X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fcombo.c;h=8dfd40d9558c6d2a17f8caf94c11e53a8c0aa639;hb=01ddc5c06877693df7d5f286d43e433fedd87103;hp=6d62f6264d8a3575843535951a6f36f43039e3c7;hpb=82b4543a9674f3f408eb2b3f2f8053841875bd1a;p=libfirm diff --git a/ir/opt/combo.c b/ir/opt/combo.c index 6d62f6264..8dfd40d95 100644 --- a/ir/opt/combo.c +++ b/ir/opt/combo.c @@ -23,16 +23,42 @@ * @author Michael Beck * @version $Id$ * - * Note that the current implementation lack the leaders/followers - * support. + * This is a slightly enhanced version of Cliff Clicks combo algorithm + * - support for commutative nodes is added, Add(a,b) and Add(b,a) ARE congruent + * - supports all Firm direct (by a data edge) identities except Mux + * (Mux can be a 2-input or 1-input identity, only 2-input is implemented yet) + * - supports Confirm nodes (handle them like Copies but do NOT remove them) + * - let Cmp nodes calculate Top like all othe data nodes: this would let + * Mux nodes to calculate Unknown instead of taking the true result + * - let Cond(Top) always select FALSE/default: This is tricky. Nodes are only reavaluated + * IFF the predecessor changed its type. Because nodes are initialized with Top + * this never happens, let all Proj(Cond) be unreachable. + * We avoid this condition by the same way we work around Phi: whenever a Block + * node is placed on the list, place its Cond nodes (and because they are Tuple + * all its Proj-nodes either on the cprop list) + * Especially, this changes the meaning of Click's example: + * + * int main() { + * int x; + * + * if (x == 2) + * printf("x == 2\n"); + * if (x == 3) + * printf("x == 3\n"); + * } + * + * Would print: + * x == 2 + * x == 3 + * + * using Click's version while is silent with our. + * - support for global congruences is implemented but not tested yet * * Note further that we use the terminology from Click's work here, which is different * in some cases from Firm terminology. Especially, Click's type is a * Firm tarval/entity, nevertheless we call it type here for "maximum compatibility". */ -#ifdef HAVE_CONFIG_H -# include "config.h" -#endif +#include "config.h" #include @@ -50,9 +76,12 @@ #include "irop.h" #include "irouts.h" #include "irgmod.h" +#include "iropt_dbg.h" #include "debug.h" +#include "array_t.h" #include "error.h" - +#include "irnodeset.h" +#include "irpass.h" #include "tv_t.h" #include "irprintf.h" @@ -61,6 +90,9 @@ /* define this to check that all type translations are monotone */ #define VERIFY_MONOTONE +/* define this to check the consistency of partitions */ +#define CHECK_PARTITIONS + typedef struct node_t node_t; typedef struct partition_t partition_t; typedef struct opcode_key_t opcode_key_t; @@ -75,9 +107,14 @@ typedef void (*compute_func)(node_t *node); struct opcode_key_t { ir_opcode code; /**< The Firm opcode. */ ir_mode *mode; /**< The mode of all nodes in the partition. */ + int arity; /**< The arity of this opcode (needed for Phi etc. */ 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; }; @@ -110,35 +147,42 @@ typedef union { */ struct node_t { ir_node *node; /**< The IR-node itself. */ - list_head node_list; /**< Double-linked list of entries. */ + list_head node_list; /**< Double-linked list of leader/follower entries. */ list_head cprop_list; /**< Double-linked partition.cprop list. */ partition_t *part; /**< points to the partition this node belongs to */ node_t *next; /**< Next node on local list (partition.touched, fallen). */ + node_t *race_next; /**< Next node on race list. */ lattice_elem_t type; /**< The associated lattice element "type". */ int max_user_input; /**< Maximum input number of Def-Use edges. */ int next_edge; /**< Index of the next Def-Use edge to use. */ + int n_followers; /**< Number of Follower in the outs set. */ unsigned on_touched:1; /**< Set, if this node is on the partition.touched set. */ 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 flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */ }; /** * A partition containing congruent nodes. */ struct partition_t { - list_head entries; /**< The head of partition node list. */ + list_head Leader; /**< The head of partition Leader node list. */ + list_head Follower; /**< The head of partition Follower node list. */ list_head cprop; /**< The head of partition.cprop list. */ + list_head cprop_X; /**< The head of partition.cprop (Cond nodes and its Projs) list. */ partition_t *wl_next; /**< Next entry in the work list if any. */ partition_t *touched_next; /**< Points to the next partition in the touched set. */ partition_t *cprop_next; /**< Points to the next partition in the cprop list. */ partition_t *split_next; /**< Points to the next partition in the list that must be split by split_by(). */ node_t *touched; /**< The partition.touched set of this partition. */ - unsigned n_nodes; /**< Number of entries in this partition. */ + unsigned n_leader; /**< Number of entries in this partition.Leader. */ unsigned n_touched; /**< Number of entries in the partition.touched. */ int max_user_inputs; /**< Maximum number of user inputs of all entries. */ unsigned on_worklist:1; /**< Set, if this partition is in the work list. */ unsigned on_touched:1; /**< Set, if this partition is on the touched set. */ unsigned on_cprop:1; /**< Set, if this partition is on the cprop list. */ + unsigned type_is_T_or_C:1;/**< Set, if all nodes in this partition have type Top or Constant. */ #ifdef DEBUG_libfirm partition_t *dbg_next; /**< Link all partitions for debugging */ unsigned nr; /**< A unique number for (what-)mapping, >0. */ @@ -152,9 +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(). */ + 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 @@ -174,50 +223,294 @@ typedef void *(*what_func)(const node_t *node, environment_t *env); /** The debug module handle. */ DEBUG_ONLY(static firm_dbg_module_t *dbg;) +/** The what reason. */ +DEBUG_ONLY(static const char *what_reason;) + /** Next partition number. */ DEBUG_ONLY(static unsigned part_nr = 0); +/** 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); + +#ifdef CHECK_PARTITIONS +/** + * Check a partition. + */ +static void check_partition(const partition_t *T) +{ + node_t *node; + unsigned n = 0; + + list_for_each_entry(node_t, node, &T->Leader, node_list) { + assert(node->is_follower == 0); + assert(node->flagged == 0); + assert(node->part == T); + ++n; + } + assert(n == T->n_leader); + + list_for_each_entry(node_t, node, &T->Follower, node_list) { + assert(node->is_follower == 1); + assert(node->flagged == 0); + assert(node->part == T); + } +} /* check_partition */ + +/** + * check that all leader nodes in the partition have the same opcode. + */ +static void check_opcode(const partition_t *Z) +{ + node_t *node; + opcode_key_t key; + int first = 1; + + list_for_each_entry(node_t, node, &Z->Leader, node_list) { + ir_node *irn = node->node; + + if (first) { + key.code = get_irn_opcode(irn); + key.mode = get_irn_mode(irn); + key.arity = get_irn_arity(irn); + key.u.proj = 0; + key.u.ent = NULL; + + switch (get_irn_opcode(irn)) { + case iro_Proj: + key.u.proj = get_Proj_proj(irn); + break; + 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((unsigned)key.code == get_irn_opcode(irn)); + assert(key.mode == get_irn_mode(irn)); + assert(key.arity == get_irn_arity(irn)); + + switch (get_irn_opcode(irn)) { + case iro_Proj: + assert(key.u.proj == get_Proj_proj(irn)); + break; + 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; + } + } + } +} /* check_opcode */ + +static void check_all_partitions(environment_t *env) +{ +#ifdef DEBUG_libfirm + partition_t *P; + node_t *node; + + for (P = env->dbg_list; P != NULL; P = P->dbg_next) { + check_partition(P); + if (! P->type_is_T_or_C) + check_opcode(P); + list_for_each_entry(node_t, node, &P->Follower, node_list) { + node_t *leader = identity(node); + + 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) +{ + +#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) +{ + do_check_list(list, offsetof(node_t, next), Z); +} /* check_list */ + +#else +#define check_partition(T) +#define check_list(list, Z) +#define check_all_partitions(env) +#endif /* CHECK_PARTITIONS */ + #ifdef DEBUG_libfirm -static INLINE lattice_elem_t get_partition_type(const partition_t *X); +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); - DB((dbg, LEVEL_2, "%s part%u (%u, %+F) {\n ", msg, part->nr, part->n_nodes, type)); - list_for_each_entry(node_t, node, &part->entries, node_list) { + DB((dbg, LEVEL_2, "%s part%u%s (%u, %+F) {\n ", + msg, part->nr, part->type_is_T_or_C ? "*" : "", + part->n_leader, type)); + list_for_each_entry(node_t, node, &part->Leader, node_list) { DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node)); first = 0; } + if (! list_empty(&part->Follower)) { + DB((dbg, LEVEL_2, "\n---\n ")); + first = 1; + list_for_each_entry(node_t, node, &part->Follower, node_list) { + DB((dbg, LEVEL_2, "%s%+F", first ? "" : ", ", node->node)); + first = 0; + } + } DB((dbg, LEVEL_2, "\n}\n")); -} +} /* dump_partition */ + +/** + * Dumps a list. + */ +static void do_dump_list(const char *msg, const node_t *node, int ofs) +{ + const node_t *p; + int first = 1; + +#define GET_LINK(p, ofs) *((const node_t **)((char *)(p) + (ofs))) + + DB((dbg, LEVEL_3, "%s = {\n ", msg)); + for (p = node; p != NULL; p = GET_LINK(p, ofs)) { + DB((dbg, LEVEL_3, "%s%+F", first ? "" : ", ", p->node)); + first = 0; + } + DB((dbg, LEVEL_3, "\n}\n")); + +#undef GET_LINK +} /* do_dump_list */ + +/** + * Dumps a race 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) +{ + 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")); for (P = env->dbg_list; P != NULL; P = P->dbg_next) dump_partition("", P); -} +} /* dump_all_partitions */ + +/** + * Sump a split 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)); + for (p = list; p != NULL; p = p->split_next) + DB((dbg, LEVEL_2, "part%u, ", p->nr)); + DB((dbg, LEVEL_2, "\n}\n")); +} /* dump_split_list */ + +/** + * Dump partition and type for a node. + */ +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); + + ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type); + return 1; +} /* dump_partition_hook */ #else #define dump_partition(msg, part) +#define dump_race_list(msg, list) +#define dump_list(msg, list) #define dump_all_partitions(env) +#define dump_split_list(list) #endif #if defined(VERIFY_MONOTONE) && defined (DEBUG_libfirm) /** * Verify that a type transition is monotone */ -static void verify_type(const lattice_elem_t old_type, const lattice_elem_t new_type) { - if (old_type.tv == new_type.tv) { +static void verify_type(const lattice_elem_t old_type, node_t *node) +{ + if (old_type.tv == node->type.tv) { /* no change */ return; } @@ -225,23 +518,22 @@ static void verify_type(const lattice_elem_t old_type, const lattice_elem_t new_ /* from Top down-to is always allowed */ return; } - if (old_type.tv == tarval_reachable) { - panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type); - } - if (new_type.tv == tarval_bottom || new_type.tv == tarval_reachable) { + if (node->type.tv == tarval_bottom || node->type.tv == tarval_reachable) { /* bottom reached */ return; } - panic("verify_type(): wrong translation from %+F to %+F", old_type, new_type); -} + panic("combo: wrong translation from %+F to %+F on node %+F", old_type, node->type, node->node); +} /* verify_type */ + #else -#define verify_type(old_type, new_type) +#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; @@ -254,7 +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 */ @@ -264,7 +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 */ @@ -274,9 +568,10 @@ static void listmap_term(listmap_t *map) { * @param map the listmap * @param id the id to search for * - * @return the asociated listmap entry for the given id + * @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; @@ -299,26 +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); +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->u.proj != o2->u.proj || o1->u.ent != o2->u.ent; + o1->arity != o2->arity || + 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; @@ -329,7 +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); @@ -346,7 +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 */ @@ -357,7 +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)) @@ -368,12 +672,14 @@ 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; X->on_worklist = 1; env->worklist = X; -} +} /* add_to_worklist */ /** * Create a new empty partition. @@ -382,22 +688,26 @@ 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->entries); + INIT_LIST_HEAD(&part->Leader); + INIT_LIST_HEAD(&part->Follower); INIT_LIST_HEAD(&part->cprop); + INIT_LIST_HEAD(&part->cprop_X); part->wl_next = NULL; part->touched_next = NULL; part->cprop_next = NULL; part->split_next = NULL; part->touched = NULL; - part->n_nodes = 0; + part->n_leader = 0; part->n_touched = 0; part->max_user_inputs = 0; part->on_worklist = 0; part->on_touched = 0; part->on_cprop = 0; + part->type_is_T_or_C = 0; #ifdef DEBUG_libfirm part->dbg_next = env->dbg_list; env->dbg_list = part; @@ -410,9 +720,10 @@ 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) { - return list_entry(X->entries.next, node_t, node_list); -} +static inline node_t *get_first_node(const partition_t *X) +{ + return list_entry(X->Leader.next, node_t, node_list); +} /* get_first_node */ /** * Return the type of a partition (assuming partition is non-empty and @@ -422,7 +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 */ @@ -437,45 +749,40 @@ 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); node->node = irn; node->part = part; node->next = NULL; + node->race_next = NULL; node->type.tv = tarval_top; node->max_user_input = 0; node->next_edge = 0; + node->n_followers = 0; node->on_touched = 0; node->on_cprop = 0; node->on_fallen = 0; + node->is_follower = 0; + node->flagged = 0; set_irn_node(irn, node); - list_add_tail(&node->node_list, &part->entries); - ++part->n_nodes; + list_add_tail(&node->node_list, &part->Leader); + ++part->n_leader; return node; } /* create_partition_node */ /** - * Pre-Walker, init all Block-Phi lists. - */ -static void init_block_phis(ir_node *irn, void *env) { - (void) env; - - if (is_Block(irn)) { - set_Block_phis(irn, NULL); - } -} - -/** - * Post-Walker, initialize all Nodes' type to U or top and place + * 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; @@ -485,31 +792,32 @@ static void create_initial_partitions(ir_node *irn, void *ctx) { if (node->max_user_input > part->max_user_inputs) part->max_user_inputs = node->max_user_input; - if (is_Phi(irn)) { - add_Block_phi(get_nodes_block(irn), irn); + if (is_Block(irn)) { + set_Block_phis(irn, NULL); } } /* create_initial_partitions */ /** - * Add a partition to the touched set if not already there. - * - * @param part the partition - * @param env the environment + * Post-Walker, collect all Block-Phi lists, set Cond. */ -static INLINE void add_to_touched(partition_t *part, environment_t *env) { - if (part->on_touched == 0) { - part->touched_next = env->touched; - env->touched = part; - part->on_touched = 1; +static void init_block_phis(ir_node *irn, void *ctx) +{ + (void) ctx; + + if (is_Phi(irn)) { + add_Block_phi(get_nodes_block(irn), irn); } -} /* add_to_touched */ +} /* init_block_phis */ /** - * Add a node to the entry.partition.touched set if not already there. + * Add a node to the entry.partition.touched set and + * node->partition to the touched set if not already there. * - * @param y a node + * @param y a node + * @param env the environment */ -static INLINE void add_to_partition_touched(node_t *y) { +static inline void add_to_touched(node_t *y, environment_t *env) +{ if (y->on_touched == 0) { partition_t *part = y->part; @@ -517,8 +825,70 @@ static INLINE void add_to_partition_touched(node_t *y) { part->touched = y; y->on_touched = 1; ++part->n_touched; + + if (part->on_touched == 0) { + part->touched_next = env->touched; + env->touched = part; + part->on_touched = 1; + } + + check_list(part->touched, part); + } +} /* add_to_touched */ + +/** + * Place a node on the cprop list. + * + * @param y the node + * @param env the environment + */ +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 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); + y->on_cprop = 1; + + DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr)); + + /* place its partition on the cprop list */ + if (Y->on_cprop == 0) { + Y->cprop_next = env->cprop; + env->cprop = Y; + Y->on_cprop = 1; + } + } + irn = y->node; + if (get_irn_mode(irn) == mode_T) { + /* mode_T nodes always produce tarval_bottom, so we must explicitly + add it's Proj's to get constant evaluation to work */ + int i; + + for (i = get_irn_n_outs(irn) - 1; i >= 0; --i) { + node_t *proj = get_irn_node(get_irn_out(irn, i)); + + add_to_cprop(proj, env); + } + } else if (is_Block(irn)) { + /* Due to the way we handle Phi's, we must place all Phis of a block on the list + * if someone placed the block. The Block is only placed if the reachability + * changes, and this must be re-evaluated in compute_Phi(). */ + ir_node *phi; + for (phi = get_Block_phis(irn); phi != NULL; phi = get_Phi_next(phi)) { + node_t *p = get_irn_node(phi); + add_to_cprop(p, env); + } } -} /* add_to_partition_touched */ +} /* add_to_cprop */ /** * Update the worklist: If Z is on worklist then add Z' to worklist. @@ -528,8 +898,9 @@ static INLINE void add_to_partition_touched(node_t *y) { * @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) { - if (Z->on_worklist || Z_prime->n_nodes < Z->n_nodes) { +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 { add_to_worklist(Z, env); @@ -537,228 +908,748 @@ static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t } /* update_worklist */ /** - * Split a partition by a local list. + * Make all inputs to x no longer be F.def_use edges. + * + * @param x the node + */ +static void move_edges_to_leader(node_t *x) +{ + ir_node *irn = x->node; + int i, j, k; + + for (i = get_irn_arity(irn) - 1; i >= 0; --i) { + node_t *pred = get_irn_node(get_irn_n(irn, i)); + ir_node *p; + int n; + + p = pred->node; + n = get_irn_n_outs(p); + for (j = 1; j <= pred->n_followers; ++j) { + if (p->out[j].pos == i && p->out[j].use == irn) { + /* found a follower edge to x, move it to the Leader */ + ir_def_use_edge edge = p->out[j]; + + /* remove this edge from the Follower set */ + p->out[j] = p->out[pred->n_followers]; + --pred->n_followers; + + /* sort it into the leader set */ + for (k = pred->n_followers + 2; k <= n; ++k) { + if (p->out[k].pos >= edge.pos) + break; + p->out[k - 1] = p->out[k]; + } + /* place the new edge here */ + p->out[k - 1] = edge; + + /* edge found and moved */ + break; + } + } + } +} /* move_edges_to_leader */ + +/** + * Split a partition that has NO followers by a local list. * - * @param Z the Z partition to split + * @param Z partition to split * @param g a (non-empty) node list * @param env the environment * * @return a new partition containing the nodes of g */ -static partition_t *split(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; - int max_input, max_arity, arity; + int max_input; dump_partition("Splitting ", Z); + dump_list("by list ", g); assert(g != NULL); /* Remove g from Z. */ for (node = g; node != NULL; node = node->next) { + assert(node->part == Z); list_del(&node->node_list); ++n; } - assert(n < Z->n_nodes); - Z->n_nodes -= n; + assert(n < Z->n_leader); + Z->n_leader -= n; - /* Move g to a new partition, Z’. */ + /* Move g to a new partition, Z'. */ Z_prime = new_partition(env); - max_arity = max_input = 0; + max_input = 0; for (node = g; node != NULL; node = node->next) { - list_add(&node->node_list, &Z_prime->entries); + list_add_tail(&node->node_list, &Z_prime->Leader); node->part = Z_prime; - arity = get_irn_arity(node->node); - if (arity > max_arity) - max_arity = arity; if (node->max_user_input > max_input) max_input = node->max_user_input; } Z_prime->max_user_inputs = max_input; - Z_prime->n_nodes = n; + Z_prime->n_leader = n; + + check_partition(Z); + check_partition(Z_prime); + + /* for now, copy the type info tag, it will be adjusted in split_by(). */ + Z_prime->type_is_T_or_C = Z->type_is_T_or_C; update_worklist(Z, Z_prime, env); dump_partition("Now ", Z); dump_partition("Created new ", Z_prime); return Z_prime; -} /* split */ +} /* split_no_followers */ /** - * Returns non-zero if the i'th input of a Phi node is live. + * Make the Follower -> Leader transition for a node. * - * @param phi a Phi-node - * @param i an input number - * - * @return non-zero if the i'th input of the given Phi node is live + * @param n the node */ -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); - lattice_elem_t type = get_node_type(pred); - - return type.tv != tarval_unreachable; - } - /* else it's the control input, always live */ - return 1; -} /* is_live_input */ +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)); + n->is_follower = 0; + move_edges_to_leader(n); + list_del(&n->node_list); + list_add_tail(&n->node_list, &n->part->Leader); + ++n->part->n_leader; +} /* follower_to_leader */ /** - * Return non-zero if a type is a constant. + * The environment for one race step. */ -static int is_constant_type(lattice_elem_t type) { - if (type.tv != tarval_bottom && type.tv != tarval_top) - return 1; - return 0; -} /* is_constant_type */ +typedef struct step_env { + node_t *initial; /**< The initial node list. */ + node_t *unwalked; /**< The unwalked node list. */ + node_t *walked; /**< The walked node list. */ + int index; /**< Next index of Follower use_def edge. */ + unsigned side; /**< side number. */ +} step_env; /** - * Place a node on the cprop list. + * Return non-zero, if a input is a real follower * - * @param y the node - * @param env the environment + * @param irn the node to check + * @param input number of the input */ -static void add_node_to_cprop(node_t *y, environment_t *env) { - /* Add y to y.partition.cprop. */ - if (y->on_cprop == 0) { - partition_t *Y = y->part; - - list_add_tail(&y->cprop_list, &Y->cprop); - y->on_cprop = 1; - - DB((dbg, LEVEL_3, "Add %+F to part%u.cprop\n", y->node, Y->nr)); +static int is_real_follower(const ir_node *irn, int input) +{ + node_t *pred; - /* place its partition on the cprop list */ - if (Y->on_cprop == 0) { - Y->cprop_next = env->cprop; - env->cprop = Y; - Y->on_cprop = 1; + switch (get_irn_opcode(irn)) { + case iro_Confirm: + if (input == 1) { + /* ignore the Confirm bound input */ + return 0; } - } - if (get_irn_mode(y->node) == mode_T) { - /* mode_T nodes always produce tarval_bottom, so we must explicitly - add it's Proj's to get constant evaluation to work */ - int i; - - for (i = get_irn_n_outs(y->node) - 1; i >= 0; --i) { - node_t *proj = get_irn_node(get_irn_out(y->node, i)); - - add_node_to_cprop(proj, env); + break; + case iro_Mux: + if (input == 0) { + /* ignore the Mux sel input */ + return 0; } - } + break; + case iro_Phi: { + /* dead inputs are not follower edges */ + ir_node *block = get_nodes_block(irn); + node_t *pred = get_irn_node(get_Block_cfgpred(block, input)); - if (is_Block(y->node)) { - /* Due to the way we handle Phi's, we must place all Phis of a block on the list - * if someone placed the block. The Block is only placed if the reachability - * changes, and this must be re-evaluated in compute_Phi(). */ - ir_node *phi; - for (phi = get_Block_phis(y->node); phi != NULL; phi = get_Phi_next(phi)) { - node_t *p = get_irn_node(phi); - add_node_to_cprop(p, env); - } + if (pred->type.tv == tarval_unreachable) + return 0; + break; } -} /* add_node_to_cprop */ - -/** - * Check whether a type is neither Top or a constant. - * Note: U is handled like Top here, R is a constant. - * - * @param type the type to check - */ -static int type_is_neither_top_nor_const(const lattice_elem_t type) { - if (is_tarval(type.tv)) { - if (type.tv == tarval_top) + case iro_Sub: + case iro_Shr: + case iro_Shl: + case iro_Shrs: + case iro_Rotl: + if (input == 1) { + /* only a Sub x,0 / Shift x,0 might be a follower */ return 0; - if (tarval_is_constant(type.tv)) + } + break; + case iro_Add: + case iro_Or: + case iro_Eor: + pred = get_irn_node(get_irn_n(irn, input)); + if (is_tarval(pred->type.tv) && tarval_is_null(pred->type.tv)) return 0; - } else { - /* is a symconst */ - return 0; + break; + case iro_Mul: + pred = get_irn_node(get_irn_n(irn, input)); + if (is_tarval(pred->type.tv) && tarval_is_one(pred->type.tv)) + return 0; + break; + case iro_And: + pred = get_irn_node(get_irn_n(irn, input)); + if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv)) + return 0; + break; + default: + assert(!"opcode not implemented yet"); + break; } return 1; -} +} /* is_real_follower */ /** - * Split the partitions if caused by the first entry on the worklist. - * - * @param env the environment + * Do one step in the race. */ -static void cause_splits(environment_t *env) { - partition_t *X, *Y, *Z; - node_t *x, *y, *e; - int i, end_idx; - ir_opcode code; - ir_node *succ; +static int step(step_env *env) +{ + node_t *n; - /* remove the first partition from the worklist */ - X = env->worklist; - env->worklist = X->wl_next; - X->on_worklist = 0; - - dump_partition("Cause_split: ", X); - end_idx = env->end_idx; - for (i = -1; i <= X->max_user_inputs; ++i) { - /* empty the touched set: already done, just clear the list */ - env->touched = NULL; + if (env->initial != NULL) { + /* Move node from initial to unwalked */ + n = env->initial; + env->initial = n->race_next; - list_for_each_entry(node_t, x, &X->entries, node_list) { - int num_edges; + n->race_next = env->unwalked; + env->unwalked = n; - if (i == -1) { - x->next_edge = 1; - } - num_edges = get_irn_n_outs(x->node); + return 0; + } - while (x->next_edge <= num_edges) { - ir_def_use_edge *edge = &x->node->out[x->next_edge]; + while (env->unwalked != NULL) { + /* let n be the first node in unwalked */ + n = env->unwalked; + while (env->index < n->n_followers) { + const ir_def_use_edge *edge = &n->node->out[1 + env->index]; - /* check if we have necessary edges */ - if (edge->pos > i) - break; + /* let m be n.F.def_use[index] */ + node_t *m = get_irn_node(edge->use); - ++x->next_edge; + assert(m->is_follower); + /* + * Some inputs, like the get_Confirm_bound are NOT + * real followers, sort them out. + */ + if (! is_real_follower(m->node, edge->pos)) { + ++env->index; + continue; + } + ++env->index; - succ = edge->use; + /* only followers from our partition */ + if (m->part != n->part) + continue; - /* ignore the "control input" for non-pinned nodes - if we are running in GCSE mode */ - if (i < end_idx && get_irn_pinned(succ) != op_pin_state_pinned) - continue; + if ((m->flagged & env->side) == 0) { + m->flagged |= env->side; - y = get_irn_node(succ); - if (is_constant_type(y->type)) { - code = get_irn_opcode(succ); - if (code == iro_Sub || code == iro_Cmp) - add_node_to_cprop(y, env); + if (m->flagged != 3) { + /* visited the first time */ + /* add m to unwalked not as first node (we might still need to + check for more follower node */ + m->race_next = n->race_next; + n->race_next = m; + return 0; } + /* else already visited by the other side and on the other list */ + } + } + /* move n to walked */ + env->unwalked = n->race_next; + n->race_next = env->walked; + env->walked = n; + env->index = 0; + } + return 1; +} /* step */ - /* Partitions of constants should not be split simply because their Nodes have unequal - functions or incongruent inputs. */ - if (type_is_neither_top_nor_const(y->type) && - (! is_Phi(y->node) || is_live_input(y->node, i))) { - Y = y->part; - add_to_touched(Y, env); - add_to_partition_touched(y); - } +/** + * Clear the flags from a list and check for + * nodes that where touched from both sides. + * + * @param list the list + */ +static int clear_flags(node_t *list) +{ + int res = 0; + node_t *n; + + for (n = list; n != NULL; n = n->race_next) { + if (n->flagged == 3) { + /* we reach a follower from both sides, this will split congruent + * inputs and make it a leader. */ + follower_to_leader(n); + res = 1; + } + n->flagged = 0; + } + return res; +} /* clear_flags */ + +/** + * Split a partition by a local list using the race. + * + * @param pX pointer to the partition to split, might be changed! + * @param gg a (non-empty) node list + * @param env the environment + * + * @return a new partition containing the nodes of gg + */ +static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) +{ + partition_t *X = *pX; + partition_t *X_prime; + list_head tmp; + step_env senv[2]; + node_t *g, *h, *node, *t; + int max_input, transitions, winner, shf; + unsigned n; + DEBUG_ONLY(static int run = 0;) + + DB((dbg, LEVEL_2, "Run %d ", run++)); + if (list_empty(&X->Follower)) { + /* if the partition has NO follower, we can use the fast + splitting algorithm. */ + return split_no_followers(X, gg, env); + } + /* else do the race */ + + dump_partition("Splitting ", X); + dump_list("by list ", gg); + + INIT_LIST_HEAD(&tmp); + + /* Remove gg from X.Leader and put into g */ + g = NULL; + for (node = gg; node != NULL; node = node->next) { + assert(node->part == X); + assert(node->is_follower == 0); + + list_del(&node->node_list); + list_add_tail(&node->node_list, &tmp); + node->race_next = g; + g = node; + } + /* produce h */ + h = NULL; + list_for_each_entry(node_t, node, &X->Leader, node_list) { + node->race_next = h; + h = node; + } + /* restore X.Leader */ + list_splice(&tmp, &X->Leader); + + senv[0].initial = g; + senv[0].unwalked = NULL; + senv[0].walked = NULL; + senv[0].index = 0; + senv[0].side = 1; + + senv[1].initial = h; + senv[1].unwalked = NULL; + senv[1].walked = NULL; + senv[1].index = 0; + senv[1].side = 2; + + /* + * Some informations on the race that are not stated clearly in Click's + * thesis. + * 1) A follower stays on the side that reach him first. + * 2) If the other side reches a follower, if will be converted to + * a leader. /This must be done after the race is over, else the + * edges we are iterating on are renumbered./ + * 3) /New leader might end up on both sides./ + * 4) /If one side ends up with new Leaders, we must ensure that + * they can split out by opcode, hence we have to put _every_ + * partition with new Leader nodes on the cprop list, as + * opcode splitting is done by split_by() at the end of + * constant propagation./ + */ + for (;;) { + if (step(&senv[0])) { + winner = 0; + break; + } + if (step(&senv[1])) { + winner = 1; + break; + } + } + assert(senv[winner].initial == NULL); + assert(senv[winner].unwalked == NULL); + + /* clear flags from walked/unwalked */ + shf = winner; + transitions = clear_flags(senv[0].unwalked) << shf; + transitions |= clear_flags(senv[0].walked) << shf; + shf ^= 1; + transitions |= clear_flags(senv[1].unwalked) << shf; + transitions |= clear_flags(senv[1].walked) << shf; + + dump_race_list("winner ", senv[winner].walked); + + /* Move walked_{winner} to a new partition, X'. */ + X_prime = new_partition(env); + max_input = 0; + n = 0; + for (node = senv[winner].walked; node != NULL; node = node->race_next) { + list_del(&node->node_list); + node->part = X_prime; + if (node->is_follower) { + list_add_tail(&node->node_list, &X_prime->Follower); + } else { + list_add_tail(&node->node_list, &X_prime->Leader); + ++n; + } + if (node->max_user_input > max_input) + max_input = node->max_user_input; + } + X_prime->n_leader = n; + X_prime->max_user_inputs = max_input; + X->n_leader -= X_prime->n_leader; + + /* for now, copy the type info tag, it will be adjusted in split_by(). */ + X_prime->type_is_T_or_C = X->type_is_T_or_C; + + /* + * Even if a follower was not checked by both sides, it might have + * loose its congruence, so we need to check this case for all follower. + */ + list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) { + if (identity(node) == node) { + follower_to_leader(node); + transitions |= 1; + } + } + + check_partition(X); + check_partition(X_prime); + + /* X' is the smaller part */ + add_to_worklist(X_prime, env); + + /* + * If there where follower to leader transitions, ensure that the nodes + * can be split out if necessary. + */ + if (transitions & 1) { + /* place winner partition on the cprop list */ + if (X_prime->on_cprop == 0) { + X_prime->cprop_next = env->cprop; + env->cprop = X_prime; + X_prime->on_cprop = 1; + } + } + if (transitions & 2) { + /* place other partition on the cprop list */ + if (X->on_cprop == 0) { + X->cprop_next = env->cprop; + env->cprop = X; + X->on_cprop = 1; + } + } + + dump_partition("Now ", X); + dump_partition("Created new ", X_prime); + + /* we have to ensure that the partition containing g is returned */ + if (winner != 0) { + *pX = X_prime; + return X; + } + + return X_prime; +} /* split */ + +/** + * Returns non-zero if the i'th input of a Phi node is live. + * + * @param phi a Phi-node + * @param i an input number + * + * @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) +{ + if (i >= 0) { + ir_node *block = get_nodes_block(phi); + ir_node *pred = get_Block_cfgpred(block, i); + lattice_elem_t type = get_node_type(pred); + + return type.tv != tarval_unreachable; + } + /* else it's the control input, always live */ + return 1; +} /* is_live_input */ + +/** + * Return non-zero if a type is a constant. + */ +static int is_constant_type(lattice_elem_t type) +{ + if (type.tv != tarval_bottom && type.tv != tarval_top) + return 1; + return 0; +} /* is_constant_type */ + +/** + * Check whether a type is neither Top or a constant. + * Note: U is handled like Top here, R is a constant. + * + * @param type the type to check + */ +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; + if (tarval_is_constant(type.tv)) + return 0; + } else { + /* is a symconst */ + return 0; + } + return 1; +} /* type_is_neither_top_nor_const */ + +/** + * Collect nodes to the touched list. + * + * @param list the list which contains the nodes that must be evaluated + * @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) +{ + node_t *x, *y; + int end_idx = env->end_idx; + + list_for_each_entry(node_t, x, list, node_list) { + int num_edges; + + if (idx == -1) { + /* leader edges start AFTER follower edges */ + x->next_edge = x->n_followers + 1; + } + num_edges = get_irn_n_outs(x->node); + + /* for all edges in x.L.def_use_{idx} */ + while (x->next_edge <= num_edges) { + const ir_def_use_edge *edge = &x->node->out[x->next_edge]; + ir_node *succ; + + /* check if we have necessary edges */ + if (edge->pos > idx) + break; + + ++x->next_edge; + + succ = edge->use; + + /* only non-commutative nodes */ + if (env->commutative && + (idx == 0 || idx == 1) && is_op_commutative(get_irn_op(succ))) + continue; + + /* ignore the "control input" for non-pinned nodes + if we are running in GCSE mode */ + if (idx < end_idx && get_irn_pinned(succ) != op_pin_state_pinned) + continue; + + y = get_irn_node(succ); + assert(get_irn_n(succ, idx) == x->node); + + /* ignore block edges touching followers */ + if (idx == -1 && y->is_follower) + continue; + + if (is_constant_type(y->type)) { + ir_opcode code = get_irn_opcode(succ); + if (code == iro_Sub || code == iro_Cmp) + add_to_cprop(y, env); + } + + /* Partitions of constants should not be split simply because their Nodes have unequal + functions or incongruent inputs. */ + if (type_is_neither_top_nor_const(y->type) && + (! is_Phi(y->node) || is_live_input(y->node, idx))) { + add_to_touched(y, env); } } + } +} /* collect_touched */ + +/** + * Collect commutative nodes to the touched list. + * + * @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) +{ + node_t *x, *y; + + list_for_each_entry(node_t, x, list, node_list) { + int num_edges; + + num_edges = get_irn_n_outs(x->node); + + x->next_edge = x->n_followers + 1; + + /* for all edges in x.L.def_use_{idx} */ + while (x->next_edge <= num_edges) { + const ir_def_use_edge *edge = &x->node->out[x->next_edge]; + ir_node *succ; + + /* check if we have necessary edges */ + if (edge->pos > 1) + break; + + ++x->next_edge; + if (edge->pos < 0) + continue; + + succ = edge->use; + + /* only commutative nodes */ + if (!is_op_commutative(get_irn_op(succ))) + continue; + + y = get_irn_node(succ); + if (is_constant_type(y->type)) { + ir_opcode code = get_irn_opcode(succ); + if (code == iro_Eor) + add_to_cprop(y, env); + } + + /* Partitions of constants should not be split simply because their Nodes have unequal + functions or incongruent inputs. */ + if (type_is_neither_top_nor_const(y->type)) { + add_to_touched(y, env); + } + } + } +} /* collect_commutative_touched */ + +/** + * Split the partitions if caused by the first entry on the worklist. + * + * @param env the environment + */ +static void cause_splits(environment_t *env) +{ + partition_t *X, *Z, *N; + int idx; + + /* remove the first partition from the worklist */ + X = env->worklist; + env->worklist = X->wl_next; + X->on_worklist = 0; + + dump_partition("Cause_split: ", X); + + if (env->commutative) { + /* handle commutative nodes first */ + + /* empty the touched set: already done, just clear the list */ + env->touched = NULL; + + collect_commutative_touched(&X->Leader, env); + collect_commutative_touched(&X->Follower, env); + + for (Z = env->touched; Z != NULL; Z = N) { + 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); + + /* beware, split might change Z */ + N = Z->touched_next; - for (Z = env->touched; Z != NULL; Z = Z->touched_next) { /* remove it from the touched set */ Z->on_touched = 0; - if (Z->n_nodes != Z->n_touched) { - DB((dbg, LEVEL_2, "Split part%d by touched\n", Z->nr)); - split(Z, Z->touched, env); + /* Empty local Z.touched. */ + 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_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_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); + } + } + + /* combine temporary leader and follower list */ + for (idx = -1; idx <= X->max_user_inputs; ++idx) { + /* empty the touched set: already done, just clear the list */ + env->touched = NULL; + + collect_touched(&X->Leader, idx, env); + collect_touched(&X->Follower, idx, env); + + for (Z = env->touched; Z != NULL; Z = N) { + node_t *e; + node_t *touched = Z->touched; + unsigned n_touched = Z->n_touched; + + assert(Z->touched != NULL); + + /* beware, split might change Z */ + N = Z->touched_next; + + /* remove it from the touched set */ + Z->on_touched = 0; + /* Empty local Z.touched. */ - for (e = Z->touched; e != NULL; e = e->next) { + for (e = touched; e != NULL; e = e->next) { + assert(e->is_follower == 0); e->on_touched = 0; } 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); + } else + assert(n_touched <= Z->n_leader); } } } /* cause_splits */ @@ -775,7 +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; @@ -783,7 +1675,7 @@ static partition_t *split_by_what(partition_t *X, what_func What, /* Let map be an empty mapping from the range of What to (local) list of Nodes. */ listmap_init(&map); - list_for_each_entry(node_t, x, &X->entries, node_list) { + list_for_each_entry(node_t, x, &X->Leader, node_list) { void *id = What(x, env); listmap_entry_t *entry; @@ -807,8 +1699,8 @@ static partition_t *split_by_what(partition_t *X, what_func What, S = iter->list; /* Add SPLIT( X, S ) to P. */ - DB((dbg, LEVEL_2, "Split part%d by what\n", X->nr)); - R = split(X, S, env); + DB((dbg, LEVEL_2, "Split part%d by WHAT = %s\n", X->nr, what_reason)); + R = split(&X, S, env); R->split_next = *P; *P = R; } @@ -821,18 +1713,21 @@ 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; key.code = get_irn_opcode(irn); key.mode = get_irn_mode(irn); + key.arity = get_irn_arity(irn); key.u.proj = 0; key.u.ent = NULL; @@ -843,6 +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; } @@ -852,14 +1768,21 @@ 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; int i = env->lambda_input; if (i >= get_irn_arity(node->node)) { - /* we are outside the allowed range */ + /* + * We are outside the allowed range: This can happen even + * if we have split by opcode first: doing so might move Followers + * to Leaders and those will have a different opcode! + * Note that in this case the partition is on the cprop list and will be + * split again. + */ return NULL; } @@ -870,19 +1793,73 @@ 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) +{ + ir_node *irn = node->node; + ir_node *skipped = skip_Proj(irn); + ir_node *pred, *left, *right; + node_t *p; + partition_t *pl, *pr; + int i = env->lambda_input; + + if (i >= get_irn_arity(node->node)) { + /* + * We are outside the allowed range: This can happen even + * if we have split by opcode first: doing so might move Followers + * to Leaders and those will have a different opcode! + * Note that in this case the partition is on the cprop list and will be + * split again. + */ + return NULL; + } + + /* ignore the "control input" for non-pinned nodes + if we are running in GCSE mode */ + if (i < env->end_idx && get_irn_pinned(skipped) != op_pin_state_pinned) + return NULL; + + if (i == -1) { + pred = get_irn_n(skipped, i); + p = get_irn_node(pred); + return p->part; + } + + if (is_op_commutative(get_irn_op(irn))) { + /* normalize partition order by returning the "smaller" on input 0, + the "bigger" on input 1. */ + left = get_binop_left(irn); + pl = get_irn_node(left)->part; + right = get_binop_right(irn); + pr = get_irn_node(right)->part; + + if (i == 0) + return pl < pr ? pl : pr; + else + return pl > pr ? pl : pr; + } else { + /* a not split out Follower */ + pred = get_irn_n(irn, i); + p = get_irn_node(pred); + + return p->part; + } +} /* lambda_commutative_partition */ + /** - * Checks whether a type is a constant. + * Returns true if a type is a constant (and NOT Top + * or Bottom). */ -static int is_type_constant(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); - /* else it is a symconst */ - return 1; -} + return is_entity(type.sym.entity_p); +} /* is_con */ /** * Implements split_by(). @@ -890,34 +1867,56 @@ static int is_type_constant(lattice_elem_t type) { * @param X the partition to split * @param env the environment */ -static void split_by(partition_t *X, environment_t *env) { - partition_t *P = NULL; +static void split_by(partition_t *X, environment_t *env) +{ + partition_t *I, *P = NULL; int input; - DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr)); + dump_partition("split_by", X); + + if (X->n_leader == 1) { + /* we have only one leader, no need to split, just check it's type */ + node_t *x = get_first_node(X); + X->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type); + return; + } + + DEBUG_ONLY(what_reason = "lambda n.(n.type)";) P = split_by_what(X, lambda_type, &P, env); + dump_split_list(P); + + /* adjust the type tags, we have split partitions by type */ + for (I = P; I != NULL; I = I->split_next) { + node_t *x = get_first_node(I); + I->type_is_T_or_C = x->type.tv == tarval_top || is_con(x->type); + } + do { partition_t *Y = P; P = P->split_next; - if (Y->n_nodes > 1) { - lattice_elem_t type = get_partition_type(Y); - + if (Y->n_leader > 1) { /* we do not want split the TOP or constant partitions */ - if (type.tv != tarval_top && !is_type_constant(type)) { + if (! Y->type_is_T_or_C) { partition_t *Q = NULL; - DB((dbg, LEVEL_2, "WHAT = lambda n.(n.opcode) on part%d\n", Y->nr)); + DEBUG_ONLY(what_reason = "lambda n.(n.opcode)";) Q = split_by_what(Y, lambda_opcode, &Q, env); + dump_split_list(Q); do { partition_t *Z = Q; Q = Q->split_next; - if (Z->n_nodes > 1) { + if (Z->n_leader > 1) { const node_t *first = get_first_node(Z); int arity = get_irn_arity(first->node); partition_t *R, *S; + what_func what = lambda_partition; + DEBUG_ONLY(char buf[64];) + + if (env->commutative && is_op_commutative(get_irn_op(first->node))) + what = lambda_commutative_partition; /* * BEWARE: during splitting by input 2 for instance we might @@ -932,10 +1931,12 @@ static void split_by(partition_t *X, environment_t *env) { partition_t *Z_prime = R; R = R->split_next; - if (Z_prime->n_nodes > 1) { + if (Z_prime->n_leader > 1) { env->lambda_input = input; - DB((dbg, LEVEL_2, "WHAT = lambda n.(n[%d].partition) on part%d\n", input, Z_prime->nr)); - S = split_by_what(Z_prime, lambda_partition, &S, env); + DEBUG_ONLY(snprintf(buf, sizeof(buf), "lambda n.(n[%d].partition)", input);) + DEBUG_ONLY(what_reason = buf;) + S = split_by_what(Z_prime, what, &S, env); + dump_split_list(S); } else { Z_prime->split_next = S; S = Z_prime; @@ -956,15 +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) { @@ -988,10 +1984,17 @@ 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) || has_Block_entity(block)) { + /* start block and labelled blocks are always reachable */ + node->type.tv = tarval_reachable; + return; + } + for (i = get_Block_n_cfgpreds(block) - 1; i >= 0; --i) { node_t *pred = get_irn_node(get_Block_cfgpred(block, i)); @@ -1009,7 +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 */ @@ -1019,14 +2023,19 @@ static void compute_Bad(node_t *node) { * * @param node the node */ -static void compute_Unknown(node_t *node) { - /* While Unknown nodes compute Top, but this is dangerous: - * a if (unknown) would lead to BOTH control flows unreachable. +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 * graph. - * For now, we compute bottom here. + * + * 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 + * (jump threading for instance) might replace them by Phib's... */ - node->type.tv = tarval_bottom; + node->type.tv = tarval_UNKNOWN; } /* compute_Unknown */ /** @@ -1034,28 +2043,58 @@ 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; } /* compute_Jmp */ +/** + * (Re-)compute the type for the Return node. + * + * @param node the 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. */ + node->type.tv = tarval_reachable; +} /* compute_Return */ + /** * (Re-)compute the type for the End 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)); @@ -1065,7 +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: @@ -1078,7 +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; @@ -1125,20 +2164,15 @@ 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)); lattice_elem_t a = l->type; lattice_elem_t b = r->type; - node_t *block = get_irn_node(get_nodes_block(sub)); ir_mode *mode; - if (block->type.tv == tarval_unreachable) { - node->type.tv = tarval_top; - return; - } - if (a.tv == tarval_top || b.tv == tarval_top) { node->type.tv = tarval_top; } else if (a.tv == tarval_bottom || b.tv == tarval_bottom) { @@ -1167,30 +2201,20 @@ static void compute_Add(node_t *node) { } } /* compute_Add */ -/** - * Returns true if a type is a constant. - */ -static int is_con(const lattice_elem_t type) { - return is_entity(type.sym.entity_p) || tarval_is_constant(type.tv); -} - /** * (Re-)compute the type for a Sub. Special case: both nodes are congruent. * * @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)); lattice_elem_t a = l->type; lattice_elem_t b = r->type; - node_t *block = get_irn_node(get_nodes_block(sub)); + tarval *tv; - if (block->type.tv == tarval_unreachable) { - node->type.tv = tarval_top; - return; - } if (a.tv == tarval_top || b.tv == tarval_top) { node->type.tv = tarval_top; } else if (is_con(a) && is_con(b)) { @@ -1205,45 +2229,97 @@ static void compute_Sub(node_t *node) { } } else if (r->part == l->part && (!mode_is_float(get_irn_mode(l->node)))) { - if (node->type.tv == tarval_top) { - /* - * BEWARE: a - a is NOT always 0 for floating Point values, as - * NaN op NaN = NaN, so we must check this here. - */ - ir_mode *mode = get_irn_mode(sub); - node->type.tv = get_mode_null(mode); + /* + * BEWARE: a - a is NOT always 0 for floating Point values, as + * NaN op NaN = NaN, so we must check this here. + */ + ir_mode *mode = get_irn_mode(sub); + tv = get_mode_null(mode); + + /* if the node was ONCE evaluated by all constants, but now + 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->type.tv != tv) + tv = tarval_bottom; + node->type.tv = tv; + } else { + node->type.tv = tarval_bottom; + } +} /* compute_Sub */ + +/** + * (Re-)compute the type for an Eor. Special case: both nodes are congruent. + * + * @param node the 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)); + lattice_elem_t a = l->type; + lattice_elem_t b = r->type; + tarval *tv; + + if (a.tv == tarval_top || b.tv == tarval_top) { + node->type.tv = tarval_top; + } else if (is_con(a) && is_con(b)) { + if (is_tarval(a.tv) && is_tarval(b.tv)) { + node->type.tv = tarval_eor(a.tv, b.tv); + } else if (is_tarval(a.tv) && tarval_is_null(a.tv)) { + node->type = b; + } else if (is_tarval(b.tv) && tarval_is_null(b.tv)) { + node->type = a; } else { node->type.tv = tarval_bottom; } + } else if (r->part == l->part) { + ir_mode *mode = get_irn_mode(eor); + tv = get_mode_null(mode); + + /* if the node was ONCE evaluated by all constants, but now + 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->type.tv != tv) + tv = tarval_bottom; + node->type.tv = tv; } else { node->type.tv = tarval_bottom; } -} /* compute_Sub */ +} /* compute_Eor */ /** * (Re-)compute the type for Cmp. * * @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)); lattice_elem_t a = l->type; lattice_elem_t b = r->type; + ir_mode *mode = get_irn_mode(get_Cmp_left(cmp)); 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 propbably do something */ - node->type.tv = tarval_b_true; } else if (r->part == l->part) { /* both nodes congruent, we can probably do something */ + if (mode_is_float(mode)) { + /* beware of NaN's */ + node->type.tv = tarval_bottom; + } else { + node->type.tv = tarval_b_true; + } + } else if (is_con(a) && is_con(b)) { node->type.tv = tarval_b_true; } else { node->type.tv = tarval_bottom; } -} /* compute_Proj_Cmp */ +} /* compute_Cmp */ /** * (Re-)compute the type for a Proj(Cmp). @@ -1251,29 +2327,38 @@ 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)); lattice_elem_t a = l->type; lattice_elem_t b = r->type; pn_Cmp pnc = get_Proj_proj(proj); + tarval *tv; if (a.tv == tarval_top || b.tv == tarval_top) { - node->type.tv = tarval_top; + node->type.tv = tarval_undefined; } else if (is_con(a) && is_con(b)) { default_compute(node); - } else if (r->part == l->part && - (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) { - if (node->type.tv == tarval_top) { - /* - * BEWARE: a == a is NOT always True for floating Point values, as - * NaN != NaN is defined, so we must check this here. - */ - node->type.tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b); - } else { - node->type.tv = tarval_bottom; - } + + /* + * BEWARE: a == a is NOT always True for floating Point values, as + * NaN != NaN is defined, so we must check this here. + * (while for some pnc we could still optimize we have to stay + * consistent with compute_Cmp, so don't do anything for floats) + */ + } else if (r->part == l->part && !mode_is_float(get_irn_mode(l->node))) { + + tv = pnc & pn_Cmp_Eq ? tarval_b_true : tarval_b_false; + + /* if the node was ONCE evaluated by all constants, but now + 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->type.tv != tv) + tv = tarval_bottom; + node->type.tv = tv; } else { node->type.tv = tarval_bottom; } @@ -1285,12 +2370,65 @@ 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); node_t *selector = get_irn_node(sel); + /* + * Note: it is crucial for the monotony that the Proj(Cond) + * are evaluates after all predecessors of the Cond selector are + * processed. + * Example + * + * if (x != 0) + * + * Due to the fact that 0 is a const, the Cmp gets immediately + * on the cprop list. It will be evaluated before x is evaluated, + * might leaving x as Top. When later x is evaluated, the Cmp + * might change its value. + * BUT if the Cond is evaluated before this happens, Proj(Cond, FALSE) + * gets R, and later changed to F if Cmp is evaluated to True! + * + * We prevent this by putting Conds in an extra cprop_X queue, which + * gets evaluated after the cprop queue is empty. + * + * 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 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) { @@ -1302,7 +2440,12 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { node->type.tv = tarval_reachable; } else { assert(selector->type.tv == tarval_top); - 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); @@ -1315,7 +2458,12 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { node->type.tv = tarval_reachable; } else { assert(selector->type.tv == tarval_top); - node->type.tv = tarval_unreachable; + if (tarval_UNKNOWN == tarval_top) { + /* any condition based on Top is "!=" */ + node->type.tv = tarval_reachable; + } else { + node->type.tv = tarval_unreachable; + } } } } else { @@ -1323,10 +2471,16 @@ 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) { - node->type.tv = tarval_unreachable; + 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 { + 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; @@ -1352,28 +2506,23 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { } /* compute_Proj_Cond */ /** - * (Re-)compute the type for a Proj-Nodes. + * (Re-)compute the type for a Proj-Node. * * @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))); ir_node *pred = get_Proj_pred(proj); - if (get_Proj_proj(proj) == pn_Start_X_initial_exec && is_Start(pred)) { - /* The initial_exec node is ALWAYS reachable. */ - node->type.tv = tarval_reachable; - return; - } - if (block->type.tv == tarval_unreachable) { /* a Proj in a unreachable Block stay Top */ node->type.tv = tarval_top; return; } - if (get_irn_node(pred)->type.tv == tarval_top) { + if (get_irn_node(pred)->type.tv == tarval_top && !is_Cond(pred)) { /* if the predecessor is Top, its Proj follow */ node->type.tv = tarval_top; return; @@ -1408,11 +2557,12 @@ static void compute_Proj(node_t *node) { } /* compute_Proj */ /** - * (Re-)compute the type for a Confirm-Nodes. + * (Re-)compute the type for a Confirm. * * @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)); @@ -1434,48 +2584,376 @@ static void compute_Confirm(node_t *node) { * * @param node the node */ -static void compute(node_t *node) { - compute_func func = (compute_func)node->node->op->ops.generic; +static void compute(node_t *node) +{ + ir_node *irn = node->node; + compute_func func; + +#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_Block(irn)) { + /* for pinned nodes, check its control input */ + if (get_irn_pinned(skip_Proj(irn)) == op_pin_state_pinned) { + node_t *block = get_irn_node(get_nodes_block(irn)); + + if (block->type.tv == tarval_unreachable) { + node->type.tv = tarval_top; + return; + } + } + } + func = (compute_func)node->node->op->ops.generic; if (func != NULL) func(node); } /* compute */ +/* + * Identity functions: Note that one might thing that identity() is just a + * synonym for equivalent_node(). While this is true, we cannot use it for the algorithm + * here, because it expects that the identity node is one of the inputs, which is NOT + * always true for equivalent_node() which can handle (and does sometimes) DAGs. + * So, we have our own implementation, which copies some parts of equivalent_node() + */ + +/** + * Calculates the Identity for Phi nodes + */ +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; + int i; + + for (i = get_Phi_n_preds(phi) - 1; i >= 0; --i) { + node_t *pred_X = get_irn_node(get_Block_cfgpred(block, i)); + + if (pred_X->type.tv == tarval_reachable) { + node_t *pred = get_irn_node(get_Phi_pred(phi, i)); + + if (n_part == NULL) + n_part = pred; + else if (n_part->part != pred->part) { + /* incongruent inputs, not a follower */ + return node; + } + } + } + /* if n_part is NULL here, all inputs path are dead, the Phi computes + * tarval_top, is in the TOP partition and should NOT being split! */ + assert(n_part != NULL); + return n_part; +} /* identity_Phi */ + +/** + * Calculates the Identity for commutative 0 neutral nodes. + */ +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)); + ir_mode *mode = get_irn_mode(op); + tarval *zero; + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return node; + + /* node: no input should be tarval_top, else the binop would be also + * Top and not being split. */ + zero = get_mode_null(mode); + if (a->type.tv == zero) + return b; + if (b->type.tv == zero) + return a; + return node; +} /* identity_comm_zero_binop */ + +/** + * Calculates the Identity for Shift nodes. + */ +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); + tarval *zero; + + /* node: no input should be tarval_top, else the binop would be also + * Top and not being split. */ + zero = get_mode_null(mode); + if (b->type.tv == zero) + return get_irn_node(get_binop_left(op)); + return node; +} /* identity_shift */ + +/** + * Calculates the Identity for Mul nodes. + */ +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)); + ir_mode *mode = get_irn_mode(op); + tarval *one; + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return node; + + /* node: no input should be tarval_top, else the binop would be also + * Top and not being split. */ + one = get_mode_one(mode); + if (a->type.tv == one) + return b; + if (b->type.tv == one) + return a; + return node; +} /* identity_Mul */ + +/** + * Calculates the Identity for Sub nodes. + */ +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); + + /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */ + if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)) + return node; + + /* node: no input should be tarval_top, else the binop would be also + * Top and not being split. */ + if (b->type.tv == get_mode_null(mode)) + return get_irn_node(get_Sub_left(sub)); + return node; +} /* identity_Sub */ + +/** + * Calculates the Identity for And nodes. + */ +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)); + tarval *neutral = get_mode_all_one(get_irn_mode(and)); + + /* node: no input should be tarval_top, else the And would be also + * Top and not being split. */ + if (a->type.tv == neutral) + return b; + if (b->type.tv == neutral) + return a; + return node; +} /* identity_And */ + +/** + * Calculates the Identity for Confirm nodes. + */ +static node_t *identity_Confirm(node_t *node) +{ + ir_node *confirm = node->node; + + /* a Confirm is always a Copy */ + return get_irn_node(get_Confirm_value(confirm)); +} /* identity_Confirm */ + +/** + * Calculates the Identity for Mux nodes. + */ +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)); + /*node_t *sel; */ + + if (t->part == f->part) + return t; + + /* for now, the 1-input identity is not supported */ +#if 0 + sel = get_irn_node(get_Mux_sel(mux)); + + /* Mux sel input is mode_b, so it is always a tarval */ + if (sel->type.tv == tarval_b_true) + return t; + if (sel->type.tv == tarval_b_false) + return f; +#endif + return node; +} /* identity_Mux */ + +/** + * Calculates the Identity for nodes. + */ +static node_t *identity(node_t *node) +{ + ir_node *irn = node->node; + + switch (get_irn_opcode(irn)) { + case iro_Phi: + return identity_Phi(node); + case iro_Mul: + return identity_Mul(node); + case iro_Add: + case iro_Or: + case iro_Eor: + return identity_comm_zero_binop(node); + case iro_Shr: + case iro_Shl: + case iro_Shrs: + case iro_Rotl: + return identity_shift(node); + case iro_And: + return identity_And(node); + case iro_Sub: + return identity_Sub(node); + case iro_Confirm: + return identity_Confirm(node); + case iro_Mux: + return identity_Mux(node); + default: + return node; + } +} /* identity */ + +/** + * 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) +{ + ir_node *l = leader->node; + int j, i, n = get_irn_n_outs(l); + + DB((dbg, LEVEL_2, "%+F is a follower of %+F\n", follower, leader->node)); + /* The leader edges must remain sorted, but follower edges can + be unsorted. */ + for (i = leader->n_followers + 1; i <= n; ++i) { + if (l->out[i].use == follower) { + ir_def_use_edge t = l->out[i]; + + for (j = i - 1; j >= leader->n_followers + 1; --j) + l->out[j + 1] = l->out[j]; + ++leader->n_followers; + l->out[leader->n_followers] = t; + break; + } + } +} /* segregate_def_use_chain_1 */ + +/** + * Node follower is a (new) follower segregate its Leader + * out edges. + * + * @param follower the follower IR node + */ +static void segregate_def_use_chain(const ir_node *follower) +{ + int i; + + for (i = get_irn_arity(follower) - 1; i >= 0; --i) { + node_t *pred = get_irn_node(get_irn_n(follower, i)); + + segregate_def_use_chain_1(follower, pred); + } +} /* segregate_def_use_chain */ + /** * Propagate constant evaluation. * * @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; node_t *fallen; - unsigned n_fallen; + unsigned n_fallen, old_type_was_T_or_C; int i; while (env->cprop != NULL) { + void *oldopcode = NULL; + /* remove the first partition X from cprop */ X = env->cprop; X->on_cprop = 0; env->cprop = X->cprop_next; + old_type_was_T_or_C = X->type_is_T_or_C; + DB((dbg, LEVEL_2, "Propagate type on part%d\n", X->nr)); fallen = NULL; n_fallen = 0; - while (! list_empty(&X->cprop)) { + for (;;) { + int cprop_empty = list_empty(&X->cprop); + int cprop_X_empty = list_empty(&X->cprop_X); + + if (cprop_empty && cprop_X_empty) { + /* both cprop lists are empty */ + break; + } + /* remove the first Node x from X.cprop */ - x = list_entry(X->cprop.next, node_t, cprop_list); + if (cprop_empty) { + /* Get a node from the cprop_X list only if + * all data nodes are processed. + * This ensures, that all inputs of the Cond + * predecessor are processed if its type is still Top. + */ + x = list_entry(X->cprop_X.next, node_t, cprop_list); + } else { + x = list_entry(X->cprop.next, node_t, cprop_list); + } + + //assert(x->part == X); list_del(&x->cprop_list); x->on_cprop = 0; + if (x->is_follower && identity(x) == x) { + /* check the opcode first */ + if (oldopcode == NULL) { + oldopcode = lambda_opcode(get_first_node(X), env); + } + if (oldopcode != lambda_opcode(x, env)) { + if (x->on_fallen == 0) { + /* different opcode -> x falls out of this partition */ + x->next = fallen; + x->on_fallen = 1; + fallen = x; + ++n_fallen; + DB((dbg, LEVEL_2, "Add node %+F to fallen\n", x->node)); + } + } + + /* x will make the follower -> leader transition */ + follower_to_leader(x); + } + /* compute a new type for x */ old_type = x->type; DB((dbg, LEVEL_3, "computing type of %+F\n", x->node)); compute(x); if (x->type.tv != old_type.tv) { - verify_type(old_type, x->type); 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 @@ -1491,23 +2969,48 @@ static void propagate(environment_t *env) { node_t *y = get_irn_node(succ); /* Add y to y.partition.cprop. */ - add_node_to_cprop(y, env); + add_to_cprop(y, env); } } } - if (n_fallen > 0 && n_fallen != X->n_nodes) { + if (n_fallen > 0 && n_fallen != X->n_leader) { DB((dbg, LEVEL_2, "Splitting part%d by fallen\n", X->nr)); - Y = split(X, fallen, env); + Y = split(&X, fallen, env); + /* + * We have split out fallen node. The type of the result + * partition is NOT set yet. + */ + Y->type_is_T_or_C = 0; } else { Y = X; } - /* remove the nodes from the fallen list */ + /* remove the flags from the fallen list */ for (x = fallen; x != NULL; x = x->next) x->on_fallen = 0; - if (Y->n_nodes > 1) - split_by(Y, env); + if (old_type_was_T_or_C) { + node_t *y, *tmp; + + /* check if some nodes will make the leader -> follower transition */ + list_for_each_entry_safe(node_t, y, tmp, &Y->Leader, node_list) { + if (y->type.tv != tarval_top && ! is_con(y->type)) { + node_t *eq_node = identity(y); + + if (eq_node != y && eq_node->part == y->part) { + DB((dbg, LEVEL_2, "Node %+F is a follower of %+F\n", y->node, eq_node->node)); + /* move to Follower */ + y->is_follower = 1; + list_del(&y->node_list); + list_add_tail(&y->node_list, &Y->Follower); + --Y->n_leader; + + segregate_def_use_chain(y->node); + } + } + } + } + split_by(Y, env); } } /* propagate */ @@ -1516,84 +3019,130 @@ 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_nodes > 1) { - DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node)); + if (part->n_leader > 1 || node->is_follower) { + if (node->is_follower) { + DB((dbg, LEVEL_2, "Replacing follower %+F\n", node->node)); + } + else + DB((dbg, LEVEL_2, "Found congruence class for %+F\n", node->node)); return get_first_node(part)->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; + reachable outputs we cannot remove the block */ + return only_one_reachable_proj(pred); } return 0; -} +} /* can_exchange */ /** * 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) { - node_t *node = get_irn_node(block); - int i, j, k, n; - ir_node **ins, **in_X; - ir_node *phi, *next; +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; + ir_node **ins, **in_X; + ir_node *phi, *next; + + n = get_Block_n_cfgpreds(block); + + if (node->type.tv == tarval_unreachable) { + env->modified = 1; + + for (i = n - 1; i >= 0; --i) { + ir_node *pred = get_Block_cfgpred(block, i); + + if (! is_Bad(pred)) { + node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred))); + + if (pred_bl->flagged == 0) { + pred_bl->flagged = 3; + + if (pred_bl->type.tv == tarval_reachable) { + /* + * We will remove an edge from block to its pred. + * This might leave the pred block as an endless loop + */ + if (! is_backedge(block, i)) + keep_alive(pred_bl->node); + } + } + } + } - (void) ctx; - if (block == get_irg_end_block(current_ir_graph) || - block == get_irg_start_block(current_ir_graph)) { /* the EndBlock is always reachable even if the analysis finds out the opposite :-) */ - return; - } - if (node->type.tv == tarval_unreachable) { - /* mark dead blocks */ - set_Block_dead(block); + if (block != get_irg_end_block(current_ir_graph)) { + /* mark dead blocks */ + set_Block_dead(block); + DB((dbg, LEVEL_1, "Removing dead %+F\n", block)); + } else { + /* the endblock is unreachable */ + set_irn_in(block, 0, NULL); + } return; } - n = get_Block_n_cfgpreds(block); - if (n == 1) { /* only one predecessor combine */ ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0)); - if (can_exchange(pred)) - exchange(block, get_nodes_block(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); + exchange(block, new_block); + node->node = new_block; + env->modified = 1; + } return; } @@ -1605,11 +3154,30 @@ static void apply_cf(ir_node *block, void *ctx) { if (node->type.tv == tarval_reachable) { in_X[k++] = pred; + } else { + DB((dbg, LEVEL_1, "Removing dead input %d from %+F (%+F)\n", i, block, pred)); + if (! is_Bad(pred)) { + node_t *pred_bl = get_irn_node(get_nodes_block(skip_Proj(pred))); + + if (pred_bl->flagged == 0) { + pred_bl->flagged = 3; + + if (pred_bl->type.tv == tarval_reachable) { + /* + * We will remove an edge from block to its pred. + * This might leave the pred block as an endless loop + */ + if (! is_backedge(block, i)) + keep_alive(pred_bl->node); + } + } + } } } if (k >= n) return; + /* fix Phi's */ NEW_ARR_A(ir_node *, ins, n); for (phi = get_Block_phis(block); phi != NULL; phi = next) { node_t *node = get_irn_node(phi); @@ -1618,12 +3186,14 @@ 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; DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, c)); + DBG_OPT_COMBO(phi, c, FS_OPT_COMBO_CONST); exchange(phi, c); + env->modified = 1; } else { j = 0; for (i = 0; i < n; ++i) { @@ -1633,37 +3203,127 @@ static void apply_cf(ir_node *block, void *ctx) { ins[j++] = get_Phi_pred(phi, i); } } - if (j <= 1) { + if (j == 1) { /* this Phi is replaced by a single predecessor */ ir_node *s = ins[0]; + node_t *phi_node = get_irn_node(phi); node->node = s; - DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", phi, s)); + DB((dbg, LEVEL_1, "%+F is replaced by %+F because of cf change\n", phi, s)); + DBG_OPT_COMBO(phi, s, FS_OPT_COMBO_FOLLOWER); exchange(phi, s); + phi_node->node = s; + env->modified = 1; } else { set_irn_in(phi, j, ins); + env->modified = 1; } } } - if (k <= 1) { + /* fix block */ + if (k == 1) { /* this Block has only one live predecessor */ ir_node *pred = skip_Proj(in_X[0]); - if (can_exchange(pred)) - exchange(block, get_nodes_block(pred)); - } else { - set_irn_in(block, k, in_X); + 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); + node->node = new_block; + env->modified = 1; + return; + } } -} + set_irn_in(block, k, in_X); + env->modified = 1; +} /* apply_cf */ + +/** + * Exchange a node by its leader. + * Beware: in rare cases the mode might be wrong here, for instance + * 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) +{ + ir_mode *mode = get_irn_mode(irn); + if (mode != get_irn_mode(leader)) { + /* 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, 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) { - node_t *node = get_irn_node(irn); +static void apply_result(ir_node *irn, void *ctx) +{ + environment_t *env = ctx; + node_t *node = get_irn_node(irn); - (void) ctx; if (is_Block(irn) || is_End(irn) || is_Bad(irn)) { /* blocks already handled, do not touch the End node */ } else { @@ -1678,15 +3338,42 @@ static void apply_result(ir_node *irn, void *ctx) { node->node = bad; DB((dbg, LEVEL_1, "%+F is unreachable\n", irn)); exchange(irn, bad); - } - else if (node->type.tv == tarval_unreachable) { - ir_node *bad = get_irg_bad(current_ir_graph); - - /* see comment above */ - set_irn_node(bad, node); - node->node = bad; - DB((dbg, LEVEL_1, "%+F is unreachable\n", irn)); - exchange(irn, bad); + env->modified = 1; + } 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); + + /* see comment above */ + set_irn_node(unk, node); + node->node = unk; + DB((dbg, LEVEL_1, "%+F computes Top\n", irn)); + exchange(irn, unk); + env->modified = 1; + } } else if (get_irn_mode(irn) == mode_X) { if (is_Proj(irn)) { @@ -1694,15 +3381,24 @@ 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, 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; + } } } } @@ -1711,42 +3407,106 @@ static void apply_result(ir_node *irn, void *ctx) { if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) { tarval *tv = node->type.tv; - if (! is_Const(irn)) { + /* + * Beware: never replace mode_T nodes by constants. Currently we must mark + * mode_T nodes with constants, but do NOT replace them. + */ + 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)); - exchange(irn, c); + DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST); + exchange_leader(irn, c); + env->modified = 1; } } 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); + /* can be replaced by a SymConst */ + 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; DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, symc)); - exchange(irn, symc); + DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST); + exchange_leader(irn, symc); + env->modified = 1; } + } else if (is_Confirm(irn)) { + /* Confirms are always follower, but do not kill them here */ } else { ir_node *leader = get_leader(node); if (leader != irn) { - DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader)); - exchange(irn, leader); + int non_strict_phi = 0; + + /* + * Beware: Do not remove Phi(Unknown, ..., x, ..., Unknown) + * as this might create non-strict programs. + */ + if (node->is_follower && is_Phi(irn) && !is_Unknown(leader)) { + int i; + + for (i = get_Phi_n_preds(irn) - 1; i >= 0; --i) { + ir_node *pred = get_Phi_pred(irn, i); + + if (is_Unknown(pred)) { + non_strict_phi = 1; + break; + } + } + } + if (! non_strict_phi) { + DB((dbg, LEVEL_1, "%+F from part%d is replaced by %+F\n", irn, node->part->nr, leader)); + if (node->is_follower) + DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER); + else + DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT); + exchange_leader(irn, leader); + env->modified = 1; + } } } } } } /* apply_result */ +/** + * Fix the keep-alives by deleting unreachable ones. + */ +static void apply_end(ir_node *end, environment_t *env) +{ + int i, j, n = get_End_n_keepalives(end); + ir_node **in = NULL; + + if (n > 0) + NEW_ARR_A(ir_node *, in, n); + + /* fix the keep alive */ + for (i = j = 0; i < n; i++) { + ir_node *ka = get_End_keepalive(end, i); + node_t *node = get_irn_node(ka); + + if (! is_Block(ka)) + node = get_irn_node(get_nodes_block(ka)); + + if (node->type.tv != tarval_unreachable && !is_Bad(ka)) + in[j++] = ka; + } + if (j != n) { + set_End_keepalives(end, j, in); + env->modified = 1; + } +} /* apply_end */ + #define SET(code) op_##code->ops.generic = (op_func)compute_##code /** * 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 */ @@ -1763,32 +3523,53 @@ static void set_compute_functions(void) { SET(Phi); SET(Add); SET(Sub); + SET(Eor); SET(SymConst); SET(Cmp); SET(Proj); SET(Confirm); + SET(Return); SET(End); + SET(Call); } /* set_compute_functions */ -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); +/** + * 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; - ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type); - return 1; -} + 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]; + + if (! ir_nodeset_contains(&set, ka)) { + add_End_keepalive(end, ka); + } + } + ir_nodeset_destroy(&set); +} /* add_memory_keeps */ -void combo(ir_graph *irg) { +void combo(ir_graph *irg) +{ environment_t env; - ir_node *initial_X; + ir_node *initial_bl; node_t *start; ir_graph *rem = current_ir_graph; + int len; current_ir_graph = irg; /* register a debug mask */ FIRM_DBG_REGISTER(dbg, "firm.opt.combo"); - firm_dbg_set_mask(dbg, SET_LEVEL_3); DB((dbg, LEVEL_1, "Doing COMBO for %+F\n", irg)); @@ -1801,11 +3582,17 @@ 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.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); /* we have our own value_of function */ set_value_of_func(get_node_tarval); @@ -1813,16 +3600,29 @@ void combo(ir_graph *irg) { set_compute_functions(); DEBUG_ONLY(part_nr = 0); + 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); add_to_worklist(env.initial, &env); - irg_walk_graph(irg, init_block_phis, create_initial_partitions, &env); + irg_walk_graph(irg, create_initial_partitions, init_block_phis, &env); + + /* set the hook: from now, every node has a partition and a type */ + DEBUG_ONLY(set_dump_node_vcgattr_hook(dump_partition_hook)); + + /* all nodes on the initial partition have type Top */ + env.initial->type_is_T_or_C = 1; /* Place the START Node's partition on cprop. Place the START Node on its local worklist. */ - initial_X = get_irg_initial_exec(irg); - start = get_irn_node(initial_X); - add_node_to_cprop(start, &env); + initial_bl = get_irg_start_block(irg); + start = get_irn_node(initial_bl); + add_to_cprop(start, &env); do { propagate(&env); @@ -1831,20 +3631,48 @@ void combo(ir_graph *irg) { } while (env.cprop != NULL || env.worklist != NULL); dump_all_partitions(&env); + check_all_partitions(&env); #if 0 - set_dump_node_vcgattr_hook(dump_partition_hook); dump_ir_block_graph(irg, "-partition"); - set_dump_node_vcgattr_hook(NULL); -#else - (void)dump_partition_hook; #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 + * apply_cf(). */ + apply_end(get_irg_end(irg), &env); irg_walk_graph(irg, NULL, apply_result, &env); - pmap_destroy(env.type2id_map); + 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_RESOURCE_PHI_LIST); + + /* remove the partition hook */ + DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL)); + + DEL_ARR_F(env.kept_memory); del_set(env.opcode2id_map); obstack_free(&env.obst, NULL); @@ -1852,3 +3680,9 @@ 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 */