X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fcombo.c;h=73c54fd2eede2b647b23cad32fbe46b1eeff6307;hb=6191bc0fe9dc62446b0e0c15e9179b13161dc63d;hp=e1dd01474af123607489aeab08eb8ae663a811dd;hpb=468f669985c53e7c071d136564147c39a0b9948f;p=libfirm diff --git a/ir/opt/combo.c b/ir/opt/combo.c index e1dd01474..73c54fd2e 100644 --- a/ir/opt/combo.c +++ b/ir/opt/combo.c @@ -63,7 +63,6 @@ #include #include "iroptimize.h" -#include "archop.h" #include "irflag.h" #include "ircons.h" #include "list.h" @@ -81,7 +80,8 @@ #include "debug.h" #include "array_t.h" #include "error.h" - +#include "irnodeset.h" +#include "irpass.h" #include "tv_t.h" #include "irprintf.h" @@ -93,10 +93,6 @@ /* define this to check the consistency of partitions */ #define CHECK_PARTITIONS - -/* allow optimization of non-strict programs */ -#define WITH_UNKNOWN - typedef struct node_t node_t; typedef struct partition_t partition_t; typedef struct opcode_key_t opcode_key_t; @@ -115,6 +111,10 @@ struct opcode_key_t { union { long proj; /**< For Proj nodes, its proj number */ ir_entity *ent; /**< For Sel Nodes, its entity */ + int intVal; /**< For Conv/Div Nodes: strict/remainderless */ + unsigned uintVal;/**< for Builtin: the kind */ + ir_node *block; /**< for Block: itself */ + void *ptr; /**< generic pointer for hash/cmp */ } u; }; @@ -160,7 +160,6 @@ struct node_t { unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */ unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */ unsigned is_follower:1; /**< Set, if this node is a follower. */ - unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */ unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */ }; @@ -197,11 +196,14 @@ typedef struct environment_t { partition_t *touched; /**< the touched set. */ partition_t *initial; /**< The initial partition. */ set *opcode2id_map; /**< The opcodeMode->id map. */ - pmap *type2id_map; /**< The type->id map. */ + ir_node **kept_memory; /**< Array of memory nodes that must be kept. */ int end_idx; /**< -1 for local and 0 for global congruences. */ int lambda_input; /**< Captured argument for lambda_partition(). */ - char modified; /**< Set, if the graph was modified. */ - char commutative; /**< Set, if commutation nodes should be handled specially. */ + unsigned modified:1; /**< Set, if the graph was modified. */ + unsigned unopt_cf:1; /**< If set, control flow is not optimized due to Unknown. */ + /* options driving the optimization */ + unsigned commutative:1; /**< Set, if commutation nodes should be handled specially. */ + unsigned opt_unknown:1; /**< Set, if non-strict programs should be optimized. */ #ifdef DEBUG_libfirm partition_t *dbg_list; /**< List of all partitions. */ #endif @@ -210,8 +212,8 @@ typedef struct environment_t { /** Type of the what function. */ typedef void *(*what_func)(const node_t *node, environment_t *env); -#define get_irn_node(follower) ((node_t *)get_irn_link(follower)) -#define set_irn_node(follower, node) set_irn_link(follower, node) +#define get_irn_node(irn) ((node_t *)get_irn_link(irn)) +#define set_irn_node(irn, node) set_irn_link(irn, node) /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */ #undef tarval_unreachable @@ -227,12 +229,8 @@ DEBUG_ONLY(static const char *what_reason;) /** Next partition number. */ DEBUG_ONLY(static unsigned part_nr = 0); -/** The tarval returned by Unknown nodes. */ -#ifdef WITH_UNKNOWN -#define tarval_UNKNOWN tarval_top -#else -#define tarval_UNKNOWN tarval_bad -#endif +/** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */ +static tarval *tarval_UNKNOWN; /* forward */ static node_t *identity(node_t *node); @@ -241,7 +239,8 @@ static node_t *identity(node_t *node); /** * Check a partition. */ -static void check_partition(const partition_t *T) { +static void check_partition(const partition_t *T) +{ node_t *node; unsigned n = 0; @@ -263,7 +262,8 @@ static void check_partition(const partition_t *T) { /** * check that all leader nodes in the partition have the same opcode. */ -static void check_opcode(const partition_t *Z) { +static void check_opcode(const partition_t *Z) +{ node_t *node; opcode_key_t key; int first = 1; @@ -285,12 +285,27 @@ static void check_opcode(const partition_t *Z) { case iro_Sel: key.u.ent = get_Sel_entity(irn); break; + case iro_Conv: + key.u.intVal = get_Conv_strict(irn); + break; + case iro_Div: + key.u.intVal = get_Div_no_remainder(irn); + break; + case iro_Block: + key.u.block = irn; + break; + case iro_Load: + key.mode = get_Load_mode(irn); + break; + case iro_Builtin: + key.u.intVal = get_Builtin_kind(irn); + break; default: break; } first = 0; } else { - assert(key.code == get_irn_opcode(irn)); + assert((unsigned)key.code == get_irn_opcode(irn)); assert(key.mode == get_irn_mode(irn)); assert(key.arity == get_irn_arity(irn)); @@ -301,6 +316,21 @@ static void check_opcode(const partition_t *Z) { case iro_Sel: assert(key.u.ent == get_Sel_entity(irn)); break; + case iro_Conv: + assert(key.u.intVal == get_Conv_strict(irn)); + break; + case iro_Div: + assert(key.u.intVal == get_Div_no_remainder(irn)); + break; + case iro_Block: + assert(key.u.block == irn); + break; + case iro_Load: + assert(key.mode == get_Load_mode(irn)); + break; + case iro_Builtin: + assert(key.u.intVal == (int) get_Builtin_kind(irn)); + break; default: break; } @@ -308,7 +338,8 @@ static void check_opcode(const partition_t *Z) { } } /* check_opcode */ -static void check_all_partitions(environment_t *env) { +static void check_all_partitions(environment_t *env) +{ #ifdef DEBUG_libfirm partition_t *P; node_t *node; @@ -323,26 +354,36 @@ static void check_all_partitions(environment_t *env) { assert(leader != node && leader->part == node->part); } } +#else + (void) env; #endif } /** * Check list. */ -static void do_check_list(const node_t *list, int ofs, const partition_t *Z) { - const node_t *e; +static void do_check_list(const node_t *list, int ofs, const partition_t *Z) +{ +#ifndef NDEBUG + const node_t *e; #define NEXT(e) *((const node_t **)((char *)(e) + (ofs))) for (e = list; e != NULL; e = NEXT(e)) { assert(e->part == Z); } #undef NEXT +#else + (void) list; + (void) ofs; + (void) Z; +#endif } /* ido_check_list */ /** * Check a local list. */ -static void check_list(const node_t *list, const partition_t *Z) { +static void check_list(const node_t *list, const partition_t *Z) +{ do_check_list(list, offsetof(node_t, next), Z); } /* check_list */ @@ -358,7 +399,8 @@ static inline lattice_elem_t get_partition_type(const partition_t *X); /** * Dump partition to output. */ -static void dump_partition(const char *msg, const partition_t *part) { +static void dump_partition(const char *msg, const partition_t *part) +{ const node_t *node; int first = 1; lattice_elem_t type = get_partition_type(part); @@ -384,7 +426,8 @@ static void dump_partition(const char *msg, const partition_t *part) { /** * Dumps a list. */ -static void do_dump_list(const char *msg, const node_t *node, int ofs) { +static void do_dump_list(const char *msg, const node_t *node, int ofs) +{ const node_t *p; int first = 1; @@ -403,21 +446,24 @@ static void do_dump_list(const char *msg, const node_t *node, int ofs) { /** * Dumps a race list. */ -static void dump_race_list(const char *msg, const node_t *list) { +static void dump_race_list(const char *msg, const node_t *list) +{ do_dump_list(msg, list, offsetof(node_t, race_next)); } /* dump_race_list */ /** * Dumps a local list. */ -static void dump_list(const char *msg, const node_t *list) { +static void dump_list(const char *msg, const node_t *list) +{ do_dump_list(msg, list, offsetof(node_t, next)); } /* dump_list */ /** * Dump all partitions. */ -static void dump_all_partitions(const environment_t *env) { +static void dump_all_partitions(const environment_t *env) +{ const partition_t *P; DB((dbg, LEVEL_2, "All partitions\n===============\n")); @@ -428,7 +474,8 @@ static void dump_all_partitions(const environment_t *env) { /** * Sump a split list. */ -static void dump_split_list(const partition_t *list) { +static void dump_split_list(const partition_t *list) +{ const partition_t *p; DB((dbg, LEVEL_2, "Split by %s produced = {\n", what_reason)); @@ -440,7 +487,8 @@ static void dump_split_list(const partition_t *list) { /** * Dump partition and type for a node. */ -static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) { +static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) +{ ir_node *irn = local != NULL ? local : n; node_t *node = get_irn_node(irn); @@ -460,7 +508,8 @@ static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) { /** * Verify that a type transition is monotone */ -static void do_verify_type(const lattice_elem_t old_type, node_t *node) { +static void verify_type(const lattice_elem_t old_type, node_t *node) +{ if (old_type.tv == node->type.tv) { /* no change */ return; @@ -476,17 +525,15 @@ static void do_verify_type(const lattice_elem_t old_type, node_t *node) { panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node); } /* verify_type */ -#define decl_verify(node) lattice_elem_t old_type = (node)->type; -#define verify_type(node) do_verify_type(old_type, node) #else -#define decl_verify(node) (void)0; -#define verify_type(node) +#define verify_type(old_type, node) #endif /** * Compare two pointer values of a listmap. */ -static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) { +static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) +{ const listmap_entry_t *e1 = elt; const listmap_entry_t *e2 = key; @@ -499,7 +546,8 @@ static int listmap_cmp_ptr(const void *elt, const void *key, size_t size) { * * @param map the listmap */ -static void listmap_init(listmap_t *map) { +static void listmap_init(listmap_t *map) +{ map->map = new_set(listmap_cmp_ptr, 16); map->values = NULL; } /* listmap_init */ @@ -509,7 +557,8 @@ static void listmap_init(listmap_t *map) { * * @param map the listmap */ -static void listmap_term(listmap_t *map) { +static void listmap_term(listmap_t *map) +{ del_set(map->map); } /* listmap_term */ @@ -521,7 +570,8 @@ static void listmap_term(listmap_t *map) { * * @return the associated listmap entry for the given id */ -static listmap_entry_t *listmap_find(listmap_t *map, void *id) { +static listmap_entry_t *listmap_find(listmap_t *map, void *id) +{ listmap_entry_t key, *entry; key.id = id; @@ -544,27 +594,32 @@ static listmap_entry_t *listmap_find(listmap_t *map, void *id) { * * @return a hash value for the given opcode map entry */ -static unsigned opcode_hash(const opcode_key_t *entry) { - return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent) + entry->arity; +static unsigned opcode_hash(const opcode_key_t *entry) +{ + return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ptr) + entry->arity; } /* opcode_hash */ /** * Compare two entries in the opcode map. */ -static int cmp_opcode(const void *elt, const void *key, size_t size) { +static int cmp_opcode(const void *elt, const void *key, size_t size) +{ const opcode_key_t *o1 = elt; const opcode_key_t *o2 = key; (void) size; return o1->code != o2->code || o1->mode != o2->mode || o1->arity != o2->arity || - o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent; + o1->u.proj != o2->u.proj || + o1->u.intVal != o2->u.intVal || /* this already checks uIntVal */ + o1->u.ptr != o2->u.ptr; } /* cmp_opcode */ /** * Compare two Def-Use edges for input position. */ -static int cmp_def_use_edge(const void *a, const void *b) { +static int cmp_def_use_edge(const void *a, const void *b) +{ const ir_def_use_edge *ea = a; const ir_def_use_edge *eb = b; @@ -575,7 +630,8 @@ static int cmp_def_use_edge(const void *a, const void *b) { /** * We need the Def-Use edges sorted. */ -static void sort_irn_outs(node_t *node) { +static void sort_irn_outs(node_t *node) +{ ir_node *irn = node->node; int n_outs = get_irn_n_outs(irn); @@ -592,7 +648,8 @@ static void sort_irn_outs(node_t *node) { * * @return the associated type of this node */ -static inline lattice_elem_t get_node_type(const ir_node *irn) { +static inline lattice_elem_t get_node_type(const ir_node *irn) +{ return get_irn_node(irn)->type; } /* get_node_type */ @@ -603,7 +660,8 @@ static inline lattice_elem_t get_node_type(const ir_node *irn) { * * @return the associated type of this node */ -static inline tarval *get_node_tarval(const ir_node *irn) { +static inline tarval *get_node_tarval(const ir_node *irn) +{ lattice_elem_t type = get_node_type(irn); if (is_tarval(type.tv)) @@ -614,7 +672,8 @@ static inline tarval *get_node_tarval(const ir_node *irn) { /** * Add a partition to the worklist. */ -static inline void add_to_worklist(partition_t *X, environment_t *env) { +static inline void add_to_worklist(partition_t *X, environment_t *env) +{ assert(X->on_worklist == 0); DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr)); X->wl_next = env->worklist; @@ -629,8 +688,9 @@ static inline void add_to_worklist(partition_t *X, environment_t *env) { * * @return a newly allocated partition */ -static inline partition_t *new_partition(environment_t *env) { - partition_t *part = obstack_alloc(&env->obst, sizeof(*part)); +static inline partition_t *new_partition(environment_t *env) +{ + partition_t *part = OALLOC(&env->obst, partition_t); INIT_LIST_HEAD(&part->Leader); INIT_LIST_HEAD(&part->Follower); @@ -660,7 +720,8 @@ static inline partition_t *new_partition(environment_t *env) { /** * Get the first node from a partition. */ -static inline node_t *get_first_node(const partition_t *X) { +static inline node_t *get_first_node(const partition_t *X) +{ return list_entry(X->Leader.next, node_t, node_list); } /* get_first_node */ @@ -672,7 +733,8 @@ static inline node_t *get_first_node(const partition_t *X) { * * @return the type of the first element of the partition */ -static inline lattice_elem_t get_partition_type(const partition_t *X) { +static inline lattice_elem_t get_partition_type(const partition_t *X) +{ const node_t *first = get_first_node(X); return first->type; } /* get_partition_type */ @@ -687,9 +749,10 @@ static inline lattice_elem_t get_partition_type(const partition_t *X) { * * @return the created node */ -static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) { +static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) +{ /* create a partition node and place it in the partition */ - node_t *node = obstack_alloc(&env->obst, sizeof(*node)); + node_t *node = OALLOC(&env->obst, node_t); INIT_LIST_HEAD(&node->node_list); INIT_LIST_HEAD(&node->cprop_list); @@ -705,7 +768,6 @@ static node_t *create_partition_node(ir_node *irn, partition_t *part, environmen node->on_cprop = 0; node->on_fallen = 0; node->is_follower = 0; - node->by_all_const = 0; node->flagged = 0; set_irn_node(irn, node); @@ -719,7 +781,8 @@ static node_t *create_partition_node(ir_node *irn, partition_t *part, environmen * Pre-Walker, initialize all Nodes' type to U or top and place * all nodes into the TOP partition. */ -static void create_initial_partitions(ir_node *irn, void *ctx) { +static void create_initial_partitions(ir_node *irn, void *ctx) +{ environment_t *env = ctx; partition_t *part = env->initial; node_t *node; @@ -737,7 +800,8 @@ static void create_initial_partitions(ir_node *irn, void *ctx) { /** * Post-Walker, collect all Block-Phi lists, set Cond. */ -static void init_block_phis(ir_node *irn, void *ctx) { +static void init_block_phis(ir_node *irn, void *ctx) +{ (void) ctx; if (is_Phi(irn)) { @@ -752,7 +816,8 @@ static void init_block_phis(ir_node *irn, void *ctx) { * @param y a node * @param env the environment */ -static inline void add_to_touched(node_t *y, environment_t *env) { +static inline void add_to_touched(node_t *y, environment_t *env) +{ if (y->on_touched == 0) { partition_t *part = y->part; @@ -777,15 +842,17 @@ static inline void add_to_touched(node_t *y, environment_t *env) { * @param y the node * @param env the environment */ -static void add_to_cprop(node_t *y, environment_t *env) { +static void add_to_cprop(node_t *y, environment_t *env) +{ ir_node *irn; /* Add y to y.partition.cprop. */ if (y->on_cprop == 0) { partition_t *Y = y->part; + ir_node *irn = y->node; - /* place Conds and its Proj nodes on the cprop_X list */ - if (is_Cond(skip_Proj(y->node))) + /* place Conds and all its Projs on the cprop_X list */ + if (is_Cond(skip_Proj(irn))) list_add_tail(&y->cprop_list, &Y->cprop_X); else list_add_tail(&y->cprop_list, &Y->cprop); @@ -831,7 +898,8 @@ static void add_to_cprop(node_t *y, environment_t *env) { * @param Z_prime the Z' partition, a previous part of Z * @param env the environment */ -static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) { +static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) +{ if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) { add_to_worklist(Z_prime, env); } else { @@ -844,7 +912,8 @@ static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t * * @param x the node */ -static void move_edges_to_leader(node_t *x) { +static void move_edges_to_leader(node_t *x) +{ ir_node *irn = x->node; int i, j, k; @@ -889,7 +958,8 @@ static void move_edges_to_leader(node_t *x) { * * @return a new partition containing the nodes of g */ -static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) { +static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) +{ partition_t *Z_prime; node_t *node; unsigned n = 0; @@ -939,7 +1009,8 @@ static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t * * @param n the node */ -static void follower_to_leader(node_t *n) { +static void follower_to_leader(node_t *n) +{ assert(n->is_follower == 1); DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node)); @@ -967,7 +1038,8 @@ typedef struct step_env { * @param irn the node to check * @param input number of the input */ -static int is_real_follower(const ir_node *irn, int input) { +static int is_real_follower(const ir_node *irn, int input) +{ node_t *pred; switch (get_irn_opcode(irn)) { @@ -1019,10 +1091,6 @@ static int is_real_follower(const ir_node *irn, int input) { if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv)) return 0; break; - case iro_Min: - case iro_Max: - /* all inputs are followers */ - return 1; default: assert(!"opcode not implemented yet"); break; @@ -1033,7 +1101,8 @@ static int is_real_follower(const ir_node *irn, int input) { /** * Do one step in the race. */ -static int step(step_env *env) { +static int step(step_env *env) +{ node_t *n; if (env->initial != NULL) { @@ -1100,7 +1169,8 @@ static int step(step_env *env) { * * @param list the list */ -static int clear_flags(node_t *list) { +static int clear_flags(node_t *list) +{ int res = 0; node_t *n; @@ -1125,7 +1195,8 @@ static int clear_flags(node_t *list) { * * @return a new partition containing the nodes of gg */ -static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { +static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) +{ partition_t *X = *pX; partition_t *X_prime; list_head tmp; @@ -1298,7 +1369,8 @@ static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { * * @return non-zero if the i'th input of the given Phi node is live */ -static int is_live_input(ir_node *phi, int i) { +static int is_live_input(ir_node *phi, int i) +{ if (i >= 0) { ir_node *block = get_nodes_block(phi); ir_node *pred = get_Block_cfgpred(block, i); @@ -1313,7 +1385,8 @@ static int is_live_input(ir_node *phi, int i) { /** * Return non-zero if a type is a constant. */ -static int is_constant_type(lattice_elem_t type) { +static int is_constant_type(lattice_elem_t type) +{ if (type.tv != tarval_bottom && type.tv != tarval_top) return 1; return 0; @@ -1325,7 +1398,8 @@ static int is_constant_type(lattice_elem_t type) { * * @param type the type to check */ -static int type_is_neither_top_nor_const(const lattice_elem_t type) { +static int type_is_neither_top_nor_const(const lattice_elem_t type) +{ if (is_tarval(type.tv)) { if (type.tv == tarval_top) return 0; @@ -1345,7 +1419,8 @@ static int type_is_neither_top_nor_const(const lattice_elem_t type) { * @param idx the index of the def_use edge to evaluate * @param env the environment */ -static void collect_touched(list_head *list, int idx, environment_t *env) { +static void collect_touched(list_head *list, int idx, environment_t *env) +{ node_t *x, *y; int end_idx = env->end_idx; @@ -1410,7 +1485,8 @@ static void collect_touched(list_head *list, int idx, environment_t *env) { * @param list the list which contains the nodes that must be evaluated * @param env the environment */ -static void collect_commutative_touched(list_head *list, environment_t *env) { +static void collect_commutative_touched(list_head *list, environment_t *env) +{ node_t *x, *y; list_for_each_entry(node_t, x, list, node_list) { @@ -1460,7 +1536,8 @@ static void collect_commutative_touched(list_head *list, environment_t *env) { * * @param env the environment */ -static void cause_splits(environment_t *env) { +static void cause_splits(environment_t *env) +{ partition_t *X, *Z, *N; int idx; @@ -1481,9 +1558,12 @@ static void cause_splits(environment_t *env) { collect_commutative_touched(&X->Follower, env); for (Z = env->touched; Z != NULL; Z = N) { - node_t *e; - node_t *touched = Z->touched; - unsigned n_touched = Z->n_touched; + node_t *e, *n; + node_t *touched = Z->touched; + node_t *touched_aa = NULL; + node_t *touched_ab = NULL; + unsigned n_touched_aa = 0; + unsigned n_touched_ab = 0; assert(Z->touched != NULL); @@ -1494,18 +1574,45 @@ static void cause_splits(environment_t *env) { Z->on_touched = 0; /* Empty local Z.touched. */ - for (e = touched; e != NULL; e = e->next) { + for (e = touched; e != NULL; e = n) { + node_t *left = get_irn_node(get_irn_n(e->node, 0)); + node_t *right = get_irn_node(get_irn_n(e->node, 1)); + assert(e->is_follower == 0); e->on_touched = 0; + n = e->next; + + /* + * Note: op(a, a) is NOT congruent to op(a, b). + * So, we must split the touched list. + */ + if (left->part == right->part) { + e->next = touched_aa; + touched_aa = e; + ++n_touched_aa; + } else { + e->next = touched_ab; + touched_ab = e; + ++n_touched_ab; + } } + assert(n_touched_aa + n_touched_ab == Z->n_touched); Z->touched = NULL; Z->n_touched = 0; - if (0 < n_touched && n_touched < Z->n_leader) { - DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr)); - split(&Z, touched, env); + if (0 < n_touched_aa && n_touched_aa < Z->n_leader) { + partition_t *Z_prime = Z; + DB((dbg, LEVEL_2, "Split part%d by touched_aa\n", Z_prime->nr)); + split(&Z_prime, touched_aa, env); } else - assert(n_touched <= Z->n_leader); + assert(n_touched_aa <= Z->n_leader); + + if (0 < n_touched_ab && n_touched_ab < Z->n_leader) { + partition_t *Z_prime = Z; + DB((dbg, LEVEL_2, "Split part%d by touched_ab\n", Z_prime->nr)); + split(&Z_prime, touched_ab, env); + } else + assert(n_touched_ab <= Z->n_leader); } } @@ -1559,7 +1666,8 @@ static void cause_splits(environment_t *env) { * @return *P */ static partition_t *split_by_what(partition_t *X, what_func What, - partition_t **P, environment_t *env) { + partition_t **P, environment_t *env) +{ node_t *x, *S; listmap_t map; listmap_entry_t *iter; @@ -1605,13 +1713,15 @@ static partition_t *split_by_what(partition_t *X, what_func What, } /* split_by_what */ /** lambda n.(n.type) */ -static void *lambda_type(const node_t *node, environment_t *env) { +static void *lambda_type(const node_t *node, environment_t *env) +{ (void)env; return node->type.tv; } /* lambda_type */ /** lambda n.(n.opcode) */ -static void *lambda_opcode(const node_t *node, environment_t *env) { +static void *lambda_opcode(const node_t *node, environment_t *env) +{ opcode_key_t key, *entry; ir_node *irn = node->node; @@ -1628,6 +1738,27 @@ static void *lambda_opcode(const node_t *node, environment_t *env) { case iro_Sel: key.u.ent = get_Sel_entity(irn); break; + case iro_Conv: + key.u.intVal = get_Conv_strict(irn); + break; + case iro_Div: + key.u.intVal = get_Div_no_remainder(irn); + break; + case iro_Block: + /* + * Some ugliness here: Two Blocks having the same + * IJmp predecessor would be congruent, which of course is wrong. + * We fix it by never letting blocks be congruent + * which cannot be detected by combo either. + */ + key.u.block = irn; + break; + case iro_Load: + key.mode = get_Load_mode(irn); + break; + case iro_Builtin: + key.u.intVal = get_Builtin_kind(irn); + break; default: break; } @@ -1637,7 +1768,8 @@ static void *lambda_opcode(const node_t *node, environment_t *env) { } /* lambda_opcode */ /** lambda n.(n[i].partition) */ -static void *lambda_partition(const node_t *node, environment_t *env) { +static void *lambda_partition(const node_t *node, environment_t *env) +{ ir_node *skipped = skip_Proj(node->node); ir_node *pred; node_t *p; @@ -1661,12 +1793,12 @@ static void *lambda_partition(const node_t *node, environment_t *env) { pred = i == -1 ? get_irn_n(skipped, i) : get_irn_n(node->node, i); p = get_irn_node(pred); - return p->part; } /* lambda_partition */ /** lambda n.(n[i].partition) for commutative nodes */ -static void *lambda_commutative_partition(const node_t *node, environment_t *env) { +static void *lambda_commutative_partition(const node_t *node, environment_t *env) +{ ir_node *irn = node->node; ir_node *skipped = skip_Proj(irn); ir_node *pred, *left, *right; @@ -1721,7 +1853,8 @@ static void *lambda_commutative_partition(const node_t *node, environment_t *env * Returns true if a type is a constant (and NOT Top * or Bottom). */ -static int is_con(const lattice_elem_t type) { +static int is_con(const lattice_elem_t type) +{ /* be conservative */ if (is_tarval(type.tv)) return tarval_is_constant(type.tv); @@ -1734,7 +1867,8 @@ static int is_con(const lattice_elem_t type) { * @param X the partition to split * @param env the environment */ -static void split_by(partition_t *X, environment_t *env) { +static void split_by(partition_t *X, environment_t *env) +{ partition_t *I, *P = NULL; int input; @@ -1823,15 +1957,10 @@ static void split_by(partition_t *X, environment_t *env) { * * @param node the node */ -static void default_compute(node_t *node) { +static void default_compute(node_t *node) +{ int i; ir_node *irn = node->node; - node_t *block = get_irn_node(get_nodes_block(irn)); - - if (block->type.tv == tarval_unreachable) { - node->type.tv = tarval_top; - return; - } /* if any of the data inputs have type top, the result is type top */ for (i = get_irn_arity(irn) - 1; i >= 0; --i) { @@ -1855,12 +1984,13 @@ static void default_compute(node_t *node) { * * @param node the node */ -static void compute_Block(node_t *node) { +static void compute_Block(node_t *node) +{ int i; ir_node *block = node->node; - if (block == get_irg_start_block(current_ir_graph)) { - /* start block is always reachable */ + if (block == get_irg_start_block(current_ir_graph) || has_Block_entity(block)) { + /* start block and labelled blocks are always reachable */ node->type.tv = tarval_reachable; return; } @@ -1882,7 +2012,8 @@ static void compute_Block(node_t *node) { * * @param node the node */ -static void compute_Bad(node_t *node) { +static void compute_Bad(node_t *node) +{ /* Bad nodes ALWAYS compute Top */ node->type.tv = tarval_top; } /* compute_Bad */ @@ -1892,7 +2023,8 @@ static void compute_Bad(node_t *node) { * * @param node the node */ -static void compute_Unknown(node_t *node) { +static void compute_Unknown(node_t *node) +{ /* While Unknown nodes should compute Top this is dangerous: * a Top input to a Cond would lead to BOTH control flows unreachable. * While this is correct in the given semantics, it would destroy the Firm @@ -1901,7 +2033,7 @@ static void compute_Unknown(node_t *node) { * It would be safe to compute Top IF it can be assured, that only Cmp * nodes are inputs to Conds. We check that first. * This is the way Frontends typically build Firm, but some optimizations - * (cond_eval for instance) might replace them by Phib's... + * (jump threading for instance) might replace them by Phib's... */ node->type.tv = tarval_UNKNOWN; } /* compute_Unknown */ @@ -1911,7 +2043,8 @@ static void compute_Unknown(node_t *node) { * * @param node the node */ -static void compute_Jmp(node_t *node) { +static void compute_Jmp(node_t *node) +{ node_t *block = get_irn_node(get_nodes_block(node->node)); node->type = block->type; @@ -1922,7 +2055,8 @@ static void compute_Jmp(node_t *node) { * * @param node the node */ -static void compute_Return(node_t *node) { +static void compute_Return(node_t *node) +{ /* The Return node is NOT dead if it is in a reachable block. * This is already checked in compute(). so we can return * Reachable here. */ @@ -1934,17 +2068,33 @@ static void compute_Return(node_t *node) { * * @param node the node */ -static void compute_End(node_t *node) { +static void compute_End(node_t *node) +{ /* the End node is NOT dead of course */ node->type.tv = tarval_reachable; } /* compute_End */ +/** + * (Re-)compute the type for a Call. + * + * @param node the node + */ +static void compute_Call(node_t *node) +{ + /* + * A Call computes always bottom, even if it has Unknown + * predecessors. + */ + node->type.tv = tarval_bottom; +} /* compute_Call */ + /** * (Re-)compute the type for a SymConst node. * * @param node the node */ -static void compute_SymConst(node_t *node) { +static void compute_SymConst(node_t *node) +{ ir_node *irn = node->node; node_t *block = get_irn_node(get_nodes_block(irn)); @@ -1954,7 +2104,6 @@ static void compute_SymConst(node_t *node) { } switch (get_SymConst_kind(irn)) { case symconst_addr_ent: - /* case symconst_addr_name: cannot handle this yet */ node->type.sym = get_SymConst_symbol(irn); break; default: @@ -1967,7 +2116,8 @@ static void compute_SymConst(node_t *node) { * * @param node the node */ -static void compute_Phi(node_t *node) { +static void compute_Phi(node_t *node) +{ int i; ir_node *phi = node->node; lattice_elem_t type; @@ -2014,7 +2164,8 @@ static void compute_Phi(node_t *node) { * * @param node the node */ -static void compute_Add(node_t *node) { +static void compute_Add(node_t *node) +{ ir_node *sub = node->node; node_t *l = get_irn_node(get_Add_left(sub)); node_t *r = get_irn_node(get_Add_right(sub)); @@ -2055,7 +2206,8 @@ static void compute_Add(node_t *node) { * * @param node the node */ -static void compute_Sub(node_t *node) { +static void compute_Sub(node_t *node) +{ ir_node *sub = node->node; node_t *l = get_irn_node(get_Sub_left(sub)); node_t *r = get_irn_node(get_Sub_right(sub)); @@ -2075,7 +2227,6 @@ static void compute_Sub(node_t *node) { } else { node->type.tv = tarval_bottom; } - node->by_all_const = 1; } else if (r->part == l->part && (!mode_is_float(get_irn_mode(l->node)))) { /* @@ -2089,7 +2240,7 @@ static void compute_Sub(node_t *node) { this breaks AND we get from the argument partitions a different result, switch to bottom. This happens because initially all nodes are in the same partition ... */ - if (node->by_all_const && node->type.tv != tv) + if (node->type.tv != tv) tv = tarval_bottom; node->type.tv = tv; } else { @@ -2102,7 +2253,8 @@ static void compute_Sub(node_t *node) { * * @param node the node */ -static void compute_Eor(node_t *node) { +static void compute_Eor(node_t *node) +{ ir_node *eor = node->node; node_t *l = get_irn_node(get_Eor_left(eor)); node_t *r = get_irn_node(get_Eor_right(eor)); @@ -2122,7 +2274,6 @@ static void compute_Eor(node_t *node) { } else { node->type.tv = tarval_bottom; } - node->by_all_const = 1; } else if (r->part == l->part) { ir_mode *mode = get_irn_mode(eor); tv = get_mode_null(mode); @@ -2131,7 +2282,7 @@ static void compute_Eor(node_t *node) { this breaks AND we get from the argument partitions a different result, switch to bottom. This happens because initially all nodes are in the same partition ... */ - if (node->by_all_const && node->type.tv != tv) + if (node->type.tv != tv) tv = tarval_bottom; node->type.tv = tv; } else { @@ -2144,7 +2295,8 @@ static void compute_Eor(node_t *node) { * * @param node the node */ -static void compute_Cmp(node_t *node) { +static void compute_Cmp(node_t *node) +{ ir_node *cmp = node->node; node_t *l = get_irn_node(get_Cmp_left(cmp)); node_t *r = get_irn_node(get_Cmp_right(cmp)); @@ -2170,7 +2322,8 @@ static void compute_Cmp(node_t *node) { * @param node the node * @param cond the predecessor Cmp node */ -static void compute_Proj_Cmp(node_t *node, ir_node *cmp) { +static void compute_Proj_Cmp(node_t *node, ir_node *cmp) +{ ir_node *proj = node->node; node_t *l = get_irn_node(get_Cmp_left(cmp)); node_t *r = get_irn_node(get_Cmp_right(cmp)); @@ -2183,7 +2336,6 @@ static void compute_Proj_Cmp(node_t *node, ir_node *cmp) { node->type.tv = tarval_undefined; } else if (is_con(a) && is_con(b)) { default_compute(node); - node->by_all_const = 1; } else if (r->part == l->part && (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) { /* @@ -2196,7 +2348,7 @@ static void compute_Proj_Cmp(node_t *node, ir_node *cmp) { this breaks AND we get from the argument partitions a different result, switch to bottom. This happens because initially all nodes are in the same partition ... */ - if (node->by_all_const && node->type.tv != tv) + if (node->type.tv != tv) tv = tarval_bottom; node->type.tv = tv; } else { @@ -2210,7 +2362,8 @@ static void compute_Proj_Cmp(node_t *node, ir_node *cmp) { * @param node the node * @param cond the predecessor Cond node */ -static void compute_Proj_Cond(node_t *node, ir_node *cond) { +static void compute_Proj_Cond(node_t *node, ir_node *cond) +{ ir_node *proj = node->node; long pnc = get_Proj_proj(proj); ir_node *sel = get_Cond_selector(cond); @@ -2236,9 +2389,38 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { * * Note that this even happens with Click's original algorithm, if * Cmp(x, 0) is evaluated to True first and later changed to False - * if x was Top first tiome and later Const ... + * if x was Top first and later changed to a Const ... * It is unclear how Click solved that problem ... + * + * However, in rare cases even this does not help, if a Top reaches + * a compare through a Phi, than Proj(Cond) is evaluated changing + * the type of the Phi to something other. + * So, we take the last resort and bind the type to R once + * it is calculated. + * + * (This might be even the way Click works around the whole problem). + * + * Finally, we may miss some optimization possibilities due to this: + * + * x = phi(Top, y) + * if (x == 0) + * + * If Top reaches the if first, than we decide for != here. + * If y later is evaluated to 0, we cannot revert this decision + * and must live with both outputs enabled. If this happens, + * we get an unresolved if (true) in the code ... + * + * In Click's version where this decision is done at the Cmp, + * the Cmp is NOT optimized away than (if y evaluated to 1 + * for instance) and we get a if (1 == 0) here ... + * + * Both solutions are suboptimal. + * At least, we could easily detect this problem and run + * cf_opt() (or even combo) again :-( */ + if (node->type.tv == tarval_reachable) + return; + if (get_irn_mode(sel) == mode_b) { /* an IF */ if (pnc == pn_Cond_true) { @@ -2250,8 +2432,12 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { node->type.tv = tarval_reachable; } else { assert(selector->type.tv == tarval_top); - /* any condition based on Top is "!=" */ - node->type.tv = tarval_unreachable; + if (tarval_UNKNOWN == tarval_top) { + /* any condition based on Top is "!=" */ + node->type.tv = tarval_unreachable; + } else { + node->type.tv = tarval_unreachable; + } } } else { assert(pnc == pn_Cond_false); @@ -2264,8 +2450,12 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { node->type.tv = tarval_reachable; } else { assert(selector->type.tv == tarval_top); - /* any condition based on Top is "!=" */ - node->type.tv = tarval_reachable; + if (tarval_UNKNOWN == tarval_top) { + /* any condition based on Top is "!=" */ + node->type.tv = tarval_reachable; + } else { + node->type.tv = tarval_unreachable; + } } } } else { @@ -2273,14 +2463,16 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { if (selector->type.tv == tarval_bottom) { node->type.tv = tarval_reachable; } else if (selector->type.tv == tarval_top) { - if (pnc == get_Cond_defaultProj(cond)) { + if (tarval_UNKNOWN == tarval_top && + pnc == get_Cond_default_proj(cond)) { /* a switch based of Top is always "default" */ node->type.tv = tarval_reachable; - } else + } else { node->type.tv = tarval_unreachable; + } } else { long value = get_tarval_long(selector->type.tv); - if (pnc == get_Cond_defaultProj(cond)) { + if (pnc == get_Cond_default_proj(cond)) { /* default switch, have to check ALL other cases */ int i; @@ -2310,7 +2502,8 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { * * @param node the node */ -static void compute_Proj(node_t *node) { +static void compute_Proj(node_t *node) +{ ir_node *proj = node->node; ir_mode *mode = get_irn_mode(proj); node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj))); @@ -2360,7 +2553,8 @@ static void compute_Proj(node_t *node) { * * @param node the node */ -static void compute_Confirm(node_t *node) { +static void compute_Confirm(node_t *node) +{ ir_node *confirm = node->node; node_t *pred = get_irn_node(get_Confirm_value(confirm)); @@ -2377,103 +2571,26 @@ static void compute_Confirm(node_t *node) { node->type = pred->type; } /* compute_Confirm */ -/** - * (Re-)compute the type for a Max. - * - * @param node the node - */ -static void compute_Max(node_t *node) { - ir_node *op = node->node; - node_t *l = get_irn_node(get_binop_left(op)); - node_t *r = get_irn_node(get_binop_right(op)); - lattice_elem_t a = l->type; - lattice_elem_t b = r->type; - - if (a.tv == tarval_top || b.tv == tarval_top) { - node->type.tv = tarval_top; - } else if (is_con(a) && is_con(b)) { - /* both nodes are constants, we can probably do something */ - if (a.tv == b.tv) { - /* this case handles SymConsts as well */ - node->type = a; - } else { - ir_mode *mode = get_irn_mode(op); - tarval *tv_min = get_mode_min(mode); - - if (a.tv == tv_min) - node->type = b; - else if (b.tv == tv_min) - node->type = a; - else if (is_tarval(a.tv) && is_tarval(b.tv)) { - if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt) - node->type.tv = a.tv; - else - node->type.tv = b.tv; - } else { - node->type.tv = tarval_bad; - } - } - } else if (r->part == l->part) { - /* both nodes congruent, we can probably do something */ - node->type = a; - } else { - node->type.tv = tarval_bottom; - } -} /* compute_Max */ - -/** - * (Re-)compute the type for a Min. - * - * @param node the node - */ -static void compute_Min(node_t *node) { - ir_node *op = node->node; - node_t *l = get_irn_node(get_binop_left(op)); - node_t *r = get_irn_node(get_binop_right(op)); - lattice_elem_t a = l->type; - lattice_elem_t b = r->type; - - if (a.tv == tarval_top || b.tv == tarval_top) { - node->type.tv = tarval_top; - } else if (is_con(a) && is_con(b)) { - /* both nodes are constants, we can probably do something */ - if (a.tv == b.tv) { - /* this case handles SymConsts as well */ - node->type = a; - } else { - ir_mode *mode = get_irn_mode(op); - tarval *tv_max = get_mode_max(mode); - - if (a.tv == tv_max) - node->type = b; - else if (b.tv == tv_max) - node->type = a; - else if (is_tarval(a.tv) && is_tarval(b.tv)) { - if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt) - node->type.tv = a.tv; - else - node->type.tv = b.tv; - } else { - node->type.tv = tarval_bad; - } - } - } else if (r->part == l->part) { - /* both nodes congruent, we can probably do something */ - node->type = a; - } else { - node->type.tv = tarval_bottom; - } -} /* compute_Min */ - /** * (Re-)compute the type for a given node. * * @param node the node */ -static void compute(node_t *node) { +static void compute(node_t *node) +{ ir_node *irn = node->node; compute_func func; - decl_verify(node); + +#ifndef VERIFY_MONOTONE + /* + * Once a node reaches bottom, the type cannot fall further + * in the lattice and we can stop computation. + * Do not take this exit if the monotony verifier is + * enabled to catch errors. + */ + if (node->type.tv == tarval_bottom) + return; +#endif if (is_no_Block(irn)) { /* for pinned nodes, check its control input */ @@ -2482,7 +2599,7 @@ static void compute(node_t *node) { if (block->type.tv == tarval_unreachable) { node->type.tv = tarval_top; - goto end; + return; } } } @@ -2490,8 +2607,6 @@ static void compute(node_t *node) { func = (compute_func)node->node->op->ops.generic; if (func != NULL) func(node); -end: - verify_type(node); } /* compute */ /* @@ -2505,7 +2620,8 @@ end: /** * Calculates the Identity for Phi nodes */ -static node_t *identity_Phi(node_t *node) { +static node_t *identity_Phi(node_t *node) +{ ir_node *phi = node->node; ir_node *block = get_nodes_block(phi); node_t *n_part = NULL; @@ -2534,7 +2650,8 @@ static node_t *identity_Phi(node_t *node) { /** * Calculates the Identity for commutative 0 neutral nodes. */ -static node_t *identity_comm_zero_binop(node_t *node) { +static node_t *identity_comm_zero_binop(node_t *node) +{ ir_node *op = node->node; node_t *a = get_irn_node(get_binop_left(op)); node_t *b = get_irn_node(get_binop_right(op)); @@ -2558,7 +2675,8 @@ static node_t *identity_comm_zero_binop(node_t *node) { /** * Calculates the Identity for Shift nodes. */ -static node_t *identity_shift(node_t *node) { +static node_t *identity_shift(node_t *node) +{ ir_node *op = node->node; node_t *b = get_irn_node(get_binop_right(op)); ir_mode *mode = get_irn_mode(b->node); @@ -2575,7 +2693,8 @@ static node_t *identity_shift(node_t *node) { /** * Calculates the Identity for Mul nodes. */ -static node_t *identity_Mul(node_t *node) { +static node_t *identity_Mul(node_t *node) +{ ir_node *op = node->node; node_t *a = get_irn_node(get_Mul_left(op)); node_t *b = get_irn_node(get_Mul_right(op)); @@ -2599,7 +2718,8 @@ static node_t *identity_Mul(node_t *node) { /** * Calculates the Identity for Sub nodes. */ -static node_t *identity_Sub(node_t *node) { +static node_t *identity_Sub(node_t *node) +{ ir_node *sub = node->node; node_t *b = get_irn_node(get_Sub_right(sub)); ir_mode *mode = get_irn_mode(sub); @@ -2618,7 +2738,8 @@ static node_t *identity_Sub(node_t *node) { /** * Calculates the Identity for And nodes. */ -static node_t *identity_And(node_t *node) { +static node_t *identity_And(node_t *node) +{ ir_node *and = node->node; node_t *a = get_irn_node(get_And_left(and)); node_t *b = get_irn_node(get_And_right(and)); @@ -2636,7 +2757,8 @@ static node_t *identity_And(node_t *node) { /** * Calculates the Identity for Confirm nodes. */ -static node_t *identity_Confirm(node_t *node) { +static node_t *identity_Confirm(node_t *node) +{ ir_node *confirm = node->node; /* a Confirm is always a Copy */ @@ -2646,7 +2768,8 @@ static node_t *identity_Confirm(node_t *node) { /** * Calculates the Identity for Mux nodes. */ -static node_t *identity_Mux(node_t *node) { +static node_t *identity_Mux(node_t *node) +{ ir_node *mux = node->node; node_t *t = get_irn_node(get_Mux_true(mux)); node_t *f = get_irn_node(get_Mux_false(mux)); @@ -2668,58 +2791,11 @@ static node_t *identity_Mux(node_t *node) { return node; } /* identity_Mux */ -/** - * Calculates the Identity for Min nodes. - */ -static node_t *identity_Min(node_t *node) { - ir_node *op = node->node; - node_t *a = get_irn_node(get_binop_left(op)); - node_t *b = get_irn_node(get_binop_right(op)); - ir_mode *mode = get_irn_mode(op); - tarval *tv_max; - - if (a->part == b->part) { - /* leader of multiple predecessors */ - return a; - } - - /* works even with NaN */ - tv_max = get_mode_max(mode); - if (a->type.tv == tv_max) - return b; - if (b->type.tv == tv_max) - return a; - return node; -} /* identity_Min */ - -/** - * Calculates the Identity for Max nodes. - */ -static node_t *identity_Max(node_t *node) { - ir_node *op = node->node; - node_t *a = get_irn_node(get_binop_left(op)); - node_t *b = get_irn_node(get_binop_right(op)); - ir_mode *mode = get_irn_mode(op); - tarval *tv_min; - - if (a->part == b->part) { - /* leader of multiple predecessors */ - return a; - } - - /* works even with NaN */ - tv_min = get_mode_min(mode); - if (a->type.tv == tv_min) - return b; - if (b->type.tv == tv_min) - return a; - return node; -} /* identity_Max */ - /** * Calculates the Identity for nodes. */ -static node_t *identity(node_t *node) { +static node_t *identity(node_t *node) +{ ir_node *irn = node->node; switch (get_irn_opcode(irn)) { @@ -2744,10 +2820,6 @@ static node_t *identity(node_t *node) { return identity_Confirm(node); case iro_Mux: return identity_Mux(node); - case iro_Min: - return identity_Min(node); - case iro_Max: - return identity_Max(node); default: return node; } @@ -2757,7 +2829,8 @@ static node_t *identity(node_t *node) { * Node follower is a (new) follower of leader, segregate Leader * out edges. */ -static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) { +static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) +{ ir_node *l = leader->node; int j, i, n = get_irn_n_outs(l); @@ -2783,7 +2856,8 @@ static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) { * * @param follower the follower IR node */ -static void segregate_def_use_chain(const ir_node *follower) { +static void segregate_def_use_chain(const ir_node *follower) +{ int i; for (i = get_irn_arity(follower) - 1; i >= 0; --i) { @@ -2798,7 +2872,8 @@ static void segregate_def_use_chain(const ir_node *follower) { * * @param env the environment */ -static void propagate(environment_t *env) { +static void propagate(environment_t *env) +{ partition_t *X, *Y; node_t *x; lattice_elem_t old_type; @@ -2870,6 +2945,7 @@ static void propagate(environment_t *env) { compute(x); if (x->type.tv != old_type.tv) { DB((dbg, LEVEL_2, "node %+F has changed type from %+F to %+F\n", x->node, old_type, x->type)); + verify_type(old_type, x); if (x->on_fallen == 0) { /* Add x to fallen. Nodes might fall from T -> const -> _|_, so check that they are @@ -2935,7 +3011,8 @@ static void propagate(environment_t *env) { * * @param irn the node */ -static ir_node *get_leader(node_t *node) { +static ir_node *get_leader(node_t *node) +{ partition_t *part = node->part; if (part->n_leader > 1 || node->is_follower) { @@ -2950,38 +3027,47 @@ static ir_node *get_leader(node_t *node) { return node->node; } /* get_leader */ +/** + * Returns non-zero if a mode_T node has only one reachable output. + */ +static int only_one_reachable_proj(ir_node *n) +{ + int i, k = 0; + + for (i = get_irn_n_outs(n) - 1; i >= 0; --i) { + ir_node *proj = get_irn_out(n, i); + node_t *node; + + /* skip non-control flow Proj's */ + if (get_irn_mode(proj) != mode_X) + continue; + + node = get_irn_node(proj); + if (node->type.tv == tarval_reachable) { + if (++k > 1) + return 0; + } + } + return 1; +} /* only_one_reachable_proj */ + /** * Return non-zero if the control flow predecessor node pred * is the only reachable control flow exit of its block. * - * @param pred the control flow exit + * @param pred the control flow exit + * @param block the destination block */ -static int can_exchange(ir_node *pred) { - if (is_Start(pred)) +static int can_exchange(ir_node *pred, ir_node *block) +{ + if (is_Start(pred) || has_Block_entity(block)) return 0; else if (is_Jmp(pred)) return 1; else if (get_irn_mode(pred) == mode_T) { - int i, k; - /* if the predecessor block has more than one reachable outputs we cannot remove the block */ - k = 0; - for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) { - ir_node *proj = get_irn_out(pred, i); - node_t *node; - - /* skip non-control flow Proj's */ - if (get_irn_mode(proj) != mode_X) - continue; - - node = get_irn_node(proj); - if (node->type.tv == tarval_reachable) { - if (++k > 1) - return 0; - } - } - return 1; + return only_one_reachable_proj(pred); } return 0; } /* can_exchange */ @@ -2990,7 +3076,8 @@ static int can_exchange(ir_node *pred) { * Block Post-Walker, apply the analysis results on control flow by * shortening Phi's and Block inputs. */ -static void apply_cf(ir_node *block, void *ctx) { +static void apply_cf(ir_node *block, void *ctx) +{ environment_t *env = ctx; node_t *node = get_irn_node(block); int i, j, k, n; @@ -3040,7 +3127,7 @@ static void apply_cf(ir_node *block, void *ctx) { /* only one predecessor combine */ ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0)); - if (can_exchange(pred)) { + if (can_exchange(pred, block)) { ir_node *new_block = get_nodes_block(pred); DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block)); DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF); @@ -3091,7 +3178,7 @@ static void apply_cf(ir_node *block, void *ctx) { if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) { /* this Phi is replaced by a constant */ tarval *tv = node->type.tv; - ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv); + ir_node *c = new_Const(tv); set_irn_node(c, node); node->node = c; @@ -3131,7 +3218,7 @@ static void apply_cf(ir_node *block, void *ctx) { /* this Block has only one live predecessor */ ir_node *pred = skip_Proj(in_X[0]); - if (can_exchange(pred)) { + if (can_exchange(pred, block)) { ir_node *new_block = get_nodes_block(pred); DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF); exchange(block, new_block); @@ -3150,25 +3237,82 @@ static void apply_cf(ir_node *block, void *ctx) { * AddP(x, NULL) is a follower of x, but with different mode. * Fix it here. */ -static void exchange_leader(ir_node *irn, ir_node *leader) { +static void exchange_leader(ir_node *irn, ir_node *leader) +{ ir_mode *mode = get_irn_mode(irn); if (mode != get_irn_mode(leader)) { - /* The conv is a no-op, so we are fre to place in + /* The conv is a no-op, so we are free to place it * either in the block of the leader OR in irn's block. * Probably placing it into leaders block might reduce * the number of Conv due to CSE. */ ir_node *block = get_nodes_block(leader); dbg_info *dbg = get_irn_dbg_info(irn); - leader = new_rd_Conv(dbg, current_ir_graph, block, leader, mode); + leader = new_rd_Conv(dbg, block, leader, mode); } exchange(irn, leader); -} +} /* exchange_leader */ + +/** + * Check, if all users of a mode_M node are dead. Use + * the Def-Use edges for this purpose, as they still + * reflect the situation. + */ +static int all_users_are_dead(const ir_node *irn) +{ + int i, n = get_irn_n_outs(irn); + + for (i = 1; i <= n; ++i) { + const ir_node *succ = irn->out[i].use; + const node_t *block = get_irn_node(get_nodes_block(succ)); + const node_t *node; + + if (block->type.tv == tarval_unreachable) { + /* block is unreachable */ + continue; + } + node = get_irn_node(succ); + if (node->type.tv != tarval_top) { + /* found a reachable user */ + return 0; + } + } + /* all users are unreachable */ + return 1; +} /* all_user_are_dead */ + +/** + * Walker: Find reachable mode_M nodes that have only + * unreachable users. These nodes must be kept later. + */ +static void find_kept_memory(ir_node *irn, void *ctx) +{ + environment_t *env = ctx; + node_t *node, *block; + + if (get_irn_mode(irn) != mode_M) + return; + + block = get_irn_node(get_nodes_block(irn)); + if (block->type.tv == tarval_unreachable) + return; + + node = get_irn_node(irn); + if (node->type.tv == tarval_top) + return; + + /* ok, we found a live memory node. */ + if (all_users_are_dead(irn)) { + DB((dbg, LEVEL_1, "%+F must be kept\n", irn)); + ARR_APP1(ir_node *, env->kept_memory, irn); + } +} /* find_kept_memory */ /** * Post-Walker, apply the analysis results; */ -static void apply_result(ir_node *irn, void *ctx) { +static void apply_result(ir_node *irn, void *ctx) +{ environment_t *env = ctx; node_t *node = get_irn_node(irn); @@ -3187,12 +3331,30 @@ static void apply_result(ir_node *irn, void *ctx) { DB((dbg, LEVEL_1, "%+F is unreachable\n", irn)); exchange(irn, bad); env->modified = 1; - } - else if (node->type.tv == tarval_top) { - /* don't kick away Unknown's, they might be still needed */ - if (! is_Unknown(irn)) { - ir_mode *mode = get_irn_mode(irn); - ir_node *unk = new_r_Unknown(current_ir_graph, mode); + } else if (node->type.tv == tarval_top) { + ir_mode *mode = get_irn_mode(irn); + + if (mode == mode_M) { + /* never kill a mode_M node */ + if (is_Proj(irn)) { + ir_node *pred = get_Proj_pred(irn); + node_t *pnode = get_irn_node(pred); + + if (pnode->type.tv == tarval_top) { + /* skip the predecessor */ + ir_node *mem = get_memop_mem(pred); + node->node = mem; + DB((dbg, LEVEL_1, "%+F computes Top, replaced by %+F\n", irn, mem)); + exchange(irn, mem); + env->modified = 1; + } + } + /* leave other nodes, especially PhiM */ + } else if (mode == mode_T) { + /* Do not kill mode_T nodes, kill their Projs */ + } else if (! is_Unknown(irn)) { + /* don't kick away Unknown's, they might be still needed */ + ir_node *unk = new_r_Unknown(current_ir_graph, mode); /* control flow should already be handled at apply_cf() */ assert(mode != mode_X); @@ -3211,17 +3373,24 @@ static void apply_result(ir_node *irn, void *ctx) { ir_node *cond = get_Proj_pred(irn); if (is_Cond(cond)) { - node_t *sel = get_irn_node(get_Cond_selector(cond)); - - if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) { - /* Cond selector is a constant and the Proj is reachable, make a Jmp */ - ir_node *jmp = new_r_Jmp(current_ir_graph, block->node); + if (only_one_reachable_proj(cond)) { + ir_node *jmp = new_r_Jmp(block->node); set_irn_node(jmp, node); node->node = jmp; DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp)); DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF); exchange(irn, jmp); env->modified = 1; + } else { + node_t *sel = get_irn_node(get_Cond_selector(cond)); + tarval *tv = sel->type.tv; + + if (is_tarval(tv) && tarval_is_constant(tv)) { + /* The selector is a constant, but more + * than one output is active: An unoptimized + * case found. */ + env->unopt_cf = 1; + } } } } @@ -3236,7 +3405,7 @@ static void apply_result(ir_node *irn, void *ctx) { */ if (! is_Const(irn) && get_irn_mode(irn) != mode_T) { /* can be replaced by a constant */ - ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv); + ir_node *c = new_Const(tv); set_irn_node(c, node); node->node = c; DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c)); @@ -3247,7 +3416,7 @@ static void apply_result(ir_node *irn, void *ctx) { } else if (is_entity(node->type.sym.entity_p)) { if (! is_SymConst(irn)) { /* can be replaced by a SymConst */ - ir_node *symc = new_r_SymConst(current_ir_graph, block->node, get_irn_mode(irn), node->type.sym, symconst_addr_ent); + ir_node *symc = new_r_SymConst(current_ir_graph, get_irn_mode(irn), node->type.sym, symconst_addr_ent); set_irn_node(symc, node); node->node = symc; @@ -3298,7 +3467,8 @@ static void apply_result(ir_node *irn, void *ctx) { /** * Fix the keep-alives by deleting unreachable ones. */ -static void apply_end(ir_node *end, environment_t *env) { +static void apply_end(ir_node *end, environment_t *env) +{ int i, j, n = get_End_n_keepalives(end); ir_node **in; @@ -3327,7 +3497,8 @@ static void apply_end(ir_node *end, environment_t *env) { /** * sets the generic functions to compute. */ -static void set_compute_functions(void) { +static void set_compute_functions(void) +{ int i; /* set the default compute function */ @@ -3351,19 +3522,41 @@ static void set_compute_functions(void) { SET(Confirm); SET(Return); SET(End); + SET(Call); +} /* set_compute_functions */ - if (op_Max != NULL) - SET(Max); - if (op_Min != NULL) - SET(Min); +/** + * Add memory keeps. + */ +static void add_memory_keeps(ir_node **kept_memory, int len) +{ + ir_node *end = get_irg_end(current_ir_graph); + int i; + ir_nodeset_t set; -} /* set_compute_functions */ + ir_nodeset_init(&set); + + /* check, if those nodes are already kept */ + for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) + ir_nodeset_insert(&set, get_End_keepalive(end, i)); + + for (i = len - 1; i >= 0; --i) { + ir_node *ka = kept_memory[i]; -void combo(ir_graph *irg) { + if (! ir_nodeset_contains(&set, ka)) { + add_End_keepalive(end, ka); + } + } + ir_nodeset_destroy(&set); +} /* add_memory_keeps */ + +void combo(ir_graph *irg) +{ environment_t env; ir_node *initial_bl; node_t *start; ir_graph *rem = current_ir_graph; + int len; current_ir_graph = irg; @@ -3381,11 +3574,14 @@ void combo(ir_graph *irg) { env.dbg_list = NULL; #endif env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4); - env.type2id_map = pmap_create(); + env.kept_memory = NEW_ARR_F(ir_node *, 0); env.end_idx = get_opt_global_cse() ? 0 : -1; env.lambda_input = 0; - env.commutative = 1; env.modified = 0; + env.unopt_cf = 0; + /* options driving the optimization */ + env.commutative = 1; + env.opt_unknown = 1; assure_irg_outs(irg); assure_cf_loop(irg); @@ -3396,7 +3592,12 @@ void combo(ir_graph *irg) { set_compute_functions(); DEBUG_ONLY(part_nr = 0); - ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST); + + if (env.opt_unknown) + tarval_UNKNOWN = tarval_top; + else + tarval_UNKNOWN = tarval_bad; /* create the initial partition and place it on the work list */ env.initial = new_partition(&env); @@ -3429,6 +3630,11 @@ void combo(ir_graph *irg) { #endif /* apply the result */ + + /* check, which nodes must be kept */ + irg_walk_graph(irg, NULL, find_kept_memory, &env); + + /* kill unreachable control flow */ irg_block_walk_graph(irg, NULL, apply_cf, &env); /* Kill keep-alives of dead blocks: this speeds up apply_result() * and fixes assertion because dead cf to dead blocks is NOT removed by @@ -3436,20 +3642,29 @@ void combo(ir_graph *irg) { apply_end(get_irg_end(irg), &env); irg_walk_graph(irg, NULL, apply_result, &env); + len = ARR_LEN(env.kept_memory); + if (len > 0) + add_memory_keeps(env.kept_memory, len); + + if (env.unopt_cf) { + DB((dbg, LEVEL_1, "Unoptimized Control Flow left")); + } + if (env.modified) { /* control flow might changed */ set_irg_outs_inconsistent(irg); set_irg_extblk_inconsistent(irg); set_irg_doms_inconsistent(irg); set_irg_loopinfo_inconsistent(irg); + set_irg_entity_usage_state(irg, ir_entity_usage_not_computed); } - ir_free_resources(irg, IR_RESOURCE_IRN_LINK); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST); /* remove the partition hook */ DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL)); - pmap_destroy(env.type2id_map); + DEL_ARR_F(env.kept_memory); del_set(env.opcode2id_map); obstack_free(&env.obst, NULL); @@ -3457,3 +3672,9 @@ void combo(ir_graph *irg) { set_value_of_func(NULL); current_ir_graph = rem; } /* combo */ + +/* Creates an ir_graph pass for combo. */ +ir_graph_pass_t *combo_pass(const char *name) +{ + return def_graph_pass(name ? name : "combo", combo); +} /* combo_pass */