+/*
+ * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
+ *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
/**
*
* @file loop_unrolling.c
* Created: 16.11.2004
* CVS-ID: $Id$
* Copyright: (c) 2004 Universität Karlsruhe
- * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#ifdef HAVE_STRING_H
#include <string.h>
#endif
-#ifdef HAVE_MALLOC_H
-#include <malloc.h>
-#endif
-#ifdef HAVE_ALLOCA_H
-#include <alloca.h>
-#endif
-# include "loop_unrolling.h"
-
-# include "irgwalk.h"
-# include "ircons.h"
-# include "irgmod.h"
-# include "irloop_t.h"
-# include "irgopt_t.h"
-# include "irnode_t.h"
-# include "irouts.h"
-# include "hashptr.h"
-# include "pset.h"
-# include "strength_red.h"
+#include "loop_unrolling.h"
+
+#include "irnode_t.h"
+#include "irgwalk.h"
+#include "ircons.h"
+#include "irgmod.h"
+#include "irloop_t.h"
+#include "irgopt_t.h"
+#include "irouts.h"
+#include "trouts.h"
+#include "hashptr.h"
+#include "pset.h"
+#include "strength_red.h"
+#include "compute_loop_info.h"
+#include "irdump.h"
+#include "irtools.h"
+#include "xmalloc.h"
/* We will unroll maximal 4-times. */
#define MAX_UNROLL 4
return c1->irn != c2->irn;
}
-static INLINE int * new_backedge_arr(struct obstack *obst, int size)
-{
- int *res = NEW_ARR_D (int, obst, size);
- memset(res, 0, sizeof(int) * size);
- return res;
-}
-
-static INLINE void new_backedge_info(ir_node *n) {
- switch(get_irn_opcode(n)) {
- case iro_Block:
- n->attr.block.cg_backedge = NULL;
- n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
- break;
- case iro_Phi:
- n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
- break;
- case iro_Filter:
- n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
- break;
- default: ;
- }
-}
-
/**
* Remember the new node in the old node by using a field all nodes have.
*/
old->link = new;
}
-/**
- * Copies the node to the new obstack. The Ins of the new node point to
- * the predecessors on the old obstack. For block/phi nodes not all
- * predecessors might be copied. n->link points to the new node.
- * For Phi and Block nodes the function allocates in-arrays with an arity
- * only for useful predecessors. The arity is determined by counting
- * the non-bad predecessors of the block.
- *
- * @param n The node to be copied
- * @param env if non-NULL, the node number attribute will be copied to the new node
- */
-static void
-copy_node (ir_node *n, void *env)
-{
- ir_node *nn;
- int new_arity;
- ir_op *op = get_irn_op(n);
- int copy_node_nr = false;
-
- /* The end node looses it's flexible in array. This doesn't matter,
- as dead node elimination builds End by hand, inlineing doesn't use
- the End node. */
- /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
-
- if (op == op_Bad)
- /* node copied already */
- return;
-
- new_arity = get_irn_arity(n);
-
- nn = new_ir_node(get_irn_dbg_info(n),
- current_ir_graph,
- NULL, /* no block yet, will be set later */
- op,
- get_irn_mode(n),
- new_arity,
- get_irn_in(n));
-
-
- /* Copy the attributes. These might point to additional data. If this
- was allocated on the old obstack the pointers now are dangling. This
- frees e.g. the memory of the graph_arr allocated in new_immBlock. */
- copy_node_attr(n, nn);
- new_backedge_info(nn);
- set_new_node(n, nn);
-
-#if DEBUG_libfirm
- if (copy_node_nr) {
- /* for easier debugging, we want to copy the node numbers too */
- nn->node_nr = n->node_nr;
- }
-#endif
-
-}
/*Check if phi in the loop head is.
*
* @param phi Must be a phi node from the loop.
static void
set_preds (set *l_n, copies_t *value, induct_var_info *info, int unroll_factor, void *env)
{
+ ir_node *loop_head;
int i, p, irn_arity;
copies_t key, *value_pred;
- // The head of the unrolling loop.
- ir_node *loop_head = get_loop_node(info->l_itervar_phi, 0);
+ if(value->copy[0] == NULL ||
+ get_irn_op(value->irn) != get_irn_op(value->copy[0]))
+ return;
+ /* The head of the unrolling loop. */
+ loop_head = get_loop_node(info->l_itervar_phi, 0);
irn_arity = get_irn_arity(value->irn);
for (i = 0; i < irn_arity; i++) {
if (copy_head) {
/* The first copy of the loop head must point to the loop head.*/
value->copy[0] = new_Block(1, &backedge_jmp);
-
- for (i = 1; i < unroll_factor - 1; ++i) {
+ if(info->l_itervar_phi->flags & do_loop){
+ value_backedge_jmp->copy[0] = new_r_Jmp(current_ir_graph, value->copy[0]);
+ for (i = 1; i < unroll_factor - 1; ++i) {
/* all other copies must point to the copy before it in the array. */
- value->copy[i] = new_Block(1, &value_backedge_jmp->copy[i-1]);
- }
+ value->copy[i] = new_Block(1, &backedge_jmp);
+ value_backedge_jmp->copy[i] = new_r_Jmp(current_ir_graph, value->copy[i]);
+ }
+ for (i = 1; i < unroll_factor - 1; ++i){
+ set_nodes_block(value_backedge_jmp->copy[i-1], get_nodes_block(backedge_jmp));
+ set_Block_cfgpred (value->copy[i], 0, value_backedge_jmp->copy[i-1]);
+ }
+ }else
+ for (i = 1; i < unroll_factor - 1; ++i) {
+ /* all other copies must point to the copy before it in the array. */
+ set_nodes_block(value_backedge_jmp->copy[i-1], get_nodes_block(backedge_jmp));
+ value->copy[i] = new_Block(1, &value_backedge_jmp->copy[i-1]);
+ }
}
else {
/* If the loop head must not be copy. block.irn is the successor of the loop head.*/
{
int p;
- /* The first copy of Phi node get the node along the backedge as predecessor. The next copies
- the copies of this node.*/
- phi->copy[0] = phi_pred->irn;
- for (p = 1; p < unroll_factor - 1; ++p) {
- /* If two phi nodes are in cycle. */
- if (phi_pred->copy[p - 1] == NULL && get_irn_op(phi_pred->irn) == op_Phi) {
- if (p & 1)
- phi->copy[p] = phi->irn;
- else
- phi->copy[p] = phi_pred->irn;
- } else
- phi->copy[p] = phi_pred->copy[p - 1];
- }
+ if(phi_pred != NULL){
+ /* The first copy of Phi node get the node along the backedge as predecessor. The next copies
+ the copies of this node.*/
+ phi->copy[0] = phi_pred->irn;
+ for (p = 1; p < unroll_factor - 1; ++p) {
+ /* If two phi nodes are in cycle. */
+ if (phi_pred->copy[p - 1] == NULL && get_irn_op(phi_pred->irn) == op_Phi) {
+ if (p & 1)
+ phi->copy[p] = phi->irn;
+ else
+ phi->copy[p] = phi_pred->irn;
+ } else
+ phi->copy[p] = phi_pred->copy[p - 1];
+ }
+ }else
+ /* The predecessors of phi are loop invariant and the copies von phi
+ must be phi it self.*/
+ for(p = 0; p < unroll_factor - 1; ++p)
+ phi->copy[p] = phi->irn;
}
/**
static void
copy_loop_body(set *l_n, induct_var_info *info, int unroll_factor)
{
- int i;
+ int i, opt;
copies_t *value, *info_op, *phi, *loop_h = NULL, key, *value_block;
-
- ir_node *loop_head = get_loop_node(info->l_itervar_phi, 0);
+ ir_node *proj_b, *cond, *proj_1, *proj_0, *loop_head;
+ proj_b = get_irn_out(info->cmp, 0);
+ cond = get_irn_out(proj_b, 0);
+ proj_0 = get_irn_out(cond, 0);
+ proj_1 = get_irn_out(cond, 1);
+
+ loop_head = get_loop_node(info->l_itervar_phi, 0);
+ /* I will create some nodes and need the optimization to
+ be set off.*/
+ opt = get_optimize();
+ set_optimize(0);
for (value = set_first(l_n); value != NULL; value = set_next(l_n)) {
if(value->irn == loop_head)
else if (is_Phi_in_loop_head(value->irn, loop_head))
phi = value;
else if (copy_loop_head){
- /* The loop head must be copied. */
- for (i = 0; i < unroll_factor - 1; i++){
- copy_node(value->irn, NULL);
- value->copy[i] = get_irn_link(value->irn);
- }
+ if(value->irn == proj_0){
+ for (i = 0; i < unroll_factor - 1; i++)
+ /* In the copies of the loop head we replace the cmp branche with a jmp node.
+ This is just for "for" loops."do" loops are handled in function
+ new_loop_head().*/
+ value->copy[i] = new_r_Jmp(current_ir_graph, get_nodes_block(value->irn));
+ }else if(value->irn != proj_b && value->irn != cond &&
+ value->irn != proj_1)
+ /* The loop head must be copied. */
+ for (i = 0; i < unroll_factor - 1; i++){
+ copy_irn_to_irg(value->irn, current_ir_graph);
+ value->copy[i] = get_irn_link(value->irn);
+ }
} else {
/* The loop head and its nodes must not be copied. */
if((value->irn->op == op_Block &&
- value->irn != loop_head) ||
- (value->irn->op != op_Block &&
- get_nodes_block(value->irn) != loop_head)) {
- for (i = 0; i<unroll_factor -1; i++){
- copy_node(value->irn, NULL);
- value->copy[i] = get_irn_link(value->irn);
- }
+ value->irn != loop_head) ||
+ (value->irn->op != op_Block &&
+ get_nodes_block(value->irn) != loop_head)) {
+ for (i = 0; i<unroll_factor -1; i++) {
+ copy_irn_to_irg(value->irn, current_ir_graph);
+ value->copy[i] = get_irn_link(value->irn);
+ }
}
}
}
+
/* Treat the loop head block */
new_loop_head(l_n, info, loop_h, unroll_factor, copy_loop_head);
+ /* I have created the nodes and I set the optimization
+ to it old value.*/
+ set_optimize(opt);
/* Similarly treat the Phis in the loop head block. info->op is the node
along the backedges. */
}
for (value = set_first(l_n); value != NULL; value = set_next(l_n)) {
+
/* Set the predecessors of the copies. */
- if (copy_loop_head)
+ if (copy_loop_head){
set_preds(l_n, value, info, unroll_factor, NULL);
- else if ((value->irn->op != op_Block) && get_nodes_block(value->irn) != loop_head)
+ }else if ((value->irn->op != op_Block) && get_nodes_block(value->irn) != loop_head)
set_preds(l_n, value, info, unroll_factor, NULL);
- if (is_Phi_in_loop_head(value->irn, loop_head))
+ if (is_Phi_in_loop_head(value->irn, loop_head) &&
+ has_backedges(value->irn))
/* Set the backedge of phis in the loop head. */
set_phi_backedge(l_n, value, info, unroll_factor);
- if ((value->irn->op != op_Block) && !is_Phi_in_loop_head(value->irn, loop_head)) {
+ if ((value->irn->op != op_Block) && !is_Phi_in_loop_head(value->irn, loop_head) &&
+ value->copy[0] != NULL ) {
ir_node *nodes_block = get_nodes_block(value->irn);
if (!copy_loop_head && nodes_block == loop_head)
/* Set the copy of the node in the accordant copy of its block. */
for(i = 0; i < unroll_factor - 1; i++){
set_nodes_block(value->copy[i], value_block->copy[i]);
- //add_End_keepalive(get_irg_end(current_ir_graph), value->copy[p]);
}
}
}
+ /* I must repair some control flow edges, when the loop head
+ have been copied.*/
+ if (copy_loop_head && !(info->l_itervar_phi->flags & do_loop)){
+ key.irn = proj_0;
+ value = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
+ key.irn = get_irn_out(proj_0, 0);
+ info_op = set_find( l_n, &key, sizeof(key), HASH_PTR(key.irn));
+ for (i = 0; i < unroll_factor - 1; i++){
+ set_Block_cfgpred(info_op->copy[i], 0, value->copy[i]);
+ }
+ }
}
+
/**
* Returns non-zero if a Proj node from the loop has a predecessor that
* can throw an exception.
* after loop block must have as well all copies of this node as predecessors.
*
* @param l_n Contains all nodes of the loop.
- * @block block A block after the loop.
+ * @param block A block after the loop.
* @param loop_in A node from the loop, that is predecessor of the end block.
* @param unroll_factor An integer 2 <= unroll_factor <= 4.
*/
*
* @param l_n Contains all nodes of the loop.
* @param loop_outs Contains nodes after the loop,that have as predecessor a node from the loop.
- * @block node A node after the loop.
+ * @param node A node after the loop.
* @param loop_in A node (Proj) from the loop, that is predecessor of *node.
* @param unroll_factor An integer 2 <= unroll_factor <= 4.
*/
ir_node *loop_head = get_loop_node(info->l_itervar_phi, 0);
ir_node *end_block = get_irg_end_block(current_ir_graph);
- for (value = set_first(l_n); value != NULL; value = set_next(l_n))
+ for (value = set_first(l_n); value != NULL; value = set_next(l_n)){
+ if(get_irn_node_nr(value->irn) == 35047)
+ printf("\nHi\n");
if (! is_Phi_in_loop_head(value->irn, loop_head) &&
(get_irn_opcode(value->irn) == iro_Proj && value->copy[0] != NULL))
for (i = 0; i < get_irn_n_outs(value->irn); ++i) {
(key.irn->op != op_Block && get_Block_dom_depth(get_nodes_block(key.irn)) >
get_Block_dom_depth(loop_head))) {
- for (p = 0; p < get_irn_arity(key.irn); p++)
+ for (p = 0; p < get_irn_arity(key.irn); p++)
if (value->irn == get_irn_n(key.irn, p)) {
- if (get_irn_op(value->irn) == op_Proj && is_exception_possible(value->irn)) {
- if (set_find(loop_outs, &key, sizeof(key), HASH_PTR(key.irn)) == NULL) {
- /* If the loop out is for exceptions in the loop. */
- if ((get_irn_op(key.irn) == op_Phi && get_irn_mode(key.irn) == mode_M) ||
+ if (get_irn_op(value->irn) == op_Proj && is_exception_possible(value->irn)) {
+ if (set_find(loop_outs, &key, sizeof(key), HASH_PTR(key.irn)) == NULL) {
+ /* If the loop out is for exceptions in the loop. */
+ if ((get_irn_op(key.irn) == op_Phi && get_irn_mode(key.irn) == mode_M) ||
get_irn_op(key.irn) == op_Call)
- new_after_loop_node(l_n,loop_outs, key.irn, value, unroll_factor);
- else
- continue;
- } else
- continue;
- } else
- set_irn_n(key.irn, p, value->copy[unroll_factor-2]);
+ new_after_loop_node(l_n,loop_outs, key.irn, value, unroll_factor);
+ else
+ continue;
+ } else
+ continue;
+ } else
+ /* Loop outs in the loop head must be not changed.*/
+ if(get_nodes_block(value->irn) != loop_head)
+ set_irn_n(key.irn, p, value->copy[unroll_factor-2]);
}
}
}
+ }
/* This function searches for loop outs associated with function call in the unrolling loop. */
new_end_block (end_block, loop_head, l_n, loop_endblock_outs, unroll_factor);
}
* Called from a post walker.
*
* @param n An IR node.
- * @param env Free environment pointer.
+ * @param env points to a result
*/
static void do_loop_unroll(ir_node *n, void *env)
{
+ int *unroll_done = env;
induct_var_info info;
copies_t *value;
set *loop_nodes, *loop_outs, *loop_endblock_outs;
- ir_node *cmp_out, *phi_init, *loop_head;
+ ir_node *phi_init, *loop_head;
ir_node *backedge_jmp;
int l_sons = 0, unroll_factor = 0;
- int cmp_typ, init, iter_end, iter_increment, diff, iter_number;
+ tarval *init, *iter_end, *iter_increment,*tarval_null, *tarval_one, *tarval_three, *tarval_two, *diff, *iter_number;
int backedge_pos;
copy_loop_head = 0;
l_sons = get_loop_n_sons(info.l_itervar_phi);
if ( l_sons != 0)
return;
+ /* "do" loops are not implementit gut for this reason
+ at the time we don't unroll them.*/
+ if(info.l_itervar_phi->flags & do_loop)
+ return;
- cmp_out = get_irn_out(info.cmp, 0);
phi_init = get_Phi_pred(info.itervar_phi, info.init_pred_pos);
- if (! is_Proj(cmp_out)) return;
- if (get_irn_op(info.increment) != op_Const ||
- get_irn_op(phi_init) != op_Const ||
- get_irn_op(info.cmp_const) != op_Const) return;
+ if ((!(info.l_itervar_phi->flags & loop_is_count_loop)) ||
+ (info.l_itervar_phi->flags & loop_is_endless) ||
+ (info.l_itervar_phi->flags & loop_is_dead) ||
+ (info.l_itervar_phi->flags & once))
+ return;
- cmp_typ = get_Proj_proj(cmp_out);
- init = get_tarval_long(get_Const_tarval(phi_init));
- iter_end = get_tarval_long(get_Const_tarval(info.cmp_const));
- iter_increment = get_tarval_long(get_Const_tarval(info.increment));
+ init = info.l_itervar_phi->loop_iter_start;
+ iter_end = info.l_itervar_phi->loop_iter_end;
+ iter_increment = info.l_itervar_phi->loop_iter_increment;
+ tarval_null = get_mode_null(get_tarval_mode(iter_increment));
+ tarval_one = get_mode_one(get_tarval_mode(init));
- if (iter_end < init){
- int p = iter_end;
- iter_end = init;
- init = p;
- }
+ if(tarval_is_double(init) || tarval_is_double(iter_end) || tarval_is_double(iter_increment))
+ return;
+
+ if((tarval_is_negative(init) && tarval_is_negative(iter_end)) ||
+ (!tarval_is_negative(init) && !tarval_is_negative(iter_end)) ||
+ (tarval_is_null(init) || tarval_is_null(iter_end)))
+ diff = tarval_abs(tarval_sub(iter_end, init));
+ else
+ diff = tarval_add(tarval_abs(iter_end),tarval_abs(init));
- iter_increment = iter_increment < 0 ? -iter_increment : iter_increment;
- diff = iter_end - init;
+ iter_increment = tarval_abs(iter_increment);
- if (diff == 0 || iter_increment == 0) return;
+ if(!(info.l_itervar_phi->flags & do_loop))
+ diff = tarval_add(diff, tarval_one);
/* Test for the value of unroll factor. */
- iter_number = diff/iter_increment;
- if ((cmp_typ == pn_Cmp_Le || cmp_typ == pn_Cmp_Ge) && (iter_end % iter_increment == 0))
- iter_number ++;
+ if (tarval_cmp(tarval_mod(diff,iter_increment), tarval_null) == pn_Cmp_Eq)
+ iter_number = tarval_div(diff, iter_increment);
+ else
+ return;
+ tarval_two = tarval_add(tarval_one, tarval_one);
+ tarval_three = tarval_add(tarval_two,tarval_one);
- if ((iter_number & 3) == 0)
- unroll_factor = 4;
- else if ((iter_number % 3) == 0)
+ if (tarval_cmp(tarval_mod(iter_number, tarval_three), tarval_null) == pn_Cmp_Eq)
unroll_factor = 3;
- else if ((iter_number & 1) == 0)
+ else if (tarval_cmp(tarval_mod(iter_number, tarval_two), tarval_null) == pn_Cmp_Eq)
unroll_factor = 2;
else return;
if (get_firm_verbosity())
printf("\nloop unrolling with factor %d \n", unroll_factor);
- //DDMG(get_irn_irg(n));
+
+ /* ok, we will do unrolling */
+ *unroll_done += 1;
/* The unroll factor must be less than 4. */
assert(unroll_factor <= MAX_UNROLL);
/* A set containing all predecessors
of the end block from the unrolling loop. */
loop_endblock_outs = new_set(set_cmp, 8);
-
/* Find all nodes of the unrolling loop. */
find_loop_nodes(info.itervar_phi, get_loop_node(info.l_itervar_phi, 0), loop_nodes);
/* Set the backedge of the loop head. */
for (value = set_first(loop_nodes); value != NULL; value = set_next(loop_nodes)) {
- if(value->irn == backedge_jmp){
-
+ if (value->irn == backedge_jmp)
set_Block_cfgpred(loop_head, backedge_pos, value->copy[unroll_factor-2]);
- }
}
set_loop_outs(loop_nodes, loop_outs, loop_endblock_outs, &info, unroll_factor);
}
void optimize_loop_unrolling(ir_graph *irg /* unroll factor, max body size */)
{
ir_graph *rem;
+ int unroll_done = 0;
if ( !get_opt_loop_unrolling()) return;
/* -- Precompute some information -- */
+ /* Call algorithm that computes the loop information */
+ // construct_backedges(irg);
+ compute_loop_info(irg);
/* Call algorithm that computes the backedges */
- construct_cf_backedges(irg);
+ // construct_cf_backedges(irg);
/* Call algorithm that computes the dominator trees. */
- compute_doms(irg);
+ assure_doms(irg);
/* Call algorithm that computes the out edges */
- compute_outs(irg);
+ assure_irg_outs(irg);
collect_phiprojs(irg);
/* -- Search expressions that can be optimized -- */
- irg_walk_graph(irg, NULL, do_loop_unroll, NULL);
+ irg_walk_graph(irg, NULL, do_loop_unroll, &unroll_done);
+
+ if (unroll_done) {
+ /* unrolling was done, all info is invalid */
+ set_irg_doms_inconsistent(irg);
+ set_irg_outs_inconsistent(irg);
+ set_irg_loopinfo_state(current_ir_graph, loopinfo_cf_inconsistent);
+ set_trouts_inconsistent();
+ set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
+ }
current_ir_graph = rem;
}
projects / libfirm / blobdiff