X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fopt_osr.c;h=6d9981d1009b563ff81926366f66e5e2e1f4a31c;hb=7213344d377bb242491dfb0c291b43157c3c81d6;hp=f7af05afb45fadda5d9290c329446c3f0c05f780;hpb=e8c648b13bff68494a19bb2a0e32a04f5f0cf40a;p=libfirm diff --git a/ir/opt/opt_osr.c b/ir/opt/opt_osr.c index f7af05afb..6d9981d10 100644 --- a/ir/opt/opt_osr.c +++ b/ir/opt/opt_osr.c @@ -28,9 +28,7 @@ * by Keith D. Cooper, L. Taylor Simpson, Christopher A. Vick. * Extended version. */ -#ifdef HAVE_CONFIG_H #include "config.h" -#endif #include "adt/pdeq.h" #include "iroptimize.h" @@ -52,7 +50,7 @@ #include "irloop_t.h" #include "array.h" #include "firmstat.h" -#include "xmalloc.h" +#include "error.h" /** The debug handle. */ DEBUG_ONLY(static firm_dbg_module_t *dbg;) @@ -87,7 +85,8 @@ typedef struct iv_env { 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 flags; /**< additional flags */ + 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; @@ -125,10 +124,12 @@ static int LFTR_cmp(const void *e1, const void *e2, size_t size) { (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; @@ -136,10 +137,16 @@ static LFTR_edge *LFTR_find(ir_node *src, iv_env *env) { key.src = src; return set_find(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, ir_opcode code, ir_node *rc, iv_env *env) { LFTR_edge key; @@ -155,29 +162,35 @@ static void LFTR_add(ir_node *src, ir_node *dst, ir_opcode code, ir_node *rc, iv */ // assert(LFTR_find(src, env) == NULL); set_insert(env->lftr_edges, &key, sizeof(key), HASH_PTR(src)); -} +} /* LFTR_add */ /** - * Gets the node_entry of a node + * 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 = get_irn_link(irn); - if (! e) { + if (e == NULL) { e = obstack_alloc(&env->obst, sizeof(*e)); memset(e, 0, sizeof(*e)); set_irn_link(irn, e); } return e; -} +} /* get_irn_ne */ /** - * Gets the scc from an IV. + * 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. @@ -189,7 +202,7 @@ static scc *get_iv_scc(ir_node *iv, iv_env *env) { */ 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. @@ -202,7 +215,7 @@ static int is_rc(ir_node *irn, ir_node *header_block) { ir_node *block = get_nodes_block(irn); return (block != header_block) && block_dominates(block, header_block); -} +} /* is_rc */ /** * Set compare function for the quad set. @@ -213,7 +226,7 @@ static int quad_cmp(const void *e1, const void *e2, size_t size) { (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. @@ -237,7 +250,7 @@ static ir_node *search(ir_opcode code, ir_node *op1, ir_node *op2, iv_env *env) if (entry) return entry->res; return NULL; -} +} /* search */ /** * Add an reduced operation. @@ -258,7 +271,7 @@ static void add(ir_opcode code, ir_node *op1, ir_node *op2, ir_node *result, iv_ set_insert(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 @@ -276,7 +289,7 @@ static ir_node *find_location(ir_node *block1, ir_node *block2) { return block2; assert(block_dominates(block2, block1)); return block1; -} +} /* find_location */ /** * Create a node that executes an op1 code op1 operation. @@ -290,26 +303,25 @@ static ir_node *find_location(ir_node *block1, ir_node *block2) { * @return the newly created node */ static ir_node *do_apply(ir_opcode code, dbg_info *db, ir_node *op1, ir_node *op2, ir_mode *mode) { - ir_graph *irg = current_ir_graph; 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, irg, block, op1, op2, mode); + result = new_rd_Mul(db, block, op1, op2, mode); break; case iro_Add: - result = new_rd_Add(db, irg, block, op1, op2, mode); + result = new_rd_Add(db, block, op1, op2, mode); break; case iro_Sub: - result = new_rd_Sub(db, irg, block, op1, op2, mode); + result = new_rd_Sub(db, block, op1, op2, mode); break; default: - assert(0); + panic("Unsupported opcode"); result = NULL; } return result; -} +} /* do_apply */ /** * The Apply operation. @@ -322,27 +334,28 @@ 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; -} +} /* apply */ /** * The Reduce operation. @@ -359,18 +372,26 @@ 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) { + /* check if we have already done this operation on the iv */ + if (result == NULL) { node_entry *e, *iv_e; - int i, n; + int i; ir_mode *mode = get_irn_mode(orig); result = exact_copy(iv); - /* 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); + 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)); @@ -382,24 +403,22 @@ static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) { /* create the LFTR edge */ LFTR_add(iv, result, code, rc, env); - n = get_irn_arity(result); - for (i = 0; i < n; ++i) { + 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(orig, o, rc, env); + o = apply(iv_e->header, orig, o, rc, env); set_irn_n(result, i, o); } - } - else { + } 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. @@ -427,7 +446,7 @@ static void update_scc(ir_node *iv, node_entry *e, iv_env *env) { node_entry *pe = get_irn_ne(pred, env); if (pe->header == header && pe->pscc == NULL) { - /* set the psc here to ensure that the node is NOT enqueued another time */ + /* set the pscc here to ensure that the node is NOT enqueued another time */ pe->pscc = pscc; waitq_put(wq, pred); } @@ -435,10 +454,12 @@ static void update_scc(ir_node *iv, node_entry *e, iv_env *env) { } while (! waitq_empty(wq)); del_waitq(wq); DB((dbg, LEVEL_2, "\n")); -} +} /* update_scc */ /** - * The Replace operation. + * 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 @@ -466,8 +487,9 @@ static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env) { return 1; } return 0; -} +} /* replace */ +#if 0 /** * check if a given node is a mul with 2, 4, 8 */ @@ -492,10 +514,14 @@ static int is_x86_shift_const(ir_node *mul) { } } 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); @@ -554,10 +580,16 @@ static int is_counter_iv(ir_node *iv, iv_env *env) { 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, *header; @@ -611,7 +643,7 @@ static int check_users_for_reg_pressure(ir_node *iv, iv_env *env) { * * 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 + * Phi, only one Add/Sub with a Const. */ if (! is_counter_iv(iv, env)) return 0; @@ -621,7 +653,7 @@ static int check_users_for_reg_pressure(ir_node *iv, iv_env *env) { * to do a linear function test replacement, so go on. */ return 1; -} +} /* check_users_for_reg_pressure */ /** * Check if a node can be replaced (+, -, *). @@ -657,15 +689,10 @@ static int check_replace(ir_node *irn, iv_env *env) { } if (iv) { - if (env->flags & osr_flag_keep_reg_pressure) { + if (env->osr_flags & osr_flag_keep_reg_pressure) { if (! check_users_for_reg_pressure(iv, env)) return 0; } - /* check for x86 constants */ - if (env->flags & osr_flag_ignore_x86_shift) - if (is_x86_shift_const(irn)) - return 0; - return replace(irn, iv, rc, env); } break; @@ -673,7 +700,7 @@ static int check_replace(ir_node *irn, iv_env *env) { break; } return 0; -} +} /* check_replace */ /** * Check which SCC's are induction variables. @@ -778,10 +805,10 @@ fail: next = e->next; e->header = NULL; } -} +} /* classify_iv */ /** - * Process a SCC for the operator strength reduction. + * Process an SCC for the operator strength reduction. * * @param pscc the SCC * @param env the environment @@ -795,7 +822,7 @@ static void process_scc(scc *pscc, iv_env *env) { ir_node *irn, *next; DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc)); - for (irn = pscc->head; irn; irn = next) { + for (irn = pscc->head; irn != NULL; irn = next) { node_entry *e = get_irn_link(irn); next = e->next; @@ -812,10 +839,13 @@ static void process_scc(scc *pscc, iv_env *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; @@ -855,10 +885,10 @@ static void remove_phi_cycle(scc *pscc, iv_env *env) { exchange(irn, out_rc); } ++env->replaced; -} +} /* remove_phi_cycle */ /** - * Process a SCC for the Phi cycle removement. + * Process a SCC for the Phi cycle remove. * * @param pscc the SCC * @param env the environment @@ -885,7 +915,7 @@ static void process_phi_only_scc(scc *pscc, iv_env *env) { if (e->next != NULL) remove_phi_cycle(pscc, env); -} +} /* process_phi_only_scc */ /** @@ -904,23 +934,22 @@ static void push(iv_env *env, ir_node *n) { env->stack[env->tos++] = n; 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) -{ +static ir_node *pop(iv_env *env) { 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. @@ -928,8 +957,7 @@ static ir_node *pop(iv_env *env) * @param irn start at this node * @param env the environment */ -static void dfs(ir_node *irn, iv_env *env) -{ +static void dfs(ir_node *irn, iv_env *env) { int i, n; node_entry *node = get_irn_ne(irn, env); @@ -941,11 +969,10 @@ 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++; @@ -953,7 +980,7 @@ 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); @@ -961,7 +988,7 @@ static void dfs(ir_node *irn, iv_env *env) ir_node *pred = get_irn_n(irn, i); 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); } @@ -986,7 +1013,7 @@ static void dfs(ir_node *irn, iv_env *env) env->process_scc(pscc, env); } } -} +} /* dfs */ /** * Do the DFS by starting at the End node of a graph. @@ -996,10 +1023,10 @@ static void dfs(ir_node *irn, iv_env *env) */ 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; + ir_node *end = get_irg_end(irg); + int i; - set_using_irn_visited(irg); + ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED); current_ir_graph = irg; inc_irg_visited(irg); @@ -1008,18 +1035,17 @@ static void do_dfs(ir_graph *irg, iv_env *env) { 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); } - clear_using_irn_visited(irg); + ir_free_resources(irg, IR_RESOURCE_IRN_VISITED); current_ir_graph = rem; -} +} /* do_dfs */ /** * Post-block-walker: assign the post-order number. @@ -1029,27 +1055,7 @@ static void assign_po(ir_node *block, void *ctx) { node_entry *e = get_irn_ne(block, env); e->POnum = env->POnum++; -} - -/** - * Follows the LFTR edges and return the last node in the chain. - * - * @param irn the node that should be followed - * @param env the IV environment - * - * @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 *followEdges(ir_node *irn, iv_env *env) { - for (;;) { - LFTR_edge *e = LFTR_find(irn, env); - if (e) - irn = e->dst; - else - return irn; - } -} +} /* assign_po */ /** * Apply one LFTR edge operation. @@ -1069,8 +1075,8 @@ static ir_node *followEdges(ir_node *irn, iv_env *env) { * 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->flags & osr_flag_lftr_with_ov_check) { - tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr; + if (env->osr_flags & osr_flag_lftr_with_ov_check) { + tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr, *tv_end; tarval_int_overflow_mode_t ovmode; scc *pscc; @@ -1081,6 +1087,11 @@ static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env /* 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; } @@ -1114,32 +1125,32 @@ static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env DB((dbg, LEVEL_4, " + %+F", tv_r)); break; case iro_Sub: - tv = tarval_sub(tv_l, tv_r); - tv_init = tarval_sub(tv_init, 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); + tv_end = tarval_add(tv, tv_incr); } else { assert(pscc->code == iro_Sub); - tv = tarval_sub(tv, tv_incr); + tv_end = tarval_sub(tv, tv_incr, NULL); } tarval_set_integer_overflow_mode(ovmode); - if (tv == tarval_bad || tv_init == tarval_bad) { + if (tv == tarval_bad || tv_init == tarval_bad || tv_end == tarval_bad) { DB((dbg, LEVEL_4, " = OVERFLOW")); return NULL; } - return new_r_Const(current_ir_graph, get_irn_n(rc, -1), get_tarval_mode(tv), tv); + return new_Const(tv); } - return do_apply(e->code, NULL, rc, e->rc, get_irn_mode(rc)); -} + 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 @@ -1147,15 +1158,17 @@ static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env * Return NULL if the transformation cannot be done safely without * an Overflow. * - * @param iv the IV node that starts the LFTR edge chain + * @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 *iv, ir_node *rc, iv_env *env) { - if (env->flags & osr_flag_lftr_with_ov_check) { +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)); @@ -1170,16 +1183,16 @@ static ir_node *applyEdges(ir_node *iv, ir_node *rc, iv_env *env) { for (; rc;) { LFTR_edge *e = LFTR_find(iv, env); - if (e) { + if (e != NULL) { rc = applyOneEdge(iv, rc, e, env); iv = e->dst; - } - else + } else break; } DB((dbg, LEVEL_3, "\n")); + *pIV = iv; return rc; -} +} /* applyEdges */ /** * Walker, finds Cmp(iv, rc) or Cmp(rc, iv) @@ -1191,7 +1204,7 @@ 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); @@ -1202,18 +1215,14 @@ static void do_lftr(ir_node *cmp, void *ctx) { if (liv && is_rc(right, liv)) { iv = left; rc = right; - nright = applyEdges(iv, rc, env); - if (nright && nright != rc) { - nleft = followEdges(iv, env); - } + nright = applyEdges(&iv, rc, env); + nleft = iv; } else if (riv && is_rc(left, riv)) { iv = right; rc = left; - nleft = applyEdges(iv, rc, env); - if (nleft && nleft != rc) { - nright = followEdges(iv, env); - } + nleft = applyEdges(&iv, rc, env); + nright = iv; } if (nleft && nright) { @@ -1222,7 +1231,7 @@ static void do_lftr(ir_node *cmp, void *ctx) { set_Cmp_right(cmp, nright); ++env->lftr_replaced; } -} +} /* do_lftr */ /** * do linear function test replacement. @@ -1232,136 +1241,214 @@ 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); -} +} /* lftr */ /** * Pre-walker: set all node links to NULL and fix the * block of Proj nodes. */ -static void clear_and_fix(ir_node *irn, void *env) -{ - (void) env; +static void clear_and_fix(ir_node *irn, void *env) { + int *moved = env; set_irn_link(irn, NULL); if (is_Proj(irn)) { - ir_node *pred = get_Proj_pred(irn); - set_nodes_block(irn, get_nodes_block(pred)); + ir_node *pred = get_Proj_pred(irn); + ir_node *pred_block = get_nodes_block(pred); + + if (get_nodes_block(irn) != pred_block) { + set_nodes_block(irn, pred_block); + *moved = 1; + } } -} +} /* clear_and_fix */ -/* Performs Operator Strength Reduction for the passed graph. */ -void opt_osr(ir_graph *irg, unsigned flags) { + +/* Remove any Phi cycles with only one real input. */ +void remove_phi_cycles(ir_graph *irg) { iv_env env; ir_graph *rem; - int edges; - - if (! get_opt_strength_red()) { - /* only kill Phi cycles */ - remove_phi_cycles(irg); - return; - } + int projs_moved; rem = current_ir_graph; current_ir_graph = irg; - FIRM_DBG_REGISTER(dbg, "firm.opt.osr"); + FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi"); - DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg)); + 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.POnum = 0; - env.quad_map = new_set(quad_cmp, 64); - env.lftr_edges = new_set(LFTR_cmp, 64); + env.quad_map = NULL; + env.lftr_edges = NULL; env.replaced = 0; env.lftr_replaced = 0; - env.flags = flags; - env.process_scc = process_scc; + 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 it's predecessors. This can improve the placement of new nodes. */ - irg_walk_graph(irg, NULL, clear_and_fix, NULL); - - /* we need dominance */ - assure_doms(irg); + projs_moved = 0; + irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved); + if (projs_moved) + set_irg_outs_inconsistent(irg); - edges = edges_assure(irg); + /* we need outs for calculating the post order */ + assure_irg_outs(irg); /* calculate the post order number for blocks. */ - irg_block_edges_walk(get_irg_start_block(irg), NULL, assign_po, &env); + 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); if (env.replaced) { - /* try linear function test replacements */ - lftr(irg, &env); - set_irg_outs_inconsistent(irg); - DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced)); + DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced)); } - del_set(env.lftr_edges); - del_set(env.quad_map); DEL_ARR_F(env.stack); obstack_free(&env.obst, NULL); - if (! edges) - edges_deactivate(irg); - current_ir_graph = rem; -} +} /* remove_phi_cycles */ -/* Remove any Phi cycles with only one real input. */ -void remove_phi_cycles(ir_graph *irg) { +/** + * 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); + } + } + } + } +} /* fix_adds_and_subs */ + +/* Performs Operator Strength Reduction for the passed graph. */ +void opt_osr(ir_graph *irg, unsigned flags) { iv_env env; ir_graph *rem; + int edges; + int projs_moved; + + if (! get_opt_strength_red()) { + /* only kill Phi cycles */ + remove_phi_cycles(irg); + return; + } rem = current_ir_graph; current_ir_graph = irg; - FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi"); + FIRM_DBG_REGISTER(dbg, "firm.opt.osr"); - DB((dbg, LEVEL_1, "Doing Phi cycle removement for %+F\n", irg)); + 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 = NULL; - env.lftr_edges = NULL; + env.quad_map = new_set(quad_cmp, 64); + env.lftr_edges = new_set(LFTR_cmp, 64); env.replaced = 0; env.lftr_replaced = 0; - env.flags = 0; - env.process_scc = process_phi_only_scc; + 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 it's predecessors. This can improve the placement of new nodes. */ - irg_walk_graph(irg, NULL, clear_and_fix, NULL); + projs_moved = 0; + irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved); + if (projs_moved) + set_irg_outs_inconsistent(irg); - /* we need outs for calculating the post order */ - assure_irg_outs(irg); + /* we need dominance */ + assure_doms(irg); - /* calculate the post order number for blocks. */ - irg_out_block_walk(get_irg_start_block(irg), NULL, assign_po, &env); + edges = edges_assure(irg); + + /* calculate the post order number for blocks by walking the out edges. */ + assure_irg_outs(irg); + 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) { + if (env.need_postpass) + irg_walk_graph(irg, NULL, fix_adds_and_subs, &env); + + /* try linear function test replacements */ + lftr(irg, &env); + (void)lftr; + set_irg_outs_inconsistent(irg); - DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced)); + DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_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); + if (! edges) + edges_deactivate(irg); + current_ir_graph = rem; -} +} /* opt_osr */