/*
- * 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 "iropt_t.h"
#include "irnode_t.h"
#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 {
- int changes; /* set, if a reassociation take place */
+ int changes; /**< set, if a reassociation take place */
+ 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;
/*
} /* get_comm_Binop_ops */
/**
- * reassociate a Sub: x - c = (-c) + x
+ * reassociate a Sub: x - c = x + (-c)
*/
static int reassoc_Sub(ir_node **in)
{
block = get_nodes_block(n);
/* handles rule R6:
- * convert x - c => (-c) + x
- *
- * As there is NO real Minus in Firm it makes no sense to do this
- * for non-real constants yet.
+ * convert x - c => x + (-c)
*/
if (get_const_class(right, block) == REAL_CONSTANT) {
ir_node *left = get_Sub_left(n);
ir_mode *mode;
dbg_info *dbi;
- ir_node *irn, *c;
+ ir_node *irn;
switch (get_const_class(left, block)) {
case REAL_CONSTANT:
/* already constant, nothing to do */
return 0;
}
+
mode = get_irn_mode(n);
dbi = get_irn_dbg_info(n);
/* Beware of SubP(P, Is) */
- c = new_r_Const(current_ir_graph, block, rmode, get_mode_null(rmode));
- irn = new_rd_Sub(dbi, current_ir_graph, block, c, right, rmode);
-
- irn = new_rd_Add(dbi, current_ir_graph, block, left, irn, get_irn_mode(n));
+ irn = new_rd_Minus(dbi, current_ir_graph, block, right, rmode);
+ 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), c, get_Sub_left(n), c));
+ get_Sub_left(n), right, get_Sub_left(n), right));
- if(n == irn)
+ if (n == irn)
return 0;
exchange(n, irn);
return m1;
} /* get_mode_from_ops */
+#ifndef NEW_REASSOC
+
/**
* reassociate a commutative Binop
*
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.
return 0;
}
- if ((c_c1 != NO_CONSTANT) & (c_c2 != NO_CONSTANT)) {
+ if ((c_c1 != NO_CONSTANT) /* & (c_c2 != NO_CONSTANT) */) {
/* handles rules R7, R8, R9, R10:
- * convert c1 .OP. (c2 .OP. x) => (c1 .OP. c2) .OP. x
+ * convert c1 .OP. (c2 .OP. x) => x .OP. (c1 .OP. c2)
*/
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) {
in[0] = c1;
in[1] = c2;
- mode = get_mode_from_ops(in[0], in[1]);
- in[0] = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
- in[1] = t2;
+ 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[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));
- 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, c1, get_irn_opname(n), c2, get_irn_opname(n), t2));
+ 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);
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
#define reassoc_Eor reassoc_commutative
/**
- * reassociate using distributive law for Mul and Add/Sub
+ * Reassociate using commutative law for Mul and distributive law for Mul and Add/Sub:
*/
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
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_Const(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.
*/
-static void do_reassociation(ir_node *n, void *env)
+static void wq_walker(ir_node *n, void *env)
{
walker_t *wenv = env;
- int res;
+ set_irn_link(n, NULL);
if (is_no_Block(n)) {
ir_node *blk = get_nodes_block(n);
which or cf_opt do not guarantee yet. */
return;
}
+ waitq_put(wenv->wq, n);
+ set_irn_link(n, wenv->wq);
+ }
+} /* wq_walker */
+
+/**
+ * The walker for the reassociation.
+ */
+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);
+
+ blk = get_nodes_block(n);
+ if (is_Block_dead(blk) || get_Block_dom_depth(blk) < 0) {
+ /* We are in a dead block, do not optimize or we may fall into an endless
+ loop. We check this here instead of requiring that all dead blocks are removed
+ which or cf_opt do not guarantee yet. */
+ continue;
+ }
+
hook_reassociate(1);
/* 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);
if (op->ops.reassociate) {
res = op->ops.reassociate(&n);
- wenv->changes |= res;
+ changed |= res;
}
} while (res == 1);
-
hook_reassociate(0);
+
+ wenv->changes |= changed;
+
+ if (changed) {
+ for (i = get_irn_arity(n) - 1; i >= 0; --i) {
+ ir_node *pred = get_irn_n(n, i);
+
+ if (get_irn_link(pred) != wenv->wq) {
+ waitq_put(wenv->wq, pred);
+ set_irn_link(pred, wenv->wq);
+ }
+ }
+ }
}
} /* do_reassociation */
+/**
+ * Returns the earliest were a,b are available.
+ * 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
+ */
+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;
+
+ /* if blk_a != blk_b, one must dominate the other */
+ if (block_dominates(blk_a, blk_b))
+ res = blk_b;
+ else
+ res = blk_a;
+ if (res == get_irg_start_block(current_ir_graph))
+ return curr_blk;
+ return res;
+} /* earliest_block */
+
+/**
+ * Checks whether a node is a Constant expression.
+ * The following trees are constant expressions:
+ *
+ * Const, SymConst, Const + SymConst
+ *
+ * Handling SymConsts as const might be not a good idea for all
+ * architectures ...
+ */
+static int is_constant_expr(ir_node *irn) {
+ ir_op *op;
+
+ switch (get_irn_opcode(irn)) {
+ case iro_Const:
+ case iro_SymConst:
+ return 1;
+ case iro_Add:
+ op = get_irn_op(get_Add_left(irn));
+ if (op != op_Const && op != op_SymConst)
+ return 0;
+ op = get_irn_op(get_Add_right(irn));
+ if (op != op_Const && op != op_SymConst)
+ return 0;
+ return 1;
+ default:
+ return 0;
+ }
+} /* is_constant_expr */
+
+/**
+ * Apply distributive Law for Mul and Add/Sub
+ */
+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);
+ ir_node *x, *blk, *curr_blk;
+ ir_node *a, *b, *irn;
+ ir_op *op;
+ ir_mode *mode;
+ dbg_info *dbg;
+
+ op = get_irn_op(left);
+ if (op != get_irn_op(right))
+ return 0;
+
+ if (op == op_Shl) {
+ x = get_Shl_right(left);
+
+ 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;
+
+transform:
+ curr_blk = get_nodes_block(n);
+
+ blk = earliest_block(a, b, curr_blk);
+
+ dbg = get_irn_dbg_info(n);
+ mode = get_irn_mode(n);
+
+ if (is_Add(n))
+ irn = new_rd_Add(dbg, current_ir_graph, blk, a, b, mode);
+ else
+ irn = new_rd_Sub(dbg, current_ir_graph, 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);
+ else
+ irn = new_rd_Shl(dbg, current_ir_graph, blk, irn, x, mode);
+
+ exchange(n, irn);
+ *node = irn;
+ return 1;
+} /* reverse_rule_distributive */
+
+/**
+ * Move Constants towards the root.
+ */
+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;
+
+ l = get_binop_left(n);
+ r = get_binop_right(n);
+
+ /* check if one is already a constant expression */
+ if (is_constant_expr(l) || is_constant_expr(r))
+ return 0;
+
+ 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) {
+ /* 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);
+ dbg = dbg == get_irn_dbg_info(r) ? dbg : NULL;
+ goto transform;
+ }
+ }
+ return 0;
+
+transform:
+ /* 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))
+ return 0;
+ /* ok */
+ in[0] = a;
+ in[1] = b;
+
+ mode = get_mode_from_ops(a, b);
+ in[0] = irn = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
+
+ /* 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));
+
+ exchange(n, irn);
+ *node = irn;
+ return 1;
+} /* move_consts_up */
+
+/**
+ * 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);
+ 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)
+ return;
+
+ do {
+ ir_op *op = get_irn_op(node);
+
+ res = 0;
+ if (is_op_commutative(op)) {
+ wenv->changes |= res = move_consts_up(&node);
+ }
+ /* 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.wq = new_waitq();
- /* now we have collected enough information, optimize */
- irg_walk_graph(irg, NULL, 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);
+ }
+ set_reassoc_running(0);
/* Handle graph state */
if (env.changes) {
set_irg_outs_inconsistent(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. */
CASE(And);
CASE(Or);
CASE(Eor);
+ CASE(Shl);
default:
/* leave NULL */;
}