X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fopt_osr.c;h=275ee2e6914dad9c7106cd2fd26a69754eadf9ff;hb=64f5f0cf11fe4a0d4753060e45639604d725652d;hp=d73974654f9e23afad7673bea3aadf436f509148;hpb=7e961375c8b2f276d3aa2ab908fdd0012008251a;p=libfirm diff --git a/ir/opt/opt_osr.c b/ir/opt/opt_osr.c index d73974654..275ee2e69 100644 --- a/ir/opt/opt_osr.c +++ b/ir/opt/opt_osr.c @@ -1,51 +1,67 @@ +/* + * Copyright (C) 1995-2011 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. + */ + /** - * Project: libFIRM - * File name: ir/opt/opt_osr. - * Purpose: Operator Strength Reduction, - * Keith D. Cooper, L. Taylor Simpson, Christopher A. Vick - * Author: Michael Beck - * Modified by: - * Created: 12.5.2006 - * CVS-ID: $Id$ - * Copyright: (c) 2006 Universität Karlsruhe - * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE. - */ -#ifdef HAVE_CONFIG_H + * @file + * @brief Operator Strength Reduction. + * @date 12.5.2006 + * @author Michael Beck + * @brief + * Implementation of the Operator Strength Reduction algorithm + * by Keith D. Cooper, L. Taylor Simpson, Christopher A. Vick. + * Extended version. + */ #include "config.h" -#endif - -#ifdef HAVE_MALLOC_H -#include -#endif -#ifdef HAVE_ALLOCA_H -#include -#endif -#include "opt_osr.h" +#include "adt/pdeq.h" +#include "iroptimize.h" #include "irgraph.h" +#include "ircons.h" #include "irop_t.h" -#include "irloop.h" #include "irdom.h" +#include "irgmod.h" #include "irflag_t.h" #include "irgwalk.h" +#include "irouts.h" +#include "iredges.h" #include "debug.h" #include "obst.h" +#include "set.h" +#include "tv.h" +#include "hashptr.h" +#include "util.h" #include "irtools.h" +#include "irloop_t.h" #include "array.h" +#include "firmstat.h" +#include "error.h" +#include "irpass_t.h" /** The debug handle. */ DEBUG_ONLY(static firm_dbg_module_t *dbg;) -/* Use the link field to mark IV's */ -#define MARK_LOOP_IV(loop) set_loop_link((loop), (loop)) -#define UNMARK_LOOP_IV(loop) set_loop_link((loop), NULL) -#define IS_LOOP_IV(loop) (get_loop_link(loop) != NULL) - /** A scc. */ typedef struct scc { - ir_node *head; /**< the head of the list */ - ir_node *header; /**< the header block of this scc */ - int is_iv; /**< true, if this scc is an IV */ + ir_node *head; /**< the head of the list */ + ir_tarval *init; /**< the init value iff only one exists. */ + ir_tarval *incr; /**< the induction variable increment if only a single const exists. */ + unsigned code; /**< == iro_Add if +incr, iro_Sub if -incr, 0 if not analysed, iro_Bad else */ } scc; /** A node entry */ @@ -61,54 +77,609 @@ typedef struct node_entry { /** The environment. */ typedef struct iv_env { - node_entry *entries; /**< the node entry array */ - struct obstack obst; /**< an obstack for allocations */ - ir_node **stack; /**< the node stack */ - int tos; /**< tos index */ - unsigned nextDFSnum; /**< the current DFS number */ - unsigned replaced; /**< number of replaced ops */ - unsigned POnum; /**< current post order number */ + struct obstack obst; /**< an obstack for allocations */ + ir_node **stack; /**< the node stack */ + size_t tos; /**< tos index */ + unsigned nextDFSnum; /**< the current DFS number */ + unsigned POnum; /**< current post order number */ + set *quad_map; /**< a map from (op, iv, rc) to node */ + set *lftr_edges; /**< the set of lftr edges */ + unsigned replaced; /**< number of replaced ops */ + unsigned lftr_replaced; /**< number of applied linear function test replacements */ + unsigned osr_flags; /**< additional flags steering the transformation */ + unsigned need_postpass; /**< set, if a post pass is needed to fix Add and Sub nodes */ + /** Function called to process a SCC. */ + void (*process_scc)(scc *pscc, struct iv_env *env); } iv_env; /** - * Check if irn is an IV. + * An entry in the (op, node, node) -> node map. */ -static scc *is_iv(ir_node *irn, iv_env *env) { - int idx = get_irn_idx(irn); - node_entry *e = &env->entries[idx]; +typedef struct quadruple_t { + unsigned code; /**< the opcode of the reduced operation */ + ir_node *op1; /**< the first operand the reduced operation */ + ir_node *op2; /**< the second operand of the reduced operation */ - return e->pscc && e->pscc->is_iv ? e->pscc : NULL; -} + ir_node *res; /**< the reduced operation */ +} quadruple_t; + +/** + * A LFTR edge. + */ +typedef struct LFTR_edge { + ir_node *src; /**< the source node */ + ir_node *dst; /**< the destination node */ + unsigned code; /**< the opcode that must be applied */ + ir_node *rc; /**< the region const that must be applied */ +} LFTR_edge; + +/* forward */ +static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env); + +/** + * Compare two LFTR edges. + */ +static int LFTR_cmp(const void *e1, const void *e2, size_t size) +{ + const LFTR_edge *l1 = (const LFTR_edge*)e1; + const LFTR_edge *l2 = (const LFTR_edge*)e2; + (void) size; + + return l1->src != l2->src; +} /* LFTR_cmp */ + +/** + * Find a LFTR edge. + * + * @param src the source node of the transition + */ +static LFTR_edge *LFTR_find(ir_node *src, iv_env *env) +{ + LFTR_edge key; + + key.src = src; + + return set_find(LFTR_edge, env->lftr_edges, &key, sizeof(key), hash_ptr(src)); +} /* LFTR_find */ + +/** + * Add a LFTR edge. + * + * @param src the source node of the edge + * @param dst the destination node of the edge + * @param code the opcode of the transformed transition + * @param rc the region const used in the transition + * @param env the environment + */ +static void LFTR_add(ir_node *src, ir_node *dst, unsigned code, ir_node *rc, iv_env *env) +{ + LFTR_edge key; + + key.src = src; + key.dst = dst; + key.code = code; + key.rc = rc; + + /* + * There might be more than one edge here. This is rather bad + * because we currently store only one. + */ + (void)set_insert(LFTR_edge, env->lftr_edges, &key, sizeof(key), hash_ptr(src)); +} /* LFTR_add */ + +/** + * Gets the node_entry of a node. + * + * @param irn the node + * @param env the environment + */ +static node_entry *get_irn_ne(ir_node *irn, iv_env *env) +{ + node_entry *e = (node_entry*)get_irn_link(irn); + + if (e == NULL) { + e = OALLOCZ(&env->obst, node_entry); + set_irn_link(irn, e); + } + return e; +} /* get_irn_ne */ + +/** + * Gets the scc from an induction variable. + * + * @param iv any node of the induction variable + * @param env the environment + */ +static scc *get_iv_scc(ir_node *iv, iv_env *env) +{ + node_entry *e = get_irn_ne(iv, env); + return e->pscc; +} /* get_iv_scc */ + +/** + * Check if irn is an IV. + * + * @param irn the node to check + * @param env the environment + * + * @returns the header if it is one, NULL else + */ +static ir_node *is_iv(ir_node *irn, iv_env *env) +{ + return get_irn_ne(irn, env)->header; +} /* is_iv */ /** * Check if irn is a region constant. + * The block or irn must strictly dominate the header block. + * + * @param irn the node to check + * @param header_block the header block of the induction variable */ -static int is_rc(ir_node *irn, ir_node *header_block) { +static int is_rc(ir_node *irn, ir_node *header_block) +{ ir_node *block = get_nodes_block(irn); - return block_dominates(block, header_block); -} + return (block != header_block) && block_dominates(block, header_block); +} /* is_rc */ + +/** + * Set compare function for the quad set. + */ +static int quad_cmp(const void *e1, const void *e2, size_t size) +{ + const quadruple_t *c1 = (const quadruple_t*)e1; + const quadruple_t *c2 = (const quadruple_t*)e2; + (void) size; + + return c1->code != c2->code || c1->op1 != c2->op1 || c1->op2 != c2->op2; +} /* quad_cmp */ + +/** + * Check if an reduced operation was already calculated. + * + * @param code the opcode of the operation + * @param op1 the first operand of the operation + * @param op2 the second operand of the operation + * @param env the environment + * + * @return the already reduced node or NULL if this operation is not yet reduced + */ +static ir_node *search(unsigned code, ir_node *op1, ir_node *op2, iv_env *env) +{ + quadruple_t key, *entry; + + key.code = code; + key.op1 = op1; + key.op2 = op2; + + entry = set_find(quadruple_t, env->quad_map, &key, sizeof(key), (code * 9) ^ hash_ptr(op1) ^ hash_ptr(op2)); + if (entry) + return entry->res; + return NULL; +} /* search */ /** - * Do the replacement operation. + * Add an reduced operation. + * + * @param code the opcode of the operation + * @param op1 the first operand of the operation + * @param op2 the second operand of the operation + * @param result the result of the reduced operation + * @param env the environment + */ +static void add(unsigned code, ir_node *op1, ir_node *op2, ir_node *result, iv_env *env) +{ + quadruple_t key; + + key.code = code; + key.op1 = op1; + key.op2 = op2; + key.res = result; + + (void)set_insert(quadruple_t, env->quad_map, &key, sizeof(key), (code * 9) ^ hash_ptr(op1) ^ hash_ptr(op2)); +} /* add */ + +/** + * Find a location where to place a bin-op whose operands are in + * block1 and block2. + * + * @param block1 the block of the first operand + * @param block2 the block of the second operand + * + * Note that we know here that such a place must exists. Moreover, this means + * that either block1 dominates block2 or vice versa. So, just return + * the "smaller" one. + */ +static ir_node *find_location(ir_node *block1, ir_node *block2) +{ + if (block_dominates(block1, block2)) + return block2; + assert(block_dominates(block2, block1)); + return block1; +} /* find_location */ + +/** + * Create a node that executes an op1 code op1 operation. + * + * @param code the opcode to execute + * @param db debug info to add to the new node + * @param op1 the first operand + * @param op2 the second operand + * @param mode the mode of the new operation + * + * @return the newly created node + */ +static ir_node *do_apply(unsigned code, dbg_info *db, ir_node *op1, ir_node *op2, ir_mode *mode) +{ + ir_node *result; + ir_node *block = find_location(get_nodes_block(op1), get_nodes_block(op2)); + + switch (code) { + case iro_Mul: + result = new_rd_Mul(db, block, op1, op2, mode); + break; + case iro_Add: + result = new_rd_Add(db, block, op1, op2, mode); + break; + case iro_Sub: + result = new_rd_Sub(db, block, op1, op2, mode); + break; + default: + panic("Unsupported opcode"); + } + return result; +} /* do_apply */ + +/** + * The Apply operation. + * + * @param orig the node that represent the original operation and determines + * the opcode, debug-info and mode of a newly created one + * @param op1 the first operand + * @param op2 the second operand + * @param env the environment + * + * @return the newly created node + */ +static ir_node *apply(ir_node *header, ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env) +{ + unsigned code = get_irn_opcode(orig); + ir_node *result = search(code, op1, op2, env); + + if (result == NULL) { + dbg_info *db = get_irn_dbg_info(orig); + ir_node *op1_header = get_irn_ne(op1, env)->header; + ir_node *op2_header = get_irn_ne(op2, env)->header; + + if (op1_header == header && is_rc(op2, op1_header)) { + result = reduce(orig, op1, op2, env); + } + else if (op2_header == header && is_rc(op1, op2_header)) { + result = reduce(orig, op2, op1, env); + } + else { + result = do_apply(code, db, op1, op2, get_irn_mode(orig)); + get_irn_ne(result, env)->header = NULL; + } + } + return result; +} /* apply */ + +/** + * The Reduce operation. + * + * @param orig the node that represent the original operation and determines + * the opcode, debug-info and mode of a newly created one + * @param iv the induction variable + * @param rc the region constant + * @param env the environment + * + * @return the reduced node + */ +static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) +{ + unsigned code = get_irn_opcode(orig); + ir_node *result = search(code, iv, rc, env); + + /* check if we have already done this operation on the iv */ + if (result == NULL) { + node_entry *e, *iv_e; + int i; + ir_mode *mode = get_irn_mode(orig); + + result = exact_copy(iv); + + if (get_irn_mode(result) != mode) { + /* + * Beware: we must always create a new induction variable with the same mode + * as the node we are replacing. Especially this means the mode might be changed + * from P to I and back. This is always possible, because we have only Phi, Add + * and Sub nodes. + * However, this might lead to AddIs(Iu,Is) which we must fix. The best way to do this + * seems to be a post-pass, or we might end with useless Conv's. + */ + set_irn_mode(result, mode); + env->need_postpass = 1; + } + add(code, iv, rc, result, env); + DB((dbg, LEVEL_3, " Created new %+F for %+F (%s %+F)\n", result, iv, + get_irn_opname(orig), rc)); + + iv_e = get_irn_ne(iv, env); + e = get_irn_ne(result, env); + e->header = iv_e->header; + + /* create the LFTR edge */ + LFTR_add(iv, result, code, rc, env); + + for (i = get_irn_arity(result) - 1; i >= 0; --i) { + ir_node *o = get_irn_n(result, i); + + e = get_irn_ne(o, env); + if (e->header == iv_e->header) + o = reduce(orig, o, rc, env); + else if (is_Phi(result) || code == iro_Mul) + o = apply(iv_e->header, orig, o, rc, env); + set_irn_n(result, i, o); + } + } else { + DB((dbg, LEVEL_3, " Already Created %+F for %+F (%s %+F)\n", result, iv, + get_irn_opname(orig), rc)); + } + return result; +} /* reduce */ + +/** + * Update the scc for a newly created IV. + */ +static void update_scc(ir_node *iv, node_entry *e, iv_env *env) +{ + scc *pscc = e->pscc; + ir_node *header = e->header; + waitq *wq = new_waitq(); + + DB((dbg, LEVEL_2, " Creating SCC for new an induction variable:\n ")); + pscc->head = NULL; + waitq_put(wq, iv); + do { + ir_node *irn = (ir_node*)waitq_get(wq); + node_entry *ne = get_irn_ne(irn, env); + int i; + + ne->pscc = pscc; + ne->next = pscc->head; + pscc->head = irn; + DB((dbg, LEVEL_2, " %+F,", irn)); + + for (i = get_irn_arity(irn) - 1; i >= 0; --i) { + ir_node *pred = get_irn_n(irn, i); + node_entry *pe = get_irn_ne(pred, env); + + if (pe->header == header && pe->pscc == NULL) { + /* set the pscc here to ensure that the node is NOT enqueued another time */ + pe->pscc = pscc; + waitq_put(wq, pred); + } + } + } while (! waitq_empty(wq)); + del_waitq(wq); + DB((dbg, LEVEL_2, "\n")); +} /* update_scc */ + +/** + * The Replace operation. We found a node representing iv (+,-,*) rc + * that can be removed by replacing the induction variable iv by a new + * one that 'applies' the operation 'irn'. * * @param irn the node that will be replaced * @param iv the induction variable * @param rc the region constant * @param env the environment */ -static void replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env) { +static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env) +{ + ir_node *result; + DB((dbg, LEVEL_2, " Replacing %+F\n", irn)); -} + result = reduce(irn, iv, rc, env); + if (result != irn) { + node_entry *e; + + hook_strength_red(get_irn_irg(irn), irn); + exchange(irn, result); + e = get_irn_ne(result, env); + if (e->pscc == NULL) { + e->pscc = OALLOCZ(&env->obst, scc); + update_scc(result, e, env); + } + ++env->replaced; + return 1; + } + return 0; +} /* replace */ + +#if 0 /** - * check if a node can be replaced + * check if a given node is a mul with 2, 4, 8 */ -static void check_replace(ir_node *irn, iv_env *env) { - ir_node *left, *right, *iv, *rc; - ir_op *op = get_irn_op(irn); - opcode code = get_op_code(op); - scc *liv, *riv; +static int is_x86_shift_const(ir_node *mul) +{ + ir_node *rc; + + if (! is_Mul(mul)) + return 0; + + /* normalization put constants on the right side */ + rc = get_Mul_right(mul); + if (is_Const(rc)) { + ir_tarval *tv = get_Const_tarval(rc); + + if (tarval_is_long(tv)) { + long value = get_tarval_long(tv); + + if (value == 2 || value == 4 || value == 8) { + /* do not reduce multiplications by 2, 4, 8 */ + return 1; + } + } + } + return 0; +} /* is_x86_shift_const */ +#endif + +/** + * Check if an IV represents a counter with constant limits. + * + * @param iv any node of the induction variable + * @param env the environment + */ +static int is_counter_iv(ir_node *iv, iv_env *env) +{ + node_entry *e = get_irn_ne(iv, env); + scc *pscc = e->pscc; + ir_node *have_init = NULL; + ir_node *have_incr = NULL; + ir_opcode code = iro_Bad; + ir_node *irn; + + if (pscc->code != 0) { + /* already analysed */ + return pscc->code != iro_Bad; + } + + pscc->code = iro_Bad; + for (irn = pscc->head; irn != NULL; irn = e->next) { + if (is_Add(irn)) { + if (have_incr != NULL) + return 0; + + have_incr = get_Add_right(irn); + if (! is_Const(have_incr)) { + have_incr = get_Add_left(irn); + if (! is_Const(have_incr)) + return 0; + } + code = iro_Add; + } else if (is_Sub(irn)) { + if (have_incr != NULL) + return 0; + + have_incr = get_Sub_right(irn); + if (! is_Const(have_incr)) + return 0; + code = iro_Sub; + } else if (is_Phi(irn)) { + int i; + + for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) { + ir_node *pred = get_Phi_pred(irn, i); + node_entry *ne = get_irn_ne(pred, env); + + if (ne->header == e->header) + continue; + if (have_init != NULL) + return 0; + have_init = pred; + if (! is_Const(pred)) + return 0; + } + } else + return 0; + e = get_irn_ne(irn, env); + } + pscc->init = get_Const_tarval(have_init); + pscc->incr = get_Const_tarval(have_incr); + pscc->code = code; + return code != iro_Bad; +} /* is_counter_iv */ + +/** + * Check the users of an induction variable for register pressure. + * + * @param iv any node of the induction variable + * @param env the environment + * + * @return non-zero if the register pressure is estimated + * to not increase, zero else + */ +static int check_users_for_reg_pressure(ir_node *iv, iv_env *env) +{ + ir_node *irn; + ir_node *have_user = NULL; + ir_node *have_cmp = NULL; + node_entry *e = get_irn_ne(iv, env); + scc *pscc = e->pscc; + + for (irn = pscc->head; irn != NULL; irn = e->next) { + foreach_out_edge(irn, edge) { + ir_node *user = get_edge_src_irn(edge); + node_entry *ne = get_irn_ne(user, env); + + if (e->header == ne->header) { + /* found user from the same IV */ + continue; + } + if (is_Cmp(user)) { + if (have_cmp != NULL) { + /* more than one cmp, for now end here */ + return 0; + } + have_cmp = user; + } else { + /* user is a real user of the IV */ + if (have_user != NULL) { + /* found the second user */ + return 0; + } + have_user = user; + } + } + e = get_irn_ne(irn, env); + } + + if (have_user == NULL) { + /* no user, ignore */ + return 1; + } + + if (have_cmp == NULL) { + /* fine, only one user, try to reduce */ + return 1; + } + /* + * We found one user AND at least one cmp. + * We should check here if we can transform the Cmp. + * + * For now our capabilities for doing linear function test + * are limited, so check if the iv has the right form: Only ONE + * Phi, only one Add/Sub with a Const. + */ + if (! is_counter_iv(iv, env)) + return 0; + + /* + * Ok, we have only one increment AND it is a Const, we might be able + * to do a linear function test replacement, so go on. + */ + return 1; +} /* check_users_for_reg_pressure */ + +/** + * Check if a node can be replaced (+, -, *). + * + * @param irn the node to check + * @param env the environment + * + * @return non-zero if irn should be Replace'd + */ +static int check_replace(ir_node *irn, iv_env *env) +{ + ir_node *left, *right, *iv, *rc; + ir_op *op = get_irn_op(irn); + unsigned code = get_op_code(op); + ir_node *liv, *riv; switch (code) { case iro_Mul: @@ -121,37 +692,48 @@ static void check_replace(ir_node *irn, iv_env *env) { liv = is_iv(left, env); riv = is_iv(right, env); - if (liv && is_rc(right, liv->header)) { + if (liv && is_rc(right, liv)) { iv = left; rc = right; } - else if (is_op_commutative(op) && - riv && is_rc(left, riv->header)) { + else if (riv && is_op_commutative(op) && + is_rc(left, riv)) { iv = right; rc = left; } if (iv) { - replace(irn, iv, rc, env); - ++env->replaced; + if (env->osr_flags & osr_flag_keep_reg_pressure) { + if (! check_users_for_reg_pressure(iv, env)) + return 0; + } + return replace(irn, iv, rc, env); } break; + default: + break; } -} + return 0; +} /* check_replace */ /** - * check which SCC's are induction variables + * Check which SCC's are induction variables. + * + * @param pscc a SCC + * @param env the environment */ -static void classify_iv(scc *pscc, iv_env *env) { +static void classify_iv(scc *pscc, iv_env *env) +{ ir_node *irn, *next, *header = NULL; - node_entry *h, *b; - int j; + node_entry *b, *h = NULL; + int j, only_phi, num_outside; + ir_node *out_rc; /* find the header block for this scc */ for (irn = pscc->head; irn; irn = next) { - node_entry *e = get_irn_link(irn); + node_entry *e = (node_entry*)get_irn_link(irn); ir_node *block = get_nodes_block(irn); next = e->next; - b = &env->entries[get_irn_idx(block)]; + b = get_irn_ne(block, env); if (header) { if (h->POnum < b->POnum) { @@ -164,19 +746,42 @@ static void classify_iv(scc *pscc, iv_env *env) { h = b; } } - pscc->header = header; + /* check if this scc contains only Phi, Add or Sub nodes */ + only_phi = 1; + num_outside = 0; + out_rc = NULL; for (irn = pscc->head; irn; irn = next) { - node_entry *e = get_irn_link(irn); + node_entry *e = get_irn_ne(irn, env); next = e->next; switch (get_irn_opcode(irn)) { - case iro_Add: case iro_Sub: + only_phi = 0; + { + ir_node *left = get_Sub_left(irn); + node_entry *left_entry = get_irn_ne(left, env); + ir_node *right = get_Sub_right(irn); + node_entry *right_entry = get_irn_ne(right, env); + + if (left_entry->pscc != e->pscc || + (right_entry->pscc != e->pscc && !is_rc(right, header))) { + /* + * Not an induction variable. + * Region constant are only allowed on right hand side. + */ + goto fail; + } + } + break; + + case iro_Add: + only_phi = 0; + /* fall through */ case iro_Phi: for (j = get_irn_arity(irn) - 1; j >= 0; --j) { - ir_node *pred = get_irn_n(irn, j); - node_entry *pe = &env->entries[get_irn_idx(pred)]; + ir_node *pred = get_irn_n(irn, j); + node_entry *pe = get_irn_ne(pred, env); if (pe->pscc != e->pscc) { /* not in the same SCC, must be a region const */ @@ -184,6 +789,12 @@ static void classify_iv(scc *pscc, iv_env *env) { /* not an induction variable */ goto fail; } + if (! out_rc) { + out_rc = pred; + ++num_outside; + } else if (out_rc != pred) { + ++num_outside; + } } } break; @@ -193,128 +804,199 @@ static void classify_iv(scc *pscc, iv_env *env) { } } /* found an induction variable */ - DB((dbg, LEVEL_2, " Found an induction variable in %+F\n", pscc->head)); - pscc->is_iv = 1; + DB((dbg, LEVEL_2, " Found an induction variable:\n ")); + if (only_phi && num_outside == 1) { + /* a phi cycle with only one real predecessor can be collapsed */ + DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n ")); + + for (irn = pscc->head; irn; irn = next) { + node_entry *e = get_irn_ne(irn, env); + next = e->next; + e->header = NULL; + exchange(irn, out_rc); + } + ++env->replaced; + return; + } + + /* set the header for every node in this scc */ + for (irn = pscc->head; irn; irn = next) { + node_entry *e = get_irn_ne(irn, env); + e->header = header; + next = e->next; + DB((dbg, LEVEL_2, " %+F,", irn)); + } + DB((dbg, LEVEL_2, "\n")); return; fail: for (irn = pscc->head; irn; irn = next) { - node_entry *e = get_irn_link(irn); + node_entry *e = get_irn_ne(irn, env); next = e->next; - check_replace(irn, env); + e->header = NULL; } -} +} /* classify_iv */ /** - * Do the replacement: We must do this as an additional step because - * of our loop-tree structure. + * Process an SCC for the operator strength reduction. + * + * @param pscc the SCC + * @param env the environment */ -static void find_replacement(ir_loop *loop, iv_env *env) { - int i, n; +static void process_scc(scc *pscc, iv_env *env) +{ + ir_node *head = pscc->head; + node_entry *e = (node_entry*)get_irn_link(head); - if (! IS_LOOP_IV(loop)) { - /* do replacements */ - n = get_loop_n_nodes(loop); - for (i = 0; i < n; ++i) { - ir_node *irn = get_loop_node(loop, i); - ir_node *left, *right, *iv, *rc; - opcode code = get_irn_opcode(irn); +#ifdef DEBUG_libfirm + { + ir_node *irn, *next; - switch (code) { - case iro_Mul: - case iro_Add: - case iro_Sub: - iv = rc = NULL; + DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc)); + for (irn = pscc->head; irn != NULL; irn = next) { + node_entry *e = (node_entry*)get_irn_link(irn); - left = get_binop_left(irn); - right = get_binop_right(irn); + next = e->next; - if (is_iv(left, env) && is_rc(right, left)) { - iv = left; rc = right; - } - else if (code != iro_Sub && - is_iv(right, env) && is_rc(left, right)) { - iv = right; rc = left; - } + DB((dbg, LEVEL_4, " %+F,", irn)); + } + DB((dbg, LEVEL_4, "\n")); + } +#endif + + if (e->next == NULL) { + /* this SCC has only a single member */ + check_replace(head, env); + } else { + classify_iv(pscc, env); + } +} /* process_scc */ + +/** + * If an SCC is a Phi only cycle, remove it. + * + * @param pscc an SCC that consists of Phi nodes only + * @param env the environment + */ +static void remove_phi_cycle(scc *pscc, iv_env *env) +{ + ir_node *irn, *next; + int j; + ir_node *out_rc; + + /* check if this scc contains only Phi nodes */ + out_rc = NULL; + for (irn = pscc->head; irn; irn = next) { + node_entry *e = get_irn_ne(irn, env); - if (iv) { - replace(irn, iv, rc, env); - ++env->replaced; + next = e->next; + if (! is_Phi(irn)) + return; + + for (j = get_irn_arity(irn) - 1; j >= 0; --j) { + ir_node *pred = get_irn_n(irn, j); + node_entry *pe = get_irn_ne(pred, env); + + if (pe->pscc != e->pscc) { + /* not in the same SCC, must be the only input */ + if (! out_rc) { + out_rc = pred; + } else if (out_rc != pred) { + return; } - break; } } } - n = get_loop_n_sons(loop); - for (i = 0; i < n; ++i) { - ir_loop *child = get_loop_son(loop, i); - find_replacement(child, env); + /* found a Phi cycle */ + DB((dbg, LEVEL_2, " Found an USELESS Phi cycle:\n ")); + + for (irn = pscc->head; irn; irn = next) { + node_entry *e = get_irn_ne(irn, env); + next = e->next; + e->header = NULL; + exchange(irn, out_rc); } -} + ++env->replaced; +} /* remove_phi_cycle */ /** - * Process a SCC given as a list. + * Process a SCC for the Phi cycle remove. + * + * @param pscc the SCC + * @param env the environment */ -static void process_scc(scc *pscc, iv_env *env) { +static void process_phi_only_scc(scc *pscc, iv_env *env) +{ ir_node *head = pscc->head; - node_entry *e = get_irn_link(head); - - pscc->is_iv = 0; + node_entry *e = (node_entry*)get_irn_link(head); #ifdef DEBUG_libfirm { ir_node *irn, *next; - DB((dbg, LEVEL_3, " SCC at %p:\n ", pscc)); + DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc)); for (irn = pscc->head; irn; irn = next) { - node_entry *e = get_irn_link(irn); + node_entry *e = (node_entry*)get_irn_link(irn); next = e->next; - DB((dbg, LEVEL_3, " %+F,", irn)); + DB((dbg, LEVEL_4, " %+F,", irn)); } - DB((dbg, LEVEL_3, "\n")); + DB((dbg, LEVEL_4, "\n")); } #endif - if (e->next == NULL) { - /* this SCC has only a single member */ - check_replace(head, env); - } - else { - classify_iv(pscc, env); - } -} + if (e->next != NULL) + remove_phi_cycle(pscc, env); +} /* process_phi_only_scc */ + /** * Push a node onto the stack. + * + * @param env the environment + * @param n the node to push */ -static void push(iv_env *env, ir_node *n) { +static void push(iv_env *env, ir_node *n) +{ + node_entry *e; + if (env->tos == ARR_LEN(env->stack)) { - int nlen = ARR_LEN(env->stack) * 2; + size_t nlen = ARR_LEN(env->stack) * 2; ARR_RESIZE(ir_node *, env->stack, nlen); } env->stack[env->tos++] = n; - env->entries[get_irn_idx(n)].in_stack = 1; -} + e = get_irn_ne(n, env); + e->in_stack = 1; +} /* push */ /** - * pop a node from the stack + * Pop a node from the stack. + * + * @param env the environment * * @return The topmost node */ static ir_node *pop(iv_env *env) { - ir_node *n = env->stack[--env->tos]; - env->entries[get_irn_idx(n)].in_stack = 0; - return n; -} + ir_node *n = env->stack[--env->tos]; + node_entry *e = get_irn_ne(n, env); + + e->in_stack = 0; + return n; +} /* pop */ +/** + * Do Tarjan's SCC algorithm and drive OSR. + * + * @param irn start at this node + * @param env the environment + */ static void dfs(ir_node *irn, iv_env *env) { int i, n; - node_entry *node = &env->entries[get_irn_idx(irn)]; + node_entry *node = get_irn_ne(irn, env); mark_irn_visited(irn); @@ -324,11 +1006,10 @@ static void dfs(ir_node *irn, iv_env *env) for (i = 0; i < n; ++i) { ir_node *pred = get_irn_n(irn, i); - if (irn_not_visited(pred)) + if (!irn_visited(pred)) dfs(pred, env); } - } - else { + } else { ir_node *block = get_nodes_block(irn); node->DFSnum = env->nextDFSnum++; @@ -336,15 +1017,15 @@ static void dfs(ir_node *irn, iv_env *env) push(env, irn); /* handle the block */ - if (irn_not_visited(block)) + if (!irn_visited(block)) dfs(block, env); n = get_irn_arity(irn); for (i = 0; i < n; ++i) { ir_node *pred = get_irn_n(irn, i); - node_entry *o = &env->entries[get_irn_idx(pred)]; + node_entry *o = get_irn_ne(pred, env); - if (irn_not_visited(pred)) { + if (!irn_visited(pred)) { dfs(pred, env); node->low = MIN(node->low, o->low); } @@ -352,100 +1033,465 @@ static void dfs(ir_node *irn, iv_env *env) node->low = MIN(o->DFSnum, node->low); } if (node->low == node->DFSnum) { - scc *pscc = obstack_alloc(&env->obst, sizeof(*pscc)); + scc *pscc = OALLOCZ(&env->obst, scc); ir_node *x; - pscc->head = NULL; do { node_entry *e; x = pop(env); - e = &env->entries[get_irn_idx(x)]; + e = get_irn_ne(x, env); e->pscc = pscc; e->next = pscc->head; pscc->head = x; - - /* link the node entry for easier access */ - set_irn_link(x, e); } while (x != irn); - process_scc(pscc, env); + env->process_scc(pscc, env); } } -} +} /* dfs */ /** - * Do the DFS by starting end the End node + * Do the DFS by starting at the End node of a graph. + * + * @param irg the graph to process + * @param env the environment */ -static void do_dfs(ir_graph *irg, iv_env *env) { - ir_graph *rem = current_ir_graph; - ir_node *end = get_irg_end(irg); - int i, n; +static void do_dfs(ir_graph *irg, iv_env *env) +{ + ir_node *end = get_irg_end(irg); + int i; + + ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); - current_ir_graph = irg; inc_irg_visited(irg); /* visit all visible nodes */ dfs(end, env); /* visit the keep-alives */ - n = get_End_n_keepalives(end); - for (i = 0; i < n; ++i) { + for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) { ir_node *ka = get_End_keepalive(end, i); - if (irn_not_visited(ka)) + if (!irn_visited(ka)) dfs(ka, env); } - current_ir_graph = rem; -} + + ir_free_resources(irg, IR_RESOURCE_IRN_VISITED); +} /* do_dfs */ /** * Post-block-walker: assign the post-order number. */ -static void assign_po(ir_node *block, void *ctx) { - iv_env *env = ctx; - int idx = get_irn_idx(block); +static void assign_po(ir_node *block, void *ctx) +{ + iv_env *env = (iv_env*)ctx; + node_entry *e = get_irn_ne(block, env); - env->entries[idx].POnum = env->POnum++; -} + e->POnum = env->POnum++; +} /* assign_po */ -/* Performs Operator Strength Reduction for the passed graph. */ -void opt_osr(ir_graph *irg) { - iv_env env; +/** + * Apply one LFTR edge operation. + * Return NULL if the transformation cannot be done safely without + * an Overflow. + * + * @param iv the induction variable + * @param rc the constant that should be translated + * @param e the LFTR edge + * @param env the IV environment + * + * @return the translated region constant or NULL + * if the translation was not possible + * + * @note + * In the current implementation only the last edge is stored, so + * only one chain exists. That's why we might miss some opportunities. + */ +static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env) +{ + if (env->osr_flags & osr_flag_lftr_with_ov_check) { + ir_tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr, *tv_end; + tarval_int_overflow_mode_t ovmode; + scc *pscc; + ir_graph *irg; + + if (! is_counter_iv(iv, env)) { + DB((dbg, LEVEL_4, " not counter IV")); + return NULL; + } + + /* overflow can only be decided for Consts */ + if (! is_Const(e->rc)) { + if (e->code == iro_Add && mode_is_reference(get_irn_mode(e->rc))) { + /* However we allow ONE Pointer Add, as pointer arithmetic with wrap + around is undefined anyway */ + return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->rc)); + } + DB((dbg, LEVEL_4, " = UNKNOWN (%+F)", e->rc)); + return NULL; + } + + tv_l = get_Const_tarval(rc); + tv_r = get_Const_tarval(e->rc); + + ovmode = tarval_get_integer_overflow_mode(); + tarval_set_integer_overflow_mode(TV_OVERFLOW_BAD); + + pscc = get_iv_scc(iv, env); + tv_incr = pscc->incr; + tv_init = pscc->init; + + /* + * Check that no overflow occurs: + * init must be transformed without overflow + * the new rc must be transformed without overflow + * rc +/- incr must be possible without overflow + */ + switch (e->code) { + case iro_Mul: + tv = tarval_mul(tv_l, tv_r); + tv_init = tarval_mul(tv_init, tv_r); + tv_incr = tarval_mul(tv_incr, tv_r); + DB((dbg, LEVEL_4, " * %+F", tv_r)); + break; + case iro_Add: + tv = tarval_add(tv_l, tv_r); + tv_init = tarval_add(tv_init, tv_r); + DB((dbg, LEVEL_4, " + %+F", tv_r)); + break; + case iro_Sub: + tv = tarval_sub(tv_l, tv_r, NULL); + tv_init = tarval_sub(tv_init, tv_r, NULL); + DB((dbg, LEVEL_4, " - %+F", tv_r)); + break; + default: + panic("Unsupported opcode"); + } - if (! get_opt_strength_red()) + if (tv == tarval_bad || tv_init == tarval_bad) { + tarval_set_integer_overflow_mode(ovmode); + DB((dbg, LEVEL_4, " = OVERFLOW")); + return NULL; + } + + if (pscc->code == iro_Add) { + tv_end = tarval_add(tv, tv_incr); + } else { + assert(pscc->code == iro_Sub); + tv_end = tarval_sub(tv, tv_incr, NULL); + } + + tarval_set_integer_overflow_mode(ovmode); + + if (tv_end == tarval_bad) { + DB((dbg, LEVEL_4, " = OVERFLOW")); + return NULL; + } + irg = get_irn_irg(iv); + return new_r_Const(irg, tv); + } + return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(e->dst)); +} /* applyOneEdge */ + +/** + * Applies the operations represented by the LFTR edges to a + * region constant and returns the value. + * Return NULL if the transformation cannot be done safely without + * an Overflow. + * + * @param pIV points to the IV node that starts the LFTR edge chain + * after translation points to the new IV + * @param rc the region constant that should be translated + * @param env the IV environment + * + * @return the translated region constant or NULL + * if the translation was not possible + */ +static ir_node *applyEdges(ir_node **pIV, ir_node *rc, iv_env *env) +{ + ir_node *iv = *pIV; + if (env->osr_flags & osr_flag_lftr_with_ov_check) { + /* overflow can only be decided for Consts */ + if (! is_counter_iv(iv, env)) { + DB((dbg, LEVEL_4, "not counter IV\n", rc)); + return NULL; + } + if (! is_Const(rc)) { + DB((dbg, LEVEL_4, " = UNKNOWN (%+F)\n", rc)); + return NULL; + } + DB((dbg, LEVEL_4, "%+F", get_Const_tarval(rc))); + } + + for (; rc;) { + LFTR_edge *e = LFTR_find(iv, env); + if (e != NULL) { + rc = applyOneEdge(iv, rc, e, env); + iv = e->dst; + } else + break; + } + DB((dbg, LEVEL_3, "\n")); + *pIV = iv; + return rc; +} /* applyEdges */ + +/** + * Walker, finds Cmp(iv, rc) or Cmp(rc, iv) + * and tries to optimize them. + */ +static void do_lftr(ir_node *cmp, void *ctx) +{ + iv_env *env = (iv_env*)ctx; + ir_node *left, *right, *liv, *riv; + ir_node *iv, *rc; + ir_node *nleft = NULL, *nright = NULL; + + if (!is_Cmp(cmp)) return; - FIRM_DBG_REGISTER(dbg, "firm.opt.osr"); - firm_dbg_set_mask(dbg, SET_LEVEL_2); + left = get_Cmp_left(cmp); + right = get_Cmp_right(cmp); - /* and dominance as well */ - assure_doms(irg); + liv = is_iv(left, env); + riv = is_iv(right, env); + if (liv && is_rc(right, liv)) { + iv = left; rc = right; - DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg)); + nright = applyEdges(&iv, rc, env); + nleft = iv; + } + else if (riv && is_rc(left, riv)) { + iv = right; rc = left; + + nleft = applyEdges(&iv, rc, env); + nright = iv; + } + + if (nleft && nright) { + DB((dbg, LEVEL_2, " LFTR for %+F\n", cmp)); + set_Cmp_left(cmp, nleft); + set_Cmp_right(cmp, nright); + ++env->lftr_replaced; + } +} /* do_lftr */ + +/** + * do linear function test replacement. + * + * @param irg the graph that should be optimized + * @param env the IV environment + */ +static void lftr(ir_graph *irg, iv_env *env) +{ + irg_walk_graph(irg, NULL, do_lftr, env); +} /* lftr */ + +/* Remove any Phi cycles with only one real input. */ +void remove_phi_cycles(ir_graph *irg) +{ + iv_env env; - env.entries = NEW_ARR_F(node_entry, get_irg_last_idx(irg)); - memset(env.entries, 0, get_irg_last_idx(irg)); + assure_irg_properties(irg, + IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE + | IR_GRAPH_PROPERTY_CONSISTENT_OUTS + | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES); + + FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi"); + + DB((dbg, LEVEL_1, "Doing Phi cycle removement for %+F\n", irg)); obstack_init(&env.obst); - env.stack = NEW_ARR_F(ir_node *, 128); - env.tos = 0; - env.nextDFSnum = 0; - env.replaced = 0; - env.POnum = 0; + env.stack = NEW_ARR_F(ir_node *, 128); + env.tos = 0; + env.nextDFSnum = 0; + env.POnum = 0; + env.quad_map = NULL; + env.lftr_edges = NULL; + env.replaced = 0; + env.lftr_replaced = 0; + env.osr_flags = 0; + env.need_postpass = 0; + env.process_scc = process_phi_only_scc; + + /* Clear all links and move Proj nodes into the + * the same block as their predecessors. + * This can improve the placement of new nodes. + */ + irg_walk_graph(irg, NULL, firm_clear_link, NULL); + + /* calculate the post order number for blocks. */ + irg_out_block_walk(get_irg_start_block(irg), NULL, assign_po, &env); + + /* calculate the SCC's and drive OSR. */ + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); + do_dfs(irg, &env); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); - /* calculate the post order number */ - irg_block_walk_graph(irg, NULL, assign_po, &env); + if (env.replaced) { + DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", + env.replaced)); + } + DEL_ARR_F(env.stack); + obstack_free(&env.obst, NULL); + + confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW); +} + +ir_graph_pass_t *remove_phi_cycles_pass(const char *name) +{ + return def_graph_pass(name ? name : "remove_phi_cycles", remove_phi_cycles); +} /* remove_phi_cycles_pass */ + +/** + * Post-walker: fix Add and Sub nodes that where results of I<->P conversions. + */ +static void fix_adds_and_subs(ir_node *irn, void *ctx) +{ + (void) ctx; + + if (is_Add(irn)) { + ir_mode *mode = get_irn_mode(irn); + + if (mode_is_int(mode)) { + ir_node *pred; + + pred = get_Add_left(irn); + if (get_irn_mode(pred) != mode) { + ir_node *block = get_nodes_block(pred); + + pred = new_r_Conv(block, pred, mode); + set_Add_left(irn, pred); + } + pred = get_Add_right(irn); + if (get_irn_mode(pred) != mode) { + ir_node *block = get_nodes_block(pred); + + pred = new_r_Conv(block, pred, mode); + set_Add_right(irn, pred); + } + } + } else if (is_Sub(irn)) { + ir_mode *mode = get_irn_mode(irn); + + if (mode_is_int(mode)) { + ir_node *left = get_Sub_left(irn); + ir_node *right = get_Sub_right(irn); + ir_mode *l_mode = get_irn_mode(left); + ir_mode *r_mode = get_irn_mode(right); + + if (mode_is_int(l_mode) && mode_is_int(r_mode)) { + if (l_mode != mode) { + ir_node *block = get_nodes_block(left); + + left = new_r_Conv(block, left, mode); + set_Sub_left(irn, left); + } + if (r_mode != mode) { + ir_node *block = get_nodes_block(right); + + right = new_r_Conv(block, right, mode); + set_Sub_right(irn, right); + } + } + } else if (mode_is_reference(mode)) { + ir_node *left = get_Sub_left(irn); + ir_node *right = get_Sub_right(irn); + ir_mode *l_mode = get_irn_mode(left); + ir_mode *r_mode = get_irn_mode(right); + if (mode_is_int(l_mode)) { + /* Usually, Sub(I*,P) is an error, hence the verifier rejects it. + * However, it is correct in this case, so add Conv to make verifier happy. */ + ir_node *block = get_nodes_block(right); + ir_node *lconv = new_r_Conv(block, left, r_mode); + assert(mode_is_reference(r_mode)); + set_Sub_left(irn, lconv); + } + } + } +} /* fix_adds_and_subs */ + +/* Performs Operator Strength Reduction for the passed graph. */ +void opt_osr(ir_graph *irg, unsigned flags) +{ + iv_env env; + + FIRM_DBG_REGISTER(dbg, "firm.opt.osr"); + + assure_irg_properties(irg, + IR_GRAPH_PROPERTY_CONSISTENT_DOMINANCE + | IR_GRAPH_PROPERTY_CONSISTENT_OUTS + | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES); + + DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg)); + + obstack_init(&env.obst); + env.stack = NEW_ARR_F(ir_node *, 128); + env.tos = 0; + env.nextDFSnum = 0; + env.POnum = 0; + env.quad_map = new_set(quad_cmp, 64); + env.lftr_edges = new_set(LFTR_cmp, 64); + env.replaced = 0; + env.lftr_replaced = 0; + env.osr_flags = flags; + env.need_postpass = 0; + env.process_scc = process_scc; + + /* Clear all links and move Proj nodes into the + * the same block as its predecessors. + * This can improve the placement of new nodes. + */ + irg_walk_graph(irg, NULL, firm_clear_link, NULL); + + irg_block_edges_walk(get_irg_start_block(irg), NULL, assign_po, &env); + + /* calculate the SCC's and drive OSR. */ + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); do_dfs(irg, &env); if (env.replaced) { - set_irg_outs_inconsistent(irg); - set_irg_loopinfo_inconsistent(irg); + if (env.need_postpass) + irg_walk_graph(irg, NULL, fix_adds_and_subs, &env); + + /* try linear function test replacements */ + lftr(irg, &env); + (void)lftr; + + DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced)); } - DB((dbg, LEVEL_1, "Replacements: %u\n\n", env.replaced)); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK); + del_set(env.lftr_edges); + del_set(env.quad_map); DEL_ARR_F(env.stack); obstack_free(&env.obst, NULL); - DEL_ARR_F(env.entries); + + confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_NONE); } + +typedef struct pass_t { + ir_graph_pass_t pass; + unsigned flags; +} pass_t; + +/** +* Wrapper for running opt_osr() as an ir_graph pass. +*/ +static int pass_wrapper(ir_graph *irg, void *context) +{ + pass_t *pass = (pass_t*)context; + opt_osr(irg, pass->flags); + return 0; +} /* pass_wrapper */ + +ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags) +{ + pass_t *pass = XMALLOCZ(pass_t); + + pass->flags = flags; + return def_graph_pass_constructor( + &pass->pass, name ? name : "osr", pass_wrapper); +} /* opt_osr_pass */