X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fopt%2Fcombo.c;h=73c54fd2eede2b647b23cad32fbe46b1eeff6307;hb=6191bc0fe9dc62446b0e0c15e9179b13161dc63d;hp=099091c7efddcfd20b701d91dc5c7c7aa6acce10;hpb=a7a93d6fd495d0860b62933318a38e86e69e1d23;p=libfirm diff --git a/ir/opt/combo.c b/ir/opt/combo.c index 099091c7e..73c54fd2e 100644 --- a/ir/opt/combo.c +++ b/ir/opt/combo.c @@ -23,18 +23,46 @@ * @author Michael Beck * @version $Id$ * + * 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 #include "iroptimize.h" -#include "archop.h" #include "irflag.h" #include "ircons.h" #include "list.h" @@ -52,14 +80,15 @@ #include "debug.h" #include "array_t.h" #include "error.h" - +#include "irnodeset.h" +#include "irpass.h" #include "tv_t.h" #include "irprintf.h" #include "irdump.h" /* define this to check that all type translations are monotone */ -#undef VERIFY_MONOTONE +#define VERIFY_MONOTONE /* define this to check the consistency of partitions */ #define CHECK_PARTITIONS @@ -82,6 +111,10 @@ struct opcode_key_t { union { long proj; /**< For Proj nodes, its proj number */ ir_entity *ent; /**< For Sel Nodes, its entity */ + int intVal; /**< For Conv/Div Nodes: strict/remainderless */ + unsigned uintVal;/**< for Builtin: the kind */ + ir_node *block; /**< for Block: itself */ + void *ptr; /**< generic pointer for hash/cmp */ } u; }; @@ -127,7 +160,6 @@ struct node_t { unsigned on_cprop:1; /**< Set, if this node is on the partition.cprop list. */ unsigned on_fallen:1; /**< Set, if this node is on the fallen list. */ unsigned is_follower:1; /**< Set, if this node is a follower. */ - unsigned by_all_const:1; /**< Set, if this node was once evaluated by all constants. */ unsigned flagged:2; /**< 2 Bits, set if this node was visited by race 1 or 2. */ }; @@ -138,6 +170,7 @@ struct partition_t { 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. */ @@ -163,11 +196,14 @@ typedef struct environment_t { partition_t *touched; /**< the touched set. */ partition_t *initial; /**< The initial partition. */ set *opcode2id_map; /**< The opcodeMode->id map. */ - pmap *type2id_map; /**< The type->id map. */ + ir_node **kept_memory; /**< Array of memory nodes that must be kept. */ int end_idx; /**< -1 for local and 0 for global congruences. */ int lambda_input; /**< Captured argument for lambda_partition(). */ - char nonstd_cond; /**< Set, if a Condb note has a non-Cmp predecessor. */ - char modified; /**< Set, if the graph was modified. */ + 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 @@ -176,8 +212,8 @@ typedef struct environment_t { /** Type of the what function. */ typedef void *(*what_func)(const node_t *node, environment_t *env); -#define get_irn_node(follower) ((node_t *)get_irn_link(follower)) -#define set_irn_node(follower, node) set_irn_link(follower, node) +#define get_irn_node(irn) ((node_t *)get_irn_link(irn)) +#define set_irn_node(irn, node) set_irn_link(irn, node) /* we do NOT use tarval_unreachable here, instead we use Top for this purpose */ #undef tarval_unreachable @@ -187,10 +223,13 @@ 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. */ +/** The tarval returned by Unknown nodes: set to either tarval_bad OR tarval_top. */ static tarval *tarval_UNKNOWN; /* forward */ @@ -200,7 +239,8 @@ static node_t *identity(node_t *node); /** * Check a partition. */ -static void check_partition(const partition_t *T) { +static void check_partition(const partition_t *T) +{ node_t *node; unsigned n = 0; @@ -219,39 +259,131 @@ static void check_partition(const partition_t *T) { } } /* check_partition */ -static void check_all_partitions(environment_t *env) { +/** + * 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; -#ifdef DEBUG_libfirm 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) { - const node_t *e; +static void do_check_list(const node_t *list, int ofs, const partition_t *Z) +{ +#ifndef NDEBUG + const node_t *e; #define NEXT(e) *((const node_t **)((char *)(e) + (ofs))) for (e = list; e != NULL; e = NEXT(e)) { assert(e->part == Z); } #undef NEXT +#else + (void) list; + (void) ofs; + (void) Z; +#endif } /* ido_check_list */ /** * Check a local list. */ -static void check_list(const node_t *list, const partition_t *Z) { +static void check_list(const node_t *list, const partition_t *Z) +{ do_check_list(list, offsetof(node_t, next), Z); } /* check_list */ @@ -262,12 +394,13 @@ static void check_list(const node_t *list, const partition_t *Z) { #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); @@ -293,7 +426,8 @@ static void dump_partition(const char *msg, const partition_t *part) { /** * Dumps a list. */ -static void do_dump_list(const char *msg, const node_t *node, int ofs) { +static void do_dump_list(const char *msg, const node_t *node, int ofs) +{ const node_t *p; int first = 1; @@ -307,46 +441,76 @@ static void do_dump_list(const char *msg, const node_t *node, int ofs) { 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) { +static void dump_race_list(const char *msg, const node_t *list) +{ do_dump_list(msg, list, offsetof(node_t, race_next)); -} +} /* dump_race_list */ /** * Dumps a local list. */ -static void dump_list(const char *msg, const node_t *list) { +static void dump_list(const char *msg, const node_t *list) +{ do_dump_list(msg, list, offsetof(node_t, next)); -} +} /* dump_list */ /** * Dump all partitions. */ -static void dump_all_partitions(const environment_t *env) { +static void dump_all_partitions(const environment_t *env) +{ const partition_t *P; DB((dbg, LEVEL_2, "All partitions\n===============\n")); 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; } @@ -354,23 +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; @@ -383,7 +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 */ @@ -393,7 +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 */ @@ -403,9 +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; @@ -428,27 +594,32 @@ static listmap_entry_t *listmap_find(listmap_t *map, void *id) { * * @return a hash value for the given opcode map entry */ -static unsigned opcode_hash(const opcode_key_t *entry) { - return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ent); +static unsigned opcode_hash(const opcode_key_t *entry) +{ + return (entry->mode - (ir_mode *)0) * 9 + entry->code + entry->u.proj * 3 + HASH_PTR(entry->u.ptr) + entry->arity; } /* opcode_hash */ /** * Compare two entries in the opcode map. */ -static int cmp_opcode(const void *elt, const void *key, size_t size) { +static int cmp_opcode(const void *elt, const void *key, size_t size) +{ const opcode_key_t *o1 = elt; const opcode_key_t *o2 = key; (void) size; return o1->code != o2->code || o1->mode != o2->mode || o1->arity != o2->arity || - o1->u.proj != o2->u.proj || o1->u.ent != o2->u.ent; + o1->u.proj != o2->u.proj || + o1->u.intVal != o2->u.intVal || /* this already checks uIntVal */ + o1->u.ptr != o2->u.ptr; } /* cmp_opcode */ /** * Compare two Def-Use edges for input position. */ -static int cmp_def_use_edge(const void *a, const void *b) { +static int cmp_def_use_edge(const void *a, const void *b) +{ const ir_def_use_edge *ea = a; const ir_def_use_edge *eb = b; @@ -459,7 +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); @@ -476,7 +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 */ @@ -487,7 +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)) @@ -498,7 +672,8 @@ static INLINE tarval *get_node_tarval(const ir_node *irn) { /** * Add a partition to the worklist. */ -static INLINE void add_to_worklist(partition_t *X, environment_t *env) { +static inline void add_to_worklist(partition_t *X, environment_t *env) +{ assert(X->on_worklist == 0); DB((dbg, LEVEL_2, "Adding part%d to worklist\n", X->nr)); X->wl_next = env->worklist; @@ -513,12 +688,14 @@ static INLINE void add_to_worklist(partition_t *X, environment_t *env) { * * @return a newly allocated partition */ -static INLINE partition_t *new_partition(environment_t *env) { - partition_t *part = obstack_alloc(&env->obst, sizeof(*part)); +static inline partition_t *new_partition(environment_t *env) +{ + partition_t *part = OALLOC(&env->obst, partition_t); INIT_LIST_HEAD(&part->Leader); INIT_LIST_HEAD(&part->Follower); INIT_LIST_HEAD(&part->cprop); + INIT_LIST_HEAD(&part->cprop_X); part->wl_next = NULL; part->touched_next = NULL; part->cprop_next = NULL; @@ -543,7 +720,8 @@ static INLINE partition_t *new_partition(environment_t *env) { /** * Get the first node from a partition. */ -static INLINE node_t *get_first_node(const partition_t *X) { +static inline node_t *get_first_node(const partition_t *X) +{ return list_entry(X->Leader.next, node_t, node_list); } /* get_first_node */ @@ -555,7 +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 */ @@ -570,9 +749,10 @@ static INLINE lattice_elem_t get_partition_type(const partition_t *X) { * * @return the created node */ -static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) { +static node_t *create_partition_node(ir_node *irn, partition_t *part, environment_t *env) +{ /* create a partition node and place it in the partition */ - node_t *node = obstack_alloc(&env->obst, sizeof(*node)); + node_t *node = OALLOC(&env->obst, node_t); INIT_LIST_HEAD(&node->node_list); INIT_LIST_HEAD(&node->cprop_list); @@ -588,7 +768,6 @@ static node_t *create_partition_node(ir_node *irn, partition_t *part, environmen node->on_cprop = 0; node->on_fallen = 0; node->is_follower = 0; - node->by_all_const = 0; node->flagged = 0; set_irn_node(irn, node); @@ -599,21 +778,11 @@ static node_t *create_partition_node(ir_node *irn, partition_t *part, environmen } /* 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); - } -} /* init_block_phis */ - -/** - * 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; @@ -623,15 +792,22 @@ 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_Block(irn)) { + set_Block_phis(irn, NULL); + } +} /* create_initial_partitions */ + +/** + * Post-Walker, collect all Block-Phi lists, set Cond. + */ +static void init_block_phis(ir_node *irn, void *ctx) +{ + (void) ctx; + if (is_Phi(irn)) { add_Block_phi(get_nodes_block(irn), irn); - } else if (is_Cond(irn)) { - /* check if all Cond's have a Cmp predecessor. */ - if (get_irn_mode(irn) == mode_b && !is_Cmp(skip_Proj(get_Cond_selector(irn)))) - env->nonstd_cond = 1; - } -} /* create_initial_partitions */ +} /* init_block_phis */ /** * Add a node to the entry.partition.touched set and @@ -640,7 +816,8 @@ static void create_initial_partitions(ir_node *irn, void *ctx) { * @param y a node * @param env the environment */ -static INLINE void add_to_touched(node_t *y, environment_t *env) { +static inline void add_to_touched(node_t *y, environment_t *env) +{ if (y->on_touched == 0) { partition_t *part = y->part; @@ -665,12 +842,20 @@ static INLINE void add_to_touched(node_t *y, environment_t *env) { * @param y the node * @param env the environment */ -static void add_to_cprop(node_t *y, environment_t *env) { +static void add_to_cprop(node_t *y, environment_t *env) +{ + ir_node *irn; + /* Add y to y.partition.cprop. */ if (y->on_cprop == 0) { partition_t *Y = y->part; + ir_node *irn = y->node; - list_add_tail(&y->cprop_list, &Y->cprop); + /* 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)); @@ -682,22 +867,23 @@ static void add_to_cprop(node_t *y, environment_t *env) { Y->on_cprop = 1; } } - if (get_irn_mode(y->node) == mode_T) { + 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(y->node) - 1; i >= 0; --i) { - node_t *proj = get_irn_node(get_irn_out(y->node, 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(y->node)) { + } 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(y->node); phi != NULL; phi = get_Phi_next(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); } @@ -712,7 +898,8 @@ static void add_to_cprop(node_t *y, environment_t *env) { * @param Z_prime the Z' partition, a previous part of Z * @param env the environment */ -static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) { +static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t *env) +{ if (Z->on_worklist || Z_prime->n_leader < Z->n_leader) { add_to_worklist(Z_prime, env); } else { @@ -725,7 +912,8 @@ static void update_worklist(partition_t *Z, partition_t *Z_prime, environment_t * * @param x the node */ -static void move_edges_to_leader(node_t *x) { +static void move_edges_to_leader(node_t *x) +{ ir_node *irn = x->node; int i, j, k; @@ -770,7 +958,8 @@ static void move_edges_to_leader(node_t *x) { * * @return a new partition containing the nodes of g */ -static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) { +static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t *env) +{ partition_t *Z_prime; node_t *node; unsigned n = 0; @@ -820,7 +1009,8 @@ static partition_t *split_no_followers(partition_t *Z, node_t *g, environment_t * * @param n the node */ -static void follower_to_leader(node_t *n) { +static void follower_to_leader(node_t *n) +{ assert(n->is_follower == 1); DB((dbg, LEVEL_2, "%+F make the follower -> leader transition\n", n->node)); @@ -848,7 +1038,8 @@ typedef struct step_env { * @param irn the node to check * @param input number of the input */ -static int is_real_follower(const ir_node *irn, int input) { +static int is_real_follower(const ir_node *irn, int input) +{ node_t *pred; switch (get_irn_opcode(irn)) { @@ -900,21 +1091,18 @@ static int is_real_follower(const ir_node *irn, int input) { if (is_tarval(pred->type.tv) && tarval_is_all_one(pred->type.tv)) return 0; break; - case iro_Min: - case iro_Max: - /* all inputs are followers */ - return 1; default: assert(!"opcode not implemented yet"); break; } return 1; -} +} /* is_real_follower */ /** * Do one step in the race. */ -static int step(step_env *env) { +static int step(step_env *env) +{ node_t *n; if (env->initial != NULL) { @@ -981,7 +1169,8 @@ static int step(step_env *env) { * * @param list the list */ -static int clear_flags(node_t *list) { +static int clear_flags(node_t *list) +{ int res = 0; node_t *n; @@ -1006,13 +1195,14 @@ static int clear_flags(node_t *list) { * * @return a new partition containing the nodes of gg */ -static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { +static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) +{ partition_t *X = *pX; partition_t *X_prime; list_head tmp; - step_env env1, env2, *winner; + step_env senv[2]; node_t *g, *h, *node, *t; - int max_input, transitions; + int max_input, transitions, winner, shf; unsigned n; DEBUG_ONLY(static int run = 0;) @@ -1049,44 +1239,60 @@ static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { /* restore X.Leader */ list_splice(&tmp, &X->Leader); - env1.initial = g; - env1.unwalked = NULL; - env1.walked = NULL; - env1.index = 0; - env1.side = 1; + senv[0].initial = g; + senv[0].unwalked = NULL; + senv[0].walked = NULL; + senv[0].index = 0; + senv[0].side = 1; - env2.initial = h; - env2.unwalked = NULL; - env2.walked = NULL; - env2.index = 0; - env2.side = 2; + 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(&env1)) { - winner = &env1; + if (step(&senv[0])) { + winner = 0; break; } - if (step(&env2)) { - winner = &env2; + if (step(&senv[1])) { + winner = 1; break; } } - assert(winner->initial == NULL); - assert(winner->unwalked == NULL); + assert(senv[winner].initial == NULL); + assert(senv[winner].unwalked == NULL); /* clear flags from walked/unwalked */ - transitions = clear_flags(env1.unwalked); - transitions |= clear_flags(env1.walked); - transitions |= clear_flags(env2.unwalked); - transitions |= clear_flags(env2.walked); + 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 ", winner->walked); + 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 = winner->walked; node != NULL; node = node->race_next) { + for (node = senv[winner].walked; node != NULL; node = node->race_next) { list_del(&node->node_list); node->part = X_prime; if (node->is_follower) { @@ -1112,7 +1318,7 @@ static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { list_for_each_entry_safe(node_t, node, t, &X_prime->Follower, node_list) { if (identity(node) == node) { follower_to_leader(node); - transitions = 1; + transitions |= 1; } } @@ -1126,20 +1332,28 @@ static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { * If there where follower to leader transitions, ensure that the nodes * can be split out if necessary. */ - if (transitions) { - /* place partitions on the cprop list */ - if (X_prime->on_cprop == 0) { + 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 == &env2) { + if (winner != 0) { *pX = X_prime; return X; } @@ -1155,7 +1369,8 @@ static partition_t *split(partition_t **pX, node_t *gg, environment_t *env) { * * @return non-zero if the i'th input of the given Phi node is live */ -static int is_live_input(ir_node *phi, int i) { +static int is_live_input(ir_node *phi, int i) +{ if (i >= 0) { ir_node *block = get_nodes_block(phi); ir_node *pred = get_Block_cfgpred(block, i); @@ -1170,7 +1385,8 @@ static int is_live_input(ir_node *phi, int i) { /** * Return non-zero if a type is a constant. */ -static int is_constant_type(lattice_elem_t type) { +static int is_constant_type(lattice_elem_t type) +{ if (type.tv != tarval_bottom && type.tv != tarval_top) return 1; return 0; @@ -1182,7 +1398,8 @@ static int is_constant_type(lattice_elem_t type) { * * @param type the type to check */ -static int type_is_neither_top_nor_const(const lattice_elem_t type) { +static int type_is_neither_top_nor_const(const lattice_elem_t type) +{ if (is_tarval(type.tv)) { if (type.tv == tarval_top) return 0; @@ -1193,7 +1410,7 @@ static int type_is_neither_top_nor_const(const lattice_elem_t type) { return 0; } return 1; -} +} /* type_is_neither_top_nor_const */ /** * Collect nodes to the touched list. @@ -1202,7 +1419,8 @@ static int type_is_neither_top_nor_const(const lattice_elem_t type) { * @param idx the index of the def_use edge to evaluate * @param env the environment */ -static void collect_touched(list_head *list, int idx, environment_t *env) { +static void collect_touched(list_head *list, int idx, environment_t *env) +{ node_t *x, *y; int end_idx = env->end_idx; @@ -1228,6 +1446,11 @@ static void collect_touched(list_head *list, int idx, environment_t *env) { 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) @@ -1242,7 +1465,7 @@ static void collect_touched(list_head *list, int idx, environment_t *env) { if (is_constant_type(y->type)) { ir_opcode code = get_irn_opcode(succ); - if (code == iro_Sub || code == iro_Eor || code == iro_Cmp) + if (code == iro_Sub || code == iro_Cmp) add_to_cprop(y, env); } @@ -1256,12 +1479,65 @@ static void collect_touched(list_head *list, int idx, environment_t *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) { +static void cause_splits(environment_t *env) +{ partition_t *X, *Z, *N; int idx; @@ -1272,6 +1548,74 @@ static void cause_splits(environment_t *env) { 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; + + /* remove it from the touched set */ + Z->on_touched = 0; + + /* 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 */ @@ -1322,7 +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; @@ -1354,7 +1699,7 @@ 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)); + 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; @@ -1368,13 +1713,15 @@ static partition_t *split_by_what(partition_t *X, what_func What, } /* split_by_what */ /** lambda n.(n.type) */ -static void *lambda_type(const node_t *node, environment_t *env) { +static void *lambda_type(const node_t *node, environment_t *env) +{ (void)env; return node->type.tv; } /* lambda_type */ /** lambda n.(n.opcode) */ -static void *lambda_opcode(const node_t *node, environment_t *env) { +static void *lambda_opcode(const node_t *node, environment_t *env) +{ opcode_key_t key, *entry; ir_node *irn = node->node; @@ -1391,6 +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; } @@ -1400,14 +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; } @@ -1418,14 +1793,68 @@ 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 */ + /** - * Returns true if a type is a constant. + * Returns true if a type is a constant (and NOT Top + * or Bottom). */ -static int is_con(const lattice_elem_t type) { +static int is_con(const lattice_elem_t type) +{ /* be conservative */ if (is_tarval(type.tv)) return tarval_is_constant(type.tv); @@ -1438,7 +1867,8 @@ static int is_con(const lattice_elem_t type) { * @param X the partition to split * @param env the environment */ -static void split_by(partition_t *X, environment_t *env) { +static void split_by(partition_t *X, environment_t *env) +{ partition_t *I, *P = NULL; int input; @@ -1451,8 +1881,9 @@ static void split_by(partition_t *X, environment_t *env) { return; } - DB((dbg, LEVEL_2, "WHAT = lambda n.(n.type) on part%d\n", X->nr)); + 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) { @@ -1469,8 +1900,9 @@ static void split_by(partition_t *X, environment_t *env) { 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; @@ -1480,6 +1912,11 @@ static void split_by(partition_t *X, environment_t *env) { 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 @@ -1496,8 +1933,10 @@ static void split_by(partition_t *X, environment_t *env) { R = R->split_next; 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; @@ -1518,15 +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) { @@ -1550,12 +1984,13 @@ static void default_compute(node_t *node) { * * @param node the node */ -static void compute_Block(node_t *node) { +static void compute_Block(node_t *node) +{ int i; ir_node *block = node->node; - if (block == get_irg_start_block(current_ir_graph)) { - /* start block is always reachable */ + if (block == get_irg_start_block(current_ir_graph) || has_Block_entity(block)) { + /* start block and labelled blocks are always reachable */ node->type.tv = tarval_reachable; return; } @@ -1577,7 +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 */ @@ -1587,7 +2023,8 @@ static void compute_Bad(node_t *node) { * * @param node the node */ -static void compute_Unknown(node_t *node) { +static void compute_Unknown(node_t *node) +{ /* While Unknown nodes should compute Top this is dangerous: * a Top input to a Cond would lead to BOTH control flows unreachable. * While this is correct in the given semantics, it would destroy the Firm @@ -1596,7 +2033,7 @@ static void compute_Unknown(node_t *node) { * It would be safe to compute Top IF it can be assured, that only Cmp * nodes are inputs to Conds. We check that first. * This is the way Frontends typically build Firm, but some optimizations - * (cond_eval for instance) might replace them by Phib's... + * (jump threading for instance) might replace them by Phib's... */ node->type.tv = tarval_UNKNOWN; } /* compute_Unknown */ @@ -1606,28 +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)); @@ -1637,7 +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: @@ -1650,7 +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; @@ -1697,7 +2164,8 @@ static void compute_Phi(node_t *node) { * * @param node the node */ -static void compute_Add(node_t *node) { +static void compute_Add(node_t *node) +{ ir_node *sub = node->node; node_t *l = get_irn_node(get_Add_left(sub)); node_t *r = get_irn_node(get_Add_right(sub)); @@ -1738,7 +2206,8 @@ static void compute_Add(node_t *node) { * * @param node the node */ -static void compute_Sub(node_t *node) { +static void compute_Sub(node_t *node) +{ ir_node *sub = node->node; node_t *l = get_irn_node(get_Sub_left(sub)); node_t *r = get_irn_node(get_Sub_right(sub)); @@ -1758,7 +2227,6 @@ static void compute_Sub(node_t *node) { } else { node->type.tv = tarval_bottom; } - node->by_all_const = 1; } else if (r->part == l->part && (!mode_is_float(get_irn_mode(l->node)))) { /* @@ -1769,9 +2237,10 @@ static void compute_Sub(node_t *node) { tv = get_mode_null(mode); /* if the node was ONCE evaluated by all constants, but now - this breakes AND we cat by partition a different result, switch to bottom. + this breaks AND we get from the argument partitions a different + result, switch to bottom. This happens because initially all nodes are in the same partition ... */ - if (node->by_all_const && node->type.tv != tv) + if (node->type.tv != tv) tv = tarval_bottom; node->type.tv = tv; } else { @@ -1784,7 +2253,8 @@ static void compute_Sub(node_t *node) { * * @param node the node */ -static void compute_Eor(node_t *node) { +static void compute_Eor(node_t *node) +{ ir_node *eor = node->node; node_t *l = get_irn_node(get_Eor_left(eor)); node_t *r = get_irn_node(get_Eor_right(eor)); @@ -1804,15 +2274,15 @@ static void compute_Eor(node_t *node) { } else { node->type.tv = tarval_bottom; } - node->by_all_const = 1; } else if (r->part == l->part) { ir_mode *mode = get_irn_mode(eor); tv = get_mode_null(mode); /* if the node was ONCE evaluated by all constants, but now - this breakes AND we cat by partition a different result, switch to bottom. + this breaks AND we get from the argument partitions a different + result, switch to bottom. This happens because initially all nodes are in the same partition ... */ - if (node->by_all_const && node->type.tv != tv) + if (node->type.tv != tv) tv = tarval_bottom; node->type.tv = tv; } else { @@ -1825,7 +2295,8 @@ static void compute_Eor(node_t *node) { * * @param node the node */ -static void compute_Cmp(node_t *node) { +static void compute_Cmp(node_t *node) +{ ir_node *cmp = node->node; node_t *l = get_irn_node(get_Cmp_left(cmp)); node_t *r = get_irn_node(get_Cmp_right(cmp)); @@ -1833,25 +2304,17 @@ static void compute_Cmp(node_t *node) { lattice_elem_t b = r->type; if (a.tv == tarval_top || b.tv == tarval_top) { -#ifdef WITH_UNKNOWN - /* - * Top is congruent to any other value, we can - * calculate the compare result. - */ - node->type.tv = tarval_b_true; -#else node->type.tv = tarval_top; -#endif - } else if (is_con(a) && is_con(b)) { - /* both nodes are constants, we can probably do something */ - node->type.tv = tarval_b_true; } else if (r->part == l->part) { /* both nodes congruent, we can probably do something */ node->type.tv = tarval_b_true; + } else if (is_con(a) && is_con(b)) { + /* both nodes are constants, we can probably do something */ + 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). @@ -1859,7 +2322,8 @@ static void compute_Cmp(node_t *node) { * @param node the node * @param cond the predecessor Cmp node */ -static void compute_Proj_Cmp(node_t *node, ir_node *cmp) { +static void compute_Proj_Cmp(node_t *node, ir_node *cmp) +{ ir_node *proj = node->node; node_t *l = get_irn_node(get_Cmp_left(cmp)); node_t *r = get_irn_node(get_Cmp_right(cmp)); @@ -1869,31 +2333,22 @@ static void compute_Proj_Cmp(node_t *node, ir_node *cmp) { tarval *tv; if (a.tv == tarval_top || b.tv == tarval_top) { -#ifdef WITH_UNKNOWN - /* see above */ - tv = new_tarval_from_long((pnc & pn_Cmp_Eq) ^ pn_Cmp_Eq, mode_b); - goto not_equal; -#else - node->type.tv = tarval_top; -#endif + node->type.tv = tarval_undefined; } else if (is_con(a) && is_con(b)) { default_compute(node); - node->by_all_const = 1; } else if (r->part == l->part && (!mode_is_float(get_irn_mode(l->node)) || pnc == pn_Cmp_Lt || pnc == pn_Cmp_Gt)) { /* * BEWARE: a == a is NOT always True for floating Point values, as * NaN != NaN is defined, so we must check this here. */ - tv = new_tarval_from_long(pnc & pn_Cmp_Eq, mode_b); -#ifdef WITH_UNKNOWN -not_equal: -#endif + tv = pnc & pn_Cmp_Eq ? tarval_b_true: tarval_b_false; /* if the node was ONCE evaluated by all constants, but now - this breakes AND we cat by partition a different result, switch to bottom. + this breaks AND we get from the argument partitions a different + result, switch to bottom. This happens because initially all nodes are in the same partition ... */ - if (node->by_all_const && node->type.tv != tv) + if (node->type.tv != tv) tv = tarval_bottom; node->type.tv = tv; } else { @@ -1907,12 +2362,65 @@ not_equal: * @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) { @@ -1924,7 +2432,12 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { node->type.tv = tarval_reachable; } else { assert(selector->type.tv == tarval_top); - 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); @@ -1937,7 +2450,12 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { node->type.tv = tarval_reachable; } else { assert(selector->type.tv == tarval_top); - 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 { @@ -1945,10 +2463,16 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { if (selector->type.tv == tarval_bottom) { node->type.tv = tarval_reachable; } else if (selector->type.tv == tarval_top) { - 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; @@ -1978,7 +2502,8 @@ static void compute_Proj_Cond(node_t *node, ir_node *cond) { * * @param node the node */ -static void compute_Proj(node_t *node) { +static void compute_Proj(node_t *node) +{ ir_node *proj = node->node; ir_mode *mode = get_irn_mode(proj); node_t *block = get_irn_node(get_nodes_block(skip_Proj(proj))); @@ -1989,7 +2514,7 @@ static void compute_Proj(node_t *node) { 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; @@ -2028,7 +2553,8 @@ static void compute_Proj(node_t *node) { * * @param node the node */ -static void compute_Confirm(node_t *node) { +static void compute_Confirm(node_t *node) +{ ir_node *confirm = node->node; node_t *pred = get_irn_node(get_Confirm_value(confirm)); @@ -2045,108 +2571,36 @@ static void compute_Confirm(node_t *node) { node->type = pred->type; } /* compute_Confirm */ -/** - * (Re-)compute the type for a Max. - * - * @param node the node - */ -static void compute_Max(node_t *node) { - ir_node *op = node->node; - node_t *l = get_irn_node(get_binop_left(op)); - node_t *r = get_irn_node(get_binop_right(op)); - lattice_elem_t a = l->type; - lattice_elem_t b = r->type; - - if (a.tv == tarval_top || b.tv == tarval_top) { - node->type.tv = tarval_top; - } else if (is_con(a) && is_con(b)) { - /* both nodes are constants, we can probably do something */ - if (a.tv == b.tv) { - /* this case handles symconsts as well */ - node->type = a; - } else { - ir_mode *mode = get_irn_mode(op); - tarval *tv_min = get_mode_min(mode); - - if (a.tv == tv_min) - node->type = b; - else if (b.tv == tv_min) - node->type = a; - else if (is_tarval(a.tv) && is_tarval(b.tv)) { - if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt) - node->type.tv = a.tv; - else - node->type.tv = b.tv; - } else { - node->type.tv = tarval_bad; - } - } - } else if (r->part == l->part) { - /* both nodes congruent, we can probably do something */ - node->type = a; - } else { - node->type.tv = tarval_bottom; - } -} /* compute_Max */ - -/** - * (Re-)compute the type for a Min. - * - * @param node the node - */ -static void compute_Min(node_t *node) { - ir_node *op = node->node; - node_t *l = get_irn_node(get_binop_left(op)); - node_t *r = get_irn_node(get_binop_right(op)); - lattice_elem_t a = l->type; - lattice_elem_t b = r->type; - - if (a.tv == tarval_top || b.tv == tarval_top) { - node->type.tv = tarval_top; - } else if (is_con(a) && is_con(b)) { - /* both nodes are constants, we can probably do something */ - if (a.tv == b.tv) { - /* this case handles symconsts as well */ - node->type = a; - } else { - ir_mode *mode = get_irn_mode(op); - tarval *tv_max = get_mode_max(mode); - - if (a.tv == tv_max) - node->type = b; - else if (b.tv == tv_max) - node->type = a; - else if (is_tarval(a.tv) && is_tarval(b.tv)) { - if (tarval_cmp(a.tv, b.tv) & pn_Cmp_Gt) - node->type.tv = a.tv; - else - node->type.tv = b.tv; - } else { - node->type.tv = tarval_bad; - } - } - } else if (r->part == l->part) { - /* both nodes congruent, we can probably do something */ - node->type = a; - } else { - node->type.tv = tarval_bottom; - } -} /* compute_Min */ - /** * (Re-)compute the type for a given node. * * @param node the node */ -static void compute(node_t *node) { +static void compute(node_t *node) +{ + ir_node *irn = node->node; compute_func func; - if (is_no_Block(node->node)) { - node_t *block = get_irn_node(get_nodes_block(node->node)); +#ifndef VERIFY_MONOTONE + /* + * Once a node reaches bottom, the type cannot fall further + * in the lattice and we can stop computation. + * Do not take this exit if the monotony verifier is + * enabled to catch errors. + */ + if (node->type.tv == tarval_bottom) + return; +#endif - if (block->type.tv == tarval_unreachable) { - node->type.tv = tarval_top; - return; + if (is_no_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; + } } } @@ -2166,7 +2620,8 @@ static void compute(node_t *node) { /** * Calculates the Identity for Phi nodes */ -static node_t *identity_Phi(node_t *node) { +static node_t *identity_Phi(node_t *node) +{ ir_node *phi = node->node; ir_node *block = get_nodes_block(phi); node_t *n_part = NULL; @@ -2195,7 +2650,8 @@ static node_t *identity_Phi(node_t *node) { /** * Calculates the Identity for commutative 0 neutral nodes. */ -static node_t *identity_comm_zero_binop(node_t *node) { +static node_t *identity_comm_zero_binop(node_t *node) +{ ir_node *op = node->node; node_t *a = get_irn_node(get_binop_left(op)); node_t *b = get_irn_node(get_binop_right(op)); @@ -2219,7 +2675,8 @@ static node_t *identity_comm_zero_binop(node_t *node) { /** * Calculates the Identity for Shift nodes. */ -static node_t *identity_shift(node_t *node) { +static node_t *identity_shift(node_t *node) +{ ir_node *op = node->node; node_t *b = get_irn_node(get_binop_right(op)); ir_mode *mode = get_irn_mode(b->node); @@ -2236,7 +2693,8 @@ static node_t *identity_shift(node_t *node) { /** * Calculates the Identity for Mul nodes. */ -static node_t *identity_Mul(node_t *node) { +static node_t *identity_Mul(node_t *node) +{ ir_node *op = node->node; node_t *a = get_irn_node(get_Mul_left(op)); node_t *b = get_irn_node(get_Mul_right(op)); @@ -2260,7 +2718,8 @@ static node_t *identity_Mul(node_t *node) { /** * Calculates the Identity for Sub nodes. */ -static node_t *identity_Sub(node_t *node) { +static node_t *identity_Sub(node_t *node) +{ ir_node *sub = node->node; node_t *b = get_irn_node(get_Sub_right(sub)); ir_mode *mode = get_irn_mode(sub); @@ -2274,12 +2733,13 @@ static node_t *identity_Sub(node_t *node) { if (b->type.tv == get_mode_null(mode)) return get_irn_node(get_Sub_left(sub)); return node; -} /* identity_Mul */ +} /* identity_Sub */ /** * Calculates the Identity for And nodes. */ -static node_t *identity_And(node_t *node) { +static node_t *identity_And(node_t *node) +{ ir_node *and = node->node; node_t *a = get_irn_node(get_And_left(and)); node_t *b = get_irn_node(get_And_right(and)); @@ -2297,7 +2757,8 @@ static node_t *identity_And(node_t *node) { /** * Calculates the Identity for Confirm nodes. */ -static node_t *identity_Confirm(node_t *node) { +static node_t *identity_Confirm(node_t *node) +{ ir_node *confirm = node->node; /* a Confirm is always a Copy */ @@ -2307,7 +2768,8 @@ static node_t *identity_Confirm(node_t *node) { /** * Calculates the Identity for Mux nodes. */ -static node_t *identity_Mux(node_t *node) { +static node_t *identity_Mux(node_t *node) +{ ir_node *mux = node->node; node_t *t = get_irn_node(get_Mux_true(mux)); node_t *f = get_irn_node(get_Mux_false(mux)); @@ -2329,58 +2791,11 @@ static node_t *identity_Mux(node_t *node) { return node; } /* identity_Mux */ -/** - * Calculates the Identity for Min nodes. - */ -static node_t *identity_Min(node_t *node) { - ir_node *op = node->node; - node_t *a = get_irn_node(get_binop_left(op)); - node_t *b = get_irn_node(get_binop_right(op)); - ir_mode *mode = get_irn_mode(op); - tarval *tv_max; - - if (a->part == b->part) { - /* leader of multiple predecessors */ - return a; - } - - /* works even with NaN */ - tv_max = get_mode_max(mode); - if (a->type.tv == tv_max) - return b; - if (b->type.tv == tv_max) - return a; - return node; -} /* identity_Min */ - -/** - * Calculates the Identity for Max nodes. - */ -static node_t *identity_Max(node_t *node) { - ir_node *op = node->node; - node_t *a = get_irn_node(get_binop_left(op)); - node_t *b = get_irn_node(get_binop_right(op)); - ir_mode *mode = get_irn_mode(op); - tarval *tv_min; - - if (a->part == b->part) { - /* leader of multiple predecessors */ - return a; - } - - /* works even with NaN */ - tv_min = get_mode_min(mode); - if (a->type.tv == tv_min) - return b; - if (b->type.tv == tv_min) - return a; - return node; -} /* identity_Max */ - /** * Calculates the Identity for nodes. */ -static node_t *identity(node_t *node) { +static node_t *identity(node_t *node) +{ ir_node *irn = node->node; switch (get_irn_opcode(irn)) { @@ -2405,10 +2820,6 @@ static node_t *identity(node_t *node) { return identity_Confirm(node); case iro_Mux: return identity_Mux(node); - case iro_Min: - return identity_Min(node); - case iro_Max: - return identity_Max(node); default: return node; } @@ -2418,7 +2829,8 @@ static node_t *identity(node_t *node) { * Node follower is a (new) follower of leader, segregate Leader * out edges. */ -static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) { +static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) +{ ir_node *l = leader->node; int j, i, n = get_irn_n_outs(l); @@ -2439,13 +2851,13 @@ static void segregate_def_use_chain_1(const ir_node *follower, node_t *leader) { } /* segregate_def_use_chain_1 */ /** - * Node follower is a (new) follower of leader, segregate Leader - * out edges. If follower is a n-congruent Input identity, all follower - * inputs congruent to follower are also leader. + * 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) { +static void segregate_def_use_chain(const ir_node *follower) +{ int i; for (i = get_irn_arity(follower) - 1; i >= 0; --i) { @@ -2460,7 +2872,8 @@ static void segregate_def_use_chain(const ir_node *follower) { * * @param env the environment */ -static void propagate(environment_t *env) { +static void propagate(environment_t *env) +{ partition_t *X, *Y; node_t *x; lattice_elem_t old_type; @@ -2481,9 +2894,27 @@ static void propagate(environment_t *env) { 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; @@ -2513,8 +2944,8 @@ static void propagate(environment_t *env) { 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 @@ -2580,7 +3011,8 @@ static void propagate(environment_t *env) { * * @param irn the node */ -static ir_node *get_leader(node_t *node) { +static ir_node *get_leader(node_t *node) +{ partition_t *part = node->part; if (part->n_leader > 1 || node->is_follower) { @@ -2595,38 +3027,47 @@ static ir_node *get_leader(node_t *node) { return node->node; } /* get_leader */ +/** + * Returns non-zero if a mode_T node has only one reachable output. + */ +static int only_one_reachable_proj(ir_node *n) +{ + int i, k = 0; + + for (i = get_irn_n_outs(n) - 1; i >= 0; --i) { + ir_node *proj = get_irn_out(n, i); + node_t *node; + + /* skip non-control flow Proj's */ + if (get_irn_mode(proj) != mode_X) + continue; + + node = get_irn_node(proj); + if (node->type.tv == tarval_reachable) { + if (++k > 1) + return 0; + } + } + return 1; +} /* only_one_reachable_proj */ + /** * Return non-zero if the control flow predecessor node pred * is the only reachable control flow exit of its block. * - * @param pred the control flow exit + * @param pred the control flow exit + * @param block the destination block */ -static int can_exchange(ir_node *pred) { - if (is_Start(pred)) +static int can_exchange(ir_node *pred, ir_node *block) +{ + if (is_Start(pred) || has_Block_entity(block)) return 0; else if (is_Jmp(pred)) return 1; else if (get_irn_mode(pred) == mode_T) { - int i, k; - /* if the predecessor block has more than one - reachable outputs we cannot remove the block */ - k = 0; - for (i = get_irn_n_outs(pred) - 1; i >= 0; --i) { - ir_node *proj = get_irn_out(pred, i); - node_t *node; - - /* skip non-control flow Proj's */ - if (get_irn_mode(proj) != mode_X) - continue; - - node = get_irn_node(proj); - if (node->type.tv == tarval_reachable) { - if (++k > 1) - return 0; - } - } - return 1; + reachable outputs we cannot remove the block */ + return only_one_reachable_proj(pred); } return 0; } /* can_exchange */ @@ -2635,7 +3076,8 @@ static int can_exchange(ir_node *pred) { * Block Post-Walker, apply the analysis results on control flow by * shortening Phi's and Block inputs. */ -static void apply_cf(ir_node *block, void *ctx) { +static void apply_cf(ir_node *block, void *ctx) +{ environment_t *env = ctx; node_t *node = get_irn_node(block); int i, j, k, n; @@ -2685,7 +3127,7 @@ static void apply_cf(ir_node *block, void *ctx) { /* only one predecessor combine */ ir_node *pred = skip_Proj(get_Block_cfgpred(block, 0)); - if (can_exchange(pred)) { + if (can_exchange(pred, block)) { ir_node *new_block = get_nodes_block(pred); DB((dbg, LEVEL_1, "Fuse %+F with %+F\n", block, new_block)); DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF); @@ -2727,6 +3169,7 @@ static void apply_cf(ir_node *block, void *ctx) { 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); @@ -2735,7 +3178,7 @@ static void apply_cf(ir_node *block, void *ctx) { if (is_tarval(node->type.tv) && tarval_is_constant(node->type.tv)) { /* this Phi is replaced by a constant */ tarval *tv = node->type.tv; - ir_node *c = new_r_Const(current_ir_graph, block, get_tarval_mode(tv), tv); + ir_node *c = new_Const(tv); set_irn_node(c, node); node->node = c; @@ -2770,27 +3213,106 @@ static void apply_cf(ir_node *block, void *ctx) { } } + /* fix block */ if (k == 1) { /* this Block has only one live predecessor */ ir_node *pred = skip_Proj(in_X[0]); - if (can_exchange(pred)) { + if (can_exchange(pred, block)) { ir_node *new_block = get_nodes_block(pred); DBG_OPT_COMBO(block, new_block, FS_OPT_COMBO_CF); exchange(block, new_block); node->node = new_block; env->modified = 1; + return; } - } else { - set_irn_in(block, k, in_X); - env->modified = 1; } -} + 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) { +static void apply_result(ir_node *irn, void *ctx) +{ environment_t *env = ctx; node_t *node = get_irn_node(irn); @@ -2809,17 +3331,39 @@ static void apply_result(ir_node *irn, void *ctx) { DB((dbg, LEVEL_1, "%+F is unreachable\n", irn)); exchange(irn, bad); env->modified = 1; - } - else if (node->type.tv == tarval_unreachable) { - /* don't kick away Unknown */ - if (! is_Unknown(irn)) { - ir_node *bad = get_irg_bad(current_ir_graph); + } 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(bad, node); - node->node = bad; - DB((dbg, LEVEL_1, "%+F is unreachable\n", irn)); - exchange(irn, bad); + set_irn_node(unk, node); + node->node = unk; + DB((dbg, LEVEL_1, "%+F computes Top\n", irn)); + exchange(irn, unk); env->modified = 1; } } @@ -2829,17 +3373,24 @@ static void apply_result(ir_node *irn, void *ctx) { ir_node *cond = get_Proj_pred(irn); if (is_Cond(cond)) { - node_t *sel = get_irn_node(get_Cond_selector(cond)); - - if (is_tarval(sel->type.tv) && tarval_is_constant(sel->type.tv)) { - /* Cond selector is a constant and the Proj is reachable, make a Jmp */ - ir_node *jmp = new_r_Jmp(current_ir_graph, block->node); + if (only_one_reachable_proj(cond)) { + ir_node *jmp = new_r_Jmp(block->node); set_irn_node(jmp, node); node->node = jmp; DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, jmp)); DBG_OPT_COMBO(irn, jmp, FS_OPT_COMBO_CF); exchange(irn, jmp); env->modified = 1; + } else { + node_t *sel = get_irn_node(get_Cond_selector(cond)); + tarval *tv = sel->type.tv; + + if (is_tarval(tv) && tarval_is_constant(tv)) { + /* The selector is a constant, but more + * than one output is active: An unoptimized + * case found. */ + env->unopt_cf = 1; + } } } } @@ -2854,24 +3405,24 @@ static void apply_result(ir_node *irn, void *ctx) { */ if (! is_Const(irn) && get_irn_mode(irn) != mode_T) { /* can be replaced by a constant */ - ir_node *c = new_r_Const(current_ir_graph, block->node, get_tarval_mode(tv), tv); + ir_node *c = new_Const(tv); set_irn_node(c, node); node->node = c; DB((dbg, LEVEL_1, "%+F is replaced by %+F\n", irn, c)); DBG_OPT_COMBO(irn, c, FS_OPT_COMBO_CONST); - exchange(irn, c); + 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)); DBG_OPT_COMBO(irn, symc, FS_OPT_COMBO_CONST); - exchange(irn, symc); + exchange_leader(irn, symc); env->modified = 1; } } else if (is_Confirm(irn)) { @@ -2880,13 +3431,33 @@ static void apply_result(ir_node *irn, void *ctx) { 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)); - if (node->is_follower) - DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_FOLLOWER); - else - DBG_OPT_COMBO(irn, leader, FS_OPT_COMBO_CONGRUENT); - exchange(irn, leader); - env->modified = 1; + 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; + } } } } @@ -2896,7 +3467,8 @@ static void apply_result(ir_node *irn, void *ctx) { /** * Fix the keep-alives by deleting unreachable ones. */ -static void apply_end(ir_node *end, environment_t *env) { +static void apply_end(ir_node *end, environment_t *env) +{ int i, j, n = get_End_n_keepalives(end); ir_node **in; @@ -2925,7 +3497,8 @@ static void apply_end(ir_node *end, environment_t *env) { /** * sets the generic functions to compute. */ -static void set_compute_functions(void) { +static void set_compute_functions(void) +{ int i; /* set the default compute function */ @@ -2947,36 +3520,48 @@ static void set_compute_functions(void) { SET(Cmp); SET(Proj); SET(Confirm); + SET(Return); SET(End); + SET(Call); +} /* set_compute_functions */ - if (op_Max != NULL) - SET(Max); - if (op_Min != NULL) - SET(Min); +/** + * Add memory keeps. + */ +static void add_memory_keeps(ir_node **kept_memory, int len) +{ + ir_node *end = get_irg_end(current_ir_graph); + int i; + ir_nodeset_t set; -} /* set_compute_functions */ + ir_nodeset_init(&set); -static int dump_partition_hook(FILE *F, ir_node *n, ir_node *local) { -#ifdef DEBUG_libfirm - ir_node *irn = local != NULL ? local : n; - node_t *node = get_irn_node(irn); + /* 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)); - ir_fprintf(F, "info2 : \"partition %u type %+F\"\n", node->part->nr, node->type); - return 1; -#endif -} + 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_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)); @@ -2989,11 +3574,14 @@ void combo(ir_graph *irg) { env.dbg_list = NULL; #endif env.opcode2id_map = new_set(cmp_opcode, iro_Last * 4); - env.type2id_map = pmap_create(); + env.kept_memory = NEW_ARR_F(ir_node *, 0); env.end_idx = get_opt_global_cse() ? 0 : -1; env.lambda_input = 0; - env.nonstd_cond = 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); @@ -3004,18 +3592,20 @@ void combo(ir_graph *irg) { set_compute_functions(); DEBUG_ONLY(part_nr = 0); - ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); + ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST); + + if (env.opt_unknown) + tarval_UNKNOWN = tarval_top; + else + tarval_UNKNOWN = tarval_bad; /* create the initial partition and place it on the work list */ env.initial = new_partition(&env); 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); -#ifdef WITH_UNKNOWN - tarval_UNKNOWN = env.nonstd_cond ? tarval_bad : tarval_top; -#else - tarval_UNKNOWN = tarval_bad; -#endif + /* 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; @@ -3036,17 +3626,29 @@ void combo(ir_graph *irg) { 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); - irg_walk_graph(irg, NULL, apply_result, &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); + + 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 */ @@ -3054,11 +3656,15 @@ void combo(ir_graph *irg) { set_irg_extblk_inconsistent(irg); set_irg_doms_inconsistent(irg); set_irg_loopinfo_inconsistent(irg); + set_irg_entity_usage_state(irg, ir_entity_usage_not_computed); } - ir_free_resources(irg, IR_RESOURCE_IRN_LINK); + ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_PHI_LIST); + + /* remove the partition hook */ + DEBUG_ONLY(set_dump_node_vcgattr_hook(NULL)); - pmap_destroy(env.type2id_map); + DEL_ARR_F(env.kept_memory); del_set(env.opcode2id_map); obstack_free(&env.obst, NULL); @@ -3066,3 +3672,9 @@ void combo(ir_graph *irg) { set_value_of_func(NULL); current_ir_graph = rem; } /* combo */ + +/* Creates an ir_graph pass for combo. */ +ir_graph_pass_t *combo_pass(const char *name) +{ + return def_graph_pass(name ? name : "combo", combo); +} /* combo_pass */