X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Freassoc.c;h=5770604aa3051ce96ae60becbda512772a2dee16;hb=0fbcef83aa6060534172bb13e71cdadb04428806;hp=7ac64902505a1ca6edd409ab3da7b795a2ee618e;hpb=a9a046c530ac1b20d37ffb19e79e2734b9467865;p=libfirm

diff --git a/ir/opt/reassoc.c b/ir/opt/reassoc.c
index 7ac649025..5770604aa 100644
--- a/ir/opt/reassoc.c
+++ b/ir/opt/reassoc.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 1995-2007 University of Karlsruhe.  All right reserved.
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
  *
  * This file is part of libFirm.
  *
@@ -23,9 +23,7 @@
  * @author  Michael Beck
  * @version $Id$
  */
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include "iropt_t.h"
 #include "irnode_t.h"
@@ -36,12 +34,15 @@
 #include "iropt_dbg.h"
 #include "irflag_t.h"
 #include "irgwalk.h"
+#include "irouts.h"
 #include "reassoc_t.h"
 #include "irhooks.h"
 #include "irloop.h"
 #include "pdeq.h"
 #include "debug.h"
 
+//#define NEW_REASSOC
+
 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
 
 typedef struct _walker_t {
@@ -63,16 +64,13 @@ typedef enum {
  * @param n     the node to be checked for constant
  * @param block a block that might be in a loop
  */
-static const_class_t get_const_class(ir_node *n, ir_node *block)
+static const_class_t get_const_class(const ir_node *n, const ir_node *block)
 {
-	ir_op *op = get_irn_op(n);
-
-	if (op == op_Const)
+	if (is_Const(n))
 		return REAL_CONSTANT;
 
-	/* although SymConst's are of course real constant, we cannot
-	   fold them, so handle them like region constants */
-	if (op == op_SymConst)
+	/* constant nodes which can't be folded are region constants */
+	if (is_irn_constlike(n))
 		return REGION_CONST;
 
 	/*
@@ -168,12 +166,13 @@ static int reassoc_Sub(ir_node **in)
 			/* already constant, nothing to do */
 			return 0;
 		}
+
 		mode = get_irn_mode(n);
 		dbi  = get_irn_dbg_info(n);
 
 		/* Beware of SubP(P, Is) */
 		irn = new_rd_Minus(dbi, current_ir_graph, block, right, rmode);
-		irn = new_rd_Add(dbi, current_ir_graph, block, left, irn, get_irn_mode(n));
+		irn = new_rd_Add(dbi, current_ir_graph, block, left, irn, mode);
 
 		DBG((dbg, LEVEL_5, "Applied: %n - %n => %n + (-%n)\n",
 			get_Sub_left(n), right, get_Sub_left(n), right));
@@ -209,6 +208,8 @@ static ir_mode *get_mode_from_ops(ir_node *op1, ir_node *op2)
 	return m1;
 }  /* get_mode_from_ops */
 
+#ifndef NEW_REASSOC
+
 /**
  * reassociate a commutative Binop
  *
@@ -241,7 +242,8 @@ static int reassoc_commutative(ir_node **node)
 
 		if ( ((c_c1 > NO_CONSTANT) & (c_t2 > NO_CONSTANT)) &&
 		     ((((c_c1 ^ c_c2 ^ c_t2) & REGION_CONST) == 0) || ((c_c1 & c_c2 & c_t2) == REGION_CONST)) ) {
-			/* All three are constant and either all are constant expressions or two of them are:
+			/* All three are constant and either all are constant expressions
+			 * or two of them are:
 			 * then applying this rule would lead into a cycle
 			 *
 			 * Note that if t2 is a constant so is c2 hence we save one test.
@@ -256,7 +258,8 @@ static int reassoc_commutative(ir_node **node)
 			ir_node *irn, *in[2];
 			ir_mode *mode, *mode_c1 = get_irn_mode(c1), *mode_c2 = get_irn_mode(c2);
 
-			/* It might happen, that c1 and c2 have different modes, for instance Is and Iu.
+			/* It might happen, that c1 and c2 have different modes, for
+			 * instance Is and Iu.
 			 * Handle this here.
 			 */
 			if (mode_c1 != mode_c2) {
@@ -290,9 +293,9 @@ static int reassoc_commutative(ir_node **node)
 			     c1, get_irn_opname(n), c2, get_irn_opname(n), t2,
 			     t2, get_irn_opname(n), c1, get_irn_opname(n), c2));
 			/*
-			 * In some rare cases it can really happen that we get the same node back.
-			 * This might be happen in dead loops, were the Phi nodes are already gone away.
-			 * So check this.
+			 * In some rare cases it can really happen that we get the same
+			 * node back. This might be happen in dead loops, were the Phi
+			 * nodes are already gone away. So check this.
 			 */
 			if (n != irn) {
 				exchange(n, irn);
@@ -304,6 +307,135 @@ static int reassoc_commutative(ir_node **node)
 	return 0;
 }  /* reassoc_commutative */
 
+#else
+
+static ir_op          *commutative_op;
+static ir_node        *commutative_block;
+static struct obstack  commutative_args;
+
+static void collect_args(ir_node *node)
+{
+	ir_node *left  = get_binop_left(node);
+	ir_node *right = get_binop_right(node);
+
+	if (get_irn_op(left) == commutative_op
+			&& (!get_irn_outs_computed(left) || get_irn_n_outs(left) == 1)) {
+		collect_args(left);
+	} else {
+		obstack_ptr_grow(&commutative_args, left);
+	}
+
+	if (get_irn_op(right) == commutative_op
+			&& (!get_irn_outs_computed(right) || get_irn_n_outs(right) == 1)) {
+		collect_args(right);
+	} else {
+		obstack_ptr_grow(&commutative_args, right);
+	}
+
+#ifndef NDEBUG
+	{
+		ir_mode *mode = get_irn_mode(node);
+		if (is_Add(node) && mode_is_reference(mode)) {
+			assert(get_irn_mode(left) == mode || get_irn_mode(right) == mode);
+		} else {
+			assert(get_irn_mode(left) == mode);
+			assert(get_irn_mode(right) == mode);
+		}
+	}
+#endif
+}
+
+static int compare_nodes(const ir_node *node1, const ir_node *node2)
+{
+	const_class_t class1 = get_const_class(node1, commutative_block);
+	const_class_t class2 = get_const_class(node2, commutative_block);
+
+	if (class1 == class2)
+		return 0;
+	// return get_irn_idx(node1) - get_irn_idx(node2);
+
+	if (class1 < class2)
+		return -1;
+
+	assert(class1 > class2);
+	return 1;
+}
+
+static int compare_node_ptr(const void *e1, const void *e2)
+{
+	const ir_node *node1  = *((const ir_node *const*) e1);
+	const ir_node *node2  = *((const ir_node *const*) e2);
+	return compare_nodes(node1, node2);
+}
+
+static int reassoc_commutative(ir_node **n)
+{
+	int       i;
+	int       n_args;
+	ir_node  *last;
+	ir_node **args;
+	ir_mode  *mode;
+	ir_node  *node = *n;
+
+	commutative_op    = get_irn_op(node);
+	commutative_block = get_nodes_block(node);
+
+	/* collect all nodes with same op type */
+	collect_args(node);
+
+	n_args = obstack_object_size(&commutative_args) / sizeof(ir_node*);
+	args   = obstack_finish(&commutative_args);
+
+	/* shortcut: in most cases there's nothing to do */
+	if (n_args == 2 && compare_nodes(args[0], args[1]) <= 0) {
+		obstack_free(&commutative_args, args);
+		return 0;
+	}
+
+	/* sort the arguments */
+	qsort(args, n_args, sizeof(ir_node*), compare_node_ptr);
+
+	/* build new tree */
+	last = args[n_args-1];
+	mode = get_irn_mode(last);
+	for (i = n_args-2; i >= 0; --i) {
+		ir_mode *mode_right;
+		ir_node *new_node;
+		ir_node *in[2];
+
+		in[0] = last;
+		in[1] = args[i];
+
+		/* AddP violates the assumption that all modes in args are equal...
+		 * we need some hacks to cope with this */
+		mode_right = get_irn_mode(in[1]);
+		if (mode_is_reference(mode_right)) {
+			assert(is_Add(node) && mode_is_reference(get_irn_mode(node)));
+			mode = get_irn_mode(in[1]);
+		}
+		if (mode_right != mode) {
+			assert(is_Add(node) && mode_is_reference(get_irn_mode(node)));
+			in[1] = new_r_Conv(current_ir_graph, commutative_block,in[1], mode);
+		}
+
+		/* TODO: produce useful debug info! */
+		new_node = new_ir_node(NULL, current_ir_graph, commutative_block,
+		                       commutative_op, mode, 2, in);
+		new_node = optimize_node(new_node);
+		last     = new_node;
+	}
+
+	/* CSE often returns the old node again, only exchange if needed */
+	if (last != node) {
+		exchange(node, last);
+		*n = last;
+		return 1;
+	}
+	return 0;
+}
+
+#endif
+
 #define reassoc_Add  reassoc_commutative
 #define reassoc_And  reassoc_commutative
 #define reassoc_Or   reassoc_commutative
@@ -338,7 +470,6 @@ static int reassoc_Mul(ir_node **node)
 			in[0] = new_rd_Mul(NULL, current_ir_graph, block, c, t1, mode);
 			in[1] = new_rd_Mul(NULL, current_ir_graph, block, c, t2, mode);
 
-			mode  = get_mode_from_ops(in[0], in[1]);
 			irn   = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
 
 			/* In some cases it might happen that the new irn is equal the old one, for
@@ -363,6 +494,40 @@ static int reassoc_Mul(ir_node **node)
 	return 0;
 }  /* reassoc_Mul */
 
+/**
+ * Reassociate Shl. We transform Shl(x, const) into Mul's if possible.
+ */
+static int reassoc_Shl(ir_node **node) {
+	ir_node *n = *node;
+	ir_node *c = get_Shl_right(n);
+	ir_node *x, *blk, *irn;
+	ir_mode *mode;
+	tarval *tv;
+
+	if (! is_Const(c))
+		return 0;
+
+	x = get_Shl_left(n);
+	mode = get_irn_mode(x);
+
+	tv = get_mode_one(mode);
+	tv = tarval_shl(tv, get_Const_tarval(c));
+
+	if (tv == tarval_bad)
+		return 0;
+
+	blk = get_nodes_block(n);
+	c   = new_r_Const(current_ir_graph, blk, mode, tv);
+	irn = new_rd_Mul(get_irn_dbg_info(n), current_ir_graph, blk, x, c, mode);
+
+	if (irn != n) {
+		exchange(n, irn);
+		*node = irn;
+		return 1;
+	}
+	return 0;
+}  /* reassoc_Shl */
+
 /**
  * The walker for the reassociation.
  */
@@ -393,7 +558,6 @@ static void do_reassociation(walker_t *wenv)
 	int i, res, changed;
 	ir_node *n, *blk;
 
-
 	while (! waitq_empty(wenv->wq)) {
 		n = waitq_get(wenv->wq);
 		set_irn_link(n, NULL);
@@ -446,7 +610,7 @@ static void do_reassociation(walker_t *wenv)
 
 /**
  * Returns the earliest were a,b are available.
- * Note that we know that we know that a, b both dominate
+ * Note that we know that a, b both dominate
  * the block of the previous operation, so one must dominate the other.
  *
  * If the earliest block is the start block, return curr_blk instead
@@ -504,38 +668,51 @@ static int reverse_rule_distributive(ir_node **node) {
 	ir_node *right = get_binop_right(n);
 	ir_node *x, *blk, *curr_blk;
 	ir_node *a, *b, *irn;
+	ir_op *op;
 	ir_mode *mode;
 	dbg_info *dbg;
 
-	if (! is_Mul(left) || !is_Mul(right))
+	op = get_irn_op(left);
+	if (op != get_irn_op(right))
 		return 0;
 
-	x = get_Mul_left(left);
-
-	if (x == get_Mul_left(right)) {
-		/* (x * a) +/- (x * b) */
-		a = get_Mul_right(left);
-		b = get_Mul_right(right);
-		goto transform;
-	} else if (x == get_Mul_right(right)) {
-		/* (x * a) +/- (b * x) */
-		a = get_Mul_right(left);
-		b = get_Mul_left(right);
-		goto transform;
-	}
+	if (op == op_Shl) {
+		x = get_Shl_right(left);
 
-	x = get_Mul_right(left);
-
-	if (x == get_Mul_right(right)) {
-		/* (a * x) +/- (b * x) */
-		a = get_Mul_left(left);
-		b = get_Mul_left(right);
-		goto transform;
-	} else if (x == get_Mul_left(right)) {
-		/* (a * x) +/- (x * b) */
-		a = get_Mul_left(left);
-		b = get_Mul_right(right);
-		goto transform;
+		if (x == get_Shl_right(right)) {
+			/* (a << x) +/- (b << x) ==> (a +/- b) << x */
+			a = get_Shl_left(left);
+			b = get_Shl_left(right);
+			goto transform;
+		}
+	} else if (op == op_Mul) {
+		x = get_Mul_left(left);
+
+		if (x == get_Mul_left(right)) {
+			/* (x * a) +/- (x * b) ==> (a +/- b) * x */
+			a = get_Mul_right(left);
+			b = get_Mul_right(right);
+			goto transform;
+		} else if (x == get_Mul_right(right)) {
+			/* (x * a) +/- (b * x) ==> (a +/- b) * x */
+			a = get_Mul_right(left);
+			b = get_Mul_left(right);
+			goto transform;
+		}
+
+		x = get_Mul_right(left);
+
+		if (x == get_Mul_right(right)) {
+			/* (a * x) +/- (b * x) ==> (a +/- b) * x */
+			a = get_Mul_left(left);
+			b = get_Mul_left(right);
+			goto transform;
+		} else if (x == get_Mul_left(right)) {
+			/* (a * x) +/- (x * b) ==> (a +/- b) * x */
+			a = get_Mul_left(left);
+			b = get_Mul_right(right);
+			goto transform;
+		}
 	}
 	return 0;
 
@@ -553,7 +730,11 @@ transform:
 		irn = new_rd_Sub(dbg, current_ir_graph, blk, a, b, mode);
 
 	blk  = earliest_block(irn, x, curr_blk);
-	irn = new_rd_Mul(dbg, current_ir_graph, blk, irn, x, mode);
+
+	if (op == op_Mul)
+		irn = new_rd_Mul(dbg, current_ir_graph, blk, irn, x, mode);
+	else
+		irn = new_rd_Shl(dbg, current_ir_graph, blk, irn, x, mode);
 
 	exchange(n, irn);
 	*node = irn;
@@ -567,7 +748,7 @@ static int move_consts_up(ir_node **node) {
 	ir_node *n = *node;
 	ir_op *op;
 	ir_node *l, *r, *a, *b, *c, *blk, *irn, *in[2];
-	ir_mode *mode;
+	ir_mode *mode, *ma, *mb;
 	dbg_info *dbg;
 
 	l = get_binop_left(n);
@@ -580,16 +761,19 @@ static int move_consts_up(ir_node **node) {
 	dbg = get_irn_dbg_info(n);
 	op = get_irn_op(n);
 	if (get_irn_op(l) == op) {
+		/* (a .op. b) .op. r */
 		a = get_binop_left(l);
 		b = get_binop_right(l);
 
 		if (is_constant_expr(a)) {
+			/* (C .op. b) .op. r ==> (r .op. b) .op. C */
 			c = a;
 			a = r;
 			blk = get_nodes_block(l);
 			dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL;
 			goto transform;
 		} else if (is_constant_expr(b)) {
+			/* (a .op. C) .op. r ==> (a .op. r) .op. C */
 			c = b;
 			b = r;
 			blk = get_nodes_block(l);
@@ -597,16 +781,19 @@ static int move_consts_up(ir_node **node) {
 			goto transform;
 		}
 	} else if (get_irn_op(r) == op) {
+		/* l .op. (a .op. b) */
 		a = get_binop_left(r);
 		b = get_binop_right(r);
 
 		if (is_constant_expr(a)) {
+			/* l .op. (C .op. b) ==> (l .op. b) .op. C */
 			c = a;
 			a = l;
 			blk = get_nodes_block(r);
 			dbg = dbg == get_irn_dbg_info(r) ? dbg : NULL;
 			goto transform;
 		} else if (is_constant_expr(b)) {
+			/* l .op. (a .op. C) ==> (a .op. l) .op. C */
 			c = b;
 			b = l;
 			blk = get_nodes_block(r);
@@ -617,7 +804,18 @@ static int move_consts_up(ir_node **node) {
 	return 0;
 
 transform:
-	/* check if a+b can be calculted in the same block is the old instruction */
+	/* In some cases a and b might be both of different integer mode, and c a SymConst.
+	 * in that case we could either
+	 * 1.) cast into unsigned mode
+	 * 2.) ignore
+	 * we implement the second here
+	 */
+	ma = get_irn_mode(a);
+	mb = get_irn_mode(b);
+	if (ma != mb && mode_is_int(ma) && mode_is_int(mb))
+		return 0;
+
+	/* check if (a .op. b) can be calculated in the same block is the old instruction */
 	if (! block_dominates(get_nodes_block(a), blk))
 		return 0;
 	if (! block_dominates(get_nodes_block(b), blk))
@@ -626,10 +824,11 @@ transform:
 	in[0] = a;
 	in[1] = b;
 
-	mode = get_mode_from_ops(in[0], in[1]);
-	in[0] = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
+	mode = get_mode_from_ops(a, b);
+	in[0] = irn = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
 
-	if (op == op_Add || op == op_Sub) {
+	/* beware: optimize_node might have changed the opcode, check again */
+	if (is_Add(irn) || is_Sub(irn)) {
 		reverse_rule_distributive(&in[0]);
 	}
 	in[1] = c;
@@ -661,7 +860,8 @@ static void reverse_rules(ir_node *node, void *env) {
 		if (is_op_commutative(op)) {
 			wenv->changes |= res = move_consts_up(&node);
 		}
-		if (op == op_Add || op == op_Sub) {
+		/* beware: move_consts_up might have changed the opcode, check again */
+		if (is_Add(node) || is_Sub(node)) {
 			wenv->changes |= res = reverse_rule_distributive(&node);
 		}
 	} while (res);
@@ -670,22 +870,27 @@ static void reverse_rules(ir_node *node, void *env) {
 /*
  * do the reassociation
  */
-void optimize_reassociation(ir_graph *irg)
+int optimize_reassociation(ir_graph *irg)
 {
 	walker_t env;
 	irg_loopinfo_state state;
+	ir_graph *rem;
 
 	assert(get_irg_phase_state(irg) != phase_building);
 	assert(get_irg_pinned(irg) != op_pin_state_floats &&
 		"Reassociation needs pinned graph to work properly");
 
-	/* reassociation needs constant folding */
-	if (!get_opt_reassociation() || !get_opt_constant_folding())
-		return;
+	rem = current_ir_graph;
+	current_ir_graph = irg;
 
 	/* we use dominance to detect dead blocks */
 	assure_doms(irg);
 
+#ifdef NEW_REASSOC
+	assure_irg_outs(irg);
+	obstack_init(&commutative_args);
+#endif
+
 	/*
 	 * Calculate loop info, so we could identify loop-invariant
 	 * code and threat it like a constant.
@@ -700,12 +905,17 @@ void optimize_reassociation(ir_graph *irg)
 	env.changes = 0;
 	env.wq      = new_waitq();
 
-	/* now we have collected enough information, optimize */
-	irg_walk_graph(irg, NULL, wq_walker, &env);
-	do_reassociation(&env);
+	/* disable some optimizations while reassoc is running to prevent endless loops */
+	set_reassoc_running(1);
+	{
+		/* now we have collected enough information, optimize */
+		irg_walk_graph(irg, NULL, wq_walker, &env);
+		do_reassociation(&env);
 
-	/* reverse those rules that do not result in collapsed constants */
-	irg_walk_graph(irg, NULL, reverse_rules, &env);
+		/* reverse those rules that do not result in collapsed constants */
+		irg_walk_graph(irg, NULL, reverse_rules, &env);
+	}
+	set_reassoc_running(0);
 
 	/* Handle graph state */
 	if (env.changes) {
@@ -713,7 +923,13 @@ void optimize_reassociation(ir_graph *irg)
 		set_irg_loopinfo_inconsistent(irg);
 	}
 
+#ifdef NEW_REASSOC
+	obstack_free(&commutative_args, NULL);
+#endif
+
 	del_waitq(env.wq);
+	current_ir_graph = rem;
+	return env.changes;
 }  /* optimize_reassociation */
 
 /* Sets the default reassociation operation for an ir_op_ops. */
@@ -728,6 +944,7 @@ ir_op_ops *firm_set_default_reassoc(ir_opcode code, ir_op_ops *ops)
 	CASE(And);
 	CASE(Or);
 	CASE(Eor);
+	CASE(Shl);
 	default:
 		/* leave NULL */;
 	}