- Fix perform_memop stuff in bechordal_main after adams gammlig commit

- Cleanup remat code a little bit
- Adjust remat cost limit to new costs in ia32 backend
- Don't spill unused livethroughs around blocks anymore in bespill_morgan
  (but only around loops), this makes the code faster smaller and the results
  slightly better as belady performs better for basic blocks

ir/be/bearch.h

@@ -209,7 +209,7 @@ typedef struct _arch_inverse_t {
 typedef enum _arch_irn_flags_t {
 	arch_irn_flags_none             = 0, /**< Node flags. */
 	arch_irn_flags_dont_spill       = 1, /**< This must not be spilled. */
-	arch_irn_flags_rematerializable = 2, /**< This should be replicated instead of spilled/reloaded. */
+	arch_irn_flags_rematerializable = 2, /**< This can be replicated instead of spilled/reloaded. */
 	arch_irn_flags_ignore           = 4, /**< Ignore node during register allocation. */
 	arch_irn_flags_modify_sp        = 8, /**< I modify the stack pointer. */
 	arch_irn_flags_last             = arch_irn_flags_modify_sp

ir/be/bechordal_main.c

@@ -329,10 +329,16 @@ static void memory_operand_walker(ir_node *irn, void *env) {
 	const arch_env_t *aenv = cenv->birg->main_env->arch_env;
 	const ir_edge_t  *edge, *ne;
 	ir_node          *block;
+	ir_node          *spill;
 
 	if (! be_is_Reload(irn))
 		return;
 
+	// only use memory operands, if the reload is only used by 1 node
+	if(get_irn_n_edges(irn) > 1)
+		return;
+
+	spill = be_get_Reload_mem(irn);
 	block = get_nodes_block(irn);
 
 	foreach_out_edge_safe(irn, edge, ne) {
@@ -344,7 +350,7 @@ static void memory_operand_walker(ir_node *irn, void *env) {
 
 		if (get_nodes_block(src) == block && arch_possible_memory_operand(aenv, src, pos)) {
 			DBG((cenv->dbg, LEVEL_3, "performing memory operand %+F at %+F\n", irn, src));
-			arch_perform_memory_operand(aenv, src, irn, pos);
+			arch_perform_memory_operand(aenv, src, spill, pos);
 		}
 	}
 

ir/be/benode.c

@@ -364,6 +364,7 @@ ir_node *be_new_Reload(const arch_register_class_t *cls, const arch_register_cla
 	init_node_attr(res, 2);
 	be_node_set_reg_class(res, 0, cls_frame);
 	be_node_set_reg_class(res, -1, cls);
+	be_node_set_flags(res, -1, arch_irn_flags_rematerializable);
 	return res;
 }
 
@@ -944,22 +945,18 @@ ir_node *be_reload(const arch_env_t *arch_env, const arch_register_class_t *cls,
 {
 	ir_node *reload;
 
-	ir_node *bl    = is_Block(insert) ? insert : get_nodes_block(insert);
-	ir_graph *irg  = get_irn_irg(bl);
+	ir_node *block  = get_nodes_block(insert);
+	ir_graph *irg  = get_irn_irg(block);
 	ir_node *frame = get_irg_frame(irg);
 	const arch_register_class_t *cls_frame = arch_get_irn_reg_class(arch_env, frame, -1);
 
 	assert(be_is_Spill(spill) || (is_Phi(spill) && get_irn_mode(spill) == mode_M));
 
-	reload = be_new_Reload(cls, cls_frame, irg, bl, frame, spill, mode);
+	reload = be_new_Reload(cls, cls_frame, irg, block, frame, spill, mode);
 
-	if(is_Block(insert)) {
-		insert = sched_skip(insert, 0, sched_skip_cf_predicator, (void *) arch_env);
-		sched_add_after(insert, reload);
-	}
+	assert(!is_Block(insert));
 
-	else
-		sched_add_before(insert, reload);
+	sched_add_before(insert, reload);
 
 	return reload;
 }

ir/be/bespill.c

@@ -35,7 +35,7 @@
 
 // only rematerialise when costs are less than REMAT_COST_LIMIT
 // TODO determine a good value here...
-#define REMAT_COST_LIMIT	80
+#define REMAT_COST_LIMIT	20
 
 typedef struct _reloader_t reloader_t;
 
@@ -45,10 +45,16 @@ struct _reloader_t {
 };
 
 typedef struct _spill_info_t {
+	/** the value that should get spilled */
 	ir_node *spilled_node;
+	/** list of places where the value should get reloaded */
 	reloader_t *reloaders;
 
+	/** the spill node, or a PhiM node */
 	ir_node *spill;
+	/** if we had the value of a phi spilled before but not the phi itself then
+	 * this field contains the spill for the phi value */
+	ir_node *old_spill;
 } spill_info_t;
 
 struct _spill_env_t {
@@ -84,6 +90,7 @@ static spill_info_t *get_spillinfo(const spill_env_t *env, ir_node *value) {
 	if (res == NULL) {
 		info.reloaders = NULL;
 		info.spill = NULL;
+		info.old_spill = NULL;
 		res = set_insert(env->spills, &info, sizeof(info), hash);
 	}
 
@@ -185,11 +192,19 @@ void be_spill_phi(spill_env_t *env, ir_node *node) {
 	pset_insert_ptr(env->mem_phis, node);
 
 	// create spillinfos for the phi arguments
-	get_spillinfo(env, node);
+	spill_info_t* spill = get_spillinfo(env, node);
 	for(i = 0, arity = get_irn_arity(node); i < arity; ++i) {
 		ir_node *arg = get_irn_n(node, i);
 		get_spillinfo(env, arg);
 	}
+
+	// if we had a spill for the phi value before, then remove this spill from
+	// schedule, as we will remove it in the insert spill/reload phase
+	if(spill->spill != NULL && !is_Phi(spill->spill)) {
+		//sched_remove(spill->spill);
+		spill->old_spill = spill->spill;
+		spill->spill = NULL;
+	}
 }
 
 /*
@@ -292,6 +307,17 @@ static void spill_phi(spill_env_t *env, spill_info_t *spillinfo) {
 
 		set_irn_n(spillinfo->spill, i, arg_info->spill);
 	}
+
+	// rewire reloads from old_spill to phi
+	if(spillinfo->old_spill != NULL) {
+		const ir_edge_t *edge, *next;
+		foreach_out_edge_safe(spillinfo->old_spill, edge, next) {
+			ir_node* reload = get_edge_src_irn(edge);
+			assert(be_is_Reload(reload) || is_Phi(reload));
+			set_irn_n(reload, get_edge_src_pos(edge), spillinfo->spill);
+		}
+		spillinfo->old_spill = NULL;
+	}
 }
 
 /**
@@ -377,15 +403,9 @@ static int is_remat_node(spill_env_t *env, ir_node *node) {
 
 	assert(!be_is_Spill(node));
 
-	if(be_is_Reload(node))
-		return 1;
-
-	// TODO why does arch_irn_is say rematerializable anyway?
-	if(be_is_Barrier(node))
-		return 0;
-
-	if(arch_irn_is(arch_env, node, rematerializable))
+	if(arch_irn_is(arch_env, node, rematerializable)) {
 		return 1;
+	}
 
 	if(be_is_StackParam(node))
 		return 1;
@@ -416,8 +436,9 @@ static int check_remat_conditions_costs(spill_env_t *env, ir_node *spilled, ir_n
 	} else {
 		costs += arch_get_op_estimated_cost(env->arch_env, spilled);
 	}
-	if(parentcosts + costs >= REMAT_COST_LIMIT)
+	if(parentcosts + costs >= REMAT_COST_LIMIT) {
 		return REMAT_COST_LIMIT;
+	}
 
 	argremats = 0;
 	for(i = 0, arity = get_irn_arity(spilled); i < arity; ++i) {
@@ -435,7 +456,6 @@ static int check_remat_conditions_costs(spill_env_t *env, ir_node *spilled, ir_n
 		}
 		argremats++;
 
-		// TODO can we get more accurate costs than +1?
 		costs += check_remat_conditions_costs(env, arg, reloader, parentcosts + costs);
 		if(parentcosts + costs >= REMAT_COST_LIMIT)
 			return REMAT_COST_LIMIT;
@@ -536,13 +556,10 @@ void be_insert_spills_reloads(spill_env_t *env) {
 		}
 
 		del_pset(values);
-	}
 
-	// reloads are placed now, but we might reuse the spill environment for further spilling decisions
-	del_set(env->spills);
-	env->spills = new_set(cmp_spillinfo, 1024);
+		si->reloaders = NULL;
+	}
 
 	be_remove_dead_nodes_from_schedule(env->chordal_env->irg);
-	//be_liveness_add_missing(env->chordal_env->lv);
 	be_liveness_recompute(env->chordal_env->lv);
 }

ir/be/bespillbelady.c

@@ -619,7 +619,7 @@ void be_spill_belady_spill_env(const be_chordal_env_t *chordal_env, spill_env_t
 	belady_env_t env;
 
 	FIRM_DBG_REGISTER(dbg, "firm.be.spill.belady");
-	//firm_dbg_set_mask(dbg, DBG_WSETS);
+	//firm_dbg_set_mask(dbg, DBG_START);
 
 	/* init belady env */
 	obstack_init(&env.ob);

ir/be/bespillmorgan.c

@@ -248,10 +248,6 @@ static bitset_t *construct_block_livethrough_unused(morgan_env_t* env, const ir_
 		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;
 
@@ -344,90 +340,42 @@ static bitset_t *construct_loop_livethrough_unused(morgan_env_t *env, const ir_l
 /*---------------------------------------------------------------------------*/
 
 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));
 
-	/*
-	 * 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) {
-		loop_unused_spills_needed = 0;
-	} else if(spills_needed > loop_unused_spills_possible) {
-		int i, spills;
-		int block_unused_spills_needed;
+	loop_unused_spills_needed = max_pressure - env->registers_available;
 
+	if(loop_unused_spills_needed < 0) {
+		loop_unused_spills_needed = 0;
+	} 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));
 	return loop_unused_spills_needed;
 }
@@ -453,6 +401,7 @@ static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *lo
 			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;
@@ -460,6 +409,7 @@ static int reduce_register_pressure_in_loop(morgan_env_t *env, const ir_loop *lo
 		}
 		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;

ir/be/ia32/bearch_ia32.c

@@ -703,7 +703,7 @@ static int ia32_possible_memory_operand(const void *self, const ir_node *irn, un
 		get_irn_arity(irn) != 5                       ||  /* must be a binary operation */
 		get_ia32_op_type(irn) != ia32_Normal          ||  /* must not already be a addressmode irn */
 		! (get_ia32_am_support(irn) & ia32_am_Source) ||  /* must be capable of source addressmode */
-        (i != 2 && i != 3)                            ||  /* a "real" operand position must be requested */
+		(i != 2 && i != 3)                            ||  /* a "real" operand position must be requested */
 		(i == 2 && ! is_ia32_commutative(irn))        ||  /* if first operand requested irn must be commutative */
 		is_ia32_use_frame(irn))                           /* must not already use frame */
 		return 0;