/*
- * 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.
*
* @author Michael Beck
* @version $Id$
*/
-#ifdef HAVE_CONFIG_H
#include "config.h"
-#endif
+#include "iroptimize.h"
#include "iropt_t.h"
#include "irnode_t.h"
#include "irgraph_t.h"
#include "iropt_dbg.h"
#include "irflag_t.h"
#include "irgwalk.h"
+#include "irouts.h"
#include "reassoc_t.h"
+#include "opt_init.h"
#include "irhooks.h"
#include "irloop.h"
#include "pdeq.h"
#include "debug.h"
+#include "irpass.h"
+
+//#define NEW_REASSOC
DEBUG_ONLY(static firm_dbg_module_t *dbg;)
-typedef struct _walker_t {
- int changes; /**< set, if a reassociation take place */
- waitq *wq; /**< a wait queue */
+typedef struct walker_t {
+ int changes; /**< set, if a reassociation take place */
+ ir_graph *irg;
+ waitq *wq; /**< a wait queue */
} walker_t;
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;
/*
/* 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_Minus(dbi, block, right, rmode);
+ irn = new_rd_Add(dbi, block, left, irn, mode);
DBG((dbg, LEVEL_5, "Applied: %n - %n => %n + (-%n)\n",
get_Sub_left(n), right, get_Sub_left(n), right));
return m1;
} /* get_mode_from_ops */
+#ifndef NEW_REASSOC
+
/**
* reassociate a commutative Binop
*
static int reassoc_commutative(ir_node **node)
{
ir_node *n = *node;
- ir_op *op = get_irn_op(n);
+ ir_op *op = get_irn_op(n);
ir_node *block = get_nodes_block(n);
ir_node *t1, *c1;
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.
*/
ir_node *irn, *in[2];
ir_mode *mode, *mode_c1 = get_irn_mode(c1), *mode_c2 = get_irn_mode(c2);
+ ir_graph *irg = get_irn_irg(c1);
- /* 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) {
if (mode_is_int(mode_c1) && mode_is_int(mode_c2)) {
/* get the bigger one */
if (get_mode_size_bits(mode_c1) > get_mode_size_bits(mode_c2))
- c2 = new_r_Conv(current_ir_graph, block, c2, mode_c1);
+ c2 = new_r_Conv(block, c2, mode_c1);
else if (get_mode_size_bits(mode_c1) < get_mode_size_bits(mode_c2))
- c1 = new_r_Conv(current_ir_graph, block, c1, mode_c2);
+ c1 = new_r_Conv(block, c1, mode_c2);
else {
/* Try to cast the real const */
if (c_c1 == REAL_CONSTANT)
- c1 = new_r_Conv(current_ir_graph, block, c1, mode_c2);
+ c1 = new_r_Conv(block, c1, mode_c2);
else
- c2 = new_r_Conv(current_ir_graph, block, c2, mode_c1);
+ c2 = new_r_Conv(block, c2, mode_c1);
}
}
}
in[1] = c2;
mode = get_mode_from_ops(in[0], in[1]);
- in[1] = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
+ in[1] = optimize_node(new_ir_node(NULL, irg, block, op, mode, 2, in));
in[0] = t2;
mode = get_mode_from_ops(in[0], in[1]);
- irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
+ irn = optimize_node(new_ir_node(NULL, irg, block, op, mode, 2, in));
DBG((dbg, LEVEL_5, "Applied: %n .%s. (%n .%s. %n) => %n .%s. (%n .%s. %n)\n",
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);
}
}
}
+ if (get_irn_op(c1) == op) {
+ ir_node *t = c1;
+ c1 = t1;
+ t1 = t;
+ }
+ if (get_irn_op(t1) == op) {
+ ir_node *l = get_binop_left(t1);
+ ir_node *r = get_binop_right(t1);
+ const_class_t c_r;
+
+ if (r == c1) {
+ ir_node *t = r;
+ r = l;
+ l = t;
+ }
+ c_r = get_const_class(r, block);
+ if (c_r != NO_CONSTANT) {
+ /*
+ * Beware: don't do the following op if a constant was
+ * placed below, else we will fall into a loop.
+ */
+ return 0;
+ }
+
+ if (l == c1) {
+ /* convert x .OP. (x .OP. y) => y .OP. (x .OP. x) */
+ ir_mode *mode_res = get_irn_mode(n);
+ ir_mode *mode_c1 = get_irn_mode(c1);
+ ir_graph *irg = get_irn_irg(c1);
+ ir_node *irn, *in[2];
+
+ in[0] = c1;
+ in[1] = c1;
+
+ in[1] = optimize_node(new_ir_node(NULL, irg, block, op, mode_c1, 2, in));
+ in[0] = r;
+
+ irn = optimize_node(new_ir_node(NULL, irg, block, op, mode_res, 2, in));
+
+ DBG((dbg, LEVEL_5, "Applied: %n .%s. (%n .%s. %n) => %n .%s. (%n .%s. %n)\n",
+ c1, get_irn_opname(n), l, get_irn_opname(n), r,
+ r, get_irn_opname(n), c1, get_irn_opname(n), c1));
+
+ if (n != irn) {
+ exchange(n, irn);
+ *node = irn;
+ return 1;
+ }
+ }
+ }
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];
+ ir_graph *irg = get_irn_irg(last);
+
+ 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(irg, commutative_block,in[1], mode);
+ }
+
+ /* TODO: produce useful debug info! */
+ new_node = new_ir_node(NULL, irg, 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
/* we can only multiplication rules on integer arithmetic */
if (mode_is_int(get_irn_mode(t1)) && mode_is_int(get_irn_mode(t2))) {
- 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);
+ ir_graph *irg = get_irn_irg(t1);
+ in[0] = new_rd_Mul(NULL, block, c, t1, mode);
+ in[1] = new_rd_Mul(NULL, block, c, t2, mode);
- irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
+ irn = optimize_node(new_ir_node(NULL, irg, block, op, mode, 2, in));
/* In some cases it might happen that the new irn is equal the old one, for
* instance in:
/**
* 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;
+static int reassoc_Shl(ir_node **node)
+{
+ ir_node *n = *node;
+ ir_node *c = get_Shl_right(n);
+ ir_node *x, *blk, *irn;
+ ir_graph *irg;
+ ir_mode *mode;
+ ir_tarval *tv;
if (! is_Const(c))
return 0;
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);
+ irg = get_irn_irg(blk);
+ c = new_r_Const(irg, tv);
+ irn = new_rd_Mul(get_irn_dbg_info(n), blk, x, c, mode);
if (irn != n) {
exchange(n, irn);
*/
static void wq_walker(ir_node *n, void *env)
{
- walker_t *wenv = env;
+ walker_t *wenv = (walker_t*)env;
set_irn_link(n, NULL);
- if (is_no_Block(n)) {
+ if (!is_Block(n)) {
ir_node *blk = get_nodes_block(n);
if (is_Block_dead(blk) || get_Block_dom_depth(blk) < 0) {
int i, res, changed;
ir_node *n, *blk;
-
while (! waitq_empty(wenv->wq)) {
- n = waitq_get(wenv->wq);
+ n = (ir_node*)waitq_get(wenv->wq);
set_irn_link(n, NULL);
blk = get_nodes_block(n);
/* reassociation must run until a fixpoint is reached. */
changed = 0;
do {
- ir_op *op = get_irn_op(n);
- ir_mode *mode = get_irn_mode(n);
+ ir_op *op = get_irn_op(n);
+ ir_mode *mode = get_irn_mode(n);
res = 0;
/* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
- if (mode_is_float(mode) && get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)
+ if (mode_is_float(mode) && get_irg_fp_model(wenv->irg) & fp_strict_algebraic)
break;
if (op->ops.reassociate) {
*
* If the earliest block is the start block, return curr_blk instead
*/
-static ir_node *earliest_block(ir_node *a, ir_node *b, ir_node *curr_blk) {
+static ir_node *earliest_block(ir_node *a, ir_node *b, ir_node *curr_blk)
+{
ir_node *blk_a = get_nodes_block(a);
ir_node *blk_b = get_nodes_block(b);
ir_node *res;
res = blk_b;
else
res = blk_a;
- if (res == get_irg_start_block(current_ir_graph))
+ if (res == get_irg_start_block(get_irn_irg(curr_blk)))
return curr_blk;
return res;
} /* earliest_block */
* Handling SymConsts as const might be not a good idea for all
* architectures ...
*/
-static int is_constant_expr(ir_node *irn) {
+static int is_constant_expr(ir_node *irn)
+{
ir_op *op;
switch (get_irn_opcode(irn)) {
/**
* Apply distributive Law for Mul and Add/Sub
*/
-static int reverse_rule_distributive(ir_node **node) {
+static int reverse_rule_distributive(ir_node **node)
+{
ir_node *n = *node;
ir_node *left = get_binop_left(n);
ir_node *right = get_binop_right(n);
x = get_Shl_right(left);
if (x == get_Shl_right(right)) {
- /* (a << x) +/- (b << x) */
+ /* (a << x) +/- (b << x) ==> (a +/- b) << x */
a = get_Shl_left(left);
b = get_Shl_left(right);
goto transform;
x = get_Mul_left(left);
if (x == get_Mul_left(right)) {
- /* (x * a) +/- (x * b) */
+ /* (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) */
+ /* (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 * 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 * x) +/- (x * b) ==> (a +/- b) * x */
a = get_Mul_left(left);
b = get_Mul_right(right);
goto transform;
mode = get_irn_mode(n);
if (is_Add(n))
- irn = new_rd_Add(dbg, current_ir_graph, blk, a, b, mode);
+ irn = new_rd_Add(dbg, blk, a, b, mode);
else
- irn = new_rd_Sub(dbg, current_ir_graph, blk, a, b, mode);
+ irn = new_rd_Sub(dbg, blk, a, b, mode);
blk = earliest_block(irn, x, curr_blk);
if (op == op_Mul)
- irn = new_rd_Mul(dbg, current_ir_graph, blk, irn, x, mode);
+ irn = new_rd_Mul(dbg, blk, irn, x, mode);
else
- irn = new_rd_Shl(dbg, current_ir_graph, blk, irn, x, mode);
+ irn = new_rd_Shl(dbg, blk, irn, x, mode);
exchange(n, irn);
*node = irn;
/**
* Move Constants towards the root.
*/
-static int move_consts_up(ir_node **node) {
+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, *ma, *mb;
dbg_info *dbg;
+ ir_graph *irg;
l = get_binop_left(n);
r = get_binop_right(n);
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);
dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL;
goto transform;
}
- } else if (get_irn_op(r) == op) {
+ }
+ 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);
if (ma != mb && mode_is_int(ma) && mode_is_int(mb))
return 0;
- /* check if a+b can be calculated in the same block is the old instruction */
+ /* 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))
in[1] = b;
mode = get_mode_from_ops(a, b);
- in[0] = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
+ irg = get_irn_irg(blk);
+ in[0] = irn = optimize_node(new_ir_node(dbg, irg, 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;
mode = get_mode_from_ops(in[0], in[1]);
- irn = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
+ irn = optimize_node(new_ir_node(dbg, irg, blk, op, mode, 2, in));
exchange(n, irn);
*node = irn;
/**
* Apply the rules in reverse order, removing code that was not collapsed
*/
-static void reverse_rules(ir_node *node, void *env) {
- walker_t *wenv = env;
- ir_mode *mode = get_irn_mode(node);
+static void reverse_rules(ir_node *node, void *env)
+{
+ walker_t *wenv = (walker_t*)env;
+ ir_graph *irg = get_irn_irg(node);
+ ir_mode *mode = get_irn_mode(node);
int res;
/* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
- if (mode_is_float(mode) && get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)
+ if (mode_is_float(mode) && get_irg_fp_model(irg) & fp_strict_algebraic)
return;
do {
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);
/*
* 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.
construct_cf_backedges(irg);
env.changes = 0;
+ env.irg = irg;
env.wq = new_waitq();
/* disable some optimizations while reassoc is running to prevent endless loops */
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 */
+/* create a pass for the reassociation */
+ir_graph_pass_t *optimize_reassociation_pass(const char *name)
+{
+ return def_graph_pass_ret(name ? name : "reassoc", optimize_reassociation);
+} /* optimize_reassociation_pass */
+
/* Sets the default reassociation operation for an ir_op_ops. */
ir_op_ops *firm_set_default_reassoc(ir_opcode code, ir_op_ops *ops)
{