X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fopt_osr.c;h=3253615ae658a8b6943c2e311f8d60db248c79d7;hb=51ae0985ed587b82ae03ab2f752b5a9917a309b2;hp=c646988d11f5507f49885e74a001947cc771b1c4;hpb=0d470055af1b3bc6c917e6f004123e5aab47de8f;p=libfirm diff --git a/ir/opt/opt_osr.c b/ir/opt/opt_osr.c index c646988d1..3253615ae 100644 --- a/ir/opt/opt_osr.c +++ b/ir/opt/opt_osr.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2007 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -26,21 +26,23 @@ * @summary * Implementation of the Operator Strength Reduction algorithm * by Keith D. Cooper, L. Taylor Simpson, Christopher A. Vick. + * Extended version. */ #ifdef HAVE_CONFIG_H #include "config.h" #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" @@ -51,13 +53,17 @@ #include "array.h" #include "firmstat.h" #include "xmalloc.h" +#include "error.h" /** The debug handle. */ DEBUG_ONLY(static firm_dbg_module_t *dbg;) /** A scc. */ typedef struct scc { - ir_node *head; /**< the head of the list */ + ir_node *head; /**< the head of the list */ + tarval *init; /**< the init value iff only one exists. */ + 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 */ @@ -117,11 +123,11 @@ static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env); static int LFTR_cmp(const void *e1, const void *e2, size_t size) { const LFTR_edge *l1 = e1; const LFTR_edge *l2 = e2; + (void) size; return l1->src != l2->src; } -#if 0 /** * Find a LFTR edge. */ @@ -132,7 +138,6 @@ static LFTR_edge *LFTR_find(ir_node *src, iv_env *env) { return set_find(env->lftr_edges, &key, sizeof(key), HASH_PTR(src)); } -#endif /** * Add a LFTR edge. @@ -167,6 +172,14 @@ static node_entry *get_irn_ne(ir_node *irn, iv_env *env) { return e; } +/** + * Gets the scc from an IV. + */ +static scc *get_iv_scc(ir_node *iv, iv_env *env) { + node_entry *e = get_irn_ne(iv, env); + return e->pscc; +} + /** * Check if irn is an IV. * @@ -198,6 +211,7 @@ static int is_rc(ir_node *irn, ir_node *header_block) { static int quad_cmp(const void *e1, const void *e2, size_t size) { const quadruple_t *c1 = e1; const quadruple_t *c2 = e2; + (void) size; return c1->code != c2->code || c1->op1 != c2->op1 || c1->op2 != c2->op2; } @@ -292,7 +306,7 @@ static ir_node *do_apply(ir_opcode code, dbg_info *db, ir_node *op1, ir_node *op result = new_rd_Sub(db, irg, block, op1, op2, mode); break; default: - assert(0); + panic("Unsupported opcode"); result = NULL; } return result; @@ -309,24 +323,25 @@ static ir_node *do_apply(ir_opcode code, dbg_info *db, ir_node *op1, ir_node *op * * @return the newly created node */ -static ir_node *apply(ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env) { +static ir_node *apply(ir_node *header, ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env) { ir_opcode code = get_irn_opcode(orig); ir_node *result = search(code, op1, op2, env); - if (! result) { + 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 != NULL && is_rc(op2, op1_header)) { + if (op1_header == header && is_rc(op2, op1_header)) { result = reduce(orig, op1, op2, env); } - else if (op2_header != NULL && is_rc(op1, op2_header)) { + 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; } + get_irn_ne(result, env)->header = NULL; + } } return result; } @@ -346,15 +361,15 @@ static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) { ir_opcode code = get_irn_opcode(orig); ir_node *result = search(code, iv, rc, env); - if (! result) { + if (result == NULL) { node_entry *e, *iv_e; int i, n; ir_mode *mode = get_irn_mode(orig); result = exact_copy(iv); - /* Beware: we must always create a new nduction variable with the same mode - as the node we are replacing. Espicially this means the mode might be changed + /* 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. */ set_irn_mode(result, mode); @@ -376,12 +391,8 @@ static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) { e = get_irn_ne(o, env); if (e->header == iv_e->header) o = reduce(orig, o, rc, env); - else if (is_Phi(result)) - o = apply(orig, o, rc, env); - else { - if (code == iro_Mul) - o = apply(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); } } @@ -392,6 +403,42 @@ static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) { return result; } +/** + * 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 = 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")); +} + /** * The Replace operation. * @@ -402,28 +449,184 @@ static ir_node *reduce(ir_node *orig, 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; - ir_loop *iv_loop = get_irn_loop(get_nodes_block(iv)); - ir_loop *irn_loop = get_irn_loop(get_nodes_block(irn)); - - /* only replace nodes that are in the same (or deeper loops) */ - if (get_loop_depth(irn_loop) >= get_loop_depth(iv_loop)) { - DB((dbg, LEVEL_2, " Replacing %+F\n", irn)); - - result = reduce(irn, iv, rc, env); - if (result != irn) { - node_entry *e, *iv_e; - - hook_strength_red(current_ir_graph, irn); - exchange(irn, result); - e = get_irn_ne(result, env); - iv_e = get_irn_ne(iv, env); - e->header = iv_e->header; + + DB((dbg, LEVEL_2, " Replacing %+F\n", irn)); + + result = reduce(irn, iv, rc, env); + if (result != irn) { + node_entry *e; + + hook_strength_red(current_ir_graph, irn); + exchange(irn, result); + e = get_irn_ne(result, env); + if (e->pscc == NULL) { + e->pscc = obstack_alloc(&env->obst, sizeof(*e->pscc)); + memset(e->pscc, 0, sizeof(*e->pscc)); + update_scc(result, e, env); } + ++env->replaced; return 1; } return 0; } +#if 0 +/** + * check if a given node is a mul with 2, 4, 8 + */ +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)) { + 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; +} +#endif + +/** + * Check if an IV represents a counter with constant limits. + */ +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; +} + +/** + * Check the users of an induction variable for register pressure. + */ +static int check_users_for_reg_pressure(ir_node *iv, iv_env *env) { + ir_node *irn, *header; + ir_node *have_user = NULL; + ir_node *have_cmp = NULL; + node_entry *e = get_irn_ne(iv, env); + scc *pscc = e->pscc; + + header = e->header; + for (irn = pscc->head; irn != NULL; irn = e->next) { + const ir_edge_t *edge; + + 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 if a node can be replaced (+, -, *). * @@ -458,21 +661,10 @@ static int check_replace(ir_node *irn, iv_env *env) { } if (iv) { - if (code == iro_Mul && env->flags & osr_flag_ignore_x86_shift) { - if (is_Const(rc)) { - 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 */ - break; - } - } - } + if (env->flags & osr_flag_keep_reg_pressure) { + if (! check_users_for_reg_pressure(iv, env)) + return 0; } - return replace(irn, iv, rc, env); } break; @@ -583,8 +775,7 @@ fail: node_entry *e = get_irn_ne(irn, env); next = e->next; - if (! check_replace(irn, env)) - e->header = NULL; + e->header = NULL; } } @@ -780,7 +971,7 @@ static void dfs(ir_node *irn, iv_env *env) scc *pscc = obstack_alloc(&env->obst, sizeof(*pscc)); ir_node *x; - pscc->head = NULL; + memset(pscc, 0, sizeof(*pscc)); do { node_entry *e; @@ -807,6 +998,8 @@ static void do_dfs(ir_graph *irg, iv_env *env) { ir_node *end = get_irg_end(irg); int i, n; + ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); + current_ir_graph = irg; inc_irg_visited(irg); @@ -822,6 +1015,8 @@ static void do_dfs(ir_graph *irg, iv_env *env) { dfs(ka, env); } + ir_free_resources(irg, IR_RESOURCE_IRN_VISITED); + current_ir_graph = rem; } @@ -835,7 +1030,6 @@ static void assign_po(ir_node *block, void *ctx) { e->POnum = env->POnum++; } -#if 0 /** * Follows the LFTR edges and return the last node in the chain. * @@ -861,7 +1055,8 @@ static ir_node *followEdges(ir_node *irn, iv_env *env) { * Return NULL if the transformation cannot be done safely without * an Overflow. * - * @param rc the IV node that should be translated + * @param iv the induction variable + * @param rc the constant that should be translated * @param e the LFTR edge * @param env the IV environment * @@ -872,10 +1067,16 @@ static ir_node *followEdges(ir_node *irn, iv_env *env) { * 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 *rc, LFTR_edge *e, iv_env *env) { +static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env) { if (env->flags & osr_flag_lftr_with_ov_check) { - tarval *tv_l, *tv_r, *tv; + tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr; tarval_int_overflow_mode_t ovmode; + scc *pscc; + + 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)) { @@ -889,26 +1090,48 @@ static ir_node *applyOneEdge(ir_node *rc, LFTR_edge *e, iv_env *env) { 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 = 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 = 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); + 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: - assert(0); + panic("Unsupported opcode"); tv = tarval_bad; } + + if (pscc->code == iro_Add) { + tv = tarval_add(tv, tv_incr); + } else { + assert(pscc->code == iro_Sub); + tv = tarval_sub(tv, tv_incr, NULL); + } + tarval_set_integer_overflow_mode(ovmode); - if (tv == tarval_bad) { + if (tv == tarval_bad || tv_init == tarval_bad) { DB((dbg, LEVEL_4, " = OVERFLOW")); return NULL; } @@ -931,10 +1154,12 @@ static ir_node *applyOneEdge(ir_node *rc, LFTR_edge *e, iv_env *env) { * if the translation was not possible */ static ir_node *applyEdges(ir_node *iv, ir_node *rc, iv_env *env) { - ir_node *irn = iv; - if (env->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; @@ -942,11 +1167,11 @@ static ir_node *applyEdges(ir_node *iv, ir_node *rc, iv_env *env) { DB((dbg, LEVEL_4, "%+F", get_Const_tarval(rc))); } - for (irn = iv; rc;) { - LFTR_edge *e = LFTR_find(irn, env); + for (; rc;) { + LFTR_edge *e = LFTR_find(iv, env); if (e) { - rc = applyOneEdge(rc, e, env); - irn = e->dst; + rc = applyOneEdge(iv, rc, e, env); + iv = e->dst; } else break; @@ -965,7 +1190,7 @@ static void do_lftr(ir_node *cmp, void *ctx) { ir_node *iv, *rc; ir_node *nleft = NULL, *nright = NULL; - if (get_irn_op(cmp) != op_Cmp) + if (!is_Cmp(cmp)) return; left = get_Cmp_left(cmp); @@ -1007,7 +1232,6 @@ static void do_lftr(ir_node *cmp, void *ctx) { static void lftr(ir_graph *irg, iv_env *env) { irg_walk_graph(irg, NULL, do_lftr, env); } -#endif /** * Pre-walker: set all node links to NULL and fix the @@ -1015,11 +1239,12 @@ static void lftr(ir_graph *irg, iv_env *env) { */ static void clear_and_fix(ir_node *irn, void *env) { + (void) env; set_irn_link(irn, NULL); if (is_Proj(irn)) { ir_node *pred = get_Proj_pred(irn); - set_irn_n(irn, -1, get_irn_n(pred, -1)); + set_nodes_block(irn, get_nodes_block(pred)); } } @@ -1027,6 +1252,7 @@ static void clear_and_fix(ir_node *irn, void *env) void opt_osr(ir_graph *irg, unsigned flags) { iv_env env; ir_graph *rem; + int edges; if (! get_opt_strength_red()) { /* only kill Phi cycles */ @@ -1061,17 +1287,19 @@ void opt_osr(ir_graph *irg, unsigned flags) { /* we need dominance */ assure_doms(irg); - assure_irg_outs(irg); + + edges = edges_assure(irg); /* calculate the post order number for blocks. */ - irg_out_block_walk(get_irg_start_block(irg), NULL, assign_po, &env); + irg_block_edges_walk(get_irg_start_block(irg), NULL, assign_po, &env); /* calculate the SCC's and drive OSR. */ do_dfs(irg, &env); if (env.replaced) { /* try linear function test replacements */ - //lftr(irg, &env); + //lftr(irg, &env); // currently buggy :-( + (void) lftr; set_irg_outs_inconsistent(irg); DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced)); @@ -1082,6 +1310,9 @@ void opt_osr(ir_graph *irg, unsigned flags) { DEL_ARR_F(env.stack); obstack_free(&env.obst, NULL); + if (! edges) + edges_deactivate(irg); + current_ir_graph = rem; } @@ -1115,7 +1346,7 @@ void remove_phi_cycles(ir_graph *irg) { */ irg_walk_graph(irg, NULL, clear_and_fix, NULL); - /* we need dominance */ + /* we need outs for calculating the post order */ assure_irg_outs(irg); /* calculate the post order number for blocks. */ @@ -1126,7 +1357,7 @@ void remove_phi_cycles(ir_graph *irg) { if (env.replaced) { set_irg_outs_inconsistent(irg); - DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced)); + DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced)); } DEL_ARR_F(env.stack);