X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fopt_osr.c;h=165fff022134d6cf3324e20d88e4b9a6f575f562;hb=7547cf525ad54f59d2dc3f39d4b257911c94989b;hp=6bcf4bc8a1026eb03e25a44b284db7a7d80b5728;hpb=e1397b01aceb38b6bb62c319007146af3b922f39;p=libfirm diff --git a/ir/opt/opt_osr.c b/ir/opt/opt_osr.c index 6bcf4bc8a..165fff022 100644 --- a/ir/opt/opt_osr.c +++ b/ir/opt/opt_osr.c @@ -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. @@ -51,16 +51,17 @@ #include "array.h" #include "firmstat.h" #include "error.h" +#include "irpass_t.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 */ - 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 */ @@ -118,7 +119,8 @@ 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) { +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; @@ -131,7 +133,8 @@ 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; @@ -148,7 +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, ir_opcode code, ir_node *rc, iv_env *env) +{ LFTR_edge key; key.src = src; @@ -160,7 +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 */ @@ -170,12 +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) { +static node_entry *get_irn_ne(ir_node *irn, iv_env *env) +{ node_entry *e = 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; @@ -187,7 +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 */ @@ -200,7 +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 */ @@ -211,7 +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); @@ -220,7 +226,8 @@ 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) { +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; @@ -238,7 +245,8 @@ 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(ir_opcode code, ir_node *op1, ir_node *op2, iv_env *env) +{ quadruple_t key, *entry; key.code = code; @@ -261,7 +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(ir_opcode code, ir_node *op1, ir_node *op2, ir_node *result, iv_env *env) +{ quadruple_t key; key.code = code; @@ -284,7 +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)); @@ -302,7 +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(ir_opcode 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)); @@ -318,7 +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 */ @@ -334,7 +344,8 @@ 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) { +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); @@ -368,7 +379,8 @@ 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) { +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); @@ -423,7 +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(); @@ -466,7 +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)); @@ -475,12 +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; @@ -493,7 +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)) @@ -502,7 +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); @@ -523,7 +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; @@ -591,7 +606,8 @@ 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) { +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; @@ -663,7 +679,8 @@ 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); @@ -708,7 +725,8 @@ 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; @@ -813,7 +831,8 @@ 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); @@ -847,7 +866,8 @@ 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; @@ -893,7 +913,8 @@ 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); @@ -924,7 +945,8 @@ 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)) { @@ -943,7 +965,8 @@ 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); @@ -957,7 +980,8 @@ 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); @@ -996,10 +1020,9 @@ 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; @@ -1021,14 +1044,13 @@ 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 */ @@ -1043,14 +1065,13 @@ 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) { +static void assign_po(ir_node *block, void *ctx) +{ iv_env *env = ctx; node_entry *e = get_irn_ne(block, env); @@ -1074,11 +1095,13 @@ 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; + 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")); @@ -1131,23 +1154,23 @@ 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) { - 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, NULL); + 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_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 */ @@ -1166,7 +1189,8 @@ 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 */ @@ -1198,7 +1222,8 @@ 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) { +static void do_lftr(ir_node *cmp, void *ctx) +{ iv_env *env = ctx; ir_node *left, *right, *liv, *riv; ir_node *iv, *rc; @@ -1239,7 +1264,8 @@ 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 */ @@ -1247,7 +1273,8 @@ 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) { +static void clear_and_fix(ir_node *irn, void *env) +{ int *moved = env; set_irn_link(irn, NULL); @@ -1264,13 +1291,10 @@ static void clear_and_fix(ir_node *irn, void *env) { /* 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; + int projs_moved; FIRM_DBG_REGISTER(dbg, "firm.opt.remove_phi"); @@ -1290,8 +1314,8 @@ 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); @@ -1316,14 +1340,18 @@ void remove_phi_cycles(ir_graph *irg) { 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) +{ + 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) { +static void fix_adds_and_subs(ir_node *irn, void *ctx) +{ (void) ctx; if (is_Add(irn)) { @@ -1375,21 +1403,12 @@ static void fix_adds_and_subs(ir_node *irn, void *ctx) { } /* 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"); DB((dbg, LEVEL_1, "Doing Operator Strength Reduction for %+F\n", irg)); @@ -1434,7 +1453,7 @@ void opt_osr(ir_graph *irg, unsigned flags) { irg_walk_graph(irg, NULL, fix_adds_and_subs, &env); /* try linear function test replacements */ - //lftr(irg, &env); // currently buggy :-( + lftr(irg, &env); (void)lftr; set_irg_outs_inconsistent(irg); @@ -1449,6 +1468,28 @@ void opt_osr(ir_graph *irg, unsigned flags) { if (! edges) edges_deactivate(irg); - - current_ir_graph = rem; } /* opt_osr */ + +struct pass_t { + ir_graph_pass_t pass; + unsigned flags; +}; + +/** +* Wrapper for running opt_osr() as an ir_graph pass. +*/ +static int pass_wrapper(ir_graph *irg, void *context) +{ + struct pass_t *pass = 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 = XMALLOCZ(struct pass_t); + + pass->flags = flags; + return def_graph_pass_constructor( + &pass->pass, name ? name : "osr", pass_wrapper); +} /* opt_osr_pass */