X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Ftailrec.c;h=e1945ea418a54fc3862e187615907984842f791a;hb=0aa3f54e9f29fbed49bb4781c28ab1157499a92e;hp=b0402da9d39cb9d12f5e4d15926df9c4c1a69d47;hpb=1ec30d95387eb392ba5a1adc7958ebd91383d59c;p=libfirm

diff --git a/ir/opt/tailrec.c b/ir/opt/tailrec.c
index b0402da9d..e1945ea41 100644
--- a/ir/opt/tailrec.c
+++ b/ir/opt/tailrec.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2011 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -22,18 +22,16 @@
  * @brief   Tail-recursion call optimization.
  * @date    08.06.2004
  * @author  Michael Beck
- * @version $Id$
  */
-#ifdef HAVE_CONFIG_H
-# include "config.h"
-#endif
+#include "config.h"
 
 #include <string.h>
 #include <assert.h>
 
+#include "debug.h"
 #include "iroptimize.h"
 #include "scalar_replace.h"
-#include "array.h"
+#include "array_t.h"
 #include "irprog_t.h"
 #include "irgwalk.h"
 #include "irgmod.h"
@@ -45,12 +43,15 @@
 #include "trouts.h"
 #include "irouts.h"
 #include "irhooks.h"
-#include "xmalloc.h"
+#include "ircons_t.h"
+#include "irpass.h"
+
+DEBUG_ONLY(static firm_dbg_module_t *dbg;)
 
 /**
  * the environment for collecting data
  */
-typedef struct _collect_t {
+typedef struct collect_t {
 	ir_node *proj_X;      /**< initial exec proj */
 	ir_node *block;       /**< old first block */
 	int     blk_idx;      /**< cfgpred index of the initial exec in block */
@@ -61,27 +62,28 @@ typedef struct _collect_t {
 /**
  * walker for collecting data, fills a collect_t environment
  */
-static void collect_data(ir_node *node, void *env) {
-	collect_t *data = env;
+static void collect_data(ir_node *node, void *env)
+{
+	collect_t *data = (collect_t*)env;
 	ir_node *pred;
-	ir_op *op;
+	ir_opcode opcode;
 
 	switch (get_irn_opcode(node)) {
 	case iro_Proj:
 		pred = get_Proj_pred(node);
 
-		op = get_irn_op(pred);
-		if (op == op_Proj) {
+		opcode = (ir_opcode)get_irn_opcode(pred);
+		if (opcode == iro_Proj) {
 			ir_node *start = get_Proj_pred(pred);
 
-			if (get_irn_op(start) == op_Start) {
+			if (is_Start(start)) {
 				if (get_Proj_proj(pred) == pn_Start_T_args) {
 					/* found Proj(ProjT(Start)) */
 					set_irn_link(node, data->proj_data);
 					data->proj_data = node;
 				}
 			}
-		} else if (op == op_Start) {
+		} else if (opcode == iro_Start) {
 			if (get_Proj_proj(node) == pn_Start_X_initial_exec) {
 				/* found ProjX(Start) */
 				data->proj_X = node;
@@ -94,7 +96,7 @@ static void collect_data(ir_node *node, void *env) {
 		/*
 		 * the first block has the initial exec as cfg predecessor
 		 */
-		if (node != get_irg_start_block(current_ir_graph)) {
+		if (node != get_irg_start_block(get_irn_irg(node))) {
 			for (i = 0; i < n_pred; ++i) {
 				if (get_Block_cfgpred(node, i) == data->proj_X) {
 					data->block   = node;
@@ -110,40 +112,46 @@ static void collect_data(ir_node *node, void *env) {
 	}
 }
 
+typedef enum tail_rec_variants {
+	TR_DIRECT,  /**< direct return value, i.e. return func(). */
+	TR_ADD,     /**< additive return value, i.e. return x +/- func() */
+	TR_MUL,     /**< multiplicative return value, i.e. return x * func() or return -func() */
+	TR_BAD,     /**< any other transformation */
+	TR_UNKNOWN  /**< during construction */
+} tail_rec_variants;
+
+typedef struct tr_env {
+	int               n_tail_calls;  /**< number of tail calls found */
+	int               n_ress;        /**< number of return values */
+	tail_rec_variants *variants;     /**< return value variants */
+	ir_node           *rets;         /**< list of returns that can be transformed */
+} tr_env;
+
+
 /**
  * do the graph reconstruction for tail-recursion elimination
  *
- * @param irg           the graph that will reconstructed
- * @param rets          linked list of all rets
- * @param n_tail_calls  number of tail-recursion calls
+ * @param irg  the graph that will reconstructed
+ * @param env  tail recursion environment
  */
-static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
+static void do_opt_tail_rec(ir_graph *irg, tr_env *env)
+{
 	ir_node *end_block = get_irg_end_block(irg);
 	ir_node *block, *jmp, *call, *calls;
 	ir_node **in;
 	ir_node **phis;
 	ir_node ***call_params;
 	ir_node *p, *n;
-	int i, j, n_params;
+	int i, j, n_params, n_locs;
 	collect_t data;
 	int rem            = get_optimize();
 	ir_entity *ent     = get_irg_entity(irg);
 	ir_type *method_tp = get_entity_type(ent);
 
-	assert(n_tail_calls);
-
-	/* we add new nodes, so the outs are inconsistent */
-	set_irg_outs_inconsistent(irg);
+	assert(env->n_tail_calls > 0);
 
 	/* we add new blocks and change the control flow */
-	set_irg_doms_inconsistent(irg);
-	set_irg_extblk_inconsistent(irg);
-
-	/* we add a new loop */
-	set_irg_loopinfo_inconsistent(irg);
-
-	/* calls are removed */
-	set_trouts_inconsistent();
+	clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE);
 
 	/* we must build some new nodes WITHOUT CSE */
 	set_optimize(0);
@@ -157,7 +165,7 @@ static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
 	irg_walk_graph(irg, NULL, collect_data, &data);
 
 	/* check number of arguments */
-	call = get_irn_link(end_block);
+	call     = (ir_node*)get_irn_link(end_block);
 	n_params = get_Call_n_params(call);
 
 	assert(data.proj_X && "Could not find initial exec from Start");
@@ -166,27 +174,30 @@ static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
 	assert((data.proj_data || n_params == 0) && "Could not find Proj(ProjT(Start)) of non-void function");
 
 	/* allocate in's for phi and block construction */
-	NEW_ARR_A(ir_node *, in, n_tail_calls + 1);
+	NEW_ARR_A(ir_node *, in, env->n_tail_calls + 1);
 
-	in[0] = data.proj_X;
+	/* build a new header block for the loop we create */
+	i = 0;
+	in[i++] = data.proj_X;
 
 	/* turn Return's into Jmp's */
-	for (i = 1, p = rets; p; p = n) {
+	for (p = env->rets; p; p = n) {
 		ir_node *block = get_nodes_block(p);
 
-		n = get_irn_link(p);
-		in[i++] = new_r_Jmp(irg, block);
+		n = (ir_node*)get_irn_link(p);
+		in[i++] = new_r_Jmp(block);
 
-		exchange(p, new_r_Bad(irg));
+		// exchange(p, new_r_Bad(irg));
 
 		/* we might generate an endless loop, so add
-		* the block to the keep-alive list */
+		 * the block to the keep-alive list */
 		add_End_keepalive(get_irg_end(irg), block);
 	}
+	assert(i == env->n_tail_calls + 1);
 
-	/* create a new block at start */
-	block = new_r_Block(irg, n_tail_calls + 1, in);
-	jmp   = new_r_Jmp(irg, block);
+	/* now create it */
+	block = new_r_Block(irg, i, in);
+	jmp   = new_r_Jmp(block);
 
 	/* the old first block is now the second one */
 	set_Block_cfgpred(data.block, data.blk_idx, jmp);
@@ -196,43 +207,42 @@ static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
 
 	/* build the memory phi */
 	i = 0;
-	in[i] = new_r_Proj(irg, get_irg_start_block(irg), get_irg_start(irg), mode_M, pn_Start_M);
+	in[i] = new_r_Proj(get_irg_start(irg), mode_M, pn_Start_M);
 	set_irg_initial_mem(irg, in[i]);
 	++i;
 
-	for (calls = call; calls; calls = get_irn_link(calls)) {
+	for (calls = call; calls != NULL; calls = (ir_node*)get_irn_link(calls)) {
 		in[i] = get_Call_mem(calls);
 		++i;
 	}
-	assert(i == n_tail_calls + 1);
+	assert(i == env->n_tail_calls + 1);
 
-	phis[0] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode_M);
+	phis[0] = new_r_Phi(block, env->n_tail_calls + 1, in, mode_M);
 
 	/* build the data Phi's */
 	if (n_params > 0) {
 		ir_node *calls;
 		ir_node *args;
-		ir_node *args_bl;
 
-		NEW_ARR_A(ir_node **, call_params, n_tail_calls);
+		NEW_ARR_A(ir_node **, call_params, env->n_tail_calls);
 
 		/* collect all parameters */
-		for (i = 0, calls = call; calls; calls = get_irn_link(calls)) {
+		for (i = 0, calls = call; calls != NULL;
+		     calls = (ir_node*)get_irn_link(calls)) {
 			call_params[i] = get_Call_param_arr(calls);
 			++i;
 		}
 
 		/* build new Proj's and Phi's */
 		args    = get_irg_args(irg);
-		args_bl = get_nodes_block(args);
 		for (i = 0; i < n_params; ++i) {
 			ir_mode *mode = get_type_mode(get_method_param_type(method_tp, i));
 
-			in[0] = new_r_Proj(irg, args_bl, args, mode, i);
-			for (j = 0; j < n_tail_calls; ++j)
+			in[0] = new_r_Proj(args, mode, i);
+			for (j = 0; j < env->n_tail_calls; ++j)
 				in[j + 1] = call_params[j][i];
 
-			phis[i + 1] = new_r_Phi(irg, block, n_tail_calls + 1, in, mode);
+			phis[i + 1] = new_r_Phi(block, env->n_tail_calls + 1, in, mode);
 		}
 	}
 
@@ -240,27 +250,147 @@ static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
 	 * ok, we are here, so we have build and collected all needed Phi's
 	 * now exchange all Projs into links to Phi
 	 */
-	for (p = data.proj_m; p; p = n) {
-		n = get_irn_link(p);
-		exchange(p, phis[0]);
-	}
+	exchange(data.proj_m, phis[0]);
 	for (p = data.proj_data; p; p = n) {
 		long proj = get_Proj_proj(p);
 
 		assert(0 <= proj && proj < n_params);
-		n = get_irn_link(p);
+		n = (ir_node*)get_irn_link(p);
 		exchange(p, phis[proj + 1]);
 	}
 
 	/* tail recursion was done, all info is invalid */
-	set_irg_doms_inconsistent(irg);
-	set_irg_outs_inconsistent(irg);
-	set_irg_extblk_inconsistent(irg);
-	set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
-	set_trouts_inconsistent();
+	clear_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE
+	                   | IR_GRAPH_PROPERTY_CONSISTENT_LOOPINFO);
 	set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
 
 	set_optimize(rem);
+
+	/* check if we need new values */
+	n_locs = 0;
+	for (i = 0; i < env->n_ress; ++i) {
+		if (env->variants[i] != TR_DIRECT) {
+			++n_locs;
+			break;
+		}
+	}
+
+	if (n_locs > 0) {
+		ir_node *start_block;
+		ir_node **in;
+		ir_mode **modes;
+
+		NEW_ARR_A(ir_node *, in, env->n_ress);
+		NEW_ARR_A(ir_mode *, modes, env->n_ress);
+		ssa_cons_start(irg, env->n_ress);
+
+		start_block = get_irg_start_block(irg);
+		set_r_cur_block(irg, start_block);
+
+		/* set the neutral elements for the iteration start */
+		for (i = 0; i < env->n_ress; ++i) {
+			ir_type *tp = get_method_res_type(method_tp, i);
+			ir_mode *mode = get_type_mode(tp);
+
+			modes[i] = mode;
+			if (env->variants[i] == TR_ADD) {
+				set_r_value(irg, i, new_r_Const(irg, get_mode_null(mode)));
+			} else if (env->variants[i] == TR_MUL) {
+				set_r_value(irg, i, new_r_Const(irg, get_mode_one(mode)));
+			}
+		}
+		mature_immBlock(start_block);
+
+		/* no: we can kill all returns */
+		for (p = env->rets; p; p = n) {
+			ir_node *block = get_nodes_block(p);
+			ir_node *jmp, *tuple;
+
+			set_r_cur_block(irg, block);
+			n = (ir_node*)get_irn_link(p);
+
+			ir_node *const call = skip_Proj(get_Return_mem(p));
+			ir_node *const mem  = get_Call_mem(call);
+
+			/* create a new jump, free of CSE */
+			set_optimize(0);
+			jmp = new_r_Jmp(block);
+			set_optimize(rem);
+
+			for (i = 0; i < env->n_ress; ++i) {
+				ir_mode *mode = modes[i];
+				if (env->variants[i] != TR_DIRECT) {
+					in[i] = get_r_value(irg, i, mode);
+				} else {
+					in[i] = new_r_Bad(irg, mode);
+				}
+			}
+			/* create a new tuple for the return values */
+			tuple = new_r_Tuple(block, env->n_ress, in);
+
+			turn_into_tuple(call, pn_Call_max+1);
+			set_Tuple_pred(call, pn_Call_M,         mem);
+			set_Tuple_pred(call, pn_Call_X_regular, jmp);
+			set_Tuple_pred(call, pn_Call_X_except,  new_r_Bad(irg, mode_X));
+			set_Tuple_pred(call, pn_Call_T_result,  tuple);
+
+			for (i = 0; i < env->n_ress; ++i) {
+				ir_node *res = get_Return_res(p, i);
+				if (env->variants[i] != TR_DIRECT) {
+					set_r_value(irg, i, res);
+				}
+			}
+
+			exchange(p, new_r_Bad(irg, mode_X));
+		}
+
+		/* finally fix all other returns */
+		end_block = get_irg_end_block(irg);
+		for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
+			ir_node *ret = get_Block_cfgpred(end_block, i);
+			ir_node *block;
+
+			/* search all Returns of a block */
+			if (! is_Return(ret))
+				continue;
+
+			block = get_nodes_block(ret);
+			set_r_cur_block(irg, block);
+			for (j = 0; j < env->n_ress; ++j) {
+				ir_node *pred = get_Return_res(ret, j);
+				ir_node *n;
+
+				switch (env->variants[j]) {
+				case TR_DIRECT:
+					continue;
+
+				case TR_ADD:
+					n = get_r_value(irg, j, modes[j]);
+					n = new_r_Add(block, n, pred, modes[j]);
+					set_Return_res(ret, j, n);
+					break;
+
+				case TR_MUL:
+					n = get_r_value(irg, j, modes[j]);
+					n = new_r_Mul(block, n, pred, modes[j]);
+					set_Return_res(ret, j, n);
+					break;
+
+				default:
+					assert(!"unexpected tail recursion variant");
+				}
+			}
+		}
+		ssa_cons_finish(irg);
+	} else {
+		ir_node *bad = new_r_Bad(irg, mode_X);
+
+		/* no: we can kill all returns */
+		for (p = env->rets; p; p = n) {
+			n = (ir_node*)get_irn_link(p);
+			exchange(p, bad);
+		}
+	}
 }
 
 /**
@@ -272,42 +402,200 @@ static void do_opt_tail_rec(ir_graph *irg, ir_node *rets, int n_tail_calls) {
  *
  * @return non-zero if it's ok to do tail recursion
  */
-static int check_lifetime_of_locals(ir_graph *irg) {
-	ir_node *irg_frame = get_irg_frame(irg);
+static int check_lifetime_of_locals(ir_graph *irg)
+{
+	ir_node *irg_frame;
 	int i;
+	ir_type *frame_tp = get_irg_frame_type(irg);
 
-	if (get_irg_outs_state(irg) != outs_consistent)
-		compute_irg_outs(irg);
-
+	irg_frame = get_irg_frame(irg);
 	for (i = get_irn_n_outs(irg_frame) - 1; i >= 0; --i) {
 		ir_node *succ = get_irn_out(irg_frame, i);
 
-		if (is_Sel(succ) && is_address_taken(succ))
-			return 0;
+		if (is_Sel(succ)) {
+			/* Check if we have compound arguments.
+			   For now, we cannot handle them, */
+			if (get_entity_owner(get_Sel_entity(succ)) != frame_tp)
+				return 0;
+
+			if (is_address_taken(succ))
+				return 0;
+		}
 	}
 	return 1;
 }
 
+/**
+ * Examine irn and detect the recursion variant.
+ */
+static tail_rec_variants find_variant(ir_node *irn, ir_node *call)
+{
+	ir_node           *a, *b;
+	tail_rec_variants va, vb, res;
+
+	if (skip_Proj(skip_Proj(irn)) == call) {
+		/* found it */
+		return TR_DIRECT;
+	}
+	switch (get_irn_opcode(irn)) {
+	case iro_Add:
+		/* try additive */
+		a = get_Add_left(irn);
+		if (get_nodes_block(a) != get_nodes_block(call)) {
+			/* we are outside, ignore */
+			va = TR_UNKNOWN;
+		} else {
+			va = find_variant(a, call);
+			if (va == TR_BAD)
+				return TR_BAD;
+		}
+		b = get_Add_right(irn);
+		if (get_nodes_block(b) != get_nodes_block(call)) {
+			/* we are outside, ignore */
+			vb = TR_UNKNOWN;
+		} else {
+			vb = find_variant(b, call);
+			if (vb == TR_BAD)
+				return TR_BAD;
+		}
+		if (va == vb) {
+			res = va;
+		}
+		else if (va == TR_UNKNOWN)
+			res = vb;
+		else if (vb == TR_UNKNOWN)
+			res = va;
+		else {
+			/* they are different but none is TR_UNKNOWN -> incompatible */
+			return TR_BAD;
+		}
+		if (res == TR_DIRECT || res == TR_ADD)
+			return TR_ADD;
+		/* not compatible */
+		return TR_BAD;
+
+	case iro_Sub:
+		/* try additive, but return value must be left */
+		a = get_Sub_left(irn);
+		if (get_nodes_block(a) != get_nodes_block(call)) {
+			/* we are outside, ignore */
+			va = TR_UNKNOWN;
+		} else {
+			va = find_variant(a, call);
+			if (va == TR_BAD)
+				return TR_BAD;
+		}
+		b = get_Sub_right(irn);
+		if (get_nodes_block(b) != get_nodes_block(call)) {
+			/* we are outside, ignore */
+			vb = TR_UNKNOWN;
+		} else {
+			vb = find_variant(b, call);
+			if (vb != TR_UNKNOWN)
+				return TR_BAD;
+		}
+		res = va;
+		if (res == TR_DIRECT || res == TR_ADD)
+			return res;
+		/* not compatible */
+		return TR_BAD;
+
+	case iro_Mul:
+		/* try multiplicative */
+		a = get_Mul_left(irn);
+		if (get_nodes_block(a) != get_nodes_block(call)) {
+			/* we are outside, ignore */
+			va = TR_UNKNOWN;
+		} else {
+			va = find_variant(a, call);
+			if (va == TR_BAD)
+				return TR_BAD;
+		}
+		b = get_Mul_right(irn);
+		if (get_nodes_block(b) != get_nodes_block(call)) {
+			/* we are outside, ignore */
+			vb = TR_UNKNOWN;
+		} else {
+			vb = find_variant(b, call);
+			if (vb == TR_BAD)
+				return TR_BAD;
+		}
+		if (va == vb) {
+			res = va;
+		}
+		else if (va == TR_UNKNOWN)
+			res = vb;
+		else if (vb == TR_UNKNOWN)
+			res = va;
+		else {
+			/* they are different but none is TR_UNKNOWN -> incompatible */
+			return TR_BAD;
+		}
+		if (res == TR_DIRECT || res == TR_MUL)
+			return TR_MUL;
+		/* not compatible */
+		return TR_BAD;
+
+	case iro_Minus:
+		/* try multiplicative */
+		a = get_Minus_op(irn);
+		res =  find_variant(a, call);
+		if (res == TR_DIRECT)
+			return TR_MUL;
+		if (res == TR_MUL || res == TR_UNKNOWN)
+			return res;
+		/* not compatible */
+		return TR_BAD;
+
+	default:
+		return TR_UNKNOWN;
+	}
+}
+
+
 /*
  * convert simple tail-calls into loops
  */
-int opt_tail_rec_irg(ir_graph *irg) {
-	ir_node *end_block;
-	int i, n_tail_calls = 0;
-	ir_node *rets = NULL;
-	ir_type *mtd_type, *call_type;
+void opt_tail_rec_irg(ir_graph *irg)
+{
+	tr_env    env;
+	ir_node   *end_block;
+	int       i, n_ress, n_tail_calls = 0;
+	ir_node   *rets = NULL;
+	ir_type   *mtd_type, *call_type;
+	ir_entity *ent;
+	ir_graph  *rem;
+
+	assure_irg_properties(irg,
+		IR_GRAPH_PROPERTY_MANY_RETURNS
+		| IR_GRAPH_PROPERTY_NO_BADS
+		| IR_GRAPH_PROPERTY_CONSISTENT_OUTS);
+
+	FIRM_DBG_REGISTER(dbg, "firm.opt.tailrec");
+
+	if (! check_lifetime_of_locals(irg)) {
+		confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
+		return;
+	}
 
-	if (! get_opt_tail_recursion() || ! get_opt_optimize())
-		return 0;
+	rem = current_ir_graph;
+	current_ir_graph = irg;
 
-	if (! check_lifetime_of_locals(irg))
-		return 0;
+	ent      = get_irg_entity(irg);
+	mtd_type = get_entity_type(ent);
+	n_ress   = get_method_n_ress(mtd_type);
 
-	/*
-	 * This tail recursion optimization works best
-	 * if the Returns are normalized.
-	 */
-	normalize_n_returns(irg);
+	env.variants = NULL;
+	env.n_ress   = n_ress;
+
+	if (n_ress > 0) {
+		NEW_ARR_A(tail_rec_variants, env.variants, n_ress);
+
+		for (i = 0; i < n_ress; ++i)
+			env.variants[i] = TR_DIRECT;
+	}
+
+	ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 
 	end_block = get_irg_end_block(irg);
 	set_irn_link(end_block, NULL);
@@ -315,7 +603,6 @@ int opt_tail_rec_irg(ir_graph *irg) {
 	for (i = get_Block_n_cfgpreds(end_block) - 1; i >= 0; --i) {
 		ir_node *ret = get_Block_cfgpred(end_block, i);
 		ir_node *call, *call_ptr;
-		ir_entity *ent;
 		int j;
 		ir_node **ress;
 
@@ -328,36 +615,24 @@ int opt_tail_rec_irg(ir_graph *irg) {
 		if (! is_Call(call))
 			continue;
 
+		/* the call must be in the same block as the return */
+		if (get_nodes_block(call) != get_nodes_block(ret))
+			continue;
+
 		/* check if it's a recursive call */
 		call_ptr = get_Call_ptr(call);
 
-		if (get_irn_op(call_ptr) != op_SymConst)
-			continue;
-
-		if (get_SymConst_kind(call_ptr) != symconst_addr_ent)
+		if (! is_SymConst_addr_ent(call_ptr))
 			continue;
 
 		ent = get_SymConst_entity(call_ptr);
 		if (!ent || get_entity_irg(ent) != irg)
 			continue;
 
-		/* ok, mem is routed to a recursive call, check return args */
-		ress = get_Return_res_arr(ret);
-		for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
-			ir_node *irn = skip_Proj(skip_Proj(ress[j]));
-
-			if (irn != call) {
-				/* not routed to a call */
-				break;
-			}
-		}
-		if (j >= 0)
-			continue;
-
 		/*
-		* Check, that the types match. At least in C
-		* this might fail.
-		*/
+		 * Check, that the types match. At least in C
+		 * this might fail.
+		 */
 		mtd_type  = get_entity_type(ent);
 		call_type = get_Call_type(call);
 
@@ -367,14 +642,33 @@ int opt_tail_rec_irg(ir_graph *irg) {
 			 * This can happen in C when no prototype is given
 			 * or K&R style is used.
 			 */
-#if 0
-			printf("Warning: Tail recursion fails because of different method and call types:\n");
-			dump_type(mtd_type);
-			dump_type(call_type);
-#endif
-			return 0;
+			DB((dbg, LEVEL_3, "  tail recursion fails because of call type mismatch: %+F != %+F\n", mtd_type, call_type));
+			continue;
 		}
 
+		/* ok, mem is routed to a recursive call, check return args */
+		ress = get_Return_res_arr(ret);
+		for (j = get_Return_n_ress(ret) - 1; j >= 0; --j) {
+			tail_rec_variants var = find_variant(ress[j], call);
+
+			if (var >= TR_BAD) {
+				/* cannot be transformed */
+				break;
+			}
+			if (var == TR_DIRECT) {
+				var = env.variants[j];
+			} else if (env.variants[j] == TR_DIRECT) {
+				env.variants[j] = var;
+			}
+			if (env.variants[j] != var) {
+				/* not compatible */
+				DB((dbg, LEVEL_3, "  tail recursion fails for %d return value of %+F\n", j, ret));
+				break;
+			}
+		}
+		if (j >= 0)
+			continue;
+
 		/* here, we have found a call */
 		set_irn_link(call, get_irn_link(end_block));
 		set_irn_link(end_block, call);
@@ -386,37 +680,45 @@ int opt_tail_rec_irg(ir_graph *irg) {
 	}
 
 	/* now, end_block->link contains the list of all tail calls */
-	if (! n_tail_calls)
-		return 0;
-
-	if (get_opt_tail_recursion_verbose() && get_firm_verbosity() > 1)
-		printf("  Performing tail recursion for graph %s and %d Calls\n",
-		get_entity_ld_name(get_irg_entity(irg)), n_tail_calls);
-
-	hook_tail_rec(irg, n_tail_calls);
-	do_opt_tail_rec(irg, rets, n_tail_calls);
+	if (n_tail_calls > 0) {
+		DB((dbg, LEVEL_2, "  Performing tail recursion for graph %s and %d Calls\n",
+		    get_entity_ld_name(get_irg_entity(irg)), n_tail_calls));
+
+		hook_tail_rec(irg, n_tail_calls);
+
+		env.n_tail_calls = n_tail_calls;
+		env.rets         = rets;
+		do_opt_tail_rec(irg, &env);
+		confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE);
+	} else {
+		confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_ALL);
+	}
+	ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
+	current_ir_graph = rem;
+}
 
-	return n_tail_calls;
+ir_graph_pass_t *opt_tail_rec_irg_pass(const char *name)
+{
+	return def_graph_pass(name ? name : "tailrec", opt_tail_rec_irg);
 }
 
 /*
  * optimize tail recursion away
  */
-void opt_tail_recursion(void) {
-	int i;
-	int n_opt_applications = 0;
-	ir_graph *irg;
+void opt_tail_recursion(void)
+{
+	size_t i, n;
 
-	if (! get_opt_tail_recursion() || ! get_opt_optimize())
-		return;
-
-	for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
-		irg = get_irp_irg(i);
+	FIRM_DBG_REGISTER(dbg, "firm.opt.tailrec");
 
-		if (opt_tail_rec_irg(irg))
-			++n_opt_applications;
+	DB((dbg, LEVEL_1, "Performing tail recursion ...\n"));
+	for (i = 0, n = get_irp_n_irgs(); i < n; ++i) {
+		ir_graph *irg = get_irp_irg(i);
+		opt_tail_rec_irg(irg);
 	}
+}
 
-	if (get_opt_tail_recursion_verbose())
-		printf("Performed tail recursion for %d of %d graphs\n", n_opt_applications, get_irp_n_irgs());
+ir_prog_pass_t *opt_tail_recursion_pass(const char *name)
+{
+	return def_prog_pass(name ? name : "tailrec", opt_tail_recursion);
 }