mode);
+		snprintf(buf, sizeof(buf), "spill_slot_type_%s", get_mode_name(ss->largest_mode));
+		res = new_type_primitive(new_id_from_str(buf), ss->largest_mode);
 		set_type_alignment_bytes(res, ss->align);
-    pmap_insert(types, ss->largest_mode, res);
-  }
-  else {
-    res = e->value;
+		pmap_insert(types, ss->largest_mode, res);
+	} else {
+		res = e->value;
 		assert(get_type_alignment_bytes(res) == (int)ss->align);
 	}
-  return res;
+
+	return res;
 }
 
 /**
@@ -767,12 +819,17 @@ static ir_type *get_spill_type(pmap *types, spill_slot_t *ss) {
  */
 static void assign_entities(ss_env_t *ssenv, int n_slots, spill_slot_t *ss[]) {
 	int i, offset, frame_align;
-	ir_type *frame = get_irg_frame_type(ssenv->cenv->irg);
+	ir_type *frame;
+
+	/* do not align the frame if no spill slots are needed */
+	if (n_slots <= 0)
+		return;
+
+	frame = get_irg_frame_type(ssenv->cenv->irg);
 
 	/* aligning by increasing frame size */
-	offset = get_type_size_bits(frame) / 8;
+	offset = get_type_size_bytes(frame);
 	offset = round_up2(offset, ALIGN_SPILL_AREA);
-	set_type_size_bytes(frame, -1);
 
 	/* create entities and assign offsets according to size and alignment*/
 	for (i = 0; i < n_slots; ++i) {