X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fopt_osr.c;h=569d9d060e8bc43ac87135fa81ac77cd987a65bb;hb=af82344c3d1ede9a28e652302a8c2bac7bb1c743;hp=135e58f3b604f8ab281211e38ffbf9ea7f04fc1c;hpb=ddfcdcb1227bd6fb6720f8485dc62995f7bfd98f;p=libfirm diff --git a/ir/opt/opt_osr.c b/ir/opt/opt_osr.c index 135e58f3b..569d9d060 100644 --- a/ir/opt/opt_osr.c +++ b/ir/opt/opt_osr.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved. + * Copyright (C) 1995-2011 University of Karlsruhe. All right reserved. * * This file is part of libFirm. * @@ -23,7 +23,7 @@ * @date 12.5.2006 * @author Michael Beck * @version $Id$ - * @summary + * @brief * Implementation of the Operator Strength Reduction algorithm * by Keith D. Cooper, L. Taylor Simpson, Christopher A. Vick. * Extended version. @@ -58,10 +58,10 @@ DEBUG_ONLY(static firm_dbg_module_t *dbg;) /** A scc. */ typedef struct scc { - 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 */ + 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 */ @@ -79,7 +79,7 @@ typedef struct node_entry { typedef struct iv_env { struct obstack obst; /**< an obstack for allocations */ ir_node **stack; /**< the node stack */ - int tos; /**< tos index */ + 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 */ @@ -96,7 +96,7 @@ typedef struct iv_env { * An entry in the (op, node, node) -> node map. */ typedef struct quadruple_t { - ir_opcode code; /**< the opcode of the reduced operation */ + 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 */ @@ -109,7 +109,7 @@ typedef struct quadruple_t { typedef struct LFTR_edge { ir_node *src; /**< the source node */ ir_node *dst; /**< the destination node */ - ir_opcode code; /**< the opcode that must be applied */ + unsigned code; /**< the opcode that must be applied */ ir_node *rc; /**< the region const that must be applied */ } LFTR_edge; @@ -119,9 +119,10 @@ 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 = e1; - const LFTR_edge *l2 = e2; +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; @@ -132,12 +133,13 @@ static int LFTR_cmp(const void *e1, const void *e2, size_t size) { * * @param src the source node of the transition */ -static LFTR_edge *LFTR_find(ir_node *src, iv_env *env) { +static LFTR_edge *LFTR_find(ir_node *src, iv_env *env) +{ LFTR_edge key; key.src = src; - return set_find(env->lftr_edges, &key, sizeof(key), HASH_PTR(src)); + return (LFTR_edge*)set_find(env->lftr_edges, &key, sizeof(key), HASH_PTR(src)); } /* LFTR_find */ /** @@ -149,7 +151,8 @@ static LFTR_edge *LFTR_find(ir_node *src, iv_env *env) { * @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) { +static void LFTR_add(ir_node *src, ir_node *dst, unsigned code, ir_node *rc, iv_env *env) +{ LFTR_edge key; key.src = src; @@ -161,7 +164,6 @@ static void LFTR_add(ir_node *src, ir_node *dst, ir_opcode code, ir_node *rc, iv * There might be more than one edge here. This is rather bad * because we currently store only one. */ -// assert(LFTR_find(src, env) == NULL); set_insert(env->lftr_edges, &key, sizeof(key), HASH_PTR(src)); } /* LFTR_add */ @@ -171,12 +173,12 @@ static void LFTR_add(ir_node *src, ir_node *dst, ir_opcode code, ir_node *rc, iv * @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); +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 = obstack_alloc(&env->obst, sizeof(*e)); - memset(e, 0, sizeof(*e)); + e = OALLOCZ(&env->obst, node_entry); set_irn_link(irn, e); } return e; @@ -188,7 +190,8 @@ static node_entry *get_irn_ne(ir_node *irn, iv_env *env) { * @param iv any node of the induction variable * @param env the environment */ -static scc *get_iv_scc(ir_node *iv, iv_env *env) { +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 */ @@ -201,7 +204,8 @@ static scc *get_iv_scc(ir_node *iv, iv_env *env) { * * @returns the header if it is one, NULL else */ -static ir_node *is_iv(ir_node *irn, iv_env *env) { +static ir_node *is_iv(ir_node *irn, iv_env *env) +{ return get_irn_ne(irn, env)->header; } /* is_iv */ @@ -212,7 +216,8 @@ static ir_node *is_iv(ir_node *irn, iv_env *env) { * @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 != header_block) && block_dominates(block, header_block); @@ -221,9 +226,10 @@ static int is_rc(ir_node *irn, ir_node *header_block) { /** * Set compare function for the quad set. */ -static int quad_cmp(const void *e1, const void *e2, size_t size) { - const quadruple_t *c1 = e1; - const quadruple_t *c2 = e2; +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; @@ -239,15 +245,16 @@ static int quad_cmp(const void *e1, const void *e2, size_t size) { * * @return the already reduced node or NULL if this operation is not yet reduced */ -static ir_node *search(ir_opcode code, ir_node *op1, ir_node *op2, iv_env *env) { +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(env->quad_map, &key, sizeof(key), - (code * 9) ^ HASH_PTR(op1) ^HASH_PTR(op2)); + entry = (quadruple_t*)set_find(env->quad_map, &key, sizeof(key), + (code * 9) ^ HASH_PTR(op1) ^HASH_PTR(op2)); if (entry) return entry->res; return NULL; @@ -262,7 +269,8 @@ static ir_node *search(ir_opcode code, ir_node *op1, ir_node *op2, iv_env *env) * @param result the result of the reduced operation * @param env the environment */ -static void add(ir_opcode code, ir_node *op1, ir_node *op2, ir_node *result, iv_env *env) { +static void add(unsigned code, ir_node *op1, ir_node *op2, ir_node *result, iv_env *env) +{ quadruple_t key; key.code = code; @@ -285,7 +293,8 @@ static void add(ir_opcode code, ir_node *op1, ir_node *op2, ir_node *result, iv_ * 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) { +static ir_node *find_location(ir_node *block1, ir_node *block2) +{ if (block_dominates(block1, block2)) return block2; assert(block_dominates(block2, block1)); @@ -303,7 +312,8 @@ 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) { +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)); @@ -319,7 +329,6 @@ static ir_node *do_apply(ir_opcode code, dbg_info *db, ir_node *op1, ir_node *op break; default: panic("Unsupported opcode"); - result = NULL; } return result; } /* do_apply */ @@ -335,8 +344,9 @@ 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 *header, ir_node *orig, ir_node *op1, ir_node *op2, iv_env *env) { - ir_opcode code = get_irn_opcode(orig); +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) { @@ -369,8 +379,9 @@ static ir_node *apply(ir_node *header, ir_node *orig, ir_node *op1, ir_node *op2 * * @return the reduced node */ -static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) { - ir_opcode code = get_irn_opcode(orig); +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 */ @@ -424,7 +435,8 @@ static ir_node *reduce(ir_node *orig, ir_node *iv, ir_node *rc, iv_env *env) { /** * Update the scc for a newly created IV. */ -static void update_scc(ir_node *iv, node_entry *e, iv_env *env) { +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(); @@ -433,7 +445,7 @@ static void update_scc(ir_node *iv, node_entry *e, iv_env *env) { pscc->head = NULL; waitq_put(wq, iv); do { - ir_node *irn = waitq_get(wq); + ir_node *irn = (ir_node*)waitq_get(wq); node_entry *ne = get_irn_ne(irn, env); int i; @@ -467,7 +479,8 @@ static void update_scc(ir_node *iv, node_entry *e, iv_env *env) { * @param rc the region constant * @param env the environment */ -static int 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)); @@ -476,12 +489,11 @@ static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env) { if (result != irn) { node_entry *e; - hook_strength_red(current_ir_graph, irn); + hook_strength_red(get_irn_irg(irn), 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)); + e->pscc = OALLOCZ(&env->obst, scc); update_scc(result, e, env); } ++env->replaced; @@ -494,7 +506,8 @@ static int replace(ir_node *irn, ir_node *iv, ir_node *rc, iv_env *env) { /** * check if a given node is a mul with 2, 4, 8 */ -static int is_x86_shift_const(ir_node *mul) { +static int is_x86_shift_const(ir_node *mul) +{ ir_node *rc; if (! is_Mul(mul)) @@ -503,7 +516,7 @@ static int is_x86_shift_const(ir_node *mul) { /* normalization put constants on the right side */ rc = get_Mul_right(mul); if (is_Const(rc)) { - tarval *tv = get_Const_tarval(rc); + ir_tarval *tv = get_Const_tarval(rc); if (tarval_is_long(tv)) { long value = get_tarval_long(tv); @@ -524,7 +537,8 @@ static int is_x86_shift_const(ir_node *mul) { * @param iv any node of the induction variable * @param env the environment */ -static int is_counter_iv(ir_node *iv, iv_env *env) { +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; @@ -592,14 +606,14 @@ static int is_counter_iv(ir_node *iv, iv_env *env) { * @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; +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; - header = e->header; for (irn = pscc->head; irn != NULL; irn = e->next) { const ir_edge_t *edge; @@ -664,10 +678,11 @@ static int check_users_for_reg_pressure(ir_node *iv, iv_env *env) { * * @return non-zero if irn should be Replace'd */ -static int check_replace(ir_node *irn, iv_env *env) { +static int check_replace(ir_node *irn, iv_env *env) +{ ir_node *left, *right, *iv, *rc; ir_op *op = get_irn_op(irn); - ir_opcode code = get_op_code(op); + unsigned code = get_op_code(op); ir_node *liv, *riv; switch (code) { @@ -709,7 +724,8 @@ static int check_replace(ir_node *irn, iv_env *env) { * @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 *b, *h = NULL; int j, only_phi, num_outside; @@ -717,7 +733,7 @@ static void classify_iv(scc *pscc, iv_env *env) { /* 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; @@ -814,9 +830,10 @@ fail: * @param pscc the SCC * @param env the environment */ -static void process_scc(scc *pscc, iv_env *env) { +static void process_scc(scc *pscc, iv_env *env) +{ ir_node *head = pscc->head; - node_entry *e = get_irn_link(head); + node_entry *e = (node_entry*)get_irn_link(head); #ifdef DEBUG_libfirm { @@ -824,7 +841,7 @@ static void process_scc(scc *pscc, iv_env *env) { DB((dbg, LEVEL_4, " SCC at %p:\n ", pscc)); for (irn = pscc->head; irn != NULL; irn = next) { - node_entry *e = get_irn_link(irn); + node_entry *e = (node_entry*)get_irn_link(irn); next = e->next; @@ -848,7 +865,8 @@ static void process_scc(scc *pscc, iv_env *env) { * @param pscc an SCC that consists of Phi nodes only * @param env the environment */ -static void remove_phi_cycle(scc *pscc, iv_env *env) { +static void remove_phi_cycle(scc *pscc, iv_env *env) +{ ir_node *irn, *next; int j; ir_node *out_rc; @@ -894,9 +912,10 @@ static void remove_phi_cycle(scc *pscc, iv_env *env) { * @param pscc the SCC * @param env the environment */ -static void process_phi_only_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); + node_entry *e = (node_entry*)get_irn_link(head); #ifdef DEBUG_libfirm { @@ -904,7 +923,7 @@ static void process_phi_only_scc(scc *pscc, iv_env *env) { 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; @@ -925,11 +944,12 @@ static void process_phi_only_scc(scc *pscc, iv_env *env) { * @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; @@ -944,7 +964,8 @@ static void push(iv_env *env, ir_node *n) { * * @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); @@ -958,7 +979,8 @@ 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); @@ -997,10 +1019,9 @@ 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; - memset(pscc, 0, sizeof(*pscc)); do { node_entry *e; @@ -1022,14 +1043,13 @@ static void dfs(ir_node *irn, iv_env *env) { * @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; +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 */ @@ -1044,15 +1064,14 @@ static void do_dfs(ir_graph *irg, iv_env *env) { } ir_free_resources(irg, IR_RESOURCE_IRN_VISITED); - - current_ir_graph = rem; } /* do_dfs */ /** * Post-block-walker: assign the post-order number. */ -static void assign_po(ir_node *block, void *ctx) { - iv_env *env = ctx; +static void assign_po(ir_node *block, void *ctx) +{ + iv_env *env = (iv_env*)ctx; node_entry *e = get_irn_ne(block, env); e->POnum = env->POnum++; @@ -1075,11 +1094,13 @@ static void assign_po(ir_node *block, void *ctx) { * 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) { +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) { - tarval *tv_l, *tv_r, *tv, *tv_init, *tv_incr, *tv_end; + 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")); @@ -1132,7 +1153,6 @@ static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env break; default: panic("Unsupported opcode"); - tv = tarval_bad; } if (pscc->code == iro_Add) { @@ -1148,7 +1168,8 @@ static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env DB((dbg, LEVEL_4, " = OVERFLOW")); return NULL; } - return new_Const(tv); + 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 */ @@ -1167,7 +1188,8 @@ static ir_node *applyOneEdge(ir_node *iv, ir_node *rc, LFTR_edge *e, iv_env *env * @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) { +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 */ @@ -1199,8 +1221,9 @@ static ir_node *applyEdges(ir_node **pIV, ir_node *rc, iv_env *env) { * 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 = ctx; +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; @@ -1240,7 +1263,8 @@ static void do_lftr(ir_node *cmp, void *ctx) { * @param irg the graph that should be optimized * @param env the IV environment */ -static void lftr(ir_graph *irg, iv_env *env) { +static void lftr(ir_graph *irg, iv_env *env) +{ irg_walk_graph(irg, NULL, do_lftr, env); } /* lftr */ @@ -1248,8 +1272,10 @@ static void lftr(ir_graph *irg, iv_env *env) { * 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) { - int *moved = env; +static void clear_and_fix(ir_node *irn, void *env) +{ + (void)env; + set_irn_link(irn, NULL); if (is_Proj(irn)) { @@ -1258,20 +1284,15 @@ static void clear_and_fix(ir_node *irn, void *env) { if (get_nodes_block(irn) != pred_block) { set_nodes_block(irn, pred_block); - *moved = 1; } } } /* clear_and_fix */ /* Remove any Phi cycles with only one real input. */ -void remove_phi_cycles(ir_graph *irg) { - iv_env env; - ir_graph *rem; - int projs_moved; - - rem = current_ir_graph; - current_ir_graph = irg; +void remove_phi_cycles(ir_graph *irg) +{ + iv_env env; FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi"); @@ -1291,13 +1312,10 @@ void remove_phi_cycles(ir_graph *irg) { 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. + * the same block as their predecessors. + * This can improve the placement of new nodes. */ - projs_moved = 0; - irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved); - if (projs_moved) - set_irg_outs_inconsistent(irg); + irg_walk_graph(irg, NULL, clear_and_fix, NULL); /* we need outs for calculating the post order */ assure_irg_outs(irg); @@ -1311,14 +1329,11 @@ void remove_phi_cycles(ir_graph *irg) { ir_free_resources(irg, IR_RESOURCE_IRN_LINK); if (env.replaced) { - set_irg_outs_inconsistent(irg); DB((dbg, LEVEL_1, "remove_phi_cycles: %u Cycles removed\n\n", env.replaced)); } DEL_ARR_F(env.stack); obstack_free(&env.obst, NULL); - - current_ir_graph = rem; } /* remove_phi_cycles */ ir_graph_pass_t *remove_phi_cycles_pass(const char *name) @@ -1329,7 +1344,8 @@ ir_graph_pass_t *remove_phi_cycles_pass(const char *name) /** * 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) { +static void fix_adds_and_subs(ir_node *irn, void *ctx) +{ (void) ctx; if (is_Add(irn)) { @@ -1376,25 +1392,28 @@ static void fix_adds_and_subs(ir_node *irn, void *ctx) { 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) { +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.osr"); @@ -1414,13 +1433,10 @@ void opt_osr(ir_graph *irg, unsigned flags) { 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. + * the same block as its predecessors. + * This can improve the placement of new nodes. */ - projs_moved = 0; - irg_walk_graph(irg, NULL, clear_and_fix, &projs_moved); - if (projs_moved) - set_irg_outs_inconsistent(irg); + irg_walk_graph(irg, NULL, clear_and_fix, NULL); /* we need dominance */ assure_doms(irg); @@ -1443,7 +1459,6 @@ void opt_osr(ir_graph *irg, unsigned flags) { lftr(irg, &env); (void)lftr; - set_irg_outs_inconsistent(irg); DB((dbg, LEVEL_1, "Replacements: %u + %u (lftr)\n\n", env.replaced, env.lftr_replaced)); } ir_free_resources(irg, IR_RESOURCE_IRN_LINK); @@ -1455,36 +1470,28 @@ void opt_osr(ir_graph *irg, unsigned flags) { if (! edges) edges_deactivate(irg); - - current_ir_graph = rem; } /* opt_osr */ -struct pass_t { +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) { - struct pass_t *pass = context; +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) { - struct pass_t *pass = xmalloc(sizeof(*pass)); - - pass->pass.kind = k_ir_prog_pass; - pass->pass.run_on_irg = pass_wrapper; - pass->pass.context = pass; - pass->pass.name = name ? name : "osr"; + pass_t *pass = XMALLOCZ(pass_t); pass->flags = flags; - - INIT_LIST_HEAD(&pass->pass.list); - - return &pass->pass; + return def_graph_pass_constructor( + &pass->pass, name ? name : "osr", pass_wrapper); } /* opt_osr_pass */